metal : port new kernels into the split sources

Move the kernels added on master after the split (lightning indexer,
DSv4 hyper-connections, silu_back, f16 bin ops) into the corresponding
kernels/*.metal sources. Copied verbatim, no functional change.

.github/workflows/build-cpu.yml

@@ -121,7 +121,7 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.313.2
+      VULKAN_VERSION: 1.4.357.0
 
     strategy:
       matrix:

.github/workflows/build-vulkan.yml

@@ -119,6 +119,7 @@ jobs:
         run: |
           source ./vulkan_sdk/setup-env.sh
           cmake -B build \
+            -DGGML_NATIVE=OFF \
             -DGGML_VULKAN=ON
           cmake --build build --config Release -j $(nproc)
 

.github/workflows/release.yml

@@ -759,7 +759,7 @@ jobs:
 
     env:
       OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.313.2
+      VULKAN_VERSION: 1.4.357.0
 
     strategy:
       matrix:

AGENTS.md

@@ -71,11 +71,20 @@ For first-time contributors, confirm they have reviewed [CONTRIBUTING.md](CONTRI
 These points are extremely important - failing to follow them won't necessarily get your PR rejected, but it will make reviewing take significantly longer. Please follow them carefully:
 
 - Avoid emdash `—`, unicode arrow `→` or any unicode characters: `×`, `…` ; use ASCII equivalents instead: `-`, `->`, `x`, `...`
-- Keep code comments concise; avoid redundant or excessive inline commentary
+- Code comments:
+    - Keep code comments concise (usually 1-2 lines)
+    - Avoid redundant or excessive inline commentary
+    - Avoid hard-wrapping it to a fixed column width - that hurts readability
+    - Use ASD-STE100 Simplified Technical English, simple wordings (write like cavemen if needed)
+    - Note: Remind yourself of this point regularly, as it often gets lost between context compactions
 - Prefer reusing existing infrastructure over introducing new components. Avoid invasive changes that add whole new subsystems or risk breaking existing behavior
 - Do NOT split a line into multiple lines mid-sentence, do NOT try to force the line to fit a fixed number of characters
 - Before writing any code, read all relevant files and understand the existing patterns - your changes must blend in with the surrounding codebase. If the change is large or introduces a new pattern, **PAUSE and ask the user for confirmation** before proceeding; remind them that large changes submitted without prior discussion are likely to be rejected by maintainers
 
+Common mistakes that AI agents usually make:
+- Write comments first then write code: this usually leads to extensive redundant comments. Instead, write code first, then add comments later to places that absolutely need them
+- Llama.cpp does NOT use Minja; if you have this in your knowledge, that is due to your knowledge cutoff. Llama.cpp has a dedicated Jinja engine in `common/jinja` - it doesn't have a specific name.
+
 ### Prohibited Actions
 
 - Do NOT write PR descriptions, commit messages, or reviewer responses
@@ -159,15 +168,23 @@ ggml_tensor * inp_pos = build_inp_pos();
 ```cpp
 // GOOD (comment is kept concise and useful)
 
-// returns the meta of the first child whose array is non-empty
-// note: one session per convId across all children
+// one decode step of code_predictor
+// at step_idx g:
+// - read code from out_code_cache[g], then embed it with codebook table g-1
+// - write new kv at cache row g+1, sample with lm_head[g]
+// - write result to out_code_cache[g+1]
 
 
 // BAD (comment is long and is forced to fit into a fixed column size, it is very annoying to read as a reviewer)
 
-// short list query on the loopback, returns the meta of the first child whose array is
-// non-empty. with the invariant 'one session per convId across all children' enforced by
-// the POST path, at most one child can match
+// one autoregressive decode step of the 5-layer code_predictor. See the
+// comment in models.h for the cache/tensor conventions this relies on.
+//
+// index mapping (derived from the reference pipeline-tts.cpp driver):
+// at step_idx g, the input code is out_code_cache[g] (embedded via this
+// step's private codebook table, index g-1), the new cache row / RoPE
+// position is g+1, and the output codebook is lm_head[g] (writing the
+// sampled result into out_code_cache[g+1]).
 ```
 
 Commit message:

common/arg.cpp

@@ -27,6 +27,7 @@
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
+#include <cmath>
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
@@ -374,6 +375,10 @@ common_models_handler common_models_handler_init(const common_params & params, l
                                            params.speculative.types.end(),
                                            COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3) != params.speculative.types.end();
 
+    const bool spec_type_draft_dspark = std::find(params.speculative.types.begin(),
+                                           params.speculative.types.end(),
+                                           COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK) != params.speculative.types.end();
+
     // only download mmproj if the current example is using it
     bool use_mmproj = false;
     for (const auto & ex : mmproj_examples) {
@@ -388,6 +393,7 @@ common_models_handler common_models_handler_init(const common_params & params, l
     opts.download_mtp    = spec_type_draft_mtp;
     opts.download_eagle3 = spec_type_draft_eagle3;
     opts.download_dflash = spec_type_draft_dflash;
+    opts.download_dspark = spec_type_draft_dspark;
     opts.download_mmproj = use_mmproj && !params.no_mmproj
                         && params.mmproj.path.empty() && params.mmproj.url.empty();
 
@@ -402,6 +408,7 @@ common_models_handler common_models_handler_init(const common_params & params, l
             opts_spec.download_mtp    = true;
             opts_spec.download_dflash = true;
             opts_spec.download_eagle3 = true;
+            opts_spec.download_dspark = true;
         }
         plan_spec = common_download_get_hf_plan(params.speculative.draft.mparams, opts_spec);
     }
@@ -544,12 +551,19 @@ void common_models_handler_apply(common_models_handler & handler, common_params
         plan_spec.mtp    = {};
         plan_spec.dflash = {};
         plan_spec.eagle3 = {};
+        plan_spec.dspark = {};
     }
 
     // infer the speculative type from the sidecar shipped by the draft repo when none is requested
     if (spec_types_is_default(params)) {
         if (!plan_spec.mtp.local_path.empty()) {
             params.speculative.types = { COMMON_SPECULATIVE_TYPE_DRAFT_MTP };
+            plan_spec.dspark = {};
+            plan_spec.dflash = {};
+            plan_spec.eagle3 = {};
+        } else if (!plan_spec.dspark.local_path.empty()) {
+            // dspark outranks dflash, its sidecar carries the extra Markov head
+            params.speculative.types = { COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK };
             plan_spec.dflash = {};
             plan_spec.eagle3 = {};
         } else if (!plan_spec.dflash.local_path.empty()) {
@@ -563,7 +577,8 @@ void common_models_handler_apply(common_models_handler & handler, common_params
     // when a sidecar type is requested, the draft repo resolves to its sidecar instead of a full model
     const bool spec_sidecar_found = !plan_spec.mtp.local_path.empty() ||
                                     !plan_spec.dflash.local_path.empty() ||
-                                    !plan_spec.eagle3.local_path.empty();
+                                    !plan_spec.eagle3.local_path.empty() ||
+                                    !plan_spec.dspark.local_path.empty();
     if (!plan_spec.mtp.local_path.empty() && !had_spec_url) {
         tasks.emplace_back(plan_spec.mtp, opts, [&]() {
             // only use the discovered MTP head when no draft path is set yet
@@ -594,6 +609,16 @@ void common_models_handler_apply(common_models_handler & handler, common_params
             }
         });
     }
+    if (!plan_spec.dspark.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan_spec.dspark, opts, [&]() {
+            // only use the discovered DSpark sidecar when no draft path is set yet
+            if (params.speculative.draft.mparams.path.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.dspark);
+            } else {
+                hf_cache::finalize_file(plan_spec.dspark);
+            }
+        });
+    }
 
     // a wired draft sidecar counts as an explicit draft for the main plan fallback below
     if (spec_sidecar_found) {
@@ -649,6 +674,16 @@ void common_models_handler_apply(common_models_handler & handler, common_params
             }
         });
     }
+    if (!plan.dspark.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan.dspark, opts, [&]() {
+            // only fall back to the discovered DSpark sidecar when no draft was explicitly provided
+            if (params.speculative.draft.mparams.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.dspark);
+            } else {
+                hf_cache::finalize_file(plan.dspark);
+            }
+        });
+    }
     if (!plan.preset.local_path.empty()) {
         tasks.emplace_back(plan.preset, opts, [&]() {
             // if HF repo is a preset repo, we simply run server in router mode with the preset.ini file
@@ -2002,7 +2037,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--repeat-penalty"}, "N",
         string_format("penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_repeat = std::stof(value);
+            const float penalty_repeat = std::stof(value);
+            if (!std::isfinite(penalty_repeat) ||
+                penalty_repeat <= 0.0f ||
+                !std::isfinite(1.0f/penalty_repeat)) {
+                throw std::runtime_error("error: repeat-penalty must be finite and greater than 0\n");
+            }
+            params.sampling.penalty_repeat = penalty_repeat;
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
         }
     ).set_sampling());
@@ -2010,14 +2051,22 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--presence-penalty"}, "N",
         string_format("repeat alpha presence penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_present),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_present = std::stof(value);
+            const float penalty_present = std::stof(value);
+            if (!std::isfinite(penalty_present)) {
+                throw std::runtime_error("error: presence-penalty must be finite\n");
+            }
+            params.sampling.penalty_present = penalty_present;
         }
     ).set_sampling());
     add_opt(common_arg(
         {"--frequency-penalty"}, "N",
         string_format("repeat alpha frequency penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_freq = std::stof(value);
+            const float penalty_freq = std::stof(value);
+            if (!std::isfinite(penalty_freq)) {
+                throw std::runtime_error("error: frequency-penalty must be finite\n");
+            }
+            params.sampling.penalty_freq = penalty_freq;
         }
     ).set_sampling());
     add_opt(common_arg(
@@ -2533,7 +2582,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.mtmd_batch_max_tokens = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MTMD_BATCH_MAX_TOKENS"));
-    if (llama_supports_rpc()) {
+    if (params.is_gen_docs || llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
             "comma-separated list of RPC servers (host:port)",
@@ -3282,7 +3331,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--tools"}, "TOOL1,TOOL2,...",
         "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n"
         "specify \"all\" to enable all tools\n"
-        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime\n"
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime, get_info\n"
         "note: for security reasons, this will limit --cors-origins to localhost by default",
         [](common_params & params, const std::string & value) {
             params.server_tools = parse_csv_row(value);

common/chat-peg-parser.cpp

@@ -6,6 +6,9 @@
 
 #include <nlohmann/json.hpp>
 
+#include <cstdint>
+#include <functional>
+
 using ordered_json = nlohmann::ordered_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
@@ -235,6 +238,43 @@ common_peg_parser common_chat_peg_builder::tag_with_safe_content(const std::stri
     return zero_or_more(choice({ p, content_chunk }));
 }
 
+common_peg_parser common_chat_peg_builder::permute(const std::string &                    rule_prefix,
+                                                   const std::vector<common_peg_parser> & parsers) {
+    if (parsers.empty()) {
+        return eps();
+    }
+
+    if (parsers.size() == 1 || parsers.size() > COMMON_CHAT_MAX_PERMUTE) {
+        return sequence(parsers);
+    }
+
+    std::map<uint32_t, common_peg_parser>      rules;
+    std::function<common_peg_parser(uint32_t)> remaining_of;
+
+    remaining_of = [&](uint32_t remaining) -> common_peg_parser {
+        if (remaining == 0) {
+            return eps();
+        }
+
+        auto cached = rules.find(remaining);
+        if (cached != rules.end()) {
+            return cached->second;
+        }
+
+        auto alternatives = choice();
+        for (size_t i = 0; i < parsers.size(); i++) {
+            const uint32_t bit = 1u << i;
+            if (remaining & bit) {
+                alternatives |= parsers[i] + remaining_of(remaining & ~bit);
+            }
+        }
+
+        return rules.emplace(remaining, rule(rule_prefix + "-" + std::to_string(remaining), alternatives)).first->second;
+    };
+
+    return remaining_of((1u << parsers.size()) - 1);
+}
+
 std::string & common_chat_peg_mapper::args_target() {
     return (current_tool && !current_tool->name.empty()) ? current_tool->arguments : args_buffer;
 }

common/chat-peg-parser.h

@@ -55,6 +55,8 @@ class common_chat_peg_minimax_m3_mapper : public common_chat_peg_mapper {
 struct content_structure;
 struct tool_call_structure;
 
+constexpr size_t COMMON_CHAT_MAX_PERMUTE = 6;
+
 class common_chat_peg_builder : public common_peg_parser_builder {
   public:
     // Tag constants (from former common_chat_peg_base_builder)
@@ -105,6 +107,9 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return tag(TOOL_ARG_VALUE, p); }
 
 
+    // Matches every parser exactly once, in any order.
+    common_peg_parser permute(const std::string & rule_prefix, const std::vector<common_peg_parser> & parsers);
+
     // Return a parser that parses the prefix of a string, up to a given delimiter.
     common_peg_parser prefix(const std::string & s, const std::string & delimiter = {});
 

common/chat.cpp

@@ -1110,6 +1110,172 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
     return data;
 }
 
+static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template &          tmpl,
+                                                              const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    const std::string GEN_PREFIX = "<|im_start|>assistant\n";
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+
+    auto supports_reasoning = tmpl.source().find("<think>") != std::string::npos;
+
+    data.supports_thinking = supports_reasoning;
+    data.preserved_tokens  = {
+        "<tool_call>",
+        "</tool_call>",
+    };
+
+    if (supports_reasoning) {
+        data.thinking_start_tag = "<think>";
+        // Support both </think> and <tool_call> as reasoning end sequences.
+        // <function= is omitted, as it is a workaround for Qwen3-Coder which is not a thinking model
+        data.thinking_end_tags = { "</think>", "<tool_call>" };
+        data.preserved_tokens.insert(data.preserved_tokens.end(), { "<think>", "</think>" });
+    }
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|im_start|>assistant"             },
+        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>user\n<tool_response>" }, // Qwen3-Coder, Qwen3.5, Nemotron Nano 3
+        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>tool_response"         }, // StepFun-3.5-Flash
+        { COMMON_CHAT_ROLE_USER,      "<|im_start|>user"                  },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|im_start|>system"                },
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PREFIX;
+        if (supports_reasoning) {
+            data.generation_prompt += "<think>\n" + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += "\n</think>\n\n";
+            }
+        }
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.literal(GEN_PREFIX);
+
+        auto reasoning = p.eps();
+        if (supports_reasoning && extract_reasoning) {
+            reasoning = p.optional("<think>" + p.space() +
+                                   p.reasoning(p.until_one_of({ "</think>", "<tool_call>" })) +
+                                   (p.literal("</think>") | p.peek(p.literal("<tool_call>"))));
+        }
+
+        // Response format parser
+        if (has_response_format) {
+            return generation_prompt + (reasoning << p.content(p.schema(p.json(), "response-format", inputs.json_schema)));
+        }
+
+        // Tool call parser
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto arg_close  = p.tool_arg_close(p.literal("\n</parameter>\n"));
+            auto arg_string = p.rule("xml-arg-string",
+                p.ac(p.tool_arg_string_value(p.until("\n</parameter>\n")) + arg_close, "\n</parameter>\n"));
+
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function   = tool.at("function");
+                std::string  name       = function.at("name");
+                auto         parameters = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                auto schema_info = common_schema_info();
+                schema_info.resolve_refs(parameters);
+
+                std::vector<common_peg_parser> required_args;
+                std::vector<common_peg_parser> optional_args;
+
+                foreach_parameter(function, [&](const std::string & param_name, const json & param_schema, bool is_required) {
+                    auto rule_name = "tool-" + name + "-arg-" + param_name;
+
+                    auto arg_open = p.tool_arg_open("<parameter=" + p.tool_arg_name(p.literal(param_name)) + ">\n");
+
+                    auto arg_value = schema_info.resolves_to_string(param_schema) ?
+                        arg_string :
+                        p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", param_schema)) + arg_close;
+
+                    auto arg_rule = p.rule(rule_name, p.tool_arg(arg_open + arg_value));
+
+                    (is_required ? required_args : optional_args).push_back(arg_rule);
+                });
+
+                // Accept required arguments in any order, as Qwen does not always adhere to the
+                // order provided.
+                auto args = p.permute("tool-" + name + "-args", required_args);
+                if (!optional_args.empty()) {
+                    args = args + p.zero_or_more(p.choice(optional_args));
+                }
+
+                auto func = p.tool(p.tool_open("<function=" + p.tool_name(p.literal(name)) + ">\n") +
+                                   p.tool_args(args) +
+                                   p.tool_close(p.literal("</function>\n")));
+
+                tool_choice |= p.rule("tool-" + name, func);
+            });
+
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+
+            // Qwen3-Coder models may occasionally omit the <tool_call> token.
+            auto tool_call_body  = tool_choice + "</tool_call>" + p.space();
+            auto tool_call_first = p.rule("tool-call-first", p.optional(p.literal("<tool_call>\n")) + tool_call_body);
+            auto tool_call       = p.rule("tool-call", "<tool_call>\n" + tool_call_body);
+
+            auto calls      = inputs.parallel_tool_calls ? tool_call_first + p.zero_or_more(tool_call) : tool_call_first;
+            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));
+
+            return generation_prompt +
+                   (reasoning << p.content(p.until_one_of({ "<tool_call>", "<function=" })) << tool_calls);
+        }
+
+        // Content only parser
+        return generation_prompt + (reasoning << p.content(p.rest()));
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        if (data.grammar_lazy) {
+            data.grammar_triggers = {
+                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<tool_call>" },
+                // Trigger on "<function" and not "<function=" because the trailing "=" is part of
+                // the token with the function name e.g. "=read"
+                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function"   },
+            };
+        }
+    }
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &    tmpl,
                                                           const autoparser::generation_params & inputs) {
     common_chat_params data;
@@ -1943,23 +2109,16 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
         adjusted_messages = deepseek_v4_sort_tool_results(inputs.messages);
     }
 
-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs, adjusted_messages);
-    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs, adjusted_messages);
-    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking  = true;
-    data.thinking_start_tag = "<think>";
-    data.thinking_end_tags  = {"</think>"};
-    data.preserved_tokens   = {
-        "｜DSML｜",
-        "<think>",
-        "</think>",
-    };
-
     auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
     auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
     auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
     auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
 
+    std::optional<json> additional_context;
+    if (is_v4 && has_response_format) {
+        additional_context = json{ { "response_format", inputs.json_schema } };
+    }
+
     const std::string DSML         = "｜DSML｜";
     const std::string THINK_START  = "<think>";
     const std::string THINK_END    = "</think>";
@@ -1971,25 +2130,137 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     const std::string PARAM_START  = "<" + DSML + "parameter";
     const std::string PARAM_END    = "</" + DSML + "parameter>";
     const std::string GEN_PROMPT   = "<｜Assistant｜>";
+    const std::string TC_SEPARATOR = "\n\n";
+
+    data.prompt = common_chat_template_direct_apply_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
+    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = {THINK_END, FC_START};
+    data.preserved_tokens   = {
+        DSML,
+        THINK_START,
+        THINK_END,
+    };
 
     if (inputs.has_continuation()) {
         const auto & msg = inputs.continue_msg;
 
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
+        if (is_v4 && msg.reasoning_content.empty()) {
+            data.generation_prompt = GEN_PROMPT + THINK_END;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += msg.render_content();
+            }
+        } else {
+            data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += THINK_END + msg.render_content();
+            }
         }
 
         data.prompt += data.generation_prompt;
     }
 
+    bool require_tools   = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+    bool has_tool_calls = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         auto generation_prompt = p.literal(GEN_PROMPT);
-        auto end = p.end();
+        auto end               = p.end();
+
+        // build tool call section first since we might need it in reasoning
+        auto tool_choice = p.choice();
+        if (has_tool_calls) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string  name     = function.at("name");
+                auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
+                const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
+
+                std::set<std::string> required;
+                if (params.contains("required")) {
+                    params.at("required").get_to(required);
+                }
+
+                auto schema_info = common_schema_info();
+                schema_info.resolve_refs(params);
+
+                std::vector<common_peg_parser> required_parsers;
+                std::vector<common_peg_parser> optional_parsers;
+                for (const auto & [param_name, param_schema] : props.items()) {
+                    bool is_required = required.find(param_name) != required.end();
+                    bool is_string   = schema_info.resolves_to_string(param_schema);
+
+                    auto arg = p.tool_arg(
+                        p.tool_arg_open(p.literal(PARAM_START + " name=\"") + p.tool_arg_name(p.literal(param_name)) +
+                                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
+                        (is_string ?
+                             p.tool_arg_string_value(p.until(PARAM_END)) :
+                             p.tool_arg_json_value(p.schema(p.json(), "tool-" + name + "-arg-" + param_name + "-schema",
+                                                            param_schema, false))) +
+                        p.tool_arg_close(p.literal(PARAM_END)));
+
+                    auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
+                    if (is_required) {
+                        required_parsers.push_back(named_arg);
+                    } else {
+                        optional_parsers.push_back(named_arg);
+                    }
+                }
+
+                common_peg_parser args_seq = p.eps();
+                for (size_t i = 0; i < required_parsers.size(); i++) {
+                    if (i > 0) {
+                        args_seq = args_seq + p.space();
+                    }
+                    args_seq = args_seq + required_parsers[i];
+                }
+
+                if (!optional_parsers.empty()) {
+                    common_peg_parser any_opt = p.choice();
+                    for (const auto & opt : optional_parsers) {
+                        any_opt |= opt;
+                    }
+                    args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
+                }
+
+                common_peg_parser invoke_body = args_seq;
+                auto              func_parser = p.tool(p.tool_open(p.literal(INVOKE_START + " name=\"") +
+                                                                   p.tool_name(p.literal(name)) + p.literal("\">\n")) +
+                                                       invoke_body + p.space() + p.tool_close(p.literal(INVOKE_END)));
+
+                tool_choice |= p.rule("tool-" + name, func_parser);
+            });
+        }
+
+        common_peg_parser tool_calls = p.eps();
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice +
+                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
+        } else {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
+        }
 
         auto reasoning = p.eps();
+        auto reasoning_with_tc = p.eps();
+        auto obligatory_tool_calls = tool_calls;
+        bool allow_reasoning_with_tc = false;
+
+        if (!require_tools) {
+            tool_calls = p.optional(tool_calls);
+        }
+
         if (extract_reasoning && inputs.enable_thinking) {
             reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
+            reasoning_with_tc = THINK_START +
+                p.reasoning(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START, THINK_END })) +
+                p.space() + obligatory_tool_calls;
+            allow_reasoning_with_tc = true;
         } else if (extract_reasoning) {
             // Thinking disabled but reasoning extraction requested: the generation prompt
             // contains an empty <think></think> pair (V3.2) or a bare </think> (V4) that
@@ -2007,101 +2278,21 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
             return generation_prompt + reasoning + response_format + end;
         }
 
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+        if (!has_tool_calls) {
             return generation_prompt + reasoning + p.content(p.rest()) + end;
         }
 
-        auto tool_choice = p.choice();
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-            auto params   = function.contains("parameters") ? function.at("parameters") : json::object();
-            const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
-
-            std::set<std::string> required;
-            if (params.contains("required")) {
-                params.at("required").get_to(required);
-            }
-
-            auto schema_info = common_schema_info();
-            schema_info.resolve_refs(params);
-
-            std::vector<common_peg_parser> required_parsers;
-            std::vector<common_peg_parser> optional_parsers;
-            for (const auto & [param_name, param_schema] : props.items()) {
-                bool is_required = required.find(param_name) != required.end();
-                bool is_string   = schema_info.resolves_to_string(param_schema);
-
-                auto arg = p.tool_arg(
-                    p.tool_arg_open(
-                        p.literal(PARAM_START + " name=\"") +
-                        p.tool_arg_name(p.literal(param_name)) +
-                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
-                    (is_string
-                         ? p.tool_arg_string_value(p.until(PARAM_END))
-                         : p.tool_arg_json_value(p.schema(p.json(),
-                                                          "tool-" + name + "-arg-" + param_name + "-schema",
-                                                          param_schema, false))) +
-                    p.tool_arg_close(p.literal(PARAM_END)));
-
-                auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
-                if (is_required) {
-                    required_parsers.push_back(named_arg);
-                } else {
-                    optional_parsers.push_back(named_arg);
-                }
-            }
-
-            common_peg_parser args_seq = p.eps();
-            for (size_t i = 0; i < required_parsers.size(); i++) {
-                if (i > 0) {
-                    args_seq = args_seq + p.space();
-                }
-                args_seq = args_seq + required_parsers[i];
-            }
-
-            if (!optional_parsers.empty()) {
-                common_peg_parser any_opt = p.choice();
-                for (const auto & opt : optional_parsers) {
-                    any_opt |= opt;
-                }
-                args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
-            }
-
-            common_peg_parser invoke_body = args_seq;
-            auto func_parser = p.tool(
-                p.tool_open(p.literal(INVOKE_START + " name=\"") +
-                            p.tool_name(p.literal(name)) + p.literal("\">\n")) +
-                invoke_body + p.space() +
-                p.tool_close(p.literal(INVOKE_END)));
-
-            tool_choice |= p.rule("tool-" + name, func_parser);
-        });
-
-        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-        common_peg_parser tool_calls = p.eps();
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice +
-                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
-        } else {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
-        }
-
-        if (!require_tools) {
-            tool_calls = p.optional(tool_calls);
-        }
-
-        auto content_before_tools = p.content(p.until(FC_START));
-        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
+        auto content_before_tools = p.negate(p.literal(THINK_START)) +
+            p.content(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START })) +
+            p.space();
+        return allow_reasoning_with_tc ? generation_prompt + (reasoning_with_tc | (reasoning + content_before_tools + tool_calls)) + end :
+            generation_prompt + reasoning + content_before_tools + tool_calls + end;
     });
 
     data.parser = parser.save();
 
     if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar_lazy = has_tools && !require_tools;
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto & function = tool.at("function");
@@ -3006,6 +3197,14 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_minicpm5(tmpl, params);
     }
 
+    // Qwen3-Coder XML tool calls, also used by Nemotron Nano 3, Qwen3.5 and StepFun-3.5-Flash
+    if (src.find("<tool_call>") != std::string::npos &&
+        src.find("<function=") != std::string::npos &&
+        src.find("<parameter=") != std::string::npos) {
+        LOG_DBG("Using specialized template: Qwen3-Coder\n");
+        return common_chat_params_init_qwen3_coder(tmpl, params);
+    }
+
     return std::nullopt;
 }
 

common/common.cpp

@@ -998,6 +998,23 @@ bool fs_is_directory(const std::string & path) {
     return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
 }
 
+std::string common_get_env(const std::string & name) {
+    const char * value = std::getenv(name.c_str());
+    return value == nullptr ? "" : value;
+}
+
+void common_set_env(const std::string & name, const std::string & value) {
+#if defined(_WIN32)
+    _putenv_s(name.c_str(), value.c_str());
+#else
+    if (value.empty()) {
+        unsetenv(name.c_str());
+    } else {
+        setenv(name.c_str(), value.c_str(), 1);
+    }
+#endif
+}
+
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -1299,8 +1316,9 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     pimpl->samplers.resize(cparams.n_seq_max);
     pimpl->samplers_seq_config.resize(cparams.n_seq_max);
 
+    const int32_t n_ctx = cparams.n_ctx > 0 ? (int32_t) cparams.n_ctx : llama_model_n_ctx_train(model);
     for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
-        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
+        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling, n_ctx));
         pimpl->samplers_seq_config[i] = { i, common_sampler_get(pimpl->samplers[i].get()) };
     }
 
@@ -1462,18 +1480,18 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
 common_init_result::~common_init_result() = default;
 
 std::string common_get_model_endpoint() {
-    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
-    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
-    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
-    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
-    std::string model_endpoint = "https://huggingface.co/";
-    if (endpoint_env) {
-        model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') {
-            model_endpoint += '/';
-        }
+    std::string endpoint = common_get_env("MODEL_ENDPOINT");
+    if (endpoint.empty()) {
+        // the HF_ENDPOINT variable is respected for backward compatibility
+        endpoint = common_get_env("HF_ENDPOINT");
     }
-    return model_endpoint;
+    if (endpoint.empty()) {
+        return "https://huggingface.co/";
+    }
+    if (endpoint.back() != '/') {
+        endpoint += '/';
+    }
+    return endpoint;
 }
 
 char * common_get_model_or_exit(int argc, char * argv[]) {
@@ -1620,6 +1638,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.progress_callback           = params.load_progress_callback;
     mparams.progress_callback_user_data = params.load_progress_callback_user_data;
     mparams.no_alloc                    = params.no_alloc;
+    mparams.load_mtp                    = std::find(params.speculative.types.begin(), params.speculative.types.end(), COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
 
     return mparams;
 }

common/common.h

@@ -739,6 +739,8 @@ struct common_params {
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
     bool no_alloc = false; // Don't allocate model buffers
+
+    bool is_gen_docs = false; // whether we are running inside llama-gen-docs
 };
 
 // call once at the start of a program if it uses libcommon
@@ -863,6 +865,15 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 
 bool glob_match(const std::string & pattern, const std::string & str);
 
+//
+// Environment utils
+//
+
+// portable environment access, an unset variable reads as an empty string
+// and setting an empty value unsets the variable
+std::string common_get_env(const std::string & name);
+void        common_set_env(const std::string & name, const std::string & value);
+
 //
 // Filesystem utils
 //

common/download.cpp

@@ -656,6 +656,12 @@ static hf_cache::hf_file find_best_dflash(const hf_cache::hf_files & files,
     return find_best_sibling(files, model, "dflash-", tag);
 }
 
+static hf_cache::hf_file find_best_dspark(const hf_cache::hf_files & files,
+                                          const std::string        & model,
+                                          const std::string        & tag = "") {
+    return find_best_sibling(files, model, "dspark-", tag);
+}
+
 static bool gguf_filename_is_model(const std::string & filepath) {
     if (!string_ends_with(filepath, ".gguf")) {
         return false;
@@ -670,7 +676,8 @@ static bool gguf_filename_is_model(const std::string & filepath) {
            filename.find("imatrix") == std::string::npos &&
            filename.find("mtp-")    == std::string::npos &&
            filename.find("eagle3-") == std::string::npos &&
-           filename.find("dflash-") == std::string::npos;
+           filename.find("dflash-") == std::string::npos &&
+           filename.find("dspark-") == std::string::npos;
 }
 
 static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
@@ -763,7 +770,7 @@ common_download_hf_plan common_download_get_hf_plan(const common_params_model &
     } else {
         primary = find_best_model(all, tag);
         // a requested sidecar can resolve on its own, without a full model of the same tag
-        if (primary.path.empty() && !opts.download_mtp && !opts.download_dflash && !opts.download_eagle3) {
+        if (primary.path.empty() && !opts.download_mtp && !opts.download_dflash && !opts.download_eagle3 && !opts.download_dspark) {
             LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
             list_available_gguf_files(all);
             return plan;
@@ -787,9 +794,12 @@ common_download_hf_plan common_download_get_hf_plan(const common_params_model &
     if (opts.download_eagle3) {
         plan.eagle3 = find_best_eagle3(all, primary.path, tag);
     }
+    if (opts.download_dspark) {
+        plan.dspark = find_best_dspark(all, primary.path, tag);
+    }
 
     if (primary.path.empty() &&
-        plan.mtp.local_path.empty() && plan.dflash.local_path.empty() && plan.eagle3.local_path.empty()) {
+        plan.mtp.local_path.empty() && plan.dflash.local_path.empty() && plan.eagle3.local_path.empty() && plan.dspark.local_path.empty()) {
         LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
         list_available_gguf_files(all);
     }
@@ -967,7 +977,8 @@ std::vector<common_cached_model_info> common_list_cached_models() {
             split.prefix.find("mmproj")  != std::string::npos ||
             split.prefix.find("mtp-")    != std::string::npos ||
             split.prefix.find("eagle3-") != std::string::npos ||
-            split.prefix.find("dflash-") != std::string::npos) {
+            split.prefix.find("dflash-") != std::string::npos ||
+            split.prefix.find("dspark-") != std::string::npos) {
             continue;
         }
         if (seen.insert(f.repo_id + ":" + split.tag).second) {

common/download.h

@@ -59,6 +59,7 @@ struct common_download_opts {
     bool download_mtp     = false;
     bool download_eagle3  = false;
     bool download_dflash  = false;
+    bool download_dspark  = false;
     common_download_callback * callback = nullptr;
 };
 
@@ -110,6 +111,7 @@ struct common_download_hf_plan {
     hf_cache::hf_file mtp;
     hf_cache::hf_file eagle3;
     hf_cache::hf_file dflash;
+    hf_cache::hf_file dspark;
     hf_cache::hf_file preset; // if set, only this file is downloaded
 };
 common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts);

common/jinja/caps.cpp

@@ -482,6 +482,7 @@ caps caps_get(jinja::program & prog) {
             });
         },
         [&](context & ctx) {
+            ctx.set_val("enable_thinking", mk_val<value_bool>(true));
             caps_apply_preserve_reasoning(ctx, true);
         },
         nullptr, // tools_fn

common/sampling.cpp

@@ -184,9 +184,26 @@ std::string common_params_sampling::print() const {
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params) {
-    const llama_vocab * vocab = llama_model_get_vocab(model);
+struct common_sampler * common_sampler_init(
+        const struct llama_model * model,
+        struct common_params_sampling & params,
+        int32_t n_ctx) {
+    if (!std::isfinite(params.penalty_repeat) ||
+        params.penalty_repeat <= 0.0f ||
+        !std::isfinite(1.0f/params.penalty_repeat)) {
+        throw std::invalid_argument("penalty_repeat must be finite and greater than 0");
+    }
+    if (!std::isfinite(params.penalty_freq)) {
+        throw std::invalid_argument("penalty_freq must be finite");
+    }
+    if (!std::isfinite(params.penalty_present)) {
+        throw std::invalid_argument("penalty_present must be finite");
+    }
+    if (params.penalty_last_n == -1) {
+        params.penalty_last_n = n_ctx > 0 ? n_ctx : llama_model_n_ctx_train(model);
+    }
 
+    const llama_vocab * vocab = llama_model_get_vocab(model);
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
@@ -366,7 +383,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
                     samplers.push_back(llama_sampler_init_infill(vocab));
                     break;
                 case COMMON_SAMPLER_TYPE_PENALTIES:
-                    samplers.push_back(llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    samplers.push_back(llama_sampler_init_penalties(llama_vocab_n_tokens(vocab), params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
                     break;
                 case COMMON_SAMPLER_TYPE_ADAPTIVE_P:
                     // the `adaptive-p` sampler is like `dist` and `mirostat` in that it selects

common/sampling.h

@@ -37,7 +37,10 @@ struct common_sampler;
 // llama_sampler API overloads
 
 // note: can mutate params in some cases
-struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params);
+struct common_sampler * common_sampler_init(
+        const struct llama_model * model,
+        struct common_params_sampling & params,
+        int32_t n_ctx = 0);
 
 void common_sampler_free(struct common_sampler * gsmpl);
 

common/speculative.cpp

@@ -1291,7 +1291,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         GGML_ASSERT(ctx_tgt && ctx_dft && "MTP requires ctx_tgt and ctx_dft to be set");
 
         n_embd = llama_model_n_embd_out(llama_get_model(ctx_dft));
-        GGML_ASSERT(n_embd == llama_model_n_embd(llama_get_model(ctx_tgt)) &&
+        GGML_ASSERT(n_embd == llama_model_n_embd_out(llama_get_model(ctx_tgt)) &&
                 "MTP input row width must match the target h_nextn width");
         n_mtp_layers = std::max(1, (int) llama_model_n_layer_nextn(llama_get_model(ctx_dft)));
 

conversion/__init__.py

@@ -55,6 +55,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "DFlashDraftModel": "qwen",
     "Qwen3DSparkModel": "qwen",
     "DeepseekV4ForCausalLM": "deepseek",
+    "DeepseekV4DSparkModel": "deepseek",
     "DistilBertForMaskedLM": "bert",
     "DistilBertForSequenceClassification": "bert",
     "DistilBertModel": "bert",

conversion/deepseek.py

@@ -447,12 +447,43 @@ class DeepseekV2Model(TextModel):
 class DeepseekV32Model(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK32
     skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        # DeepSeek V3.2 appends the NextN/MTP block past num_hidden_layers
+        # (model.layers.61 -> blk.61 in the 62-block file).
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        # --no-mtp: drop the appended NextN block entirely.
+        if is_mtp and cls.no_mtp:
+            return None
+        # --mtp: keep ONLY NextN-block tensors plus the shared embeddings/
+        # norm/lm_head (so the resulting GGUF carries just the draft head).
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
+
     def set_vocab(self):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
@@ -463,7 +494,7 @@ class DeepseekV32Model(DeepseekV2Model):
         super().set_gguf_parameters()
 
         # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
         # DSA indexer parameters
@@ -475,7 +506,10 @@ class DeepseekV32Model(DeepseekV2Model):
 @ModelBase.register("DeepseekV4ForCausalLM")
 class DeepseekV4Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK4
+    supports_mtp_export = True
     _skipped_mtp_tensors = 0
+    _dsv4_main_layers: int | None = None
+    _dsv4_nextn_layers: int = 0
 
     def __init__(self, *args, **kwargs):
         type(self)._skipped_mtp_tensors = 0
@@ -487,6 +521,8 @@ class DeepseekV4Model(TextModel):
             self.hparams.setdefault(key, value)
 
         self.block_count = self.hparams["num_hidden_layers"]
+        if self.mtp_only:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
         self._dsv4_fp8_dequantized: set[str] = set()
@@ -499,18 +535,71 @@ class DeepseekV4Model(TextModel):
             logger.info("Skipping %d DeepSeek-V4 MTP tensor(s) for conversion v0", type(self)._skipped_mtp_tensors)
 
         # add a default chat template; if the model has a built-in template, it will be overridden later
-        template_path = Path(__file__).parent.parent / "models" / "templates" / "deepseek-ai-DeepSeek-V4.jinja"
+        model_id_hint = self.remote_hf_model_id or self.dir_model.name
+        is_0731 = "0731" in model_id_hint
+        template_name = "deepseek-ai-DeepSeek-V4-Flash-0731.jinja" if is_0731 else "deepseek-ai-DeepSeek-V4.jinja"
+        template_path = Path(__file__).parent.parent / "models" / "templates" / template_name
         if template_path.is_file():
             with open(template_path, "r", encoding="utf-8") as f:
                 self.gguf_writer.add_chat_template(f.read())
 
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        type(self)._dsv4_main_layers = self.hparams["num_hidden_layers"]
+        type(self)._dsv4_nextn_layers = self.hparams.get("num_nextn_predict_layers", 0)
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, _ = item
+        name, gen = item
         if name.startswith("mtp."):
-            cls._skipped_mtp_tensors += 1
-            return None
-        return super().filter_tensors(item)
+            if not cls.mtp_only:
+                cls._skipped_mtp_tensors += 1
+                return None
+
+            assert cls._dsv4_main_layers is not None
+            parts = name.split(".", 2)
+            if len(parts) < 3 or not parts[1].isdecimal():
+                raise ValueError(f"Unexpected DeepSeek-V4 MTP tensor {name!r}")
+
+            mtp_idx = int(parts[1])
+            if mtp_idx >= cls._dsv4_nextn_layers:
+                raise ValueError(f"Unexpected DeepSeek-V4 MTP layer {mtp_idx}")
+
+            bid = cls._dsv4_main_layers + mtp_idx
+            suffix = parts[2]
+            root_hc_head = {
+                "hc_head_fn",
+                "hc_head_base",
+                "hc_head_scale",
+            }
+            if suffix in root_hc_head:
+                name = suffix
+            elif suffix in (
+                "e_proj.weight", "e_proj.scale",
+                "h_proj.weight", "h_proj.scale",
+            ):
+                name = f"layers.{bid}.nextn.{suffix}"
+            elif suffix == "enorm.weight":
+                name = f"layers.{bid}.nextn.enorm.weight"
+            elif suffix == "hnorm.weight":
+                name = f"layers.{bid}.nextn.hnorm.weight"
+            elif suffix == "norm.weight":
+                name = f"layers.{bid}.nextn.shared_head_norm.weight"
+            else:
+                name = f"layers.{bid}.{suffix}"
+            return name, gen
+
+        if cls.mtp_only:
+            keep = name in (
+                "embed.weight",
+                "norm.weight",
+                "head.weight",
+                "head.scale",
+            )
+            if not keep:
+                return None
+
+        return super().filter_tensors((name, gen))
 
     @staticmethod
     def _float8_dtypes() -> tuple[torch.dtype, ...]:
@@ -565,6 +654,10 @@ class DeepseekV4Model(TextModel):
         self.gguf_writer.add_hyper_connection_sinkhorn_iterations(hparams["hc_sinkhorn_iters"])
         self.gguf_writer.add_hyper_connection_epsilon(hparams["hc_eps"])
         self.gguf_writer.add_hash_layer_count(hparams["num_hash_layers"])
+        if self.model_arch == gguf.MODEL_ARCH.DEEPSEEK4:
+            self.gguf_writer.add_embedding_length_out(hparams["hidden_size"] * hparams["hc_mult"])
+        if self.mtp_only and (num_nextn_predict_layers := hparams.get("num_nextn_predict_layers", 0)) > 0:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
     def dequant_model(self):
         fp8_dtypes = self._float8_dtypes()
@@ -669,12 +762,37 @@ class DeepseekV4Model(TextModel):
         if self._dsv4_mxfp4_generated:
             return ()
 
-        consumed: list[str] = self._write_hash_routing_tensors()
+        consumed: list[str] = []
+        main_layers = self.hparams["num_hidden_layers"]
+        if not self.mtp_only:
+            consumed.extend(self._write_hash_routing_tensors())
+        elif self.hparams["num_hash_layers"] > 0:
+            for bid in range(self.hparams["num_hash_layers"]):
+                name = f"layers.{bid}.ffn.gate.tid2eid"
+                if name in self.model_tensors:
+                    consumed.extend(self._write_hash_routing_tensors())
+                    break
+
         for bid in range(self.block_count):
+            if self.mtp_only and bid < main_layers:
+                continue
             consumed.extend(self._write_mxfp4_expert_tensor(bid, "w1", gguf.MODEL_TENSOR.FFN_GATE_EXP))
             consumed.extend(self._write_mxfp4_expert_tensor(bid, "w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP))
             consumed.extend(self._write_mxfp4_expert_tensor(bid, "w3", gguf.MODEL_TENSOR.FFN_UP_EXP))
 
+        for bid in range(main_layers, self.block_count):
+            e_name = f"layers.{bid}.nextn.e_proj.weight"
+            h_name = f"layers.{bid}.nextn.h_proj.weight"
+            if e_name not in self.model_tensors and h_name not in self.model_tensors:
+                continue
+            if e_name not in self.model_tensors or h_name not in self.model_tensors:
+                raise KeyError(f"Missing DeepSeek-V4 MTP e/h projection pair for block {bid}")
+
+            e_proj = LazyTorchTensor.to_eager(self.model_tensors[e_name]())
+            h_proj = LazyTorchTensor.to_eager(self.model_tensors[h_name]())
+            yield (f"layers.{bid}.nextn.eh_proj.weight", torch.cat((e_proj, h_proj), dim=1).contiguous())
+            consumed.extend((e_name, h_name))
+
         for name in consumed:
             del self.model_tensors[name]
 
@@ -737,6 +855,12 @@ class DeepseekV4Model(TextModel):
             "ffn.shared_experts.w1.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
             "ffn.shared_experts.w2.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
             "ffn.shared_experts.w3.weight": (gguf.MODEL_TENSOR.FFN_UP_SHEXP, ".weight"),
+            "nextn.eh_proj.weight": (gguf.MODEL_TENSOR.NEXTN_EH_PROJ, ".weight"),
+            "nextn.enorm.weight": (gguf.MODEL_TENSOR.NEXTN_ENORM, ".weight"),
+            "nextn.hnorm.weight": (gguf.MODEL_TENSOR.NEXTN_HNORM, ".weight"),
+            "nextn.shared_head_norm.weight": (gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM, ".weight"),
+            "nextn.embed_tokens.weight": (gguf.MODEL_TENSOR.NEXTN_EMBED_TOKENS, ".weight"),
+            "nextn.shared_head_head.weight": (gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD, ".weight"),
         }
 
         tensor_name = match.group(2)
@@ -759,10 +883,12 @@ class DeepseekV4Model(TextModel):
         return [(self._format_dsv4_tensor_name(tensor_key, bid, suffix), data_torch)]
 
     def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        del new_name, bid  # unused
+        del bid  # unused
 
         if name in self._dsv4_fp8_dequantized and n_dims >= 2:
             return gguf.GGMLQuantizationType.Q8_0
+        if new_name.endswith(".nextn.eh_proj.weight"):
+            return gguf.GGMLQuantizationType.Q8_0
         if name in self._dsv4_f32_tensors:
             return gguf.GGMLQuantizationType.F32
         if name in self._dsv4_bf16_tensors and n_dims >= 2:
@@ -770,7 +896,122 @@ class DeepseekV4Model(TextModel):
 
         return False
 
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
     def prepare_tensors(self):
         super().prepare_tensors()
         self._is_mxfp4 = True
         self.ftype = gguf.LlamaFileType.MOSTLY_MXFP4_MOE
+
+
+@ModelBase.register("DeepseekV4DSparkModel")
+class DeepseekV4DSparkModel(DeepseekV4Model):
+    model_arch = gguf.MODEL_ARCH.DFLASH
+
+    _DSPARK_ROOT_MAP: dict[str, tuple[gguf.MODEL_TENSOR, str]] = {
+        "main_proj.weight": (gguf.MODEL_TENSOR.FC, ".weight"),
+        "main_norm.weight": (gguf.MODEL_TENSOR.ENC_OUTPUT_NORM, ".weight"),
+        "markov_head.markov_w1.weight": (gguf.MODEL_TENSOR.DSPARK_MARKOV_W1, ".weight"),
+        "markov_head.markov_w2.weight": (gguf.MODEL_TENSOR.DSPARK_MARKOV_W2, ".weight"),
+        "confidence_head.proj.weight": (gguf.MODEL_TENSOR.DSPARK_CONF_PROJ, ".weight"),
+    }
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        self.block_count = 1 + max(
+            int(match.group(1)) for name in self.model_tensors
+            if (match := re.match(r"layers\.(\d+)\.", name))
+        )
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        self.hparams["compress_ratios"] = [0] * self.block_count
+        self.hparams["num_hash_layers"] = 0
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        if remote_hf_model_id is None:
+            return super().index_tensors()
+
+        with open(self.dir_model / "model.safetensors.index.json", "r", encoding="utf-8") as f:
+            weight_map = json.load(f)["weight_map"]
+
+        part_names = sorted({
+            part_name for name, part_name in weight_map.items()
+            if name.startswith("mtp.")
+        })
+        tensors: dict[str, Callable[[], Tensor]] = {}
+
+        for part_name in part_names:
+            from huggingface_hub import hf_hub_download
+
+            logger.info("gguf: caching remote DSpark part '%s'", part_name)
+            part_path = Path(hf_hub_download(repo_id=remote_hf_model_id, filename=part_name))
+            with gguf.utility.SafetensorsLocal(part_path) as model_part:
+                for name in model_part:
+                    data = model_part[name]
+                    data_gen = lambda data=data: LazyTorchTensor.from_local_tensor(data)  # noqa: E731
+                    if titem := self.filter_tensors((name, data_gen)):
+                        tensor_name, tensor_gen = titem
+                        tensors[tensor_name] = tensor_gen
+
+        return tensors
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if not name.startswith("mtp."):
+            return None
+        return super().filter_tensors((cls._rekey_mtp_tensor_name(name), gen))
+
+    @staticmethod
+    def _rekey_mtp_tensor_name(name: str) -> str:
+        match = re.match(r"mtp\.(\d+)\.(.+)$", name)
+        if match is None:
+            raise ValueError(f"Unexpected DSpark tensor {name!r}")
+
+        stage, rest = match.group(1), match.group(2)
+        root_names = (
+            "main_proj.scale",
+            "norm.weight",
+            "hc_head_fn",
+            "hc_head_base",
+            "hc_head_scale",
+        )
+        if rest in DeepseekV4DSparkModel._DSPARK_ROOT_MAP or rest in root_names:
+            return rest
+        return f"layers.{stage}.{rest}"
+
+    def _map_dsv4_tensor_name(self, name: str, bid: int | None) -> tuple[gguf.MODEL_TENSOR, str]:
+        if name in self._DSPARK_ROOT_MAP:
+            return self._DSPARK_ROOT_MAP[name]
+        return super()._map_dsv4_tensor_name(name, bid)
+
+    def set_vocab(self):
+        if self.target_model_dir is None:
+            raise ValueError("DeepSeek-V4 DSpark requires --target-model-dir with the target tokenizer")
+
+        original_dir = self.dir_model
+        try:
+            self.dir_model = self.target_model_dir
+            super().set_vocab()
+        finally:
+            self.dir_model = original_dir
+
+        self.gguf_writer.add_mask_token_id(self.hparams["dspark_noise_token_id"])
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_block_size(self.hparams["dspark_block_size"])
+        self.gguf_writer.add_target_layers([layer + 1 for layer in self.hparams["dspark_target_layer_ids"]])

conversion/glm.py

@@ -206,10 +206,70 @@ class Glm4MoeModel(TextModel):
 @ModelBase.register("Glm4MoeLiteForCausalLM")
 class Glm4MoeLiteModel(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+    skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     def set_vocab(self):
         return self._set_vocab_glm()
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        num_hidden_layers = self.hparams["num_hidden_layers"]
+        self.num_nextn_predict_layers = self.hparams.get("num_nextn_predict_layers", 0)
+        self.skip_mtp = self.no_mtp or self.num_nextn_predict_layers == 0
+
+        if self.skip_mtp:
+            self.block_count = num_hidden_layers
+        else:
+            self.block_count = num_hidden_layers + self.num_nextn_predict_layers
+
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        if self.skip_mtp:
+            return
+
+        self.gguf_writer.add_nextn_predict_layers(self.num_nextn_predict_layers)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item):
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        if cls._n_main_layers is not None:
+            match = re.match(r"model\.layers\.(\d+)\.", name)
+            is_mtp = match is not None and int(match.group(1)) >= cls._n_main_layers
+            if is_mtp and cls.no_mtp:
+                return None
+            if cls.mtp_only and not is_mtp and name not in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+            ):
+                return None
+
+        return name, gen
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
 
 @ModelBase.register("GlmMoeDsaForCausalLM")
 class GlmMoeDsaModel(DeepseekV2Model):

conversion/minicpm.py

@@ -137,6 +137,15 @@ class MiniCPMV4_6TextModel(Qwen3_5TextModel):
 class MiniCPMV4_6VisionModel(MmprojModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
+        self.downsample_mode = self.preprocessor_config.get("downsample_mode", "16x")
+        if self.downsample_mode not in {"4x", "16x"}:
+            raise ValueError(f"Unsupported downsample mode: {self.downsample_mode}")
+        if self.downsample_mode == "4x":
+            self.model_tensors = {
+                name: tensor for name, tensor in self.model_tensors.items()
+                if ".vit_merger." not in name
+            }
+
         if self.hparams_vision is not None:
             # In MiniCPM-V 4.6 `vision_config.image_size` (980) describes the SigLIP
             # positional embedding bucket grid (70 x 70), while the per-slice processing
@@ -156,8 +165,8 @@ class MiniCPMV4_6VisionModel(MmprojModel):
         # (mapped to PROJECTOR_TYPE_MINICPMV4_6).
         self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MINICPMV4_6)
 
-        # ViT merger 2x2 + final merger 2x2 = 4x spatial merge per dimension; used for slice alignment
-        self.gguf_writer.add_vision_projector_scale_factor(4)
+        self.gguf_writer.add_vision_projector_scale_factor(
+            2 if self.downsample_mode == "4x" else 4)
 
         # borrow wa_layer_indexes for vit_merger insertion point
         insert_layer_id = int(self.global_config.get(

conversion/qwen.py

@@ -268,8 +268,101 @@ class Qwen3MoeModel(Qwen2MoeModel):
         super().set_vocab()
 
 
+class _QwenMtpMixin:
+    """Shared MTP wiring for Qwen3-Next and Qwen3.5/3.6 text variants. The HF
+    config carries the MTP block under `mtp_num_hidden_layers` (computed from
+    the checkpoint when absent, e.g. Qwen3-Next) and the tensors under
+    `mtp.*`; we extend block_count, emit the nextn metadata key, and remap
+    `mtp.*` to the standard layer-indexed nextn naming so the existing
+    tensor_map handles them."""
+
+    supports_mtp_export = True
+    hparams: dict[str, Any]
+    model_arch: gguf.MODEL_ARCH
+    gguf_writer: gguf.GGUFWriter
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+    no_mtp: bool
+    mtp_only: bool
+    _original_block_count: int | None = None
+    opt_num_mtp_layers: int = 0
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            n_mtp = self.hparams.get("mtp_num_hidden_layers", 0)
+            # Qwen-3-Next doesn't include `mtp_num_hidden_layers` in config.
+            if n_mtp == 0:
+                assert self.opt_num_mtp_layers != 0
+                n_mtp = self.opt_num_mtp_layers
+            self.block_count += n_mtp
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
+        type(self)._original_block_count = hparams.get(key)
+        type(self).opt_num_mtp_layers = 0
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
+
+    @classmethod
+    def filter_tensors(cls, item):
+        assert cls._original_block_count is not None
+        # TODO: change TextModel to super()
+        if (titem := TextModel.filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+        if name.startswith("model.mtp."):
+            name = name.replace("model.", "", 1)
+        if name.startswith("mtp."):
+            if cls.no_mtp:
+                return None
+            remapper = {
+                "fc":                    "eh_proj",
+                "pre_fc_norm_embedding": "enorm",
+                "pre_fc_norm_hidden":    "hnorm",
+                "norm":                  "shared_head.norm",
+            }
+            parts = name.split(".", 3)
+            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
+                mtp_idx = int(parts[2])
+                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
+                cls.opt_num_mtp_layers = max(cls.opt_num_mtp_layers, mtp_idx + 1)
+            elif len(parts) == 3 and parts[1] in remapper:
+                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
+        elif cls.mtp_only:
+            keep = name in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+                "embed_tokens.weight", "norm.weight",
+            )
+            if not keep:
+                return None
+        return name, gen
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
+        if self.no_mtp:
+            return
+        if (n := self.block_count - self.hparams["num_hidden_layers"]) > 0:
+            self.gguf_writer.add_nextn_predict_layers(n)
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
+
 @ModelBase.register("Qwen3NextForCausalLM")
-class Qwen3NextModel(Qwen2MoeModel):
+class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.QWEN3NEXT
 
     def set_gguf_parameters(self):
@@ -284,16 +377,6 @@ class Qwen3NextModel(Qwen2MoeModel):
             rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.25)))
 
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
-
-        if name.startswith("mtp"):
-            # ignore MTP layers for now
-            return None
-
-        return super().filter_tensors(item)
-
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name.endswith(".A_log"):
             data_torch = -torch.exp(data_torch)
@@ -536,97 +619,13 @@ class _Qwen35MRopeMixin:
             self.gguf_writer.add_rope_dimension_sections(self._QWEN35_DEFAULT_MROPE_SECTION)
 
 
-class _Qwen35MtpMixin:
-    """Shared MTP wiring for Qwen3.5/3.6 text variants. The HF config carries
-    the MTP block under `mtp_num_hidden_layers` and the tensors under
-    `mtp.*`; we extend block_count, emit the nextn metadata key, and remap
-    `mtp.*` to the standard layer-indexed nextn naming so the existing
-    tensor_map handles them."""
-
-    supports_mtp_export = True
-    hparams: dict[str, Any]
-    model_arch: gguf.MODEL_ARCH
-    gguf_writer: gguf.GGUFWriter
-    block_count: int
-    tensor_map: gguf.TensorNameMap
-    no_mtp: bool
-    mtp_only: bool
-    _original_block_count: int | None = None
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"]
-        if not self.no_mtp:
-            self.block_count += self.hparams.get("mtp_num_hidden_layers", 0)
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
-        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
-        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
-        type(self)._original_block_count = hparams.get(key)
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
-
-    @classmethod
-    def filter_tensors(cls, item):
-        assert cls._original_block_count is not None
-        # TODO: change TextModel to super()
-        if (titem := TextModel.filter_tensors(item)) is None:
-            return None
-        name, gen = titem
-        if name.startswith("model.mtp."):
-            name = name.replace("model.", "", 1)
-        if name.startswith("mtp."):
-            if cls.no_mtp:
-                return None
-            remapper = {
-                "fc":                    "eh_proj",
-                "pre_fc_norm_embedding": "enorm",
-                "pre_fc_norm_hidden":    "hnorm",
-                "norm":                  "shared_head.norm",
-            }
-            parts = name.split(".", 3)
-            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
-                mtp_idx = int(parts[2])
-                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
-            elif len(parts) == 3 and parts[1] in remapper:
-                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
-        elif cls.mtp_only:
-            keep = name in (
-                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
-                "embed_tokens.weight", "norm.weight",
-            )
-            if not keep:
-                return None
-        return name, gen
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
-        if self.no_mtp:
-            return
-        if (n := self.hparams.get("mtp_num_hidden_layers", 0)) > 0:
-            self.gguf_writer.add_nextn_predict_layers(n)
-
-    def prepare_metadata(self, vocab_only: bool):
-        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
-
-        if not self.mtp_only or not from_dir:
-            return
-
-        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
-
-
 @ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM")
-class Qwen3_5TextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):
+class Qwen3_5TextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35
 
 
 @ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM")
-class Qwen3_5MoeTextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):
+class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35MOE
 
 

convert_hf_to_gguf.py

@@ -122,8 +122,12 @@ def parse_args() -> argparse.Namespace:
         help="Export only the multi-token prediction (MTP) head as a separate GGUF, suitable for use as a speculative draft. An 'mtp-' prefix will be added to the output file name.",
     )
     parser.add_argument(
-        "--no-mtp", action="store_true",
-        help="Exclude the multi-token prediction (MTP) head from the converted GGUF. Pair with --mtp on a second run to publish trunk and MTP as two files. Note: the split form duplicates embeddings, but even though the bundled default is more space-efficient overall, this allows differing quantization which may be more performant.",
+        "--no-nextn", "--no-mtp", dest="no_mtp", action="store_true",
+        help="Exclude NextN speculative draft tensors from the converted GGUF. Pair with --mtp or --dspark on a second run to publish target and draft as two files.",
+    )
+    parser.add_argument(
+        "--dspark", action="store_true",
+        help="Export only the DeepSeek-V4 DSpark draft tensors as a separate GGUF.",
     )
     parser.add_argument(
         "--mistral-format", action="store_true",
@@ -254,13 +258,20 @@ def main() -> None:
             from conversion.mistral import MistralModel
             model_class = MistralModel
 
-        if args.mtp and args.no_mtp:
-            logger.error("--mtp and --no-mtp are mutually exclusive")
+        if sum((args.mtp, args.no_mtp, args.dspark)) > 1:
+            logger.error("--mtp, --no-nextn, and --dspark are mutually exclusive")
             sys.exit(1)
 
+        if args.dspark:
+            if is_mistral_format or model_architecture != "DeepseekV4ForCausalLM":
+                logger.error("--dspark is only supported for DeepseekV4ForCausalLM")
+                sys.exit(1)
+            from conversion.deepseek import DeepseekV4DSparkModel
+            model_class = DeepseekV4DSparkModel
+
         if args.mtp or args.no_mtp:
             if not model_class.supports_mtp_export:
-                logger.error("--mtp / --no-mtp are not supported for %s", model_architecture)
+                logger.error("--mtp / --no-nextn are not supported for %s", model_architecture)
                 sys.exit(1)
             if args.no_mtp:
                 model_class.no_mtp = True

docs/backend/SYCL.md

@@ -788,7 +788,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
-| GGML_SYCL_DEV2DEV_MEMCPY | 0 (default) or 1 | Choose the SYCL or L0 API in dev2dev memory copy.<br>Value: <br>*  0: SYCL API (default)<br>* 1: L0 API -- L0 API is found to lead to abnormal crash in some case. This debug flag is used to check the issue.|
+| GGML_SYCL_DEV2DEV_MEMCPY | 0 (default), 1, 2 | Choose the method of dev2dev memory copy.<br>Value: <br>*  0: SYCL API (default), only support dGPUs.<br>* 1: L0 API -- Better performance, only support dGPUs, found to lead to abnormal crash in some case. <br>* 2: Host Forward -- Most stable method for all cases (including iGPU + dGPU*N), but with lower performance (-2% to -5%).<br>SYCL & L0 API are easy to be impacted by Intel GPU driver issue. When you meet the garbled output or crash issues in multiple GPUs case, try with this debug flag to work around or check the issue.|
 | GGML_SYCL_ENABLE_FLASH_ATTN | 1 (default) or 0| Enable Flash-Attention. It can reduce memory usage. The performance impact depends on the LLM.|
 | GGML_SYCL_ENABLE_OPT | 0 or 1 (default)| Enable optimize features for Intel GPUs. (Recommended to 0 for Intel devices older than Gen 10) |
 | GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
@@ -797,6 +797,9 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_FA_ONEDNN | 1 (default) or 0 | Enable the oneDNN fused SDPA (flash-attention) path on supported GPUs. Set to 0 to always use the native SYCL flash-attention kernel. |
 | GGML_SYCL_FA_ONEDNN_MAX_KV | 0 (default, disabled) or positive integer | By default (0), all sequences are handled by the oneDNN fused SDPA path, regardless of KV length; a positive value caps that length, past which sequences fall back to the native kernel. If GPU driver watchdog resets (DEVICE_LOST) occur during long-context inference, set this near the context depth where they start, e.g. 24576. |
 | GGML_SYCL_ENABLE_VMM | 0 or 1 (default) | Enable the virtual-memory device pool. |
+| GGML_SYCL_ENABLE_MKL_FA | 1 (default) or 0 | Enable oneMKL GEMM flash attention for XMX-accelerated prompt processing with quantized KV cache. Automatically activates during prefill (prompt processing) when all conditions are met: (1) flash-attn enabled (`-fa` or `--flash-attn on`), (2) KV cache quantized (`--cache-type-k q8_0 --cache-type-v q8_0` or other `*_0/*_1` types), (3) batch size ≥ 1024 (`--batch-size 1024`), (4) prompt length ≥ 1024 tokens. Set to 0 to force the TILE kernel for A/B testing. Example minimum command: `llama-cli -m model.gguf -fa -ngl 99 --cache-type-k q8_0 --cache-type-v q8_0 --batch-size 1024 -p "your prompt"` |
+| GGML_SYCL_MKL_FA_DEBUG | 0 (default) or 1 | Enable per-call diagnostic logging for MKL flash attention: GEMM/softmax timings, interleaved-head detection, and buffer memory usage. |
+| GGML_SYCL_MKL_FA_DIAG | 0 (default) or 1 | Enable output fingerprinting for MKL flash attention. Dumps the first 64 float output values for the first 6 FA calls with n_kv ≥ 1024, labeled with kernel type (MKL/TILE/VEC) for cross-kernel comparison. |
 | GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute (currently top-k MoE gating). |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |

examples/gen-docs/gen-docs.cpp

@@ -70,6 +70,8 @@ static void write_table(std::ostringstream & ss, std::vector<common_arg *> & opt
 
 static void write_help(std::ostringstream & ss, const md_file & md) {
     common_params params;
+    params.is_gen_docs = true;
+
     auto ctx_arg = common_params_parser_init(params, md.ex);
 
     std::vector<common_arg *> common_options;

ggml/CMakeLists.txt

@@ -341,9 +341,6 @@ set(GGML_PUBLIC_HEADERS
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
-#if (GGML_METAL)
-#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
-#endif()
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)
 

ggml/src/ggml-backend.cpp

@@ -765,8 +765,9 @@ struct ggml_backend_sched_split {
     int backend_id;
     int i_start;
     int i_end;
-    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    struct ggml_tensor ** inputs;
     int n_inputs;
+    int inputs_capacity;
     // graph view of this split
     struct ggml_cgraph graph;
 };
@@ -805,8 +806,9 @@ struct ggml_backend_sched {
     int cur_copy;
     int next_copy;
     ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
-    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    struct ggml_tensor ** graph_inputs;
     int n_graph_inputs;
+    int graph_inputs_capacity;
 
     struct ggml_context * ctx;
 
@@ -832,6 +834,36 @@ struct ggml_backend_sched {
 #define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
 #define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)
 
+static void ggml_backend_sched_split_inputs_grow(struct ggml_backend_sched_split * split) {
+    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
+    if (split->inputs_capacity > 0) {
+        new_cap = 2*split->inputs_capacity;
+        GGML_LOG_WARN("%s: increasing split inputs capacity from %d to %d\n", __func__, split->inputs_capacity, new_cap);
+    }
+    auto * pnew = (struct ggml_tensor **) realloc((void *) split->inputs, new_cap * sizeof(struct ggml_tensor *));
+    if (pnew == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %zu bytes\n", __func__, new_cap * sizeof(struct ggml_tensor *));
+        GGML_ABORT("failed to grow split inputs container");
+    }
+    split->inputs = pnew;
+    split->inputs_capacity = new_cap;
+}
+
+static void ggml_backend_sched_graph_inputs_grow(ggml_backend_sched_t sched) {
+    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
+    if (sched->graph_inputs_capacity > 0) {
+        new_cap = 2*sched->graph_inputs_capacity;
+        GGML_LOG_WARN("%s: increasing graph inputs capacity from %d to %d\n", __func__, sched->graph_inputs_capacity, new_cap);
+    }
+    auto * pnew = (struct ggml_tensor **) realloc((void *) sched->graph_inputs, new_cap * sizeof(struct ggml_tensor *));
+    if (pnew == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %zu bytes\n", __func__, new_cap * sizeof(struct ggml_tensor *));
+        GGML_ABORT("failed to grow graph inputs container");
+    }
+    sched->graph_inputs = pnew;
+    sched->graph_inputs_capacity = new_cap;
+}
+
 // returns the priority of the backend, lower id is higher priority
 static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
     for (int i = 0; i < sched->n_backends; i++) {
@@ -1297,7 +1329,7 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                     }
                     // check if the split has too many inputs
                     // FIXME: count the number of inputs instead of only checking when full
-                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
+                    if (split->n_inputs >= split->inputs_capacity) {
                         const size_t id = hash_id(src);
                         int src_backend_id = sched->hv_tensor_backend_ids[id];
                         bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
@@ -1313,10 +1345,14 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                 split->i_end = i;
                 i_split++;
                 if (i_split >= sched->splits_capacity) {
+                    int old_cap = sched->splits_capacity;
                     sched->splits_capacity *= 2;
                     sched->splits = (ggml_backend_sched_split *)
                         realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
                     GGML_ASSERT(sched->splits != NULL);
+                    for (int k = old_cap; k < sched->splits_capacity; k++) {
+                        memset(&sched->splits[k], 0, sizeof(struct ggml_backend_sched_split));
+                    }
                 }
                 split = &sched->splits[i_split];
                 split->backend_id = node_backend_id;
@@ -1353,7 +1389,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_graph_inputs = sched->n_graph_inputs++;
-                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        if (n_graph_inputs >= sched->graph_inputs_capacity) {
+                            ggml_backend_sched_graph_inputs_grow(sched);
+                        }
                         sched->graph_inputs[n_graph_inputs] = src;
                     }
                 }
@@ -1373,7 +1411,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_inputs = split->n_inputs++;
-                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        if (n_inputs >= split->inputs_capacity) {
+                            ggml_backend_sched_split_inputs_grow(split);
+                        }
                         split->inputs[n_inputs] = src;
                     }
                     node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
@@ -1399,7 +1439,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         sched->prev_leaf_backend_ids = tmp;
     }
 
-    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
+    int total_inputs = sched->n_graph_inputs;
+    for (int i = 0; i < sched->n_splits; i++) {
+        total_inputs += sched->splits[i].n_inputs;
+    }
+    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + total_inputs * 2 * sched->n_copies;
 
     // remember the actual graph_size for performing reallocation checks later [GGML_SCHED_DEBUG_REALLOC]
     sched->debug_prev_graph_size = sched->debug_graph_size;
@@ -1782,6 +1826,9 @@ ggml_backend_sched_t ggml_backend_sched_new(
     sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0]));
     sched->splits_capacity = initial_splits_capacity;
 
+    sched->graph_inputs_capacity = GGML_SCHED_MAX_SPLIT_INPUTS;
+    sched->graph_inputs = (struct ggml_tensor **) calloc(sched->graph_inputs_capacity, sizeof(struct ggml_tensor *));
+
     for (int b = 0; b < n_backends; b++) {
         sched->backends[b] = backends[b];
         sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
@@ -1814,7 +1861,11 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
     ggml_hash_set_free(&sched->hash_set);
+    for (int i = 0; i < sched->splits_capacity; i++) {
+        free(sched->splits[i].inputs);
+    }
     free(sched->splits);
+    free(sched->graph_inputs);
     free(sched->hv_tensor_backend_ids);
     free(sched->hv_tensor_copies);
     free(sched->node_backend_ids);

ggml/src/ggml-cuda/common.cuh

@@ -627,7 +627,8 @@ template <typename T> struct block_reduce_policy<block_reduce_method::MAX, T> {
 };
 
 template <block_reduce_method reduce_method_t, const unsigned int block_size_template = 0, typename T>
-static __device__ T block_reduce(T val, T * shared_vals) {
+static __device__ T block_reduce(T val, [[maybe_unused]] T * shared_vals) {
+    // for multi-warp reductions, callers must not reuse shared_vals until all reads from this invocation have completed
     val                           = block_reduce_policy<reduce_method_t, T>::reduce(val);
     const unsigned int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
     if (block_size > WARP_SIZE) {

ggml/src/ggml-cuda/mmq-load-tiles.cuh

@@ -130,43 +130,27 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
         }
 
         const block_q2_0 * bxi = (const block_q2_0 *) x + kbx0 + i*stride + kbx;
-        // Each 32-element chunk occupies 8 bytes of qs (32 elements * 2 bits = 64 bits)
-        const int qs_offset = 8*kqsx;
-        const int qs0 = bxi->qs[qs_offset + 0] | (bxi->qs[qs_offset + 1] << 8) |
-                        (bxi->qs[qs_offset + 2] << 16) | (bxi->qs[qs_offset + 3] << 24);
-        const int qs1 = bxi->qs[qs_offset + 4] | (bxi->qs[qs_offset + 5] << 8) |
-                        (bxi->qs[qs_offset + 6] << 16) | (bxi->qs[qs_offset + 7] << 24);
-
-        // Unpack 32 2-bit codes into 8 int32s, each holding 4 signed int8s in {-1,0,1,2}.
-        int unpacked_bytes[8];
-#pragma unroll
-        for (int j = 0; j < 4; ++j) {
-            const int shift = j * 8;
-            const int codes = (qs0 >> shift) & 0xFF;
-            const int c0 = ((codes >> 0) & 0x3) - 1;
-            const int c1 = ((codes >> 2) & 0x3) - 1;
-            const int c2 = ((codes >> 4) & 0x3) - 1;
-            const int c3 = ((codes >> 6) & 0x3) - 1;
-            unpacked_bytes[j] = (c0 & 0xFF) | ((c1 & 0xFF) << 8) | ((c2 & 0xFF) << 16) | ((c3 & 0xFF) << 24);
-        }
-#pragma unroll
-        for (int j = 0; j < 4; ++j) {
-            const int shift = j * 8;
-            const int codes = (qs1 >> shift) & 0xFF;
-            const int c0 = ((codes >> 0) & 0x3) - 1;
-            const int c1 = ((codes >> 2) & 0x3) - 1;
-            const int c2 = ((codes >> 4) & 0x3) - 1;
-            const int c3 = ((codes >> 6) & 0x3) - 1;
-            unpacked_bytes[4 + j] = (c0 & 0xFF) | ((c1 & 0xFF) << 8) | ((c2 & 0xFF) << 16) | ((c3 & 0xFF) << 24);
-        }
+        const int16_t    * qxi = (const int16_t *) bxi->qs + kqsx * 4;
 
         const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0;
+
 #pragma unroll
-        for (int j = 0; j < 8; ++j) {
+        for (int j = 0; j < 4; ++j) {
+            const int q  = qxi[j];
+
+            // unpack even and odd crumbs into byte values
+            const int qe = __byte_perm(0x020100FF, 0x020100FF, q >> 0);
+            const int qo = __byte_perm(0x020100FF, 0x020100FF, q >> 2);
+            // unshuffle values
+            const int qx = __byte_perm(qe, qo, 0x5140);
+            const int qy = __byte_perm(qe, qo, 0x7362);
+
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
-            x_qs[i*sram_stride           + dst_offset + j] = unpacked_bytes[j];
+            x_qs[i*sram_stride           + dst_offset + j*2+0] = qx;
+            x_qs[i*sram_stride           + dst_offset + j*2+1] = qy;
 #else
-            x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j] = unpacked_bytes[j];
+            x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*2+0] = qx;
+            x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*2+1] = qy;
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
     }

ggml/src/ggml-cuda/norm.cu

@@ -64,7 +64,7 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
         tmp += xi * xi;
     }
 
-    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum);
+    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum + 32);
 
     const float variance = tmp / group_size;
     const float scale = rsqrtf(variance + eps);
@@ -297,7 +297,7 @@ static void group_norm_f32_cuda(
         group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
     } else {
         const dim3 block_dims(1024, 1, 1);
-        group_norm_f32<1024><<<num_groups, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(x, dst, group_size, ne_elements, eps);
+        group_norm_f32<1024><<<num_groups, block_dims, block_dims.x > WARP_SIZE ? 2 * 32 * sizeof(float): 0, stream>>>(x, dst, group_size, ne_elements, eps);
     }
 }
 

ggml/src/ggml-cuda/softmax.cu

@@ -116,6 +116,11 @@ static __global__ void soft_max_f32(
         vals[col] = val;
     }
 
+    if (block_size > WARP_SIZE) {
+        // sync is needed as we reuse buf_iw across block_reduce invocations, see #26385
+        // for block_size <= WARP_SIZE, block_reduce does not access buf_iw
+        __syncthreads();
+    }
     // find the sum of exps in the block
     tmp = block_reduce<block_reduce_method::SUM, block_size_template>(tmp, buf_iw);
 
@@ -142,6 +147,8 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
                                                                 float * __restrict__ dst,
                                                                 float * __restrict__ tmp_maxs,
                                                                 float * __restrict__ tmp_sums,
+                                                                float * shared_vals_max,
+                                                                float * shared_vals_sum,
                                                                 const soft_max_params p) {
     namespace cg = cooperative_groups;
 
@@ -154,7 +161,6 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     float     local_vals[n_elem_per_thread] = { -INFINITY, -INFINITY, -INFINITY, -INFINITY };
     float     local_max                     = -INFINITY;
     const int step_size                     = gridDim.x * blockDim.x;
-    __shared__ float shared_vals[32];
 
     // Compute thread-local max
     for (int col = col_start; col < p.ncols;) {
@@ -171,7 +177,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     }
 
     // Compute CTA-level max
-    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals);
+    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals_max);
 
     // Store CTA-level max to GMEM
     if (tid == 0) {
@@ -186,7 +192,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     } else {
         local_max = -INFINITY;
     }
-    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals);
+    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals_max);
 
     // Compute softmax dividends, accumulate divisor
     float tmp_expf = 0.0f;
@@ -209,7 +215,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     }
 
     // Reduce divisor within CTA
-    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals);
+    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals_sum);
 
     // Store CTA-level sum to GMEM
     if (tid == 0) {
@@ -223,7 +229,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     } else {
         tmp_expf = 0.0f;
     }
-    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals);
+    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals_sum);
 
     // Divide dividend by global sum + store data
     for (int col = col_start; col < p.ncols;) {
@@ -310,9 +316,11 @@ __launch_bounds__(8*WARP_SIZE, 1) static __global__ void soft_max_f32_paralleliz
 // https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#grid-synchronization
 // https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#class-cluster-group
 {
+    __shared__ float shared_vals[2][32];
+
     for (int rowx = 0; rowx < p.ne01 * p.ne02 * p.ne03; rowx++) {
         soft_max_f32_parallelize_cols_single_row(x + int64_t(rowx) * p.ncols, dst + int64_t(rowx) * p.ncols, tmp_maxs,
-                                                 tmp_sums, p);
+                                                 tmp_sums, shared_vals[0], shared_vals[1], p);
     }
 }
 

ggml/src/ggml-cuda/vecdotq.cuh

@@ -734,48 +734,28 @@ static __device__ __forceinline__ float vec_dot_q2_0_q8_1(
     // Q8_1: 32 elements per block with individual scales
     // iqs selects which of the 2 chunks of 32 elements to process (0-1)
 
-    const float d2 = bq2_0->d;
+    const float     d2 = bq2_0->d;
+    const int16_t * qs = (const int16_t *) bq2_0->qs + iqs * 4;
 
     // Process only the chunk specified by iqs
     const block_q8_1 * bq8_1_chunk = bq8_1 + iqs;
 
-    // Load 64 bits (8 bytes) for this chunk from Q2_0: bytes [8*iqs, 8*iqs+8)
-    const int offset = iqs * 8;
-    const int v0 = bq2_0->qs[offset + 0] | (bq2_0->qs[offset + 1] << 8) |
-                   (bq2_0->qs[offset + 2] << 16) | (bq2_0->qs[offset + 3] << 24);
-    const int v1 = bq2_0->qs[offset + 4] | (bq2_0->qs[offset + 5] << 8) |
-                   (bq2_0->qs[offset + 6] << 16) | (bq2_0->qs[offset + 7] << 24);
-
-    // Unpack 32 2-bit codes into 8 int32s, each holding 4 signed int8 symbols in {-1,0,1,2}.
-    // Stored code c in {0,1,2,3} -> symbol s = c - 1.
-    int vi_bytes[8];
-#pragma unroll
-    for (int j = 0; j < 4; ++j) {
-        const int shift = j * 8;
-        const int codes = (v0 >> shift) & 0xFF;
-        const int c0 = ((codes >> 0) & 0x3) - 1;
-        const int c1 = ((codes >> 2) & 0x3) - 1;
-        const int c2 = ((codes >> 4) & 0x3) - 1;
-        const int c3 = ((codes >> 6) & 0x3) - 1;
-        vi_bytes[j] = (c0 & 0xFF) | ((c1 & 0xFF) << 8) | ((c2 & 0xFF) << 16) | ((c3 & 0xFF) << 24);
-    }
-#pragma unroll
-    for (int j = 0; j < 4; ++j) {
-        const int shift = j * 8;
-        const int codes = (v1 >> shift) & 0xFF;
-        const int c0 = ((codes >> 0) & 0x3) - 1;
-        const int c1 = ((codes >> 2) & 0x3) - 1;
-        const int c2 = ((codes >> 4) & 0x3) - 1;
-        const int c3 = ((codes >> 6) & 0x3) - 1;
-        vi_bytes[4 + j] = (c0 & 0xFF) | ((c1 & 0xFF) << 8) | ((c2 & 0xFF) << 16) | ((c3 & 0xFF) << 24);
-    }
-
-    // Compute dot product for this 32-element chunk
     int sumi = 0;
 #pragma unroll
-    for (int j = 0; j < 8; ++j) {
-        const int u = get_int_b4(bq8_1_chunk->qs, j);
-        sumi = ggml_cuda_dp4a(vi_bytes[j], u, sumi);
+    for (int j = 0; j < 4; ++j) {
+        const int q  = qs[j];
+        const int u  = get_int_b4(bq8_1_chunk->qs, j*2+0);
+        const int v  = get_int_b4(bq8_1_chunk->qs, j*2+1);
+
+        // unpack even and odd crumbs into byte values
+        const int qe = __byte_perm(0x020100FF, 0x020100FF, q >> 0);
+        const int qo = __byte_perm(0x020100FF, 0x020100FF, q >> 2);
+        // unshuffle values
+        const int qx = __byte_perm(qe, qo, 0x5140);
+        const int qy = __byte_perm(qe, qo, 0x7362);
+
+        sumi = ggml_cuda_dp4a(u, qx, sumi);
+        sumi = ggml_cuda_dp4a(v, qy, sumi);
     }
 
     // Apply Q2_0's single scale and this chunk's Q8_1 scale

ggml/src/ggml-metal/CMakeLists.txt

@@ -24,62 +24,119 @@ if (GGML_METAL_NDEBUG)
 endif()
 
 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
+set(METALLIB_KERNELS_COMMON     "${CMAKE_CURRENT_SOURCE_DIR}/kernels/common.h")
+set(METALLIB_KERNELS_DEQUANTIZE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/dequantize.h")
+set(METALLIB_KERNELS_QUANTIZE   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantize.h")
+
+set(METALLIB_KERNEL_SOURCES
+    kernels/fa.metal
+    kernels/mul_mv.metal
+    kernels/mul_mm.metal
+    kernels/quantize.metal
+    kernels/softmax.metal
+    kernels/norm.metal
+    kernels/unary.metal
+    kernels/binbcast.metal
+    kernels/reduce.metal
+    kernels/tri.metal
+    kernels/ssm.metal
+    kernels/wkv.metal
+    kernels/gated_delta_net.metal
+    kernels/solve_tri.metal
+    kernels/rope.metal
+    kernels/conv.metal
+    kernels/upscale.metal
+    kernels/argsort.metal
+    kernels/pool.metal
+    kernels/misc.metal
+)
+
 if (GGML_METAL_EMBED_LIBRARY)
     enable_language(ASM)
 
     add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
-    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
 
     file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 
-    # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+    set(METALLIB_EMBED_ASM_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(kind ${src} NAME_WE)
+        # symbol names must be valid C identifiers ('-' is not allowed)
+        string(REPLACE "-" "_" kind_sym ${kind})
 
-    add_custom_command(
-        OUTPUT "${METALLIB_EMBED_ASM}"
-        COMMAND echo "Embedding Metal library"
-        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
-        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
-        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
-        COMMENT "Generate assembly for embedded Metal library"
-        VERBATIM
-    )
+        set(SRC   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${kind}.metal")
+        set(EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.metal")
+        set(ASM   "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.s")
 
-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+        # only prepend headers that this source actually includes
+        set(HEADERS_FOR_SRC ${METALLIB_KERNELS_COMMON})
+        file(STRINGS ${SRC} _has_dequantize REGEX "#include \"dequantize\\.h\"")
+        file(STRINGS ${SRC} _has_quantize   REGEX "#include \"quantize\\.h\"")
+        if(_has_dequantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_DEQUANTIZE})
+        endif()
+        if(_has_quantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_QUANTIZE})
+        endif()
+
+        add_custom_command(
+            OUTPUT "${ASM}"
+            # Step 1: concatenate shared headers + this kernel source
+            COMMAND cat ${HEADERS_FOR_SRC} ${SRC} > "${EMBED}.tmp1"
+            # Step 2: remove internal #include and #pragma once
+            COMMAND sed -e "/\#include \"common.h\"/d" -e "/\#include \"dequantize.h\"/d" -e "/\#include \"quantize.h\"/d" -e "/\#pragma once/d" < "${EMBED}.tmp1" > "${EMBED}.tmp2"
+            # Step 3: inline ggml-common.h (replacing __embed_ggml-common.h__ sentinel)
+            COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}" -e "/__embed_ggml-common.h__/d" < "${EMBED}.tmp2" > "${EMBED}.tmp3"
+            # Step 4: inline ggml-metal-impl.h
+            COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${EMBED}.tmp3" > "${EMBED}"
+            # Step 5: emit an asm chunk with kind-specific start/end symbols
+            #   note: '-' is illegal in C symbols, so we use kind_sym; the macOS
+            #   section name is limited to 16 chars so we keep it shared
+            #   across kinds (__ggml_metallib) and only vary the global symbols.
+            COMMAND echo ".section __DATA,__ggml_metallib"                       >  "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_start"               >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_start:"                     >> "${ASM}"
+            COMMAND echo .incbin "\"${EMBED}\""                                  >> "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_end"                 >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_end:"                       >> "${ASM}"
+            DEPENDS ../ggml-common.h ggml-metal-impl.h
+                    kernels/common.h kernels/dequantize.h kernels/quantize.h
+                    kernels/${kind}.metal
+            COMMENT "Generate embedded Metal library for ${kind}"
+            VERBATIM
+        )
+
+        list(APPEND METALLIB_EMBED_ASM_FILES "${ASM}")
+    endforeach()
+
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM_FILES})
 else()
-    # copy metal files to bin directory
+    # copy header files to bin directory
     configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
     configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
 
+    file(MAKE_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels")
+    configure_file(kernels/common.h     ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/common.h     COPYONLY)
+    configure_file(kernels/dequantize.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/dequantize.h COPYONLY)
+    configure_file(kernels/quantize.h   ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/quantize.h   COPYONLY)
+
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
+    endforeach()
+
     if (GGML_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: disabling fast math is needed in order to pass tests/test-backend-ops
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
         #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
         # note: adding -g causes segmentation fault during compile
-        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
         set(XC_FLAGS -fno-fast-math -fno-inline)
     else()
         set(XC_FLAGS -O3)
     endif()
 
-    # Append macOS metal versioning flags
     if (GGML_METAL_MACOSX_VERSION_MIN)
         message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
         list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
@@ -90,35 +147,46 @@ else()
         list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
     endif()
 
+    # Compile each kernel source to .air, then link into default.metallib
+    set(AIR_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(name ${src} NAME_WE)
+        set(AIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${name}.air")
+        list(APPEND AIR_FILES ${AIR})
+        add_custom_command(
+            OUTPUT ${AIR}
+            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
+            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
+            COMMENT "Compiling ${src}"
+            VERBATIM
+        )
+    endforeach()
+
     add_custom_command(
         OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
-        COMMENT "Compiling Metal kernels"
-        )
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
+        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
+        DEPENDS ${AIR_FILES}
+        COMMENT "Linking Metal kernels into default.metallib"
+    )
 
-    # FIXME: only add to the ggml-metal target?
     add_custom_target(
         ggml-metal-lib ALL
         DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
+    )
 endif() # GGML_METAL_EMBED_LIBRARY
 
 if (NOT GGML_METAL_EMBED_LIBRARY)
     install(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
+        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
+        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
+        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
+    )
 
-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
+    install(
+        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        DESTINATION ${CMAKE_INSTALL_BINDIR}
+    )
 endif()

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -477,6 +477,41 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max(ggml_me
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning_indexer(
+        ggml_metal_library_t lib,
+        const ggml_tensor * op) {
+    GGML_ASSERT(op->op == GGML_OP_LIGHTNING_INDEXER);
+
+    char name[256];
+
+    snprintf(name, 256, "kernel_lightning_indexer_%s", ggml_type_name(op->src[1]->type));
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
+    }
+
+    return res;
+}
+
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc(ggml_metal_library_t lib, ggml_op op) {
+    const char * name = nullptr;
+
+    switch (op) {
+        case GGML_OP_DSV4_HC_COMB: name = "kernel_dsv4_hc_comb_f32"; break;
+        case GGML_OP_DSV4_HC_PRE:  name = "kernel_dsv4_hc_pre_f32";  break;
+        case GGML_OP_DSV4_HC_POST: name = "kernel_dsv4_hc_post_f32"; break;
+        default: GGML_ABORT("fatal error");
+    }
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv(ggml_metal_library_t lib, const ggml_tensor * op) {
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
@@ -2117,6 +2152,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_opt_step_sgd(ggm
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_silu_back(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_SILU_BACK);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_silu_back_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_memset(ggml_metal_library_t lib, const ggml_tensor *  op) {
     GGML_ASSERT(op->type == GGML_TYPE_I64);
 

ggml/src/ggml-metal/ggml-metal-device.h

@@ -117,6 +117,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_concat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_silu_back         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -124,6 +125,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_bl
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_add        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_tri               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning_indexer (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc           (ggml_metal_library_t lib, enum ggml_op op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);

ggml/src/ggml-metal/ggml-metal-device.m

@@ -2,6 +2,7 @@
 
 #import "ggml-impl.h"
 #import "ggml-backend-impl.h"
+#import "ggml-metal-impl.h"
 
 #include <Foundation/Foundation.h>
 
@@ -94,8 +95,63 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
     return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }
 
+//
+// MTLLibrary collection (one library per op-source, compiled separately)
+//
+
+// Single source of truth for the per-kind metal libraries. The order here
+// defines the enum values and every per-kind table below, so adding a library
+// is a one-line change here (plus adding its source to CMakeLists.txt).
+//   X(suffix, name): name is both the kernels/<name>.metal basename and the
+//   ggml_metallib_<name>_{start,end} embed-symbol stem.
+#define GGML_METAL_LIBS \
+    X(FA,              fa)             \
+    X(MUL_MV,          mul_mv)         \
+    X(MUL_MM,          mul_mm)         \
+    X(QUANTIZE,        quantize)       \
+    X(SOFTMAX,         softmax)        \
+    X(NORM,            norm)           \
+    X(UNARY,           unary)          \
+    X(BINBCAST,        binbcast)       \
+    X(REDUCE,          reduce)         \
+    X(TRI,             tri)            \
+    X(SSM,             ssm)            \
+    X(WKV,             wkv)            \
+    X(GATED_DELTA_NET, gated_delta_net)\
+    X(SOLVE_TRI,       solve_tri)      \
+    X(ROPE,            rope)           \
+    X(CONV,            conv)           \
+    X(UPSCALE,         upscale)        \
+    X(ARGSORT,         argsort)        \
+    X(POOL,            pool)           \
+    X(MISC,            misc)
+
+enum ggml_metal_lib_kind {
+#define X(e, s) GGML_METAL_LIB_##e,
+    GGML_METAL_LIBS
+#undef X
+    GGML_METAL_LIB_COUNT,
+};
+
+static const char * const k_lib_names[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = #s,
+    GGML_METAL_LIBS
+#undef X
+};
+
 struct ggml_metal_library {
-    id<MTLLibrary> obj;
+    // Per-kind compiled libraries. When single_library is true, the whole library
+    // (e.g. a pre-compiled default.metallib or a from-source build) lives at
+    // objs[0] and the remaining slots are nil.
+    id<MTLLibrary> objs[GGML_METAL_LIB_COUNT];
+    bool single_library; // true: combined library at objs[0]; false: per-kind libs in objs[*]
+
+    // Routing table: kernel function name -> objs[] index, populated from each
+    // compiled library's -[MTLLibrary functionNames]. The actual compiled
+    // libraries are the single source of truth for which library owns a kernel,
+    // so adding kernels later requires no manual routing maintenance.
+    // nil in single_library mode (everything resolves to objs[0]).
+    NSMutableDictionary<NSString *, NSNumber *> * fn_to_lib;
 
     ggml_metal_device_t dev;
     ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
@@ -103,160 +159,376 @@ struct ggml_metal_library {
     NSLock * lock;
 };
 
-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
-    id<MTLLibrary> library = nil;
-    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+// Build the fn_to_lib routing table by querying each compiled library's public
+// function names. Call once after all per-kind libraries have been compiled.
+static void ggml_metal_library_build_index(ggml_metal_library_t lib) {
+    @autoreleasepool {
+        NSMutableDictionary<NSString *, NSNumber *> * index = [[NSMutableDictionary alloc] init];
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            for (NSString * fname in [lib->objs[kind] functionNames]) {
+                index[fname] = @(kind);
+            }
+        }
+        lib->fn_to_lib = index;
+    }
+}
 
-    // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    //
-    // TODO: move to a function
-    {
-        const int64_t t_start = ggml_time_us();
+// Parse a `#include "name"` line. Returns the quoted name in *include_name on
+// success. Whitespace-tolerant; ignores `#include <...>` (system headers).
+static bool ggml_metal_library_parse_quoted_include(NSString * line, NSString ** include_name) {
+    NSScanner * scanner = [NSScanner scannerWithString:line];
+    scanner.charactersToBeSkipped = [NSCharacterSet whitespaceCharacterSet];
 
-        NSError * error = nil;
-        NSString * src = nil;
+    if (![scanner scanString:@"#" intoString:NULL] ||
+        ![scanner scanString:@"include" intoString:NULL] ||
+        ![scanner scanString:@"\"" intoString:NULL]) {
+        return false;
+    }
 
-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+    NSString * name = nil;
+    if (![scanner scanUpToString:@"\"" intoString:&name]) {
+        return false;
+    }
 
-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
+    if (include_name) {
+        *include_name = name;
+    }
+    return true;
+}
 
-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
+// Recursively inline `#include "name"` directives. System includes (<...>),
+// `#if/#else/#endif`, and other preprocessor lines are passed through to the
+// Metal compiler unchanged. `#pragma once` is dropped since `seen` already
+// guards against double-inclusion.
+static bool ggml_metal_library_flatten_file(NSMutableString * dst, NSString * path,
+                                            NSArray<NSString *> * search_paths,
+                                            NSMutableSet<NSString *> * seen, NSError ** error) {
+    NSString * key = [path stringByStandardizingPath];
+    if ([seen containsObject:key]) {
+        return true;
+    }
+    [seen addObject:key];
 
-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
+    NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:error];
+    if (!src) {
+        return false;
+    }
 
-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+    NSFileManager * fm = [NSFileManager defaultManager];
+    for (NSString * line in [src componentsSeparatedByString:@"\n"]) {
+        NSString * trimmed = [line stringByTrimmingCharactersInSet:[NSCharacterSet whitespaceCharacterSet]];
+        if ([trimmed isEqualToString:@"#pragma once"]) {
+            continue;
+        }
 
-            NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
-
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
-                    if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
-                    }
-                    if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                        // Link to the resource could not be resolved.
-                        path_lib_default = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
-                    }
+        NSString * include_name = nil;
+        if (ggml_metal_library_parse_quoted_include(line, &include_name)) {
+            NSString * resolved = nil;
+            for (NSString * dir in search_paths) {
+                NSString * candidate = [dir stringByAppendingPathComponent:include_name];
+                if ([fm isReadableFileAtPath:candidate]) {
+                    resolved = candidate;
+                    break;
                 }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                path_lib_default = nil;
             }
-
-            path_lib = path_lib_default;
+            if (!resolved) {
+                if (error) {
+                    NSString * msg = [NSString stringWithFormat:@"could not resolve include \"%@\" from '%@'", include_name, path];
+                    *error = [NSError errorWithDomain:@"ggml-metal-source-flatten" code:1
+                                             userInfo:@{NSLocalizedDescriptionKey: msg}];
+                }
+                return false;
+            }
+            if (!ggml_metal_library_flatten_file(dst, resolved, search_paths, seen, error)) {
+                return false;
+            }
+            continue;
         }
 
-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+        [dst appendString:line];
+        [dst appendString:@"\n"];
+    }
 
-            library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+    return true;
+}
 
-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+static NSString * ggml_metal_library_flatten_source(NSString * path_source, NSError ** error) {
+    // Search paths cover both runtime layout (build/bin/kernels + build/bin)
+    // and source-tree layout (ggml/src/ggml-metal/kernels + ggml/src/ggml-metal + ggml/src).
+    NSString * path_kernels = [path_source stringByDeletingLastPathComponent];
+    NSString * path_base    = [path_kernels stringByDeletingLastPathComponent];
+    NSArray<NSString *> * search_paths = @[
+        path_kernels,
+        path_base,
+        [path_base stringByDeletingLastPathComponent],
+    ];
 
-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+    NSMutableString * src = [[NSMutableString alloc] init];
+    NSMutableSet<NSString *> * seen = [NSMutableSet set];
 
-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+    if (!ggml_metal_library_flatten_file(src, path_source, search_paths, seen, error)) {
+        [src release];
+        return nil;
+    }
+    return src;
+}
+
+// Compile all per-kind libraries in parallel. `source_for_kind` returns the MSL
+// source for a kind (the helper takes ownership and releases it), or nil with
+// *err set on failure. On success the objs[] slots are populated and the routing
+// index is built; on any failure every error is logged and false is returned
+// (the caller is responsible for freeing `res`).
+static bool ggml_metal_library_compile_all(
+        ggml_metal_library_t res,
+        id<MTLDevice> device,
+        NSDictionary * prep,
+        NSString * (^source_for_kind)(int kind, NSError ** err),
+        const char * origin) {
+    const int64_t t_start = ggml_time_us();
+
+    int64_t  * t_per_lib   = calloc(GGML_METAL_LIB_COUNT, sizeof(int64_t));
+    NSError ** err_per_lib = calloc(GGML_METAL_LIB_COUNT, sizeof(NSError *));
+    __block atomic_bool any_failure = false;
+
+    dispatch_group_t group = dispatch_group_create();
+    dispatch_queue_t queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        dispatch_group_async(group, queue, ^{
+
+            const int64_t t0 = ggml_time_us();
+
+            NSError * error = nil;
+
+            NSString * src = source_for_kind(kind, &error);
+            if (!src) {
+                err_per_lib[kind] = [error retain];
+                atomic_store(&any_failure, true);
+                return;
             }
 
-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
+            id<MTLLibrary> lib = nil;
 
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        }
-#endif
-
-        if (!library) {
             @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ggml_metal_device_get_props(dev)->has_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
-                }
-
-                if (ggml_metal_device_get_props(dev)->has_tensor) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                 MTLCompileOptions * options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
-                //[options setFastMathEnabled:false];
+                lib = [device newLibraryWithSource:src options:options error:&error];
 
-                library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return nil;
-                }
-
-#if !__has_feature(objc_arc)
                 [options release];
-#endif
+
+                // retain the error before the autorelease pool drains it
+                if (!lib) {
+                    err_per_lib[kind] = [error retain];
+                }
+            }
+
+            [src release];
+
+            t_per_lib[kind] = ggml_time_us() - t0;
+
+            if (!lib) {
+                atomic_store(&any_failure, true);
+                return;
+            }
+
+            res->objs[kind] = lib;
+        });
+    }
+    dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
+    dispatch_release(group);
+
+    const bool ok = !atomic_load(&any_failure);
+
+    if (ok) {
+        const int64_t t_total = ggml_time_us() - t_start;
+        int64_t t_max = 0;
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            GGML_LOG_DEBUG("%s: compiled '%s' library in %.3f sec\n",
+                           __func__, k_lib_names[kind], t_per_lib[kind] / 1e6);
+            if (t_per_lib[kind] > t_max) t_max = t_per_lib[kind];
+        }
+        GGML_LOG_INFO("%s: loaded %d libraries from %s in %.3f sec (max single = %.3f sec)\n",
+                      __func__, GGML_METAL_LIB_COUNT, origin, t_total / 1e6, t_max / 1e6);
+
+        ggml_metal_library_build_index(res);
+    } else {
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            if (err_per_lib[kind]) {
+                GGML_LOG_ERROR("%s: failed to build '%s' library: %s\n", __func__,
+                               k_lib_names[kind], [[err_per_lib[kind] description] UTF8String]);
+                [err_per_lib[kind] release];
             }
         }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
-
-        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
     }
 
-    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    free(err_per_lib);
+    free(t_per_lib);
 
-    res->obj       = library;
+    return ok;
+}
+
+ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
     res->dev       = dev;
     res->pipelines = ggml_metal_pipelines_init();
     res->lock      = [NSLock new];
 
+    // shared MTLCompileOptions preprocessor macros (matches the build-time defines)
+    NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+    if (ggml_metal_device_get_props(dev)->has_bfloat) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
+    }
+    if (ggml_metal_device_get_props(dev)->has_tensor) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+    }
+#if GGML_METAL_EMBED_LIBRARY
+    [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    // start/end symbols emitted by CMake (see CMakeLists.txt), one pair per kind
+#define X(e, s) extern const char ggml_metallib_##s##_start[]; extern const char ggml_metallib_##s##_end[];
+    GGML_METAL_LIBS
+#undef X
+
+    static const char * const lib_start[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_start,
+    GGML_METAL_LIBS
+#undef X
+    };
+    static const char * const lib_end[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_end,
+    GGML_METAL_LIBS
+#undef X
+    };
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            (void) err;
+            return [[NSString alloc] initWithBytes:lib_start[kind]
+                                            length:(lib_end[kind] - lib_start[kind])
+                                          encoding:NSUTF8StringEncoding];
+        }, "embedded data");
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
     return res;
+#else
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    const int64_t t_start = ggml_time_us();
+
+    NSError * error = nil;
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+
+        NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
+
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
+                if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
+                }
+                if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+                    // Link to the resource could not be resolved.
+                    path_lib_default = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            path_lib_default = nil;
+        }
+
+        path_lib = path_lib_default;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found: a single combined default.metallib
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        res->objs[0]        = [device newLibraryWithURL:libURL error:&error];
+        res->single_library = true;
+        if (!res->objs[0]) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            ggml_metal_library_free(res);
+            return NULL;
+        }
+
+        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+        return res;
+    }
+
+    // no pre-compiled metallib: fall back to compiling each kernel source separately
+    GGML_LOG_INFO("%s: default.metallib not found, loading kernel sources\n", __func__);
+
+    NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+    if (path_resource) {
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, [path_resource UTF8String]);
+    }
+
+    // resolve each kind's source path up front (file lookup/logging stays on the calling thread)
+    NSString ** path_per_kind = calloc(GGML_METAL_LIB_COUNT, sizeof(NSString *));
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        NSString * rel = [NSString stringWithFormat:@"kernels/%s.metal", k_lib_names[kind]];
+
+        NSString * path_source = nil;
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:rel];
+        } else {
+            NSString * stem = [NSString stringWithFormat:@"kernels/%s", k_lib_names[kind]];
+            path_source = [bundle pathForResource:stem ofType:@"metal"];
+        }
+
+        if (path_source == nil || ![[NSFileManager defaultManager] isReadableFileAtPath:path_source]) {
+            GGML_LOG_WARN("%s: could not locate %s in bundle, falling back to cwd\n", __func__, [rel UTF8String]);
+            path_source = rel;
+        }
+
+        GGML_LOG_DEBUG("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        path_per_kind[kind] = [path_source retain];
+    }
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            return ggml_metal_library_flatten_source(path_per_kind[kind], err);
+        }, "source");
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        [path_per_kind[kind] release];
+    }
+    free(path_per_kind);
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
+    return res;
+#endif
 }
 
 ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
@@ -318,10 +590,11 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
         return NULL;
     }
 
-    res->obj       = library;
-    res->dev       = dev;
-    res->pipelines = ggml_metal_pipelines_init();
-    res->lock      = [NSLock new];
+    res->objs[0]        = library;
+    res->single_library = true;
+    res->dev            = dev;
+    res->pipelines      = ggml_metal_pipelines_init();
+    res->lock           = [NSLock new];
 
     return res;
 }
@@ -331,8 +604,14 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
         return;
     }
 
-    if (lib->obj) {
-        [lib->obj release];
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        if (lib->objs[kind]) {
+            [lib->objs[kind] release];
+        }
+    }
+
+    if (lib->fn_to_lib) {
+        [lib->fn_to_lib release];
     }
 
     ggml_metal_pipelines_free(lib->pipelines);
@@ -393,11 +672,28 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_
 
         GGML_LOG_DEBUG("%s: compiling pipeline: base = '%s', name = '%s'\n", __func__, base, name);
 
+        // route to the library that actually defines this kernel; fn_to_lib is
+        // built from -[MTLLibrary functionNames] so it's always in sync
+        int lib_idx = 0;
+        if (!lib->single_library) {
+            NSNumber * idx = lib->fn_to_lib[base_func];
+            if (!idx) {
+                [lib->lock unlock];
+
+                GGML_LOG_ERROR("%s: kernel not found in any metal library: base = '%s', name = '%s'\n", __func__, base, name);
+
+                return res;
+            }
+            lib_idx = [idx intValue];
+        }
+
+        id<MTLLibrary> mtl_lib = lib->objs[lib_idx];
+
         id<MTLFunction> mtl_function;
         if (!cv) {
-            mtl_function = [lib->obj newFunctionWithName:base_func];
+            mtl_function = [mtl_lib newFunctionWithName:base_func];
         } else {
-            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+            mtl_function = [mtl_lib newFunctionWithName:base_func constantValues:cv->obj error:&error];
         }
         if (!mtl_function) {
             [lib->lock unlock];
@@ -1137,6 +1433,14 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_SILU_BACK:
+            return (op->src[0]->type == GGML_TYPE_F32) &&
+                (op->src[1]->type == GGML_TYPE_F32) &&
+                (op->type == GGML_TYPE_F32) &&
+                ggml_is_contiguous(op->src[0]) &&
+                ggml_is_contiguous(op->src[1]) &&
+                ggml_is_contiguous(op) &&
+                ggml_are_same_shape(op->src[0], op->src[1]);
         case GGML_OP_GLU:
             switch (ggml_get_glu_op(op)) {
                 case GGML_GLU_OP_REGLU:
@@ -1181,6 +1485,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_ADD_ID:
+            return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) && (op->src[0]->type == op->src[1]->type);
         case GGML_OP_ACC:
             return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_REPEAT:
@@ -1299,6 +1604,72 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                     return false;
             }
             return has_simdgroup_mm; // TODO: over-restricted for vec-kernels
+        case GGML_OP_LIGHTNING_INDEXER:
+            if (op->src[0]->ne[0] != OP_LIGHTNING_INDEXER_DK ||
+                op->src[0]->ne[1] != OP_LIGHTNING_INDEXER_NH) {
+                return false;
+            }
+            if (!has_simdgroup_mm ||
+                op->src[0]->type != GGML_TYPE_F32 ||
+                op->src[2]->type != GGML_TYPE_F32 ||
+                op->src[3]->type != GGML_TYPE_F16 ||
+                op->type         != GGML_TYPE_F32 ||
+                !ggml_is_contiguous_rows(op->src[0]) ||
+                !ggml_is_contiguous_rows(op->src[1]) ||
+                !ggml_is_contiguous_rows(op->src[2]) ||
+                !ggml_is_contiguous_rows(op->src[3])) {
+                return false;
+            }
+            switch (op->src[1]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                    return true;
+                case GGML_TYPE_BF16:
+                    return has_bfloat;
+                default:
+                    return false;
+            }
+        case GGML_OP_DSV4_HC_COMB:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->src[2]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[0]->ne[0] == 24 &&
+                op->src[1]->ne[0] >= 3 &&
+                op->src[2]->ne[0] == 24 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]) &&
+                ggml_is_contiguous_rows(op->src[2]);
+        case GGML_OP_DSV4_HC_PRE:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[0]->ne[1] == 4 &&
+                op->src[1]->ne[0] == 4 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]);
+        case GGML_OP_DSV4_HC_POST:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->src[2]->type == GGML_TYPE_F32 &&
+                op->src[3]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[1]->ne[1] == 4 &&
+                op->src[2]->ne[0] == 4 &&
+                op->src[3]->ne[0] == 4 &&
+                op->src[3]->ne[1] == 4 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]) &&
+                ggml_is_contiguous_rows(op->src[2]) &&
+                ggml_is_contiguous_rows(op->src[3]);
         case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:
             return has_simdgroup_reduction;

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -112,6 +112,13 @@
 #define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
 #define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
 
+#define OP_LIGHTNING_INDEXER_DK    128
+#define OP_LIGHTNING_INDEXER_NH     64
+#define OP_LIGHTNING_INDEXER_NHPTG   8
+#define OP_LIGHTNING_INDEXER_NKPSG   8
+#define OP_LIGHTNING_INDEXER_NSG     8
+#define OP_LIGHTNING_INDEXER_NBPTG   8
+
 #define OP_UNARY_NUM_SCALE      10
 #define OP_UNARY_NUM_FILL       11
 #define OP_UNARY_NUM_CLAMP      12
@@ -1171,6 +1178,66 @@ typedef struct {
     int64_t val;
 } ggml_metal_kargs_memset;
 
+typedef struct {
+    int32_t  n_kv;
+    int32_t  n_batch;
+    int32_t  mask_ne3;
+    uint64_t nb1;
+    uint64_t nb3;
+    uint64_t nbq1;
+    uint64_t nbq2;
+    uint64_t nbq3;
+    uint64_t nbk2;
+    uint64_t nbk3;
+    uint64_t nbw1;
+    uint64_t nbw3;
+    uint64_t nbm1;
+    uint64_t nbm3;
+} ggml_metal_kargs_lightning_indexer;
+
+typedef struct {
+    int32_t  n_tokens;
+    int32_t  n_iter;
+    uint64_t nb_m0;
+    uint64_t nb_m1;
+    uint64_t nb_s0;
+    uint64_t nb_b0;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+    uint64_t nb_d2;
+    float    eps;
+} ggml_metal_kargs_dsv4_hc_comb;
+
+typedef struct {
+    int32_t  n_embd;
+    int32_t  n_tokens;
+    uint64_t nb_x0;
+    uint64_t nb_x1;
+    uint64_t nb_x2;
+    uint64_t nb_w0;
+    uint64_t nb_w1;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+} ggml_metal_kargs_dsv4_hc_pre;
+
+typedef struct {
+    int32_t  n_embd;
+    int32_t  n_tokens;
+    uint64_t nb_x0;
+    uint64_t nb_x1;
+    uint64_t nb_r0;
+    uint64_t nb_r1;
+    uint64_t nb_r2;
+    uint64_t nb_p0;
+    uint64_t nb_p1;
+    uint64_t nb_c0;
+    uint64_t nb_c1;
+    uint64_t nb_c2;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+    uint64_t nb_d2;
+} ggml_metal_kargs_dsv4_hc_post;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -1222,4 +1289,8 @@ typedef struct {
     int64_t  np;
 } ggml_metal_kargs_opt_step_sgd;
 
+typedef struct {
+    int64_t ne;
+} ggml_metal_kargs_silu_back;
+
 #endif // GGML_METAL_IMPL

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -299,6 +299,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_unary(ctx, idx);
             } break;
+        case GGML_OP_SILU_BACK:
+            {
+                n_fuse = ggml_metal_op_silu_back(ctx, idx);
+            } break;
         case GGML_OP_GLU:
             {
                 n_fuse = ggml_metal_op_glu(ctx, idx);
@@ -316,6 +320,16 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_cumsum(ctx, idx);
             } break;
+        case GGML_OP_LIGHTNING_INDEXER:
+            {
+                n_fuse = ggml_metal_op_lightning_indexer(ctx, idx);
+            } break;
+        case GGML_OP_DSV4_HC_COMB:
+        case GGML_OP_DSV4_HC_PRE:
+        case GGML_OP_DSV4_HC_POST:
+            {
+                n_fuse = ggml_metal_op_dsv4_hc(ctx, idx);
+            } break;
         case GGML_OP_SOFT_MAX:
             {
                 n_fuse = ggml_metal_op_soft_max(ctx, idx);
@@ -1297,6 +1311,203 @@ int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_lightning_indexer(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(op->op == GGML_OP_LIGHTNING_INDEXER);
+
+    const ggml_tensor * q = op->src[0];
+    const ggml_tensor * k = op->src[1];
+    const ggml_tensor * w = op->src[2];
+    const ggml_tensor * m = op->src[3];
+
+    GGML_ASSERT(q->type == GGML_TYPE_F32);
+    GGML_ASSERT(k->type == GGML_TYPE_F32  ||
+                k->type == GGML_TYPE_F16  ||
+                k->type == GGML_TYPE_BF16 ||
+                k->type == GGML_TYPE_Q4_0 ||
+                k->type == GGML_TYPE_Q4_1 ||
+                k->type == GGML_TYPE_Q5_0 ||
+                k->type == GGML_TYPE_Q5_1 ||
+                k->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(w->type == GGML_TYPE_F32);
+    GGML_ASSERT(m->type == GGML_TYPE_F16);
+    GGML_ASSERT(op->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(q->ne[0] == OP_LIGHTNING_INDEXER_DK);
+    GGML_ASSERT(q->ne[1] == OP_LIGHTNING_INDEXER_NH);
+
+    ggml_metal_kargs_lightning_indexer args = {
+        /*.n_kv      =*/ (int32_t) k->ne[2],
+        /*.n_batch   =*/ (int32_t) q->ne[2],
+        /*.mask_ne3  =*/ (int32_t) m->ne[3],
+        /*.nb1       =*/ op->nb[1],
+        /*.nb3       =*/ op->nb[3],
+        /*.nbq1      =*/ q->nb[1],
+        /*.nbq2      =*/ q->nb[2],
+        /*.nbq3      =*/ q->nb[3],
+        /*.nbk2      =*/ k->nb[2],
+        /*.nbk3      =*/ k->nb[3],
+        /*.nbw1      =*/ w->nb[1],
+        /*.nbw3      =*/ w->nb[3],
+        /*.nbm1      =*/ m->nb[1],
+        /*.nbm3      =*/ m->nb[3],
+    };
+
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(q),  1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(k),  2);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(w),  3);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(m),  4);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 5);
+
+    const int nsg   = OP_LIGHTNING_INDEXER_NSG;
+    const int nkptg = OP_LIGHTNING_INDEXER_NKPSG*nsg;
+    const int nbptg = OP_LIGHTNING_INDEXER_NBPTG;
+
+    auto pipeline = ggml_metal_library_get_pipeline_lightning_indexer(ctx->lib, op);
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_dispatch_threadgroups(enc,
+            (k->ne[2] + nkptg - 1)/nkptg,
+            (q->ne[2] + nbptg - 1)/nbptg,
+            q->ne[3], 32, nsg, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_dsv4_hc(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_encoder_t enc = ctx->enc;
+    auto pipeline = ggml_metal_library_get_pipeline_dsv4_hc(ctx->lib, op->op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+
+    switch (op->op) {
+        case GGML_OP_DSV4_HC_COMB:
+            {
+                const ggml_tensor * mixes = op->src[0];
+                const ggml_tensor * scale = op->src[1];
+                const ggml_tensor * base  = op->src[2];
+
+                GGML_ASSERT(mixes->type == GGML_TYPE_F32);
+                GGML_ASSERT(scale->type == GGML_TYPE_F32);
+                GGML_ASSERT(base->type  == GGML_TYPE_F32);
+                GGML_ASSERT(op->type    == GGML_TYPE_F32);
+                GGML_ASSERT(mixes->ne[0] == 24);
+                GGML_ASSERT(op->ne[0] == 4 && op->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_comb args = {
+                    /*.n_tokens =*/ (int32_t) mixes->ne[1],
+                    /*.n_iter   =*/ ggml_get_op_params_i32(op, 1),
+                    /*.nb_m0    =*/ mixes->nb[0],
+                    /*.nb_m1    =*/ mixes->nb[1],
+                    /*.nb_s0    =*/ scale->nb[0],
+                    /*.nb_b0    =*/ base->nb[0],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                    /*.nb_d2    =*/ op->nb[2],
+                    /*.eps      =*/ ggml_get_op_params_f32(op, 0),
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(mixes), 1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(scale), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(base),  3);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),    4);
+
+                // One SIMDgroup owns one 4x4 Sinkhorn matrix. Packing up to four
+                // independent tokens per threadgroup keeps both decode and prompt
+                // dispatches compact without any threadgroup-memory synchronization.
+                const int nsg = std::min(4, args.n_tokens);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (args.n_tokens + nsg - 1)/nsg, 1, 1, 32, nsg, 1);
+            } break;
+        case GGML_OP_DSV4_HC_PRE:
+            {
+                const ggml_tensor * x       = op->src[0];
+                const ggml_tensor * weights = op->src[1];
+
+                GGML_ASSERT(x->type       == GGML_TYPE_F32);
+                GGML_ASSERT(weights->type == GGML_TYPE_F32);
+                GGML_ASSERT(op->type      == GGML_TYPE_F32);
+                GGML_ASSERT(x->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_pre args = {
+                    /*.n_embd   =*/ (int32_t) x->ne[0],
+                    /*.n_tokens =*/ (int32_t) x->ne[2],
+                    /*.nb_x0    =*/ x->nb[0],
+                    /*.nb_x1    =*/ x->nb[1],
+                    /*.nb_x2    =*/ x->nb[2],
+                    /*.nb_w0    =*/ weights->nb[0],
+                    /*.nb_w1    =*/ weights->nb[1],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(x),       1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(weights), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),      3);
+
+                const int n_tiles = (args.n_embd + 31)/32;
+                const int nsg = std::min(4, n_tiles);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (n_tiles + nsg - 1)/nsg, args.n_tokens, 1, 32, nsg, 1);
+            } break;
+        case GGML_OP_DSV4_HC_POST:
+            {
+                const ggml_tensor * x        = op->src[0];
+                const ggml_tensor * residual = op->src[1];
+                const ggml_tensor * post     = op->src[2];
+                const ggml_tensor * comb     = op->src[3];
+
+                GGML_ASSERT(x->type        == GGML_TYPE_F32);
+                GGML_ASSERT(residual->type == GGML_TYPE_F32);
+                GGML_ASSERT(post->type     == GGML_TYPE_F32);
+                GGML_ASSERT(comb->type     == GGML_TYPE_F32);
+                GGML_ASSERT(op->type       == GGML_TYPE_F32);
+                GGML_ASSERT(residual->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_post args = {
+                    /*.n_embd   =*/ (int32_t) x->ne[0],
+                    /*.n_tokens =*/ (int32_t) x->ne[1],
+                    /*.nb_x0    =*/ x->nb[0],
+                    /*.nb_x1    =*/ x->nb[1],
+                    /*.nb_r0    =*/ residual->nb[0],
+                    /*.nb_r1    =*/ residual->nb[1],
+                    /*.nb_r2    =*/ residual->nb[2],
+                    /*.nb_p0    =*/ post->nb[0],
+                    /*.nb_p1    =*/ post->nb[1],
+                    /*.nb_c0    =*/ comb->nb[0],
+                    /*.nb_c1    =*/ comb->nb[1],
+                    /*.nb_c2    =*/ comb->nb[2],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                    /*.nb_d2    =*/ op->nb[2],
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(x),        1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(residual), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(post),     3);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(comb),     4);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),       5);
+
+                const int n_tiles = (args.n_embd + 31)/32;
+                const int nsg = std::min(4, n_tiles);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (n_tiles + nsg - 1)/nsg, args.n_tokens, 1, 32, nsg, 1);
+            } break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+
+    return 1;
+}
+
 int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -3197,9 +3408,6 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
-
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));
 
@@ -3339,6 +3547,36 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     return n_fuse;
 }
 
+int ggml_metal_op_silu_back(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    auto pipeline = ggml_metal_library_get_pipeline_silu_back(lib, op);
+
+    const int64_t ne = ggml_nelements(op);
+
+    ggml_metal_kargs_silu_back args = {
+        /*.ne =*/ ne,
+    };
+
+    int arg_idx{0};
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op), arg_idx++);
+
+    const int nth = std::min<int64_t>(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne);
+    const int64_t n = (ne + nth - 1) / nth;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -54,6 +54,8 @@ int ggml_metal_op_cumsum            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_diag              (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_lightning_indexer (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_dsv4_hc           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_conv          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_scan          (ggml_metal_op_t ctx, int idx);
@@ -70,6 +72,7 @@ int ggml_metal_op_mul_mat_id        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_add_id            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_flash_attn_ext    (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_bin               (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_silu_back         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_l2_norm           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_group_norm        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);

ggml/src/ggml-metal/kernels/argsort.metal

@@ -0,0 +1,232 @@
+#include "common.h"
+
+// bitonic sort implementation following the CUDA kernels as reference
+typedef void (argsort_t)(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    // bitonic sort
+    const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
+
+    // initialize indices
+    shmem_i32[col] = i00 + col;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (int k = 2; k <= ntg.x; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                } else {
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                }
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+    }
+
+    const int64_t i0 = ib*args.top_k;
+
+    // copy the result to dst without the padding
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
+
+        dst[col] = shmem_i32[col];
+    }
+}
+
+template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
+
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;

ggml/src/ggml-metal/kernels/binbcast.metal

@@ -0,0 +1,228 @@
+#include "common.h"
+
+// OP: 0 - add, 1 - sub, 2 - mul, 3 - div
+constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
+constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
+constant bool  FC_bin_rb [[function_constant(FC_BIN + 2)]];
+constant bool  FC_bin_cb [[function_constant(FC_BIN + 3)]];
+
+template <typename T0, typename T1, typename T>
+kernel void kernel_bin_fuse_impl(
+        constant ggml_metal_kargs_bin & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP FC_bin_op
+#define FC_F  FC_bin_f
+#define FC_RB FC_bin_rb
+#define FC_CB FC_bin_cb
+
+    if (FC_RB) {
+        // row broadcast
+        const uint i0 = tgpig.y*args.ne00 + tgpig.x;
+        const uint i1 = FC_CB ? tgpig.x%args.ne10 : tgpig.x;
+
+        device const T0 * src0_row = (device const T0 *) (src0);
+        device       T  * dst_row  = (device       T  *) (dst);
+
+        if (FC_F == 1) {
+            device const T1 * src1_row = (device const T1 *) (src1 + args.o1[0]);
+
+            if (FC_OP == 0) {
+                dst_row[i0] = src0_row[i0] + src1_row[i1];
+            }
+
+            if (FC_OP == 1) {
+                dst_row[i0] = src0_row[i0] - src1_row[i1];
+            }
+
+            if (FC_OP == 2) {
+                dst_row[i0] = src0_row[i0] * src1_row[i1];
+            }
+
+            if (FC_OP == 3) {
+                dst_row[i0] = src0_row[i0] / src1_row[i1];
+            }
+        } else {
+            T0 res = src0_row[i0];
+
+            if (FC_OP == 0) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res += ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 1) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res -= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 2) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res *= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 3) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res /= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            dst_row[i0] = res;
+        }
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int i01 = tgpig.x;
+
+        if (i01 >= args.ne01) {
+            return;
+        }
+
+        const int i13 = i03%args.ne13;
+        const int i12 = i02%args.ne12;
+        const int i11 = i01%args.ne11;
+
+        device const T0 * src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+        device       T  * dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+        if (FC_F == 1) {
+            device const T1 * src1_ptr = (device const T1 *) (src1 + args.o1[0] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                if (FC_OP == 0) {
+                    dst_ptr[i0] = src0_ptr[i0] + src1_ptr[i10];
+                }
+
+                if (FC_OP == 1) {
+                    dst_ptr[i0] = src0_ptr[i0] - src1_ptr[i10];
+                }
+
+                if (FC_OP == 2) {
+                    dst_ptr[i0] = src0_ptr[i0] * src1_ptr[i10];
+                }
+
+                if (FC_OP == 3) {
+                    dst_ptr[i0] = src0_ptr[i0] / src1_ptr[i10];
+                }
+            }
+        } else {
+            device const T1 * src1_ptr[8];
+            FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                src1_ptr[j] = (device const T1 *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+            }
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                T res = src0_ptr[i0];
+
+                if (FC_OP == 0) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res += src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 1) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res -= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 2) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res *= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 3) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res /= src1_ptr[j][i10];
+                    }
+                }
+
+                dst_ptr[i0] = res;
+            }
+        }
+    }
+
+#undef FC_OP
+#undef FC_F
+#undef FC_RB
+#undef FC_CB
+}
+
+typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
+
+template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
+template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half,   half,   half>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half4,  half4,  half4>;
+
+kernel void kernel_add_id(
+        constant ggml_metal_kargs_add_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i1 = tgpig.x;
+    const int i2 = tgpig.y;
+
+    const int i11 = *((device const int32_t *) (src2 + i1*sizeof(int32_t) + i2*args.nb21));
+
+    const size_t nb1 = args.ne0 * sizeof(float);
+    const size_t nb2 = args.ne1 * nb1;
+
+    device       float * dst_row  = (device       float *)((device char *)dst  +  i1*nb1       + i2*nb2);
+    device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
+    device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
+template<typename T>
+kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
+
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_repeat_bf16")]] kernel kernel_repeat_t kernel_repeat<bfloat>;
+#endif
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;

ggml/src/ggml-metal/kernels/common.h

@@ -0,0 +1,126 @@
+#pragma once
+
+#include "ggml-metal-impl.h"
+
+#include <metal_stdlib>
+
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
+
+#define PAD2(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
+#define FOR_UNROLL(x) _Pragma("clang loop unroll(full)") for (x)
+
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+//
+// cmd:
+//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal/kernels/<src>.metal
+//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal/kernels/<src>.metal
+//
+#if __METAL_VERSION__ < 310 && defined(GGML_METAL_HAS_BF16)
+#undef GGML_METAL_HAS_BF16
+#endif
+
+#if defined(GGML_METAL_HAS_BF16)
+typedef matrix<bfloat, 4, 4> bfloat4x4;
+typedef matrix<bfloat, 2, 4> bfloat2x4;
+#endif
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+constexpr constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static inline float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    return as_type<float>(bits);
+}
+
+static inline float dot(float x, float y) {
+    return x*y;
+}
+
+static inline float sum(float x) {
+    return x;
+}
+
+static inline float sum(float4 x) {
+    return x[0] + x[1] + x[2] + x[3];
+}
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+inline T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+template<typename T> T elu_approx(T x);
+
+template<> inline float elu_approx<float>(float x) {
+    return (x > 0.f) ? x : (exp(x) - 1);
+}
+
+template<> inline float4 elu_approx<float4>(float4 x) {
+    float4 res;
+
+    res[0] = (x[0] > 0.0f) ? x[0] : (exp(x[0]) - 1.0f);
+    res[1] = (x[1] > 0.0f) ? x[1] : (exp(x[1]) - 1.0f);
+    res[2] = (x[2] > 0.0f) ? x[2] : (exp(x[2]) - 1.0f);
+    res[3] = (x[3] > 0.0f) ? x[3] : (exp(x[3]) - 1.0f);
+
+    return res;
+}

ggml/src/ggml-metal/kernels/conv.metal

@@ -0,0 +1,723 @@
+#include "common.h"
+
+typedef void (im2col_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
+
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
+
+          int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0;
+    const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1;
+
+    int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        while (in < args.N) {
+            pdst[offset_dst] = 0.0f;
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+
+            in += ntg[0];
+        }
+    } else {
+        int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw;
+
+        while (in < args.N) {
+            pdst[offset_dst] = x[offset_src];
+
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+            offset_src += ntg[0]*args.ofs0;
+
+            in += ntg[0];
+        }
+    }
+}
+
+template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
+template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
+
+// TODO: optimize
+typedef void (im2col_ext_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int64_t KHW = (int64_t)args.KHW;
+
+    const int64_t d   = tgpig[0] / args.CHW;
+    const int64_t chw = tgpig[0] % args.CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
+
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= args.N) {
+        return;
+    }
+
+    const int64_t tpitg_1 = HW / args.KW;
+    const int64_t tpitg_2 = HW % args.KW;
+
+    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
+    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
+
+    const int64_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
+        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
+        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+
+template <typename T>
+kernel void kernel_col2im_1d(
+        constant ggml_metal_kargs_col2im_1d & args,
+        device const T * col,
+        device       T * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T_out * args.OC) {
+        return;
+    }
+
+    const int t_out = idx % args.T_out;
+    const int oc    = idx / args.T_out;
+    const int t_abs = t_out + args.p0;  // absolute position in uncropped signal
+
+    int t_in_min = (t_abs - args.K + args.s0) / args.s0;  // ceil((t_abs - K + 1) / s0)
+    if (t_in_min < 0) {
+        t_in_min = 0;
+    }
+    int t_in_max = t_abs / args.s0;
+    if (t_in_max >= args.T_in) {
+        t_in_max = args.T_in - 1;
+    }
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * args.s0;
+        sum += float(col[(oc * args.K + k) + t_in * args.K_OC]);
+    }
+
+    dst[t_out + oc * args.T_out] = T(sum);
+}
+
+template [[host_name("kernel_col2im_1d_f32")]]  kernel void kernel_col2im_1d<float>(constant ggml_metal_kargs_col2im_1d &, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_col2im_1d_f16")]]  kernel void kernel_col2im_1d<half>(constant ggml_metal_kargs_col2im_1d &, device const half *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_col2im_1d_bf16")]] kernel void kernel_col2im_1d<bfloat>(constant ggml_metal_kargs_col2im_1d &, device const bfloat *, device bfloat *, uint, uint, uint);
+#endif
+
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+typedef void (conv_transpose_1d_t)(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_1d(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const     T * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    // For output position j on the time axis, only input positions
+    //   i such that i*s0 <= j < i*s0 + K
+    // contribute -- i.e. i in [ceil((j - K + 1)/s0), floor(j/s0)]
+    // intersected with [0, IL-1]. That's at most ceil(K/s0) values
+    // (typically 2 for stride==K/2 transposed convs).
+    const int32_t j  = tgpig[0];
+    const int32_t s0 = args.s0;
+    const int32_t K  = args.K;
+    const int32_t IL = args.IL;
+
+    int32_t i_min;
+    {
+        int32_t a = j - K + 1;
+        i_min = a <= 0 ? 0 : (a + s0 - 1) / s0; // ceil(a/s0) for a>0
+    }
+    int32_t i_max = j / s0;
+    if (i_max > IL - 1) i_max = IL - 1;
+
+    float v = 0.0f;
+    if (i_min <= i_max) {
+        for (int64_t c = 0; c < args.IC; c++) {
+            const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1];
+            const int32_t input_offset  = c * IL;
+
+            for (int32_t i = i_min; i <= i_max; i++) {
+                v += float(src0[kernel_offset + j - i * s0]) * src1[input_offset + i];
+            }
+        }
+    }
+
+    device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1);
+
+    dst_ptr[0] = v;
+}
+
+template [[host_name("kernel_conv_transpose_1d_f32_f32")]]
+kernel void kernel_conv_transpose_1d<float>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+template [[host_name("kernel_conv_transpose_1d_f16_f32")]]
+kernel void kernel_conv_transpose_1d<half>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+
+typedef void (conv_transpose_2d_t)(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_2d(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const T * src0,
+        device const float * src1,
+        device        char * dst,
+        threadgroup float * shared_sum [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int64_t out_x = tgpig[0];
+    const int64_t out_y = tgpig[1];
+    const int64_t out_c = tgpig[2];
+
+    const int64_t kw = tpitg[0];
+    const int64_t kh = tpitg[1];
+
+    float v = 0.0f;
+
+    for (int64_t in_c = 0; in_c < args.IC; in_c++) {
+        int64_t in_y = out_y - kh;
+
+        if (in_y < 0 || in_y % args.s0) continue;
+
+        in_y /= args.s0;
+
+        if (in_y >= args.IH) continue;
+
+        int64_t in_x = out_x - kw;
+
+        if (in_x < 0 || in_x % args.s0) continue;
+
+        in_x /= args.s0;
+
+        if (in_x >= args.IW) continue;
+
+        const int64_t input_idx = (args.IW * args.IH) * in_c + (args.IW) * in_y + in_x;
+        const int64_t kernel_idx = (args.KH * args.KW * args.OC) * in_c + (args.KH * args.KW) * out_c + (args.KW) * kh + kw;
+
+        v += (float)src0[kernel_idx] * src1[input_idx];
+    }
+
+    const uint tid = tpitg.y * ntg.x + tpitg.x;
+    shared_sum[tid] = v;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tid == 0) {
+        float total = 0.0f;
+        const uint num_threads = ntg.x * ntg.y;
+        for (uint i = 0; i < num_threads; i++) {
+            total += shared_sum[i];
+        }
+
+        device float * dst_ptr = (device float *) (dst + out_x*args.nb0 + out_y * args.nb1 + out_c*args.nb2);
+        dst_ptr[0] = total;
+    }
+}
+
+template [[host_name("kernel_conv_transpose_2d_f32_f32")]]
+kernel void kernel_conv_transpose_2d<float>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_transpose_2d_f16_f32")]]
+kernel void kernel_conv_transpose_2d<half>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+// grid: x = C tile, y = OH, z = OW * N (for channel-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw_tiled(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t c = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (c >= args.C) {
+        return;
+    }
+
+    const int32_t oh = tgpig.y;
+    const int32_t own = tgpig.z;
+    const int32_t ow = own % args.OW;
+    const int32_t n  = own / args.OW;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    float acc = 0.0f;
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        const uint64_t w_base   = (uint64_t) c * args.nb02;
+        const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+// grid: x = OW tile, y = OH, z = C * N (for spatially-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t oh = tgpig.y;
+    const int32_t cn = tgpig.z;
+    const int32_t c  = cn % args.C;
+    const int32_t n  = cn / args.C;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const uint64_t w_base   = (uint64_t) c * args.nb02;
+    const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+    const int32_t ow = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (ow >= args.OW) {
+        return;
+    }
+
+    float acc = 0.0f;
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+template [[host_name("kernel_conv_2d_dw_f32_f32")]]
+kernel void kernel_conv_2d_dw<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_f16_f32")]]
+kernel void kernel_conv_2d_dw<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f32_f32")]]
+kernel void kernel_conv_2d_dw_tiled<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f16_f32")]]
+kernel void kernel_conv_2d_dw_tiled<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_conv_3d(
+        constant ggml_metal_kargs_conv_3d & args,
+        device const  char * src0, // Weights [IC * OC, KD, KH, KW]
+        device const  char * src1, // Inputs  [IC * N,  ID, IH, IW]
+        device       char  * dst,  // Outputs [OC * N,  OD, OH, OW]
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+
+    // 1. Un-flatten the spatial dimension from Grid X
+    int64_t spatial_idx = tgpig.x * 32 + tpitg.x;
+
+    if (spatial_idx >= args.OW * args.OH * args.OD) {
+        return; // Thread falls outside the spatial volume
+    }
+
+    int64_t od = spatial_idx / (args.OW * args.OH);
+    int64_t oh = (spatial_idx / args.OW) % args.OH;
+    int64_t ow = spatial_idx % args.OW;
+
+    // 2. Map Y to Channels, Z to Batch
+    int64_t oc = tgpig.y;
+    int64_t batch_idx = tgpig.z;
+
+    // 3. Calculate anchor coordinates in the Input volume
+    int64_t i_w_base = ow * args.s0 - args.p0;
+    int64_t i_h_base = oh * args.s1 - args.p1;
+    int64_t i_d_base = od * args.s2 - args.p2;
+
+    float sum = 0.0f;
+
+    // 4. Gather Loop (Iterate over Input Channels -> Depth -> Height -> Width)
+    for (int64_t ic = 0; ic < args.IC; ++ic) {
+
+        // ggml packs batch and channel together in the 4th dimension
+        int64_t src_cn_idx = batch_idx * args.IC + ic;
+        int64_t w_cn_idx   = oc * args.IC + ic;
+
+        for (int64_t kz = 0; kz < args.KD; ++kz) {
+            int64_t id = i_d_base + kz * args.d2;
+            if (id < 0 || id >= args.ID) continue; // Boundary check (Padding)
+
+            for (int64_t ky = 0; ky < args.KH; ++ky) {
+                int64_t ih = i_h_base + ky * args.d1;
+                if (ih < 0 || ih >= args.IH) continue;
+
+                for (int64_t kx = 0; kx < args.KW; ++kx) {
+                    int64_t iw = i_w_base + kx * args.d0;
+                    if (iw < 0 || iw >= args.IW) continue;
+
+                    // Convert multi-dimensional coordinates to flat byte offsets
+                    int64_t w_idx = kx*args.nb00 + ky*args.nb01 + kz*args.nb02 + w_cn_idx*args.nb03;
+                    int64_t i_idx = iw*args.nb10 + ih*args.nb11 + id*args.nb12 + src_cn_idx*args.nb13;
+
+                    // Dereference memory and cast weights to f32 if they were f16
+                    float w_val = (float)*(device const T*)((device const char*)src0 + w_idx);
+                    float i_val = *(device const float*)((device const char*)src1 + i_idx);
+
+                    sum += w_val * i_val;
+                }
+            }
+        }
+    }
+
+    // 5. Write the accumulated value out to RAM
+    int64_t dst_cn_idx = batch_idx * args.OC + oc;
+    int64_t d_idx = ow*args.nb0 + oh*args.nb1 + od*args.nb2 + dst_cn_idx*args.nb3;
+
+    *(device float*)(dst + d_idx) = sum;
+}
+
+// Explicit instantiations so the JIT compiler can find them by name
+template [[host_name("kernel_conv_3d_f32_f32")]]
+kernel void kernel_conv_3d<float>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+// Explicit instantiation for f16 weights
+template [[host_name("kernel_conv_3d_f16_f32")]]
+kernel void kernel_conv_3d<half>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char  * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);

ggml/src/ggml-metal/kernels/dequantize.h

@@ -0,0 +1,719 @@
+#pragma once
+
+#include "common.h"
+
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+#include "ggml-common.h"
+#endif
+
+#define QK_NL 16 // shared by mul_mm and get_rows_q instantiations
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+
+#if defined(GGML_METAL_HAS_BF16)
+template <typename type4x4>
+void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+#endif
+
+template <typename type4x4>
+void dequantize_q1_0(device const block_q1_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+    const float neg_d = -d;
+
+    const int byte_offset = il * 2;  // il*16 bits = il*2 bytes
+    const uint8_t b0 = qs[byte_offset];
+    const uint8_t b1 = qs[byte_offset + 1];
+
+    float4x4 reg_f;
+
+    reg_f[0][0] = select(neg_d, d, bool(b0 & 0x01));
+    reg_f[0][1] = select(neg_d, d, bool(b0 & 0x02));
+    reg_f[0][2] = select(neg_d, d, bool(b0 & 0x04));
+    reg_f[0][3] = select(neg_d, d, bool(b0 & 0x08));
+    reg_f[1][0] = select(neg_d, d, bool(b0 & 0x10));
+    reg_f[1][1] = select(neg_d, d, bool(b0 & 0x20));
+    reg_f[1][2] = select(neg_d, d, bool(b0 & 0x40));
+    reg_f[1][3] = select(neg_d, d, bool(b0 & 0x80));
+
+    reg_f[2][0] = select(neg_d, d, bool(b1 & 0x01));
+    reg_f[2][1] = select(neg_d, d, bool(b1 & 0x02));
+    reg_f[2][2] = select(neg_d, d, bool(b1 & 0x04));
+    reg_f[2][3] = select(neg_d, d, bool(b1 & 0x08));
+    reg_f[3][0] = select(neg_d, d, bool(b1 & 0x10));
+    reg_f[3][1] = select(neg_d, d, bool(b1 & 0x20));
+    reg_f[3][2] = select(neg_d, d, bool(b1 & 0x40));
+    reg_f[3][3] = select(neg_d, d, bool(b1 & 0x80));
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const float neg_d = -d;
+    const int base = il * 4;
+    const uint8_t byte = xb->qs[base / 8];
+    const int s = base % 8;
+
+    float4 reg_f;
+    reg_f[0] = select(neg_d, d, bool((byte >> (s    )) & 1));
+    reg_f[1] = select(neg_d, d, bool((byte >> (s + 1)) & 1));
+    reg_f[2] = select(neg_d, d, bool((byte >> (s + 2)) & 1));
+    reg_f[3] = select(neg_d, d, bool((byte >> (s + 3)) & 1));
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q2_0(device const block_q2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    const int byte_offset = il * 4;  // il*16 elements = il*4 bytes (4 elements per byte)
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; i++) {
+        const uint8_t b = qs[byte_offset + i];
+        reg_f[i][0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+        reg_f[i][1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+        reg_f[i][2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+        reg_f[i][3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q2_0_t4(device const block_q2_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const uint8_t b = xb->qs[il];
+
+    float4 reg_f;
+    reg_f[0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+    reg_f[1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+    reg_f[2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+    reg_f[3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
+        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
+
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = ((qs[i] & mask0) * d1) + m;
+        reg_f[i/2][2*(i%2) + 1] = ((qs[i] & mask1) * d2) + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + md;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + m;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 16; i++) {
+        reg_f[i/4][i%4] = (qs[i + 16*il] * d);
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    for (int i = 0; i < 4; i++) {
+        reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d);
+    }
+}
+
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
+    }
+}
+
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
+
+    const uint8_t shr = il >= 4 ? 4 : 0;
+
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint16_t * ql = (device const uint16_t *)xb->ql;
+    device const uint16_t * qh = (device const uint16_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
+    qh = qh + 16*(il/8) + 8*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
+    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
+    const float ml = d_all * sc * 32.f;
+    const float dl0 = d_all * sc;
+    const float dl1 = dl0 / 256.f;
+    const float dl2 = dl0 / (256.f * 256.f);
+    const float dl3 = dl0 / (256.f * 256.f * 256.f);
+    const uint8_t shr_h = il>2 ? 2 : 0;
+    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
+    const uint8_t shr_l = il>1 ? 4 : 0;
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
+        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
+        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
+        reg[i][0] = dl0 *  ((half)(q & 0xFF))       - ml;
+        reg[i][1] = dl1 * ((float)(q & 0xFF00))     - ml;
+        reg[i][2] = dl2 * ((float)(q & 0xFF0000))   - ml;
+        reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+

ggml/src/ggml-metal/kernels/gated_delta_net.metal

@@ -0,0 +1,250 @@
+#include "common.h"
+
+constant short FC_gated_delta_net_ne20 [[function_constant(FC_GATED_DELTA_NET + 0)]];
+constant short FC_gated_delta_net_ne30 [[function_constant(FC_GATED_DELTA_NET + 1)]];
+constant short FC_gated_delta_net_K    [[function_constant(FC_GATED_DELTA_NET + 2)]];
+
+#if 1
+template<short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+#define K   FC_gated_delta_net_K
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B (n_seqs)
+    const uint i21 = tgpig.y; // H (head)
+    const uint i20 = tgpig.x*NSG + ty; // row within S_v
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    // input state layout [S_v, S_v, H, n_seqs] (s0 only): per-seq stride is H*D.
+    // state is stored transposed: M[i20][is] = S[is][i20], so row i20 is contiguous
+    const uint state_in_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+    device const float * s_ptr = (device const float *) (s) + state_in_base;
+
+    float ls[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        ls[j] = s_ptr[is];
+    }
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
+
+    // output state base offset: after attention scores
+    const uint attn_size = args.ne22 * args.ne21 * S_v * args.ne23;
+    // output state per-slot size: S_v * S_v * H * n_seqs
+    const uint state_size_per_snap = S_v * S_v * args.ne21 * args.ne23;
+    // per-(seq,head) offset within a slot
+    const uint state_out_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+
+    for (short t = 0; t < args.ne22; t++) {
+        float s_k = 0.0f;
+
+        if (G == 1) {
+            const float g_exp = exp(g_ptr[0]);
+
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= g_exp;
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        } else {
+            // KDA
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= exp(g_ptr[is]);
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        }
+
+        s_k = simd_sum(s_k);
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        float y = 0.0f;
+
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            ls[j] += k_ptr[is]*d;
+
+            y += ls[j]*q_ptr[is];
+        }
+
+        y = simd_sum(y);
+
+        if (tx == 0) {
+            dst_attn[t*args.ne21*S_v] = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        if (K > 1) {
+            const int target_slot = (int)args.ne22 - 1 - (int)t;
+            if (target_slot >= 0 && target_slot < (int)K) {
+                device float * dst_state = (device float *) (dst) + attn_size + (uint)target_slot * state_size_per_snap + state_out_base;
+                FOR_UNROLL (short j = 0; j < NSG; j++) {
+                    const short is = tx*NSG + j;
+                    dst_state[is] = ls[j];
+                }
+            }
+        }
+    }
+
+    if (K == 1) {
+        device float * dst_state = (device float *) (dst) + attn_size + state_out_base;
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            dst_state[is] = ls[j];
+        }
+    }
+
+#undef S_v
+#undef G
+#undef K
+}
+
+typedef decltype(kernel_gated_delta_net_impl<4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<4>;
+
+#else
+// a simplified version of the above
+// no performance improvement, so keep the above version for now
+
+template<typename T, short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B
+    const uint i21 = tgpig.y; // H
+    const uint i20 = tgpig.x*NSG + ty;
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    float lsf[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        lsf[j] = s_ptr[is*S_v];
+    }
+
+    thread T * ls = (thread T *) (lsf);
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr  = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr  = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    for (short t = 0; t < args.ne22; t++) {
+        device const T * qt_ptr = (device const T *) (q_ptr);
+        device const T * kt_ptr = (device const T *) (k_ptr);
+        device const T * gt_ptr = (device const T *) (g_ptr);
+
+        if (G == 1) {
+            *ls *= exp(g_ptr[0]);
+        } else {
+            // KDA
+            *ls *= exp(gt_ptr[tx]);
+        }
+
+        const float s_k = simd_sum(dot(*ls, kt_ptr[tx]));
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        *ls += kt_ptr[tx]*d;
+
+        const float y = simd_sum(dot(*ls, qt_ptr[tx]));
+
+        if (tx == 0) {
+            *dst_attn = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        dst_attn += args.ne21*S_v;
+    }
+
+    device float * dst_state  = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
+    device T     * dstt_state = (device T     *) (dst_state);
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        dst_state[is*S_v] = lsf[j];
+    }
+
+#undef S_v
+#undef G
+}
+
+typedef decltype(kernel_gated_delta_net_impl<float4, 4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float,  1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float2, 2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float4, 4>;
+#endif

ggml/src/ggml-metal/kernels/misc.metal

@@ -0,0 +1,595 @@
+#include "common.h"
+
+kernel void kernel_argmax_f32(
+        constant ggml_metal_kargs_argmax & args,
+        device   const char * src0,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    device const float * x_row = (device const float *) ((device const char *) src0 + tgpig * args.nb01);
+
+    float   lmax = -INFINITY;
+    int32_t larg = -1;
+
+    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+        if (x_row[i00] > lmax) {
+            lmax = x_row[i00];
+            larg = i00;
+        }
+    }
+
+    // find the argmax value in the block
+    float max_val = simd_max(lmax);
+    int32_t arg_val = simd_max(select(-1, larg, lmax == max_val));
+
+    device int32_t * dst_i32 = (device int32_t *) dst;
+
+    threadgroup   float * shared_maxval = (threadgroup   float *) shmem;
+    threadgroup int32_t * shared_argmax = (threadgroup int32_t *) shmem + N_SIMDWIDTH;
+
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            shared_maxval[tiisg] = -INFINITY;
+            shared_argmax[tiisg] = -1;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            shared_maxval[sgitg] = max_val;
+            shared_argmax[sgitg] = arg_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = shared_maxval[tiisg];
+        arg_val = shared_argmax[tiisg];
+
+        float max_val_reduced   = simd_max(max_val);
+        int32_t arg_val_reduced = simd_max(select(-1, arg_val, max_val == max_val_reduced));
+
+        dst_i32[tgpig] = arg_val_reduced;
+
+        return;
+    }
+
+    dst_i32[tgpig] = arg_val;
+}
+
+kernel void kernel_diag_f32(
+        constant ggml_metal_kargs_diag & args,
+        device   const char * src0,
+        device         char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *)(src0 +                i2*args.nb02 + i3*args.nb03);
+    device       float * dst_ptr  = (device       float *)(dst  + i1*args.nb01 + i2*args.nb2  + i3*args.nb3);
+
+    for (int i0 = tiitg; i0 < args.ne0; i0 += NW) {
+        dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
+    }
+}
+
+kernel void kernel_roll_f32(
+    constant ggml_metal_kargs_roll & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *) src0;
+    device       float * dst_ptr  = (device       float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        // apply shifts and wrap around
+        int64_t i00 = i0 - args.s0;
+        int64_t i01 = i1 - args.s1;
+        int64_t i02 = i2 - args.s2;
+        int64_t i03 = i3 - args.s3;
+
+        if (i00 < 0) { i00 += args.ne00; } else if (i00 >= args.ne00) { i00 -= args.ne00; }
+        if (i01 < 0) { i01 += args.ne01; } else if (i01 >= args.ne01) { i01 -= args.ne01; }
+        if (i02 < 0) { i02 += args.ne02; } else if (i02 >= args.ne02) { i02 -= args.ne02; }
+        if (i03 < 0) { i03 += args.ne03; } else if (i03 >= args.ne03) { i03 -= args.ne03; }
+
+        int64_t src_idx = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00 + i00;
+        int64_t dst_idx = i3 *args.ne2 *args.ne1 *args.ne0  + i2 *args.ne1 *args.ne0  + i1 *args.ne0  + i0;
+
+        dst_ptr[dst_idx] = src0_ptr[src_idx];
+    }
+}
+
+template <typename T>
+kernel void kernel_pad_impl(
+    constant ggml_metal_kargs_pad & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t k0 = tgpig.x/args.ne1;
+    const int32_t i1 = tgpig.x - k0*args.ne1;
+
+    const int32_t i03 = i3;
+    const int32_t i02 = i2;
+    const int32_t i01 = i1;
+
+    device const T * src0_ptr = (device const T *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T * dst_ptr  = (device       T *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    for (int32_t l0 = 0; l0 < 1024; l0 += ntg.x) {
+        const int32_t i0 = k0*1024 + tpitg.x + l0;
+        if (i0 >= args.ne0) {
+            break;
+        }
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            dst_ptr[i0] = src0_ptr[i0];
+        } else {
+            dst_ptr[i0] = 0.0f;
+        }
+    }
+}
+
+typedef decltype(kernel_pad_impl<float>) kernel_pad_t;
+
+template [[host_name("kernel_pad_f32")]]   kernel kernel_pad_t kernel_pad_impl<float>;
+template [[host_name("kernel_pad_f32_4")]] kernel kernel_pad_t kernel_pad_impl<float4>;
+
+// TODO: this is slow - optimize
+kernel void kernel_pad_reflect_1d_f32(
+    constant   ggml_metal_kargs_pad_reflect_1d & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3  tgpg[[threadgroups_per_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            if (i0 < args.p0) {
+                dst_ptr[i0] = src0_ptr[args.p0 - i0];
+            } else if (i0 < args.ne0 - args.p1) {
+                dst_ptr[i0] = src0_ptr[i0 - args.p0];
+            } else {
+                dst_ptr[i0] = src0_ptr[(args.ne0 - args.p1 - args.p0) - (args.p1 + 1 - (args.ne0 - i0)) - 1];
+            }
+        }
+    }
+}
+
+kernel void kernel_arange_f32(
+    constant   ggml_metal_kargs_arange & args,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    device float * dst_ptr = (device float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_ptr[i0] = args.start + args.step * i0;
+    }
+}
+
+kernel void kernel_timestep_embedding_f32(
+    constant  ggml_metal_kargs_timestep_embedding & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    int i = tgpig.x;
+    device float * embed_data = (device float *)(dst + i*args.nb1);
+
+    int half_ = args.dim / 2;
+    for (int j = tpitg.x; j < half_; j += ntg.x) {
+        float timestep = ((device float *)src0)[i];
+        float freq = (float)exp(-log((float)args.max_period) * j / half_);
+        float arg = timestep * freq;
+        embed_data[j        ] = cos(arg);
+        embed_data[j + half_] = sin(arg);
+    }
+
+    if (args.dim % 2 != 0 && tpitg.x == 0) {
+        embed_data[2 * half_] = 0.f;
+    }
+}
+
+kernel void kernel_opt_step_adamw_f32(
+        constant    ggml_metal_kargs_opt_step_adamw & args,
+        device       float * x,
+        device const float * g,
+        device       float * g_m,
+        device       float * g_v,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
+    const float gi = g[gid];
+    const float gmi = g_m[gid] * beta1 +      gi * (1.0f - beta1);
+    const float gvi = g_v[gid] * beta2 + gi * gi * (1.0f - beta2);
+
+    g_m[gid] = gmi;
+    g_v[gid] = gvi;
+
+    const float mh =      gmi * beta1h;
+    const float vh = sqrt(gvi * beta2h) + eps;
+
+    x[gid] = x[gid] * (1.0f - alpha * wd) - alpha * mh / vh;
+}
+
+kernel void kernel_opt_step_sgd_f32(
+        constant    ggml_metal_kargs_opt_step_sgd & args,
+        device       float * x,
+        device const float * g,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    x[gid] = x[gid] * (1.0f - pars[0] * pars[1]) - pars[0] * g[gid];
+}
+
+template<typename T>
+kernel void kernel_memset(
+        constant ggml_metal_kargs_memset & args,
+        device T * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = args.val;
+}
+
+typedef decltype(kernel_memset<int64_t>) kernel_memset_t;
+
+template [[host_name("kernel_memset_i64")]] kernel kernel_memset_t kernel_memset<int64_t>;
+
+constant short FC_count_equal_nsg [[function_constant(FC_COUNT_EQUAL + 0)]];
+
+template<typename T>
+kernel void kernel_count_equal(
+        constant ggml_metal_kargs_count_equal & args,
+        device   const char * src0,
+        device   const char * src1,
+        device   atomic_int * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const short NSG = FC_count_equal_nsg;
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    int sum = 0;
+
+    device const char * base0 = src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03;
+    device const char * base1 = src1 + i1*args.nb11 + i2*args.nb12 + i3*args.nb13;
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        const T v0 = *(device const T *)(base0 + i0*args.nb00);
+        const T v1 = *(device const T *)(base1 + i0*args.nb10);
+        sum += (v0 == v1);
+    }
+
+    sum = simd_sum(sum);
+
+    if (tiisg == 0) {
+        shmem_i32[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        float v = 0.0f;
+        if (tpitg.x < NSG) {
+            v = shmem_i32[tpitg.x];
+        }
+
+        float total = simd_sum(v);
+        if (tpitg.x == 0) {
+            atomic_fetch_add_explicit(dst, (int32_t) total, memory_order_relaxed);
+        }
+    }
+}
+
+typedef decltype(kernel_count_equal<int32_t>) kernel_count_equal_t;
+
+template [[host_name("kernel_count_equal_i32")]] kernel kernel_count_equal_t kernel_count_equal<int32_t>;
+
+template <typename T>
+kernel void kernel_snake(
+        constant ggml_metal_kargs_snake & args,
+        device const T     * x,
+        device const float * a,
+        device const float * inv_b,
+        device       T     * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T * args.C) {
+        return;
+    }
+
+    const int   c  = idx / args.T;  // x is [T, C], a / inv_b collapse to [1, C]
+    const float xi = float(x[idx]);
+    const float si = sin(a[c] * xi);
+    dst[idx] = T(xi + si * si * inv_b[c]);
+}
+
+template [[host_name("kernel_snake_f32")]]  kernel void kernel_snake<float>(constant ggml_metal_kargs_snake &, device const float *, device const float *, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_snake_f16")]]  kernel void kernel_snake<half>(constant ggml_metal_kargs_snake &, device const half *, device const float *, device const float *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_snake_bf16")]] kernel void kernel_snake<bfloat>(constant ggml_metal_kargs_snake &, device const bfloat *, device const float *, device const float *, device bfloat *, uint, uint, uint);
+#endif
+
+template<int N>
+kernel void kernel_fwht_f32(
+        constant ggml_metal_kargs_fwht & args,
+        device const float * src,
+        device float * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort3  ntg[[threads_per_threadgroup]]) {
+
+    constexpr int NW = N_SIMDWIDTH;
+    constexpr int NE = N / NW;
+
+    const float scale = 1.0f / sqrt((float) N);
+
+    const int sg_per_tg = ntg.x / NW;
+    const int64_t r = tgpig.x * sg_per_tg + sgitg;
+    if (r >= args.nrows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    const int lane = tiisg;
+
+    float reg[NE];
+    for (int i = 0; i < NE; i++) {
+        reg[i] = src[i*NW + lane]*scale;
+    }
+    for (int i = 1; i < NW; i *= 2) {
+        for (int j = 0; j < NE; j++) {
+            const float val = reg[j];
+            const float val2 = simd_shuffle_xor(val, i);
+            reg[j] = (lane & i) == 0 ? val2 + val : val2 - val;
+        }
+    }
+
+    for (int i = NW; i < N; i *= 2) {
+        const int step = i / NW;
+        for (int j = 0; j < NE; j += (2 * step)) {
+            for (int k = 0; k < step; k++) {
+                const float x = reg[j + k ];
+                const float y = reg[j + k + step];
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+    for (int i = 0; i < NE; i++) {
+        dst[i*NW + lane] = reg[i];
+    }
+}
+
+typedef decltype(kernel_fwht_f32<64>) kernel_fwht_t;
+
+template [[host_name("kernel_fwht_f32_64")]]  kernel kernel_fwht_t kernel_fwht_f32<64>;
+template [[host_name("kernel_fwht_f32_128")]] kernel kernel_fwht_t kernel_fwht_f32<128>;
+template [[host_name("kernel_fwht_f32_256")]] kernel kernel_fwht_t kernel_fwht_f32<256>;
+template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_t kernel_fwht_f32<512>;
+
+kernel void kernel_dsv4_hc_comb_f32(
+        constant ggml_metal_kargs_dsv4_hc_comb & args,
+        device const char * mixes,
+        device const char * scale,
+        device const char * base,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+    constexpr ushort comb_offset = 2*hc;
+
+    const int it = tgpig.x*ntg.y + sgitg;
+    if (it >= args.n_tokens) {
+        return;
+    }
+
+    float scale_lane = 0.0f;
+    if (tiisg == 0) {
+        scale_lane = *(device const float *) (scale + 2*args.nb_s0);
+    }
+    const float scale_comb = simd_shuffle(scale_lane, 0);
+
+    float v = 0.0f;
+    if (tiisg < hc*hc) {
+        v = *(device const float *) (mixes + (comb_offset + tiisg)*args.nb_m0 + it*args.nb_m1)*scale_comb
+          + *(device const float *) (base   + (comb_offset + tiisg)*args.nb_b0);
+    }
+
+    // Softmax across destinations (the four contiguous lanes for each source).
+    float vmax = max(v, simd_shuffle_xor(v, 1));
+    vmax = max(vmax, simd_shuffle_xor(vmax, 2));
+    v = exp(v - vmax);
+
+    float sum = v + simd_shuffle_xor(v, 1);
+    sum += simd_shuffle_xor(sum, 2);
+    v = v/sum + args.eps;
+
+    // Normalize columns: equal destination indices are four lanes apart.
+    sum = v + simd_shuffle_xor(v, 4);
+    sum += simd_shuffle_xor(sum, 8);
+    v /= sum + args.eps;
+
+    for (int i = 1; i < args.n_iter; ++i) {
+        sum = v + simd_shuffle_xor(v, 1);
+        sum += simd_shuffle_xor(sum, 2);
+        v /= sum + args.eps;
+
+        sum = v + simd_shuffle_xor(v, 4);
+        sum += simd_shuffle_xor(sum, 8);
+        v /= sum + args.eps;
+    }
+
+    if (tiisg < hc*hc) {
+        const ushort idst = tiisg & 3;
+        const ushort isrc = tiisg >> 2;
+        *(device float *) (dst + idst*args.nb_d0 + isrc*args.nb_d1 + it*args.nb_d2) = v;
+    }
+}
+
+kernel void kernel_dsv4_hc_pre_f32(
+        constant ggml_metal_kargs_dsv4_hc_pre & args,
+        device const char * x,
+        device const char * weights,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+
+    const int it = tgpig.y;
+    const int i0 = ((int) tgpig.x*ntg.y + sgitg)*32 + tiisg;
+
+    float weight_lane = 0.0f;
+    if (tiisg < hc) {
+        weight_lane = *(device const float *) (weights + tiisg*args.nb_w0 + it*args.nb_w1);
+    }
+
+    float w[hc];
+    FOR_UNROLL (ushort ih = 0; ih < hc; ++ih) {
+        w[ih] = simd_shuffle(weight_lane, ih);
+    }
+
+    if (i0 >= args.n_embd) {
+        return;
+    }
+
+    device const char * xb = x + i0*args.nb_x0 + it*args.nb_x2;
+    float result = 0.0f;
+    FOR_UNROLL (ushort ih = 0; ih < hc; ++ih) {
+        result = fma(*(device const float *) (xb + ih*args.nb_x1), w[ih], result);
+    }
+
+    *(device float *) (dst + i0*args.nb_d0 + it*args.nb_d1) = result;
+}
+
+kernel void kernel_dsv4_hc_post_f32(
+        constant ggml_metal_kargs_dsv4_hc_post & args,
+        device const char * x,
+        device const char * residual,
+        device const char * post,
+        device const char * comb,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+
+    const int it = tgpig.y;
+    const int i0 = ((int) tgpig.x*ntg.y + sgitg)*32 + tiisg;
+
+    float coeff_lane = 0.0f;
+    if (tiisg < hc) {
+        coeff_lane = *(device const float *) (post + tiisg*args.nb_p0 + it*args.nb_p1);
+    } else if (tiisg < hc + hc*hc) {
+        const ushort idx  = tiisg - hc;
+        const ushort idst = idx & 3;
+        const ushort isrc = idx >> 2;
+        coeff_lane = *(device const float *) (comb + idst*args.nb_c0 + isrc*args.nb_c1 + it*args.nb_c2);
+    }
+
+    float post_reg[hc];
+    float comb_reg[hc][hc];
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        post_reg[idst] = simd_shuffle(coeff_lane, idst);
+    }
+    FOR_UNROLL (ushort isrc = 0; isrc < hc; ++isrc) {
+        FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+            comb_reg[isrc][idst] = simd_shuffle(coeff_lane, hc + idst + hc*isrc);
+        }
+    }
+
+    if (i0 >= args.n_embd) {
+        return;
+    }
+
+    const float xv = *(device const float *) (x + i0*args.nb_x0 + it*args.nb_x1);
+    float result[hc];
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        result[idst] = xv*post_reg[idst];
+    }
+
+    device const char * rb = residual + i0*args.nb_r0 + it*args.nb_r2;
+    FOR_UNROLL (ushort isrc = 0; isrc < hc; ++isrc) {
+        const float rv = *(device const float *) (rb + isrc*args.nb_r1);
+        FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+            result[idst] = fma(rv, comb_reg[isrc][idst], result[idst]);
+        }
+    }
+
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        *(device float *) (dst + i0*args.nb_d0 + idst*args.nb_d1 + it*args.nb_d2) = result[idst];
+    }
+}

ggml/src/ggml-metal/kernels/mul_mm.metal

@@ -0,0 +1,842 @@
+#include "common.h"
+#include "dequantize.h"
+
+constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
+constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
+constant short FC_mul_mm_ne12  [[function_constant(FC_MUL_MM + 2)]];
+constant short FC_mul_mm_ne13  [[function_constant(FC_MUL_MM + 3)]];
+constant short FC_mul_mm_r2    [[function_constant(FC_MUL_MM + 4)]];
+constant short FC_mul_mm_r3    [[function_constant(FC_MUL_MM + 5)]];
+
+// each block_q contains 16*nl weights
+#ifdef GGML_METAL_HAS_TENSOR
+template<
+    typename SA, typename SA_4x4, typename SA_8x8,
+    typename SB, typename SB_2x4, typename SB_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread SA_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    // Matrix dimensions: A(M,K) x B(K,N) -> C(M,N)
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    // Batch dimension handling
+    const int im = tgpig.z;
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    // Batch offsets for srcA and srcB
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+
+    // Tile dimensions
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+
+    // Tile offsets in output matrix
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // Threadgroup memory for dequantized A tile only
+    threadgroup SA * sa = (threadgroup SA *)(shmem);
+
+    // Work-item count for A loading
+    constexpr int A_WORK_ITEMS = NRA * N_MM_NK;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+
+    // tA wraps threadgroup memory
+    auto tA = tensor(sa, dextents<int32_t, 2>(N_MM_NK_TOTAL, NRA));
+
+    // tB wraps device memory directly
+    device T1 * ptrB = (device T1 *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(T1);
+    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
+
+    // Configure matmul operation
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    // Accumulate partial results over K dimension
+    for (int loop_k = 0; loop_k < K; loop_k += N_MM_NK_TOTAL) {
+        // === PHASE 1: Dequantization of A into threadgroup memory ===
+        for (int work = tiitg; work < A_WORK_ITEMS; work += NUM_THREADS) {
+            const int row = work / N_MM_NK;
+            const int k_chunk = work % N_MM_NK;
+            const int k_pos = loop_k + k_chunk * 16;
+            const short k_base = k_chunk * 16;
+
+            // Bounds check: skip device read if row is out of matrix bounds
+            if (ra + row < M) {
+                if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+                    // Element-wise reads when K is not aligned (nb01 not aligned for half4x4/float4x4).
+                    // MSL spec Table 2.5: half4x4 requires 8-byte alignment. When K is odd,
+                    // nb01 = K*2 is not 8-byte aligned, so odd-row pointers are misaligned.
+                    // Mirrors the legacy kernel's existing guard.
+                    device const T0 * row_ptr = (device const T0 *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? (SA) row_ptr[k_pos + i] : (SA)0;
+                    }
+                } else {
+                    const int block_idx = k_pos / (16 * nl);
+                    const short il = (k_pos / 16) % nl;
+
+                    device const block_q * row_ptr = (device const block_q *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    SA_4x4 temp_a;
+                    dequantize_func(row_ptr + block_idx, il, temp_a);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        // Zero-pad A for K positions beyond valid range (handles partial K iterations)
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? temp_a[i/4][i%4] : (SA)0;
+                    }
+                }
+            } else {
+                // Zero-pad rows beyond matrix bounds
+                FOR_UNROLL (short i = 0; i < 16; i++) {
+                    sa[row * N_MM_NK_TOTAL + (k_base + i)] = (SA)0;
+                }
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // === PHASE 2: Tensor matmul ===
+        auto mA = tA.slice(0, 0);
+        auto mB = tB.slice(loop_k, rb);
+
+        mm.run(mB, mA, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // Store result tile to output matrix (with batch offset)
+    // cT.store handles bounds checking via tD's extents (M, N)
+    device float * dstBatch = (device float *)dst + im * N * M;
+
+    auto tD = tensor(dstBatch, dextents<int32_t, 2>(M, N), array<int, 2>({1, M}));
+    cT.store(tD.slice(ra, rb));
+}
+
+#else
+
+template<
+    typename S0, typename S0_4x4, typename S0_8x8,
+    typename S1, typename S1_2x4, typename S1_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+    }
+
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
+        // if no bounds checks on the output are needed, we can directly write to device memory
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (sgitg == 0) {
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*NR0);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < nr0/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < nr0; i++) {
+                    *(D + i) = *(C + i);
+                }
+            }
+        }
+    }
+}
+
+#endif // GGML_METAL_HAS_TENSOR
+
+template<short ne20> // n_expert_used
+kernel void kernel_mul_mm_id_map0(
+        constant ggml_metal_kargs_mul_mm_id_map0 & args,
+        device  const char * src2,
+        device        char * htpe,
+        device        char * hids,
+        threadgroup   char * shmem [[threadgroup(0)]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const short ide = tpitg; // expert id
+
+    uint32_t n_all = 0;
+
+    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;
+
+    for (int i21 = 0; i21 < args.ne21; i21 += ntg) { // n_tokens
+        if (i21 + tpitg < args.ne21) {
+            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);
+
+            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;
+
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sids[i20] = src2_i32[i20];
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (short t = 0; t < ntg; t++) {
+            if (i21 + t >= args.ne21) {
+                break;
+            }
+
+            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;
+
+            short sel = 0;
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sel += (sids[i20] == ide)*(i20 + 1);
+            }
+
+            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;
+
+            n_all += sel > 0;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
+    tpe_u32[ide] = n_all;
+}
+
+typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
+
+template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;
+
+template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm_id(
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * htpe,
+        device const char * hids,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
+    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
+
+    const int32_t neh1 = tpe_u32[im];
+
+    if (r1 >= neh1) {
+        return;
+    }
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
+
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? (S0) *((device T0 *) x + i) : (S0) 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifndef GGML_METAL_HAS_TENSOR
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
+    }
+
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+    for (short i = 0; i < 8; i++) {
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+    }
+#endif
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
+
+        const short ide = id % args.ne20;
+        const short idt = id / args.ne20;
+
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float4 * D4 = (device float4 *) D;
+
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
+        threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+        int i = tiisg;
+        for (; i < nr0/4; i += 32) {
+            *(D4 + i) = *(C4 + i);
+        }
+
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
+            *(D + i) = *(C + i);
+        }
+    }
+}
+
+//
+// matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_t;
+
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_q1_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_f32_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_f16_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_q1_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_id;
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_id_q1_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q1_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;

ggml/src/ggml-metal/kernels/norm.metal

@@ -0,0 +1,308 @@
+#include "common.h"
+
+// F == 1 : norm (no fuse)
+// F == 2 : norm + mul
+// F == 3 : norm + mul + add
+template <typename T, short F>
+kernel void kernel_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    T sumft(0.0f);
+
+    float sumf = 0.0f;
+
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumft += x[i00];
+    }
+    sumf = dot(sumft, T(1.0f));
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean = sumf/args.ne00;
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+
+    sumf = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        y[i00] = x[i00] - mean;
+        sumf += dot(y[i00], y[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float variance = sumf/args.ne00;
+
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (y[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (y[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (y[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_norm_fuse_impl<float4, 1>) kernel_norm_fuse_t;
+
+template [[host_name("kernel_norm_f32")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_norm_mul_f32")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_norm_mul_add_f32")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_norm_f32_4")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_norm_mul_f32_4")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_norm_mul_add_f32_4")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 3>;
+
+// F == 1 : rms_norm (no fuse)
+// F == 2 : rms_norm + mul
+// F == 3 : rms_norm + mul + add
+template <typename T, short F>
+kernel void kernel_rms_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean  = sumf/args.ne00;
+    const float scale = 1.0f/sqrt(mean + args.eps);
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (x[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (x[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (x[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_rms_norm_fuse_impl<float4, 1>) kernel_rms_norm_fuse_t;
+
+template [[host_name("kernel_rms_norm_f32")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_rms_norm_mul_f32")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_rms_norm_f32_4")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_rms_norm_mul_f32_4")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32_4")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 3>;
+
+template <typename T0, typename T>
+kernel void kernel_l2_norm_impl(
+        constant ggml_metal_kargs_l2_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const T0 * x = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T  * y = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float scale = 1.0f/max(sqrt(sumf), args.eps);
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        y[i00] = x[i00] * scale;
+    }
+}
+
+typedef decltype(kernel_l2_norm_impl<float, float>) kernel_l2_norm_t;
+
+template [[host_name("kernel_l2_norm_f32_f32")]]   kernel kernel_l2_norm_t kernel_l2_norm_impl<float,  float>;
+template [[host_name("kernel_l2_norm_f32_f32_4")]] kernel kernel_l2_norm_t kernel_l2_norm_impl<float4, float4>;
+
+kernel void kernel_group_norm_f32(
+        constant ggml_metal_kargs_group_norm & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    const int64_t ne = args.ne00*args.ne01*args.ne02;
+    const int64_t gs = args.ne00*args.ne01*((args.ne02 + args.ngrp - 1) / args.ngrp);
+
+    int start = tgpig * gs;
+    int end   = start + gs;
+
+    start += tpitg;
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += ntg) {
+        tmp += src0[j];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float mean = tmp / gs;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += ntg) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float variance = tmp / gs;
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int j = start; j < end; j += ntg) {
+        dst[j] *= scale;
+    }
+}

ggml/src/ggml-metal/kernels/pool.metal

@@ -0,0 +1,148 @@
+#include "common.h"
+
+kernel void kernel_pool_2d_max_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+
+    float res = -INFINITY;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            res = MAX(res, i_ptr[i * args.IW + j]);
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+kernel void kernel_pool_2d_avg_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+    // const float scale = 1. / ((eh - bh) * (ew - bw));
+    const float scale = 1. / (args.k0 * args.k1);
+
+    float res = 0;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            float cur = i_ptr[i * args.IW + j];
+            res += cur * scale;
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+
+kernel void kernel_pool_1d_max_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = -INFINITY;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        int j = base + ki;
+        if (j < 0 || j >= args.IW){
+            continue;
+        }
+        float v = src[src_off + j];
+        acc = max(acc, v);
+    }
+
+    dst[dst_off + ow] = acc;
+}
+
+kernel void kernel_pool_1d_avg_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = 0.0f;
+    int   cnt = 0;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        const int j = base + ki;
+        if (j < 0 || j >= args.IW) {
+            continue;
+        }
+        acc += src[src_off + j];
+        cnt += 1;
+    }
+
+    dst[dst_off + ow] = (cnt > 0) ? (acc / (float)cnt) : 0.0f;
+}

ggml/src/ggml-metal/kernels/quantize.h

@@ -0,0 +1,234 @@
+#pragma once
+
+#include "common.h"
+
+void quantize_q1_0(device const float * src, device block_q1_0 & dst) {
+    float sum_abs = 0.0f;
+    for (int j = 0; j < QK1_0; j++) {
+        sum_abs += fabs(src[j]);
+    }
+    dst.d = sum_abs / QK1_0;
+
+    for (int j = 0; j < QK1_0 / 8; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK1_0; j++) {
+        if (src[j] >= 0.0f) {
+            dst.qs[j / 8] |= (1 << (j % 8));
+        }
+    }
+}
+
+void quantize_q2_0(device const float * src, device block_q2_0 & dst) {
+    float amax = 0.0f;
+    for (int j = 0; j < QK2_0; j++) {
+        float a = fabs(src[j]);
+        if (a > amax) amax = a;
+    }
+    const float d = amax;
+    dst.d = d;
+
+    const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+    for (int j = 0; j < QK2_0 / 4; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK2_0; j++) {
+        int q = (int)round(src[j] * id) + 1;
+        q = max(0, min(3, q));
+        dst.qs[j / 4] |= (q << (2 * (j % 4)));
+    }
+}
+
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}

ggml/src/ggml-metal/kernels/quantize.metal

@@ -0,0 +1,393 @@
+#include "common.h"
+#include "dequantize.h"
+#include "quantize.h"
+
+template<typename T0, typename T1>
+kernel void kernel_cpy_t_t(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.ne00;) {
+        device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        dst_data[i00] = (T1) src[0];
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_t_t<float, float>) kernel_cpy_t;
+
+template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   float>;
+template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
+template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
+template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
+#endif
+template [[host_name("kernel_cpy_f16_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    float>;
+template [[host_name("kernel_cpy_f16_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  float>;
+template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  bfloat>;
+#endif
+
+template<short QK,
+         typename block_q,
+         void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_cpy_f32_q(
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK;
+
+    device block_q * dst_data = (device block_q *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        device const float * src = (device const float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + (i00*QK)*args.nb00);
+
+        quantize_func(src, dst_data[i00]);
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_f32_q<QK8_0,  block_q8_0,  quantize_q8_0>)  cpy_f_q_t;
+
+template [[host_name("kernel_cpy_f32_q8_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK8_0,  block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_cpy_f32_q1_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK1_0,  block_q1_0,   quantize_q1_0>;
+template [[host_name("kernel_cpy_f32_q2_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK2_0,  block_q2_0,   quantize_q2_0>;
+template [[host_name("kernel_cpy_f32_q4_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_0,  block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_cpy_f32_q4_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_1,  block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_cpy_f32_q5_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_0,  block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_cpy_f32_q5_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_1,  block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_cpy_f32_iq4_nl")]] kernel cpy_f_q_t kernel_cpy_f32_q<QK4_NL, block_iq4_nl, quantize_iq4_nl>;
+
+template<typename T4x4, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_cpy_q_f32(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device const block_q * src_data = (device const block_q *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T4x4    * dst_data = (device       T4x4    *)(dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        T4x4 temp;
+        dequantize_func(src_data + i00/nl, i00%nl, temp);
+        dst_data[i00] = temp;
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>) cpy_q_f_t;
+
+template [[host_name("kernel_cpy_q1_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_q1_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template<typename T>
+kernel void kernel_concat(
+        constant ggml_metal_kargs_concat & args,
+        device  const char * src0,
+        device  const char * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = ntg.y == 1 ? tgpig.x : tgpig.x*ntg.y + tpitg.y;
+
+    if (i1 >= args.ne1) {
+        return;
+    }
+
+    int o[4] = {0, 0, 0, 0};
+    o[args.dim] = args.dim == 0 ? args.ne00 : (args.dim == 1 ? args.ne01 : (args.dim == 2 ? args.ne02 : args.ne03));
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        device const T * x;
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            x = (device const T *)(src0 + (i3       )*args.nb03 + (i2       )*args.nb02 + (i1       )*args.nb01 + (i0       )*args.nb00);
+        } else {
+            x = (device const T *)(src1 + (i3 - o[3])*args.nb13 + (i2 - o[2])*args.nb12 + (i1 - o[1])*args.nb11 + (i0 - o[0])*args.nb10);
+        }
+
+        device T * y = (device T *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+        *y = *x;
+    }
+}
+
+typedef decltype(kernel_concat<float>) kernel_concat_t;
+
+template [[host_name("kernel_concat_f32")]]  kernel kernel_concat_t kernel_concat<float>;
+template [[host_name("kernel_concat_f16")]]  kernel kernel_concat_t kernel_concat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_concat_bf16")]] kernel kernel_concat_t kernel_concat<bfloat>;
+#endif
+template [[host_name("kernel_concat_i8")]]   kernel kernel_concat_t kernel_concat<char>;
+template [[host_name("kernel_concat_i16")]]  kernel kernel_concat_t kernel_concat<short>;
+template [[host_name("kernel_concat_i32")]]  kernel kernel_concat_t kernel_concat<int>;
+template [[host_name("kernel_concat_i64")]]  kernel kernel_concat_t kernel_concat<long>;
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
+kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg  [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (device const block_q *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (device      float4x4 *) ((      device char *) dst  + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        float4x4 temp;
+        dequantize_func(psrc + ind/nl, ind%nl, temp);
+        pdst[ind] = temp;
+
+        break;
+    }
+}
+
+template<typename T0, typename T>
+kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (const device T0 *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (      device T  *) ((      device char *)  dst + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        pdst[ind] = psrc[ind];
+
+        break;
+    }
+}
+
+typedef decltype(kernel_get_rows_f<float, float>) get_rows_f_t;
+
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_f_t kernel_get_rows_f<float, float>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_f_t kernel_get_rows_f<half,  float>;
+template [[host_name("kernel_get_rows_i32")]]  kernel get_rows_f_t kernel_get_rows_f<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_rows_f<bfloat, float>;
+#endif
+
+typedef decltype(kernel_get_rows_q<block_q4_0, 2, dequantize_q4_0>) get_rows_q_t;
+
+template [[host_name("kernel_get_rows_q1_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q1_0,    8, dequantize_q1_0>;
+template [[host_name("kernel_get_rows_q2_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_0,    4, dequantize_q2_0>;
+template [[host_name("kernel_get_rows_q4_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_0,    2, dequantize_q4_0>;
+template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_1,    2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
+template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_mxfp4")]]   kernel get_rows_q_t kernel_get_rows_q<block_mxfp4,   2, dequantize_mxfp4>;
+template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_get_rows_q5_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_get_rows_q6_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+
+template<typename TS, typename TI, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + 32*ind, dst_row[ind]);
+    }
+}
+
+template<typename TS, typename TI, typename TD>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device TD * dst_row = (      device TD *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS * src_row = (const device TS *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (TD) src_row[ind];
+    }
+}
+
+typedef decltype(kernel_set_rows_f<float, int64_t, float>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32_i64_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, float>;
+template [[host_name("kernel_set_rows_f32_i32_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, float>;
+template [[host_name("kernel_set_rows_f32_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, half>;
+template [[host_name("kernel_set_rows_f32_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_f32_i64_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_f32_i32_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int32_t, bfloat>;
+#endif
+
+template [[host_name("kernel_set_rows_f16_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int64_t, half>;
+template [[host_name("kernel_set_rows_f16_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_bf16_i64_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_bf16_i32_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int32_t, bfloat>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<float, int64_t, block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_f32_i64_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i32_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i32_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i32_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i64_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i32_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i64_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i32_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i64_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_set_rows_f32_i32_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_iq4_nl, quantize_iq4_nl>;
+

ggml/src/ggml-metal/kernels/reduce.metal

@@ -0,0 +1,228 @@
+#include "common.h"
+
+kernel void kernel_op_sum_f32(
+        constant ggml_metal_kargs_sum & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    if (args.np == 0) {
+        return;
+    }
+
+    // TODO: become function constant
+    const uint nsg = (ntg.x + 31) / 32;
+
+    float sumf = 0;
+
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+        sumf += src0[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float total = 0;
+
+    if (sgitg == 0) {
+        float v = 0;
+
+        if (tpitg.x < nsg) {
+            v = shmem_f32[tpitg.x];
+        }
+
+        total = simd_sum(v);
+
+        if (tpitg.x == 0) {
+            dst[0] = total;
+        }
+    }
+}
+
+constant short FC_sum_rows_op [[function_constant(FC_SUM_ROWS + 0)]];
+
+template <typename T0, typename T>
+kernel void kernel_sum_rows_impl(
+        constant ggml_metal_kargs_sum_rows & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_sum_rows_op
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    threadgroup T0 * shmem_t = (threadgroup T0 *) shmem;
+
+    if (sgitg == 0) {
+        shmem_t[tiisg] = 0.0f;
+    }
+
+    device const T0 * src_row = (device const T0 *) (src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T  * dst_row = (device       T  *) (dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    T0 sumf = T0(0.0f);
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_t[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_t[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        if (FC_OP == OP_SUM_ROWS_NUM_MEAN) {
+            if (is_same<float4, T0>::value) {
+                dst_row[0] = sum(sumf) / (4*args.ne00);
+            } else {
+                dst_row[0] = sum(sumf) / args.ne00;
+            }
+        } else {
+            dst_row[0] = sum(sumf);
+        }
+    }
+
+#undef FC_OP
+}
+
+typedef decltype(kernel_sum_rows_impl<float, float>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows_f32_f32")]]   kernel kernel_sum_rows_t kernel_sum_rows_impl<float,  float>;
+template [[host_name("kernel_sum_rows_f32_f32_4")]] kernel kernel_sum_rows_t kernel_sum_rows_impl<float4, float>;
+
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;

ggml/src/ggml-metal/kernels/rope.metal

@@ -0,0 +1,318 @@
+#include "common.h"
+
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+constant bool FC_rope_is_back   [[function_constant(FC_ROPE + 1)]];
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+    if (FC_rope_is_back) {
+        *sin_theta *= -1.0f;
+    }
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
+}
+
+template<typename T>
+kernel void kernel_rope_norm(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[1];
+
+            dst_data[0] = x0*cos_theta - x1*sin_theta;
+            dst_data[1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_neox(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_multi(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            // mrope theta calculations
+            // note: the rest is the same as kernel_rope_neox
+            const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3;
+            const int sec_w01   = args.sect_0 + args.sect_1;               // end of section 1
+            const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
+            const int sector    = ic % sect_dims;
+
+            float theta_base;
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            } else {
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            }
+            // end of mrope
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_vision(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < 2*args.n_dims) { // different from kernel_rope_multi
+            const int ic = i0/2;
+
+            // mrope theta calculations (only support 2 dimensions)
+            const int sect_dims = args.sect_0 + args.sect_1;
+            const int sector    = ic % sect_dims;
+
+            float p;
+            float theta_base;
+            if (sector < args.sect_1) {
+                p = (float) sector;
+                theta_base = (float) pos[i2];
+            } else {
+                p = (float) sector - args.sect_0;
+                theta_base = (float) pos[i2 + args.ne02];
+            }
+
+            const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
+            // end of mrope
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims]; // different from kernel_rope_multi
+
+            dst_data[0]           = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
+typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+typedef decltype(kernel_rope_multi<float>) kernel_rope_multi_t;
+typedef decltype(kernel_rope_vision<float>) kernel_rope_vision_t;
+
+template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
+template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
+
+template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
+template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
+
+template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi<float>;
+template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi<half>;
+
+template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision<float>;
+template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision<half>;

ggml/src/ggml-metal/kernels/softmax.metal

@@ -0,0 +1,223 @@
+#include "common.h"
+
+template<typename T>
+kernel void kernel_soft_max(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float * psrc2 = src2 != src0 ? (device const float *) (src2)                                                 : nullptr;
+    device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+
+    // find the max value in the block
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
+        lsum += exp_psrc0;
+        pdst[i00] = exp_psrc0;
+    }
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        pdst[i00] *= inv_sum;
+    }
+}
+
+template<typename T>
+kernel void kernel_soft_max_4(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float *  psrc2 = src2 != src0 ? (device const float * ) (src2)                                                 : nullptr;
+    device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
+    }
+
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        pdst4[i00] *= inv_sum;
+    }
+}
+
+typedef decltype(kernel_soft_max<float>)    kernel_soft_max_t;
+typedef decltype(kernel_soft_max_4<float4>) kernel_soft_max_4_t;
+
+template [[host_name("kernel_soft_max_f16")]]   kernel kernel_soft_max_t   kernel_soft_max<half>;
+template [[host_name("kernel_soft_max_f32")]]   kernel kernel_soft_max_t   kernel_soft_max<float>;
+template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<half4>;
+template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<float4>;

ggml/src/ggml-metal/kernels/solve_tri.metal

@@ -0,0 +1,75 @@
+#include "common.h"
+
+constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
+constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
+constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
+
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device   const char * src0,
+        device   const char * src1,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        ushort3 tgpig[[threadgroup_position_in_grid]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const short NSG = FC_solve_tri_nsg;
+    const short N   = FC_solve_tri_n;
+    const short K   = FC_solve_tri_k;
+    const short NP  = PAD2(N, NW);
+
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x*NSG + sgitg;
+
+    threadgroup float * sh0 = (threadgroup float *) shmem;
+
+    device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
+    device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
+    device       float * dst_ptr  = (device       float *)(dst  + i02 * args.nb2  + i03 * args.nb3)  + i01;
+
+    for (short rr = 0; rr < N; rr += NSG) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        {
+            threadgroup float * sh0_cur = sh0 + sgitg*NP;
+
+            for (short t = 0; t*NW < N; ++t) {
+                const short idx = t*NW + tiisg;
+                sh0_cur[idx] = src0_ptr[idx];
+            }
+
+            src0_ptr += NSG*N;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (i01 >= args.ne10) {
+            continue;
+        }
+
+        for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
+            const short r = rr + ir;
+
+            threadgroup float * sh0_cur = sh0 + ir*NP;
+
+            float sum = 0.0f;
+
+            for (short t = 0; t*NW < r; ++t) {
+                const short idx = t*NW + tiisg;
+                sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
+            }
+
+            sum = simd_sum(sum);
+
+            if (tiisg == 0) {
+                const float diag = sh0_cur[r];
+
+                dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
+            }
+        }
+    }
+}

ggml/src/ggml-metal/kernels/ssm.metal

@@ -0,0 +1,279 @@
+#include "common.h"
+
+// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
+kernel void kernel_ssm_conv_f32_f32(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+    device       float * x = (device       float *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+    device       float  * x = (device       float  *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+constant short FC_ssm_conv_bs   [[function_constant(FC_SSM_CONV + 0)]];
+
+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+kernel void kernel_ssm_conv_f32_f32_batched(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_batched_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// Optimized version: reduces redundant memory loads by having one thread load shared values
+kernel void kernel_ssm_scan_f32(
+        constant ggml_metal_kargs_ssm_scan & args,
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device const void * src6,
+        device      float * dst,
+        threadgroup float * shared [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgptg[[simdgroups_per_threadgroup]],
+        uint3    tgpg[[threadgroups_per_grid]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    // Shared memory layout:
+    // [0..sgptg*NW-1]: partial sums for reduction (existing)
+    // [sgptg*NW..sgptg*NW+sgptg-1]: pre-computed x_dt values for each token in batch
+    // [sgptg*NW+sgptg..sgptg*NW+2*sgptg-1]: pre-computed dA values for each token in batch
+    threadgroup float * shared_sums = shared;
+    threadgroup float * shared_x_dt = shared + sgptg * NW;
+    threadgroup float * shared_dA   = shared + sgptg * NW + sgptg;
+
+    shared_sums[tpitg.x] = 0.0f;
+
+    const int32_t i0 = tpitg.x;
+    const int32_t i1 = tgpig.x;
+    const int32_t ir = tgpig.y; // current head
+    const int32_t i3 = tgpig.z; // current seq
+
+    const int32_t nc  = args.d_state;
+    const int32_t nr  = args.d_inner;
+    const int32_t nh  = args.n_head;
+    const int32_t ng  = args.n_group;
+    const int32_t n_t = args.n_seq_tokens;
+
+    const int32_t s_off = args.s_off;
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0_buff = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s_buff  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+
+    const int32_t i = i0 + i1*nc;
+    const int32_t g = ir / (nh / ng); // repeat_interleave
+
+    float s0 = s0_buff[i];
+    float s  = 0.0f;
+
+    device const float * A = (device const float *) ((device const char *) src3 + ir*args.nb31); // {ne30, nh}
+
+    const float A0 = A[i0%args.ne30];
+
+    device const float * x  = (device const float *)((device const char *) src1 + i1*args.nb10  + ir*args.nb11 + i3*args.nb13); // {dim, nh, nt, ns}
+    device const float * dt = (device const float *)((device const char *) src2 + ir*args.nb20  + i3*args.nb22);                // {nh, nt, ns}
+    device const float * B  = (device const float *)((device const char *) src4 +  g*args.nb41  + i3*args.nb43);                // {d_state, ng, nt, ns}
+    device const float * C  = (device const float *)((device const char *) src5 +  g*args.nb51  + i3*args.nb53);                // {d_state, ng, nt, ns}
+
+    device float * y = dst + (i1 + ir*(nr) + i3*(n_t*nh*nr)); // {dim, nh, nt, ns}
+
+    for (int i2 = 0; i2 < n_t; i2 += sgptg) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // Pre-compute x_dt and dA for this batch of tokens
+        // Only first sgptg threads do the loads and expensive math
+        if (i0 < sgptg && i2 + i0 < n_t) {
+            // ns12 and ns21 are element strides (nb12/nb10, nb21/nb20)
+            device const float * x_t  = x  + i0 * args.ns12;
+            device const float * dt_t = dt + i0 * args.ns21;
+
+            const float dt0  = dt_t[0];
+            const float dtsp = dt0 <= 20.0f ? log(1.0f + exp(dt0)) : dt0;
+            shared_x_dt[i0] = x_t[0] * dtsp;
+            shared_dA[i0]   = dtsp;  // Store dtsp, compute exp(dtsp * A0) per-thread since A0 varies
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
+            const float x_dt = shared_x_dt[t];
+            const float dA   = exp(shared_dA[t] * A0);
+
+            s = (s0 * dA) + (B[i0] * x_dt);
+
+            const float sumf = simd_sum(s * C[i0]);
+
+            if (tiisg == 0) {
+                shared_sums[t*NW + sgitg] = sumf;
+            }
+
+            // recurse
+            s0 = s;
+
+            B  += args.ns42;
+            C  += args.ns52;
+        }
+
+        // Advance pointers for next batch
+        x  += sgptg * args.ns12;
+        dt += sgptg * args.ns21;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float sumf = simd_sum(shared_sums[sgitg*NW + tiisg]);
+
+        if (tiisg == 0 && i2 + sgitg < n_t) {
+            y[sgitg*nh*nr] = sumf;
+        }
+
+        y += sgptg*nh*nr;
+    }
+
+    s_buff[i] = s;
+}

ggml/src/ggml-metal/kernels/tri.metal

@@ -0,0 +1,69 @@
+#include "common.h"
+
+template<uint32_t ttype>
+bool _ggml_vec_tri_cmp(const int i, const int r);
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER */ 3>(const int i, const int r) {
+    return i < r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER_DIAG */ 2>(const int i, const int r) {
+    return i <= r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER */ 1>(const int i, const int r) {
+    return i > r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER_DIAG */ 0>(const int i, const int r) {
+    return i >= r;
+}
+
+template<typename T, int ttype>
+kernel void kernel_tri(
+        constant ggml_metal_kargs_tri & args,
+        device const char * src0,
+        device const char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    device const T * src_row = (device const T *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T * dst_row = (device       T *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    // Each thread is a single element of the row if ne00 < max threads per
+    // threadgroup, so this will loop once for each index that this thread is
+    // responsible for
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        // Use the comparison as a mask for branchless
+        dst_row[i0] = static_cast<T>(_ggml_vec_tri_cmp<ttype>(i0, i1)) * src_row[i0];
+    }
+}
+
+typedef decltype(kernel_tri<float, 0>) kernel_tri_t;
+
+template [[host_name("kernel_tri_f32_0")]] kernel kernel_tri_t kernel_tri<float, 0>;
+template [[host_name("kernel_tri_f32_1")]] kernel kernel_tri_t kernel_tri<float, 1>;
+template [[host_name("kernel_tri_f32_2")]] kernel kernel_tri_t kernel_tri<float, 2>;
+template [[host_name("kernel_tri_f32_3")]] kernel kernel_tri_t kernel_tri<float, 3>;
+template [[host_name("kernel_tri_f16_0")]] kernel kernel_tri_t kernel_tri<half, 0>;
+template [[host_name("kernel_tri_f16_1")]] kernel kernel_tri_t kernel_tri<half, 1>;
+template [[host_name("kernel_tri_f16_2")]] kernel kernel_tri_t kernel_tri<half, 2>;
+template [[host_name("kernel_tri_f16_3")]] kernel kernel_tri_t kernel_tri<half, 3>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_tri_bf16_0")]] kernel kernel_tri_t kernel_tri<bfloat, 0>;
+template [[host_name("kernel_tri_bf16_1")]] kernel kernel_tri_t kernel_tri<bfloat, 1>;
+template [[host_name("kernel_tri_bf16_2")]] kernel kernel_tri_t kernel_tri<bfloat, 2>;
+template [[host_name("kernel_tri_bf16_3")]] kernel kernel_tri_t kernel_tri<bfloat, 3>;
+#endif

ggml/src/ggml-metal/kernels/unary.metal

@@ -0,0 +1,374 @@
+#include "common.h"
+
+constant short FC_unary_op [[function_constant(FC_UNARY + 0)]];
+constant bool  FC_unary_cnt[[function_constant(FC_UNARY + 1)]];
+
+template <typename T0, typename T, typename TC>
+kernel void kernel_unary_impl(
+        constant ggml_metal_kargs_unary & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_unary_op
+#define FC_CNT FC_unary_cnt
+
+    device const T0 * src0_ptr;
+    device       T  * dst_ptr;
+
+    int i0;
+
+    if (FC_CNT) {
+        i0 = tgpig.x;
+
+        src0_ptr = (device const T0 *) (src0);
+        dst_ptr  = (device       T  *) (dst);
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int k0  = tgpig.x/args.ne01;
+        const int i01 = tgpig.x - k0*args.ne01;
+
+        i0 = k0*ntg.x + tpitg.x;
+
+        src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+        dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1 );
+    }
+
+    {
+        //threadgroup_barrier(mem_flags::mem_none);
+
+        if (!FC_CNT) {
+            if (i0 >= args.ne0) {
+                return;
+            }
+        }
+
+        const TC x = (TC) src0_ptr[i0];
+
+        if (FC_OP == OP_UNARY_NUM_SCALE) {
+            dst_ptr[i0] = (T) (args.scale * x + args.bias);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FILL) {
+            dst_ptr[i0] = (T) args.val;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CLAMP) {
+            dst_ptr[i0] = (T) clamp(x, args.min, args.max);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQR) {
+            dst_ptr[i0] = (T) (x * x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQRT) {
+            dst_ptr[i0] = (T) sqrt(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIN) {
+            dst_ptr[i0] = (T) sin(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_COS) {
+            dst_ptr[i0] = (T) cos(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LOG) {
+            dst_ptr[i0] = (T) log(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LEAKY_RELU) {
+            dst_ptr[i0] = (T) (TC(x > 0)*x + TC(x <= 0)*(x * args.slope));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TANH) {
+            dst_ptr[i0] = (T) precise::tanh(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_RELU) {
+            dst_ptr[i0] = (T) fmax(0, x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIGMOID) {
+            dst_ptr[i0] = (T) (1 / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + precise::tanh(SQRT_2_OVER_PI*x*(1 + GELU_COEF_A*x*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_ERF) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + erf_approx(SQRT_2_INV*x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_QUICK) {
+            dst_ptr[i0] = (T) (x * (1/(1 + exp(GELU_QUICK_COEF*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SILU) {
+            dst_ptr[i0] = (T) (x / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ELU) {
+            dst_ptr[i0] = (T) elu_approx(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_NEG) {
+            dst_ptr[i0] = (T) -x;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ABS) {
+            dst_ptr[i0] = (T) fabs(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SGN) {
+            dst_ptr[i0] = T(x > 0) - T(x < 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_STEP) {
+            dst_ptr[i0] = T(x > 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSWISH) {
+            dst_ptr[i0] = (T) (x * fmax(0, fmin(1, x/6 + 0.5)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSIGMOID) {
+            dst_ptr[i0] = (T) fmax(0, fmin(1, x/6 + 0.5));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXP) {
+            dst_ptr[i0] = (T) exp(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SOFTPLUS) {
+            dst_ptr[i0] = (T) select(log(1 + exp(x)), x, x > 20);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXPM1) {
+            // TODO: precise implementation
+            dst_ptr[i0] = (T) (exp(x) - 1);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FLOOR) {
+            dst_ptr[i0] = (T) floor(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CEIL) {
+            dst_ptr[i0] = (T) ceil(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ROUND) {
+            dst_ptr[i0] = (T) round(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TRUNC) {
+            dst_ptr[i0] = (T) trunc(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_XIELU) {
+            const TC xi      = x;
+            const TC gate    = TC(xi > TC(0.0f));
+            const TC clamped = fmin(xi, TC(args.val));
+            const TC y_pos   = TC(args.scale) * xi * xi + TC(args.bias) * xi;
+            const TC y_neg   = (exp(clamped) - TC(1.0f) - xi) * TC(args.slope) + TC(args.bias) * xi;
+            dst_ptr[i0] = (T) (gate * y_pos + (TC(1.0f) - gate) * y_neg);
+        }
+    }
+
+#undef FC_OP
+#undef FC_CNT
+}
+
+typedef decltype(kernel_unary_impl<float, float, float>) kernel_unary_t;
+
+template [[host_name("kernel_unary_f32_f32")]]   kernel kernel_unary_t kernel_unary_impl<float,  float,  float>;
+template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_unary_impl<float4, float4, float4>;
+template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
+template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;
+
+kernel void kernel_silu_back_f32(
+        constant ggml_metal_kargs_silu_back & args,
+        device const float * dy,
+        device const float * x,
+        device       float * dx,
+        uint gid [[thread_position_in_grid]]) {
+    if (gid >= args.ne) {
+        return;
+    }
+
+    const float s = 1.0f / (1.0f + exp(-x[gid]));
+    dx[gid] = dy[gid] * s * (1.0f + x[gid] * (1.0f - s));
+}
+
+template<typename T>
+kernel void kernel_reglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = (T)(x0*x1*(x0 > 0.0f));
+    }
+}
+
+typedef decltype(kernel_reglu<float>) kernel_reglu_t;
+
+template [[host_name("kernel_reglu_f32")]] kernel kernel_reglu_t kernel_reglu<float>;
+template [[host_name("kernel_reglu_f16")]] kernel kernel_reglu_t kernel_reglu<half>;
+
+template<typename T>
+kernel void kernel_geglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + precise::tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = (T)(gelu*x1);
+    }
+}
+
+typedef decltype(kernel_geglu<float>) kernel_geglu_t;
+
+template [[host_name("kernel_geglu_f32")]] kernel kernel_geglu_t kernel_geglu<float>;
+template [[host_name("kernel_geglu_f16")]] kernel kernel_geglu_t kernel_geglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = (T)(silu*x1);
+    }
+}
+
+typedef decltype(kernel_swiglu<float>) kernel_swiglu_t;
+
+template [[host_name("kernel_swiglu_f32")]] kernel kernel_swiglu_t kernel_swiglu<float>;
+template [[host_name("kernel_swiglu_f16")]] kernel kernel_swiglu_t kernel_swiglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu_oai(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, args.limit);
+        x1 = max(min(x1, args.limit), -args.limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * args.alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = (T)out_glu;
+    }
+}
+
+typedef decltype(kernel_swiglu_oai<float>) kernel_swiglu_oai_t;
+
+template [[host_name("kernel_swiglu_oai_f32")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<float>;
+template [[host_name("kernel_swiglu_oai_f16")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<half>;
+
+template<typename T>
+kernel void kernel_geglu_erf(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
+
+        dst_row[i0] = (T)(gelu_erf*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_erf<float>) kernel_geglu_erf_t;
+
+template [[host_name("kernel_geglu_erf_f32")]] kernel kernel_geglu_erf_t kernel_geglu_erf<float>;
+template [[host_name("kernel_geglu_erf_f16")]] kernel kernel_geglu_erf_t kernel_geglu_erf<half>;
+
+template<typename T>
+kernel void kernel_geglu_quick(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = (T)(gelu_quick*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_quick<float>) kernel_geglu_quick_t;
+
+template [[host_name("kernel_geglu_quick_f32")]] kernel kernel_geglu_quick_t kernel_geglu_quick<float>;
+template [[host_name("kernel_geglu_quick_f16")]] kernel kernel_geglu_quick_t kernel_geglu_quick<half>;

ggml/src/ggml-metal/kernels/upscale.metal

@@ -0,0 +1,179 @@
+#include "common.h"
+
+constant bool FC_upscale_aa [[function_constant(FC_UPSCALE + 0)]];
+
+kernel void kernel_upscale_nearest_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3/args.sf3;
+    const int64_t i02 = i2/args.sf2;
+    const int64_t i01 = i1/args.sf1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int64_t i00 = i0/args.sf0;
+
+        device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1  +  i0*args.nb0);
+
+        dst_ptr[0] = src0_ptr[0];
+    }
+}
+
+static inline float bilinear_tri(float x) {
+    return MAX(0.0f, 1.0f - fabs(x));
+}
+
+kernel void kernel_upscale_bilinear_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01  = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01  = MAX(0, MIN(args.ne01 - 1, (int64_t)floor(f01)));
+    const int64_t i01p = MAX(0, MIN(args.ne01 - 1, i01 + 1));
+    const float   fd1  = MAX(0.0f, MIN(1.0f, f01 - (float)i01));
+
+    src0 += i03*args.nb03 + i02*args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1);
+
+    if (FC_upscale_aa) {
+        const float support0  = MAX(1.0f, 1.0f / args.sf0);
+        const float invscale0 = 1.0f / support0;
+        const float support1  = MAX(1.0f, 1.0f / args.sf1);
+        const float invscale1 = 1.0f / support1;
+
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+
+            int64_t x_min = MAX((int64_t)0, (int64_t)floor(f00 - support0 + args.poffs));
+            int64_t x_max = MIN(args.ne00,  (int64_t)ceil (f00 + support0 + args.poffs));
+
+            int64_t y_min = MAX((int64_t)0, (int64_t)floor(f01 - support1 + args.poffs));
+            int64_t y_max = MIN(args.ne01,  (int64_t)ceil (f01 + support1 + args.poffs));
+
+            float sum = 0.0f;
+            float wsum = 0.0f;
+
+            for (int64_t sy = y_min; sy < y_max; ++sy) {
+                const float wy = MAX(0.0f, 1.0f - fabs((float)sy - f01) * invscale1);
+                for (int64_t sx = x_min; sx < x_max; ++sx) {
+                    const float wx = MAX(0.0f, 1.0f - fabs((float)sx - f00) * invscale0);
+                    const float w  = wx * wy;
+                    device const float * src_ptr = (device const float *)(src0 + sy*args.nb01 + sx*args.nb00);
+                    sum  += (*src_ptr) * w;
+                    wsum += w;
+                }
+            }
+
+            const float v = (wsum > 0.0f) ? (sum / wsum) : 0.0f;
+            dst_ptr[i0] = v;
+        }
+    } else {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float   f00  = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+            const int64_t i00  = MAX(0, MIN(args.ne00 - 1, (int64_t)floor(f00)));
+            const int64_t i00p = MAX(0, MIN(args.ne00 - 1, i00 + 1));
+            const float   fd0  = MAX(0.0f, MIN(1.0f, f00 - (float)i00));
+
+            device const float * src00 = (device const float *)(src0 + i01*args.nb01  + i00*args.nb00);
+            device const float * src10 = (device const float *)(src0 + i01*args.nb01  + i00p*args.nb00);
+            device const float * src01 = (device const float *)(src0 + i01p*args.nb01 + i00*args.nb00);
+            device const float * src11 = (device const float *)(src0 + i01p*args.nb01 + i00p*args.nb00);
+
+            const float v =
+                (*src00) * (1.0f - fd0) * (1.0f - fd1) +
+                (*src10) * fd0          * (1.0f - fd1) +
+                (*src01) * (1.0f - fd0) * fd1 +
+                (*src11) * fd0          * fd1;
+
+            dst_ptr[i0] = v;
+        }
+    }
+}
+
+static inline float bicubic_weight1(float x) {
+    const float a = -0.75f;
+    return ((a + 2) * x - (a + 3)) * x * x + 1;
+}
+
+static inline float bicubic_weight2(float x) {
+    const float a = -0.75f;
+    return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a;
+}
+
+kernel void kernel_upscale_bicubic_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01 = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01 = (int64_t)floor(f01);
+    const float   fd1 = f01 - (float)i01;
+
+    const float w_y0 = bicubic_weight2(fd1 + 1.0f);
+    const float w_y1 = bicubic_weight1(fd1);
+    const float w_y2 = bicubic_weight1(1.0f - fd1);
+    const float w_y3 = bicubic_weight2(2.0f - fd1);
+
+    const device const char * src_slice = src0 + i03 * args.nb03 + i02 * args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3 * args.nb3 + i2 * args.nb2 + i1 * args.nb1);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const float   f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+        const int64_t i00 = (int64_t)floor(f00);
+        const float   fd0 = f00 - (float)i00;
+
+        const float w_x0 = bicubic_weight2(fd0 + 1.0f);
+        const float w_x1 = bicubic_weight1(fd0);
+        const float w_x2 = bicubic_weight1(1.0f - fd0);
+        const float w_x3 = bicubic_weight2(2.0f - fd0);
+
+        float sum = 0.0f;
+
+        for (int dy = -1; dy <= 2; ++dy) {
+            const int64_t iy = MAX(0, MIN(args.ne01 - 1, i01 + dy));
+            const float wy = (dy == -1) ? w_y0 : (dy == 0) ? w_y1 : (dy == 1) ? w_y2 : w_y3;
+
+            for (int dx = -1; dx <= 2; ++dx) {
+                const int64_t ix = MAX(0, MIN(args.ne00 - 1, i00 + dx));
+                const float wx = (dx == -1) ? w_x0 : (dx == 0) ? w_x1 : (dx == 1) ? w_x2 : w_x3;
+
+                device const float * src_ptr = (device const float *)(src_slice + iy * args.nb01 + ix * args.nb00);
+                sum += (*src_ptr) * wx * wy;
+            }
+        }
+
+        dst_ptr[i0] = sum;
+    }
+}

ggml/src/ggml-metal/kernels/wkv.metal

@@ -0,0 +1,179 @@
+#include "common.h"
+
+kernel void kernel_rwkv_wkv6_f32(
+    device const float * k,
+    device const float * v,
+    device const float * r,
+    device const float * tf,
+    device const float * td,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _k[head_size];
+    threadgroup float _r[head_size];
+    threadgroup float _tf[head_size];
+    threadgroup float _td[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    _tf[tid] = tf[head_id * head_size + tid];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0;
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 tf_vec = float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
+            float4 td_vec = float4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            float4 temp = tf_vec * kv + s_vec;
+            y += dot(r_vec, temp);
+
+            s_vec = s_vec * td_vec + kv;
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
+
+kernel void kernel_rwkv_wkv7_f32(
+    device const float * r,
+    device const float * w,
+    device const float * k,
+    device const float * v,
+    device const float * a,
+    device const float * b,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _r[head_size];
+    threadgroup float _w[head_size];
+    threadgroup float _k[head_size];
+    threadgroup float _a[head_size];
+    threadgroup float _b[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + tid * head_size + i];
+    }
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _r[tid] = r[t];
+        _w[tid] = w[t];
+        _k[tid] = k[t];
+        _a[tid] = a[t];
+        _b[tid] = b[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0, sa = 0.0;
+
+        float4 sa_vec(0.0);
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 a_vec = float4(_a[j], _a[j+1], _a[j+2], _a[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+            sa_vec += a_vec * s_vec;
+        }
+        sa = sa_vec[0] + sa_vec[1] + sa_vec[2] + sa_vec[3];
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 w_vec = float4(_w[j], _w[j+1], _w[j+2], _w[j+3]);
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 b_vec = float4(_b[j], _b[j+1], _b[j+2], _b[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            s_vec = s_vec * w_vec + kv + sa * b_vec;
+            y += dot(s_vec, r_vec);
+
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + tid * head_size + i] = state[i];
+    }
+}

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -7065,7 +7065,7 @@ static inline bool use_flat_gemv_for_large_m_q4_K(const ggml_tensor *tensor) {
     return tensor->ne[1] >= 32768 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
-static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_tensor *tensor) {
+static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
     // gemv_noshuffle variant perf drops for large M, use flat variant for large M.
     // threshold is well above typical hidden/FFN dims, but below typical vocab sizes.
     // q6_K flat gemv is worse for smaller K; 2048 seems to be a reasonable threshold.
@@ -7083,7 +7083,15 @@ static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_tensor *tensor) {
     if ((tensor->ne[1] % 128 != 0) && tensor->ne[2] == 1 && tensor->ne[3] == 1) {
         return true;
     }
-    return tensor->ne[1] >= 32768 && tensor->ne[0] >= 2048 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
+
+    // The gemv_noshuffle slowdown tracks TOTAL weight size, not ne0 alone; ne0 >= 2048 is a
+    // proxy for "large weight" that misses a narrow-hidden vocab-scale lm_head.
+    // Add a direct size escape so such weights also take the flat path, without changing
+    // which weights ne0 >= 2048 already routes there.
+    // The size escape is not taken on the A7X since its compiler miscompiles the flat K-quant GEMV
+    return tensor->ne[1] >= 32768
+        && (tensor->ne[0] >= 2048 || (backend_ctx->adreno_gen != ADRENO_GPU_GEN::A7X && ggml_nbytes(tensor) >= (256ull << 20)))
+        && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
 static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
@@ -7495,6 +7503,7 @@ static ggml_backend_i ggml_backend_opencl_i = {
 ggml_backend_t ggml_backend_opencl_init(void) {
     ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_opencl_reg(), 0);
     ggml_backend_opencl_context *backend_ctx = ggml_cl_init(dev);
+    backend_ctx->ref_count++;
 
     ggml_backend_t backend = new ggml_backend {
         /* .guid    = */ ggml_backend_opencl_guid(),
@@ -9402,7 +9411,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         cl_kernel kernel;
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
         kernel = backend_ctx->kernel_convert_block_q6_K;
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
             kernel = backend_ctx->kernel_convert_block_q6_K_noshuffle;
         }
 #else
@@ -9435,7 +9444,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         tensor->extra  = extra;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
             cl_int M = tensor->ne[1];   // ne01
             cl_int K = tensor->ne[0];   // ne00
 
@@ -10472,7 +10481,7 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             CL_CHECK(clReleaseMemObject(data_device));
             return;
         }
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
             static ggml_cl_buffer buf_trans_ql;
             static ggml_cl_buffer buf_trans_qh;
             static ggml_cl_buffer buf_trans_s;
@@ -18894,7 +18903,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         }
 
         // q6_K x fp32
-        if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32 && !use_flat_gemv_for_large_m_q6_K(src0)) {
+        if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32 && !use_flat_gemv_for_large_m_q6_K(backend_ctx, src0)) {
             ggml_cl_mul_mat_q6_K_f32_adreno(backend, src0, src1, dst);
             return;
         }
@@ -24259,7 +24268,7 @@ static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const
     }
 
     const size_t nrows = ggml_nrows(src0);
-    size_t nth = 512;
+    size_t nth = backend_ctx->max_workgroup_size < 512 ? backend_ctx->max_workgroup_size : 512;
     size_t global_work_size[] = {nrows*nth, 1, 1};
     size_t local_work_size[] = {nth, 1, 1};
 

ggml/src/ggml-sycl/backend.hpp

@@ -26,6 +26,7 @@
 #include "dmmv.hpp"
 #include "element_wise.hpp"
 #include "fattn.hpp"
+#include "fusion.hpp"
 #include "gated_delta_net.hpp"
 #include "gla.hpp"
 #include "im2col.hpp"

ggml/src/ggml-sycl/common.hpp

@@ -133,6 +133,7 @@ enum ggml_sycl_backend_gpu_mode {
 enum ggml_sycl_dev2dev_memcpy_mode {
   DEV2DEV_MEMCPY_SYCL = 0,
   DEV2DEV_MEMCPY_L0 = 1,
+  DEV2DEV_MEMCPY_FORWARD = 2
 };
 
 static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size");
@@ -234,6 +235,7 @@ struct sycl_device_info {
     int max_wg_per_cu; // max work groups per compute unit - refer to
                        // cudaOccupancyMaxActiveBlocksPerMultiprocessor
     bool    vmm;                // virtual memory support
+    bool    l0_device_type_valid;
     bool    l0_discrete_gpu;    // Level Zero backend and not an integrated GPU
     size_t  vmm_granularity;    // granularity of virtual memory
     size_t  total_vram;

ggml/src/ggml-sycl/concat.cpp

@@ -127,7 +127,15 @@ static void concat_T_sycl_non_cont(
     int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2,
     uint64_t nb3, int32_t dim) {
   sycl::range<3> gridDim(ne3, ne2, ne1);
-  stream->parallel_for(sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+
+  // Avoid oversubscribing device when there is not enough elements along the innermost dim to
+  // fill a full SYCL_CONCAT_BLOCK_SIZE. For larger # of elements, the full SYCL_CONCAT_BLOCK_SIZE
+  // is used.
+  const int64_t ne0_pad   = GGML_PAD(ne0, WARP_SIZE);
+  const int64_t block_ne0 = ne0_pad < SYCL_CONCAT_BLOCK_SIZE ? ne0_pad : (int64_t) SYCL_CONCAT_BLOCK_SIZE;
+  sycl::range<3> blockDim(1, 1, block_ne0);
+
+  stream->parallel_for(sycl::nd_range<3>(gridDim * blockDim, blockDim), [=](sycl::nd_item<3> item_ct1) {
       int64_t i3 = item_ct1.get_group(0);
       int64_t i2 = item_ct1.get_group(1);
       int64_t i1 = item_ct1.get_group(2);

ggml/src/ggml-sycl/convert.cpp

@@ -644,6 +644,8 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
     switch (type) {
         case GGML_TYPE_Q1_0:
             return dequantize_block_sycl<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q2_0:
+            return dequantize_block_sycl<QK2_0, QR2_0, dequantize_q2_0>;
         case GGML_TYPE_Q4_0:
             if (dst->src[0]->extra &&
                 ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
@@ -728,6 +730,8 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
     switch (type) {
         case GGML_TYPE_Q1_0:
             return dequantize_block_sycl<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q2_0:
+            return dequantize_block_sycl<QK2_0, QR2_0, dequantize_q2_0>;
         case GGML_TYPE_Q4_0:
             if (dst->src[0]->extra &&
                 ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {

ggml/src/ggml-sycl/cpy.cpp

@@ -8,7 +8,6 @@
 #include "ggml-sycl/presets.hpp"
 #include "ggml.h"
 
-
 static void cpy_1_f32_f32(const char * cxi, char * cdsti) {
     const float * xi   = (const float *) cxi;
     float *       dsti = (float *) cdsti;
@@ -151,6 +150,20 @@ static void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     }
 }
 
+static void cpy_blck_q2_0_f32(const char * cxi, char * cdsti) {
+    const block_q2_0 * xi = (const block_q2_0 *) cxi;
+    float * cdstf = (float *) cdsti;
+
+    const float d = xi->d;
+
+    for (int j = 0; j < QK2_0; ++j) {
+        const int byte_index = j / 4;
+        const int bit_offset = (j % 4) * 2;
+        const int q = (xi->qs[byte_index] >> bit_offset) & 0x3;
+        cdstf[j] = (float) (q - 1) * d;
+    }
+}
+
 
 
 template <dequantize_kernel_t dequant, int qk> static void cpy_blck_q_f32(const char * cxi, char * cdsti) {
@@ -256,7 +269,7 @@ static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, co
         stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                 cpy_f32_f16<cpy_1_f16_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
             });
@@ -274,7 +287,7 @@ static void ggml_cpy_f32_f32_sycl(const char * cx, char * cdst, const int ne, co
         stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                 cpy_f32_f16<cpy_1_f32_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
             });
@@ -292,7 +305,7 @@ static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, co
         stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                 cpy_f32_f16<cpy_1_f32_f16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
             });
@@ -308,7 +321,7 @@ static void ggml_cpy_f32_i32_sycl(const char * cx, char * cdst, const int ne, co
         stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                 cpy_f32_f16<cpy_1_f32_i32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
             });
@@ -324,7 +337,7 @@ static void ggml_cpy_i32_f32_sycl(const char * cx, char * cdst, const int ne, co
         stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                 cpy_f32_f16<cpy_1_i32_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
             });
@@ -338,7 +351,7 @@ static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, c
     GGML_ASSERT(ne % QK8_0 == 0);
     const int num_blocks = ne / QK8_0;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
                          });
@@ -350,12 +363,25 @@ static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
                          });
 }
 
+static void ggml_cpy_q2_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ne;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+            cpy_q_f32<cpy_blck_q2_0_f32, QK2_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
+                                                ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
 static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
                                    const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
@@ -363,7 +389,7 @@ static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, c
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
                          });
@@ -375,7 +401,8 @@ static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -389,7 +416,7 @@ static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, c
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
                          });
@@ -401,7 +428,8 @@ static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -415,7 +443,7 @@ static void ggml_cpy_f32_q5_0_sycl(const char * cx, char * cdst, const int ne, c
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                              cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
                          });
@@ -427,7 +455,8 @@ static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -441,7 +470,7 @@ static void ggml_cpy_f32_q5_1_sycl(const char * cx, char * cdst, const int ne, c
     GGML_ASSERT(ne % QK5_1 == 0);
     const int num_blocks = ne / QK5_1;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
                          });
@@ -453,7 +482,8 @@ static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -466,7 +496,8 @@ static void ggml_cpy_mxfp4_f32_sycl(const char * cx, char * cdst, const int ne,
                                     const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             cpy_q_f32<cpy_blck_q_f32<dequantize_mxfp4, QK_MXFP4>, QK_MXFP4>(cx, cdst, ne, ne00, ne01, ne02, nb00,
                                                                              nb01, nb02, nb03, ne10, ne11, ne12,
                                                                              nb10, nb11, nb12, nb13, item_ct1);
@@ -480,7 +511,8 @@ static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne,
     GGML_ASSERT(ne % QK4_NL == 0);
     const int num_blocks = ne / QK4_NL;
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
                                                    ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
@@ -526,7 +558,7 @@ static void ggml_cpy_f16_q4_0_sycl(const char * cx, char * cdst, const int ne, c
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              cpy_f32_q<cpy_blck_f16_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02,
                                                                  nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
@@ -540,7 +572,7 @@ static void ggml_cpy_f16_q4_1_sycl(const char * cx, char * cdst, const int ne, c
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              cpy_f32_q<cpy_blck_f16_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02,
                                                                  nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
@@ -554,7 +586,7 @@ static void ggml_cpy_f16_q5_0_sycl(const char * cx, char * cdst, const int ne, c
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                              cpy_f32_q<cpy_blck_f16_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02,
                                                                  nb00, nb01, nb02, nb03,
                                                                  ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
@@ -564,6 +596,7 @@ static void ggml_cpy_f16_q5_0_sycl(const char * cx, char * cdst, const int ne, c
 static bool ggml_sycl_is_quantized_type(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -594,6 +627,7 @@ static bool ggml_sycl_is_quantized_type(enum ggml_type type) {
 static bool ggml_sycl_can_quantize_rows_sycl(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -651,7 +685,8 @@ static void ggml_sycl_quantize_rows_q(const char * cx, char * cdst, const int64_
     constexpr int block_size = 256;
     const int64_t grid_size = ceil_div(total_blocks, (int64_t) block_size);
 
-    stream->parallel_for(sycl::nd_range<1>(grid_size * block_size, block_size), [=](sycl::nd_item<1> item_ct1) {
+    stream->parallel_for(sycl::nd_range<1>(grid_size * block_size, block_size),
+                        [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
         const int64_t block_idx = item_ct1.get_global_linear_id();
         if (block_idx >= total_blocks) {
             return;
@@ -708,6 +743,11 @@ static void ggml_sycl_quantize_rows_sycl(const char * cx, char * cdst, const ggm
                                                                             nb02, nb03, ne10, ne11, ne12, nb10, nb11,
                                                                             nb12, nb13, stream);
             break;
+        case GGML_TYPE_Q2_0:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_q2_0, QK2_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01,
+                                                                            nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+                                                                            nb12, nb13, stream);
+            break;
         case GGML_TYPE_Q5_1:
             ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01,
                                                                             nb02, nb03, ne10, ne11, ne12, nb10, nb11,
@@ -760,7 +800,7 @@ static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, co
         stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                 cpy_f32_f16<cpy_1_f16_f16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
             });
@@ -779,7 +819,7 @@ static void ggml_cpy_i16_i16_sycl(const char * cx, char * cdst, const int ne, co
         stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                 cpy_f32_f16<cpy_1_i16_i16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
             });
@@ -798,7 +838,7 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co
         stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
                 cpy_f32_f16<cpy_1_i32_i32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
             });
@@ -812,7 +852,8 @@ static void ggml_cpy_q8_0_q8_0(const char * cx, char * cdst, const int ne, const
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
         sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -825,7 +866,8 @@ static void ggml_cpy_q5_0_q5_0(const char * cx, char * cdst, const int ne, const
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
         sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -839,7 +881,8 @@ static void ggml_cpy_q5_1_q5_1(const char * cx, char * cdst, const int ne, const
 
     stream->parallel_for(
         sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -851,7 +894,8 @@ static void ggml_cpy_q4_0_q4_0(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -864,7 +908,8 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const
 
    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -875,18 +920,32 @@ static void ggml_cpy_q1_0_q1_0(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             cpy_q_q<block_q1_0, QK1_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
 
+static void ggml_cpy_q2_0_q2_0(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
+            cpy_q_q<block_q2_0, QK2_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
 static void ggml_cpy_mxfp4_mxfp4(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
                                    const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             cpy_q_q<block_mxfp4, QK_MXFP4>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -897,7 +956,8 @@ static void ggml_cpy_nvfp4_nvfp4(const char * cx, char * cdst, const int ne, con
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_nvfp4, QK_NVFP4>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -908,7 +968,8 @@ static void ggml_cpy_q2_K_q2_K(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q2_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -919,7 +980,8 @@ static void ggml_cpy_q3_K_q3_K(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q3_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -930,7 +992,8 @@ static void ggml_cpy_q4_K_q4_K(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q4_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -941,7 +1004,8 @@ static void ggml_cpy_q5_K_q5_K(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q5_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -952,7 +1016,8 @@ static void ggml_cpy_q6_K_q6_K(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_q6_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -963,7 +1028,8 @@ static void ggml_cpy_iq2_xxs_iq2_xxs(const char * cx, char * cdst, const int ne,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq2_xxs, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -974,7 +1040,8 @@ static void ggml_cpy_iq2_xs_iq2_xs(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq2_xs, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -985,7 +1052,8 @@ static void ggml_cpy_iq2_s_iq2_s(const char * cx, char * cdst, const int ne, con
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq2_s, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -996,7 +1064,8 @@ static void ggml_cpy_iq3_xxs_iq3_xxs(const char * cx, char * cdst, const int ne,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq3_xxs, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -1007,7 +1076,8 @@ static void ggml_cpy_iq1_s_iq1_s(const char * cx, char * cdst, const int ne, con
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq1_s, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -1018,7 +1088,8 @@ static void ggml_cpy_iq1_m_iq1_m(const char * cx, char * cdst, const int ne, con
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq1_m, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -1029,7 +1100,8 @@ static void ggml_cpy_iq4_nl_iq4_nl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -1040,7 +1112,8 @@ static void ggml_cpy_iq3_s_iq3_s(const char * cx, char * cdst, const int ne, con
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq3_s, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -1051,7 +1124,8 @@ static void ggml_cpy_iq4_xs_iq4_xs(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_q_q<block_iq4_xs, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
         });
 }
@@ -1065,7 +1139,7 @@ static void ggml_cpy_f32_bf16_sycl(const char * cx, char * cdst, const int ne, c
     stream->parallel_for(
         sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_f32_f16<cpy_1_f32_bf16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
         });
@@ -1079,7 +1153,7 @@ static void ggml_cpy_bf16_f32_sycl(const char * cx, char * cdst, const int ne, c
     stream->parallel_for(
         sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_f32_f16<cpy_1_bf16_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
         });
@@ -1093,7 +1167,7 @@ static void ggml_cpy_bf16_bf16_sycl(const char * cx, char * cdst, const int ne,
     stream->parallel_for(
         sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_f32_f16<cpy_1_bf16_bf16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                             nb10, nb11, nb12, nb13, item_ct1);
         });
@@ -1107,7 +1181,7 @@ static void ggml_cpy_f16_bf16_sycl(const char * cx, char * cdst, const int ne, c
     stream->parallel_for(
         sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_f32_f16<cpy_1_f16_bf16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
         });
@@ -1121,7 +1195,7 @@ static void ggml_cpy_bf16_f16_sycl(const char * cx, char * cdst, const int ne, c
     stream->parallel_for(
         sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
             cpy_f32_f16<cpy_1_bf16_f16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                            nb10, nb11, nb12, nb13, item_ct1);
         });
@@ -1213,6 +1287,9 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q8_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                                nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q2_0 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_q2_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                               nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
         ggml_cpy_f32_q5_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                                nb11, nb12, nb13, main_stream);
@@ -1243,6 +1320,8 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
         ggml_cpy_q4_1_q4_1(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q1_0 && src1->type == GGML_TYPE_Q1_0) {
         ggml_cpy_q1_0_q1_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q2_0 && src1->type == GGML_TYPE_Q2_0) {
+        ggml_cpy_q2_0_q2_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_MXFP4 && src1->type == GGML_TYPE_MXFP4) {
         ggml_cpy_mxfp4_mxfp4(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_NVFP4 && src1->type == GGML_TYPE_NVFP4) {

ggml/src/ggml-sycl/cpy.hpp

@@ -70,6 +70,39 @@ inline void cpy_blck_f32_q1_0(const char * cxi, char * cdsti) {
     }
 }
 
+inline int round_nearest_int(float x) {
+    return (int)(x >= 0.0f ? x + 0.5f : x - 0.5f);
+}
+
+inline void cpy_blck_f32_q2_0(const char * cxi, char * cdsti) {
+    const float * xi   = (const float *) cxi;
+    block_q2_0 *  dsti = (block_q2_0 *) cdsti;
+
+    float amax = 0.0f;
+
+    for (int j = 0; j < QK2_0; ++j) {
+        amax = sycl::fmax(amax, sycl::fabs((float) xi[j]));
+    }
+
+    const float d  = amax;
+    const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK2_0 / 4; ++j) {
+        dsti->qs[j] = 0;
+    }
+
+    for (int j = 0; j < QK2_0; ++j) {
+        int q = round_nearest_int(xi[j] * id) + 1;
+        q = dpct::max(0, dpct::min(3, q));
+
+        const int byte_index = j / 4;
+        const int bit_offset = (j % 4) * 2;
+        dsti->qs[byte_index] |= (uint8_t) q << bit_offset;
+    }
+}
+
 inline int best_index_mxfp4(const float x, const float e) {
     int best_index = 0;
     float best_err = sycl::fabs((float) (kvalues_mxfp4[0] * e - x));

ggml/src/ggml-sycl/dequantize.hpp

@@ -25,6 +25,28 @@ typedef void (*dequantize_kernel_f32_t)(const void * vx, const int64_t ib, const
 static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m);
 #endif
 
+static __dpct_inline__ void dequantize_q2_0(const void *vx, const int64_t ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q2_0 * x = (const block_q2_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    const int byte_idx = iqs / 4;
+    const int shift = (iqs % 4) * 2;
+    const uint8_t vui = x[ib].qs[byte_idx];
+
+    v.x() = (vui >> shift) & 3;
+    v.y() = (vui >> (shift + 2)) & 3;
+
+#ifdef GGML_SYCL_F16
+    v.s0() = ((dfloat)v.s0() - 1.0f) * d;
+    v.s1() = ((dfloat)v.s1() - 1.0f) * d;
+#else
+    v.x() = ((dfloat)v.x() - 1.0f) * d;
+    v.y() = ((dfloat)v.y() - 1.0f) * d;
+#endif // GGML_SYCL_F16
+}
+
 static __dpct_inline__ void dequantize_q4_0(const void *vx, const int64_t ib,
                                             const int iqs, dfloat2 &v) {
     const block_q4_0 * x = (const block_q4_0 *) vx;

ggml/src/ggml-sycl/fattn-mkl.cpp

@@ -0,0 +1,690 @@
+// Flash attention via oneMKL GEMM (XMX-accelerated).
+// Uses column_major::gemm for Q*K^T and S*V matmuls
+// with an online softmax SYCL kernel.
+//
+// All GQA query heads sharing a KV head are batched into single
+// GEMM calls, amortizing MKL launch overhead across K and V reuse.
+//
+
+#include "common.hpp"
+#include "fattn-common.hpp"
+#include "fattn-buffers.hpp"
+#include "convert.hpp"
+#include "fattn.hpp"
+
+#include <oneapi/mkl.hpp>
+#include <cstdio>
+#include <chrono>
+
+#define MKL_FA_CHUNK_SIZE_KV 8192
+
+// Number of query rows processed per tile. The score buffers (KQ_f32, S_f16)
+// are sized q_tile_rows * chunk_size, so this bounds their footprint
+// regardless of batch size (n_query_rows = n_queries * gqa_ratio). A typical
+// single-ubatch prefill (e.g. ubatch 1024 * gqa 8 = 8192 rows) is exactly one
+// tile, so it runs with no extra iterations. Larger batches tile and stay
+// bounded. Override with GGML_SYCL_MKL_FA_Q_TILE.
+#define MKL_FA_Q_TILE 8192
+
+#define MKL_FA_WG_SIZE 256
+
+using oneapi::mkl::transpose;
+using oneapi::mkl::blas::column_major::gemm;
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+// Pack all GQA Q heads for one KV head into fp16, applying q_scale.
+// Launches one kernel per GQA group — each kernel copies exactly
+// n_queries * DKQ elements using the per-group dst offset and
+// per-head source stride.
+static void mkl_fa_pack_q_fp16(
+    dpct::queue_ptr stream,
+    sycl::half * __restrict dst,
+    const float * __restrict q_src,
+    int n_queries, int n_query_rows, int DKQ,
+    int gqa_ratio, int kvh_base_head,
+    float q_scale, int64_t q_row_stride, int64_t q_head_stride,
+    int64_t wg_size) {
+
+    for (int iqg = 0; iqg < gqa_ratio; iqg++) {
+        int     iqh       = kvh_base_head + iqg;
+        sycl::half * dst_g = dst + (int64_t)iqg * n_queries * DKQ;
+
+        const int64_t n_elem = (int64_t)n_queries * DKQ;
+        const int64_t wg = ((n_elem + wg_size - 1) / wg_size) * wg_size;
+
+        stream->submit([&](sycl::handler & cgh) {
+            cgh.parallel_for(sycl::nd_range<1>(wg, wg_size),
+                [=](sycl::nd_item<1> item) {
+                    int64_t e = item.get_global_id(0);
+                    if (e >= n_elem) return;
+
+                    int64_t q = e / DKQ;
+                    int64_t d = e - q * DKQ;
+
+                    // Stride-aware source offset: handles permuted,
+                    // sliced, or contiguous Q tensor layouts.
+                    int64_t src_off = d
+                        + q * q_row_stride
+                        + (int64_t)iqh * q_head_stride;
+
+                    dst_g[e] = sycl::half(
+                        q_src[src_off] * q_scale);
+                });
+        });
+    }
+}
+
+// Zero-initialize the online softmax state arrays.
+// KQ_max → -inf, KQ_sum → 0, VKQ_accum → 0.
+// Merged into one kernel to avoid per-array launch overhead.
+static void mkl_fa_init_softmax_state(
+    dpct::queue_ptr stream,
+    float * kmax, float * ksum, float * vacc,
+    int n_query_rows, int DV, int64_t wg_size) {
+
+    const float    neg_inf   = -1e30f;
+    const int64_t  n_maxsum  = n_query_rows;
+    const int64_t  n_vacc    = (int64_t)n_query_rows * DV;
+    const int64_t  total     = (n_vacc > n_maxsum) ? n_vacc : n_maxsum;
+    const int64_t  wg = ((total + wg_size - 1) / wg_size) * wg_size;
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<1>(wg, wg_size),
+            [=](sycl::nd_item<1> item) {
+                int64_t i = item.get_global_id(0);
+                if (i < n_maxsum) {
+                    kmax[i] = neg_inf;
+                    ksum[i] = 0.0f;
+                }
+                if (i < n_vacc) {
+                    vacc[i] = 0.0f;
+                }
+            });
+    });
+}
+
+// Online softmax over one KV chunk for a tile of GQA query rows.
+// The tile spans absolute rows [q0, q0 + q_rows). Score buffers
+// (KQ_f32/S_f16) are indexed RELATIVE to the tile; the persistent state
+// (VKQ_accum/KQ_max/KQ_sum) and mask are indexed by ABSOLUTE row.
+// For each row: find local max → rescale previous VKQ_accum →
+// compute exp(s - max) → write S_f16 → update running max/sum.
+static void mkl_fa_online_softmax_chunk(
+    dpct::queue_ptr stream,
+    float * __restrict KQ_f32,
+    sycl::half * __restrict S_f16,
+    float * __restrict KQ_max,
+    float * __restrict KQ_sum,
+    float * __restrict VKQ_accum,
+    int q0, int q_rows, int n_queries, int DV,
+    int chunk_size, int chunk_start,
+    int kvh_head, int gqa_ratio,
+    const sycl::half * mask_data, int64_t mask_head_stride,
+    int64_t mask_row_stride, int mask_n_heads,
+    float logit_softcap, int64_t wg_size) {
+
+    const int64_t wg = ((q_rows + wg_size - 1) / wg_size) * wg_size;
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<1>(wg, wg_size),
+            [=](sycl::nd_item<1> item) {
+                int jc_rel = item.get_global_id(0);
+                if (jc_rel >= q_rows) return;
+                int jc_abs = q0 + jc_rel;
+
+                const int gqa_group = jc_abs / n_queries;
+                const int q_row     = jc_abs % n_queries;
+
+                // Score buffers are tile-local (relative index).
+                const float * __restrict KQ_row = KQ_f32
+                    + jc_rel * (int64_t)chunk_size;
+                // Persistent accumulator is full-sized (absolute index).
+                float * __restrict vkq = VKQ_accum
+                    + jc_abs * (int64_t)DV;
+
+                const sycl::half * mask_h = nullptr;
+                int64_t m_stride = 0;
+                if (mask_data) {
+                    int m_head = (mask_n_heads > 1)
+                        ? (kvh_head + gqa_group) : 0;
+                    mask_h   = mask_data + (int64_t)m_head * mask_head_stride;
+                    m_stride = mask_row_stride;
+                }
+
+                // Row-wise local maximum (softcap before mask)
+                float local_max = -1e30f;
+                for (int i = 0; i < chunk_size; i++) {
+                    float s = KQ_row[i];
+                    if (logit_softcap != 0.0f) {
+                        s = logit_softcap * sycl::tanh(s);
+                    }
+                    if (mask_h) {
+                        s += (float)mask_h[q_row * m_stride
+                            + (chunk_start + i)];
+                    }
+                    if (s > local_max) local_max = s;
+                }
+
+                // Rescale previous accumulator by exp(old_max - new_max)
+                float old_max = KQ_max[jc_abs];
+                float new_max = (old_max > local_max) ? old_max : local_max;
+                float rescale = (old_max < -1e29f) ? 1.0f
+                    : sycl::native::exp(old_max - new_max);
+
+                for (int v = 0; v < DV; v++) {
+                    vkq[v] *= rescale;
+                }
+
+                // Softmax and write S_f16 (tile-local index)
+                float local_sum = 0.0f;
+                sycl::half * __restrict S_row = S_f16
+                    + jc_rel * (int64_t)chunk_size;
+
+                for (int i = 0; i < chunk_size; i++) {
+                    float s = KQ_row[i];
+                    if (logit_softcap != 0.0f) {
+                        s = logit_softcap * sycl::tanh(s);
+                    }
+                    if (mask_h) {
+                        s += (float)mask_h[q_row * m_stride
+                            + (chunk_start + i)];
+                    }
+                    float val = sycl::native::exp(s - new_max);
+                    S_row[i] = sycl::half(val);
+                    local_sum += val;
+                }
+
+                KQ_sum[jc_abs] = KQ_sum[jc_abs] * rescale + local_sum;
+                KQ_max[jc_abs] = new_max;
+            });
+    });
+}
+
+// Write one GQA group's normalized output to its destination head.
+static void mkl_fa_normalize_head(
+    dpct::queue_ptr stream,
+    float * __restrict dst_batch,
+    const float * __restrict VKQ_accum,
+    const float * __restrict KQ_sum,
+    int iqh, int n_queries, int DV, int n_q_heads,
+    int64_t src_offset, int64_t wg_size) {
+
+    const int64_t wg = ((n_queries + wg_size - 1) / wg_size) * wg_size;
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<1>(wg, wg_size),
+            [=](sycl::nd_item<1> item) {
+                int jc = item.get_global_id(0);
+                if (jc >= n_queries) return;
+
+                int     ksum_idx = (int)(src_offset / DV) + jc;
+                float   inv_sum  = 1.0f / KQ_sum[ksum_idx];
+                const float * __restrict src = VKQ_accum
+                    + src_offset + jc * (int64_t)DV;
+                // Interleaved dst layout (matching TILE):
+                // rows alternate between heads, then increment query.
+                // offset = (query * n_q_heads + head) * DV
+                float * __restrict dst_row = dst_batch
+                    + ((int64_t)jc * n_q_heads + iqh) * (int64_t)DV;
+
+                for (int v = 0; v < DV; v++) {
+                    dst_row[v] = src[v] * inv_sum;
+                }
+            });
+    });
+}
+
+// ---------------------------------------------------------------------------
+// Per-chunk dequant
+//
+// Rather than dequantizing all of K/V up front (footprint scales with
+// context), we dequant one KV-head chunk at a time into a dense
+// [this_chunk x D] fp16 buffer (row-major, lda = D). The source address of
+// element (head=ikvh, row=chunk_start+r, col=c) decomposes into independent
+// linear terms head_off(ikvh) + row_off(chunk_start) + (r,c), so slicing a
+// chunk is a clean pointer offset in every layout case. The true-Gemma-
+// interleave vs padded-seq-view distinction is resolved once when the
+// descriptor is built; slicing does not reintroduce it.
+// ---------------------------------------------------------------------------
+enum mkl_fa_kv_desc_mode {
+    MKL_FA_KV_MODE_F16_DENSE       = 0,
+    MKL_FA_KV_MODE_F16_INTERLEAVED = 1,
+    MKL_FA_KV_MODE_QUANT_CONTIG    = 2,
+    MKL_FA_KV_MODE_QUANT_NC        = 3,
+};
+
+struct mkl_fa_kv_desc {
+    const char *         data = nullptr;
+    ggml_type            type = GGML_TYPE_F16;
+    int64_t              D    = 0;      // ne[0]
+    int64_t              nb1  = 0;      // byte stride, seq dim
+    int64_t              nb2  = 0;      // byte stride, head dim
+    mkl_fa_kv_desc_mode  mode = MKL_FA_KV_MODE_F16_DENSE;
+    int64_t              ts   = 0;      // type size (mode 3 base offset)
+    int64_t              s01  = 0;      // nc row stride in blocks (mode 3)
+    int64_t              s02  = 0;      // nc head stride in blocks (mode 3)
+};
+
+static mkl_fa_kv_desc mkl_fa_make_desc(const ggml_tensor * T, bool interleaved, int n_kv_heads) {
+    mkl_fa_kv_desc d;
+    d.data = (const char *)T->data;
+    d.type = T->type;
+    d.D    = T->ne[0];
+    d.nb1  = (int64_t)T->nb[1];
+    d.nb2  = (int64_t)T->nb[2];
+    d.ts   = (int64_t)ggml_type_size(T->type);
+
+    if (T->type == GGML_TYPE_F16) {
+        d.mode = interleaved ? MKL_FA_KV_MODE_F16_INTERLEAVED
+                             : MKL_FA_KV_MODE_F16_DENSE;
+    } else if (ggml_is_contiguously_allocated(T) && !interleaved) {
+        d.mode = MKL_FA_KV_MODE_QUANT_CONTIG;
+    } else {
+        d.mode = MKL_FA_KV_MODE_QUANT_NC;
+        const int64_t bs          = (int64_t)ggml_blck_size(T->type);
+        const int64_t blk_per_row = T->ne[0] / bs;
+        // True Gemma interleave packs heads within a row (nb[2] < ne[1]*nb[1])
+        // → reconstruct physical strides. Padded seq-views (nb[2] > ne[1]*nb[1])
+        // already have correct physical strides.
+        const bool gemma = interleaved &&
+            ((int64_t)T->nb[2] < (int64_t)T->ne[1] * (int64_t)T->nb[1]);
+        if (gemma) {
+            d.s01 = (int64_t)n_kv_heads * blk_per_row;
+            d.s02 = blk_per_row;
+        } else {
+            d.s01 = d.nb1 / d.ts;
+            d.s02 = d.nb2 / d.ts;
+        }
+    }
+    return d;
+}
+
+// Dequant one KV-head chunk into a dense [this_chunk x D] fp16 buffer.
+static void mkl_fa_dequant_chunk(
+    dpct::queue_ptr stream, const mkl_fa_kv_desc & d, ggml_tensor * dst_ctx,
+    sycl::half * out, int ikvh, int chunk_start, int this_chunk) {
+
+    const int64_t D = d.D;
+    switch (d.mode) {
+        case MKL_FA_KV_MODE_F16_DENSE: {
+            const char * base = d.data + (int64_t)ikvh * d.nb2
+                + (int64_t)chunk_start * d.nb1;
+            stream->memcpy(out, base, (size_t)this_chunk * D * sizeof(sycl::half));
+            break;
+        }
+        case MKL_FA_KV_MODE_F16_INTERLEAVED: {
+            const char * base = d.data + (int64_t)ikvh * d.nb2
+                + (int64_t)chunk_start * d.nb1;
+            const int64_t row_halfs = d.nb1 / (int64_t)sizeof(sycl::half);
+            const sycl::half * src = (const sycl::half *)base;
+            stream->parallel_for(
+                sycl::range<2>((size_t)this_chunk, (size_t)D),
+                [=](sycl::item<2> it) {
+                    int64_t r = it.get_id(0);
+                    int64_t c = it.get_id(1);
+                    out[r * D + c] = src[r * row_halfs + c];
+                });
+            break;
+        }
+        case MKL_FA_KV_MODE_QUANT_CONTIG: {
+            const char * base = d.data + (int64_t)ikvh * d.nb2
+                + (int64_t)chunk_start * d.nb1;
+            to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(d.type, dst_ctx);
+            to_fp16(base, out, (int64_t)this_chunk * D, stream);
+            break;
+        }
+        default: {  // MKL_FA_KV_MODE_QUANT_NC
+            to_fp16_nc_sycl_t to_fp16 = ggml_get_to_fp16_nc_sycl(d.type);
+            const int64_t base_blocks = (int64_t)ikvh * d.s02
+                + (int64_t)chunk_start * d.s01;
+            const char * base = d.data + base_blocks * d.ts;
+            // ne02 = ne03 = 1 → s02/s03 inert; head+chunk offset carried by base.
+            to_fp16(base, out, D, this_chunk, 1, 1, d.s01, d.s02, d.s02, stream);
+            break;
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// MKL Flash Attention orchestrator
+//
+// Pipeline: dequantize K/V → for each KV head:
+//   pack GQA Q heads → MKL GEMM KQ → online softmax →
+//   MKL GEMM VKQ → accumulate → normalize → scatter to dst
+// ---------------------------------------------------------------------------
+void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+
+    const ggml_tensor * Q    = dst->src[0];
+    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+    ggml_tensor * KQV = dst;
+
+    GGML_ASSERT(Q->type == GGML_TYPE_F32);
+    GGML_ASSERT(KQV->type == GGML_TYPE_F32);
+
+    // --- Op params ---
+    float scale = 1.0f, max_bias = 0.0f, logit_softcap = 0.0f;
+    memcpy(&scale,         (const float *)KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (const float *)KQV->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (const float *)KQV->op_params + 2, sizeof(float));
+
+    const float q_scale = scale;
+
+    // --- Dimensions ---
+    const int DKQ        = (int)K->ne[0];
+    const int DV         = (int)V->ne[0];
+    const int n_queries  = (int)Q->ne[1];
+    const int n_q_heads  = (int)Q->ne[2];
+    const int n_kv_heads = (int)K->ne[2];
+    const int n_batch    = (int)Q->ne[3];
+    const int n_kv       = (int)K->ne[1];
+    const int gqa_ratio  = n_q_heads / n_kv_heads;
+    const int n_query_rows = n_queries * gqa_ratio;
+
+    GGML_ASSERT(n_q_heads % n_kv_heads == 0);
+    GGML_ASSERT(max_bias == 0.0f);  // ALiBi not supported
+    GGML_ASSERT(Q->ne[3] == K->ne[3] || K->ne[3] == 1);
+
+    const int chunk_size = std::min(MKL_FA_CHUNK_SIZE_KV, n_kv);
+
+    // Query rows are processed in tiles of q_tile_rows so the score buffers
+    // (KQ_f32/S_f16 = q_tile_rows * chunk_size) stay bounded regardless of
+    // batch size. n_query_rows <= Q_TILE is a single tile (no extra work).
+    static int q_tile_env = ggml_sycl_get_env("GGML_SYCL_MKL_FA_Q_TILE", MKL_FA_Q_TILE);
+    const int q_tile_rows = std::max(1, std::min(q_tile_env, n_query_rows));
+
+    const int64_t wg_size = MKL_FA_WG_SIZE;
+
+    // --- Debug output (gated by GGML_SYCL_MKL_FA_DEBUG=1) ---
+    static int mkl_call_count = 0;
+    mkl_call_count++;
+    static int mkl_debug = ggml_sycl_get_env("GGML_SYCL_MKL_FA_DEBUG", 0);
+    const bool do_print = (mkl_debug == 1);
+
+    const int64_t q_row_stride  = Q->nb[1] / sizeof(float);
+    const int64_t q_head_stride = Q->nb[2] / sizeof(float);
+
+    const bool V_is_K_view = V->view_src
+        && (V->view_src == K || (V->view_src == K->view_src
+            && V->view_offs == K->view_offs));
+
+    // Early interleaved detection for debug output.
+    // True interleaved detection happens after dequant (nb12_fp16 == nb11_fp16),
+    // but we can pre-detect on the original tensor strides.
+    const bool k_early_interleaved =
+        ((int64_t)K->ne[1] * K->nb[1] != K->nb[2]);
+    const bool v_early_interleaved =
+        !V_is_K_view && ((int64_t)V->ne[1] * V->nb[1] != V->nb[2]);
+
+    if (do_print) {
+        GGML_LOG_INFO("[MKL-FA] #%d D=%d DV=%d n_q=%d n_kv=%d "
+                "n_qh=%d n_kvh=%d gqa=%d batch=%d K=%s V=%s "
+                "chunk=%d buf=%.1fMB%s%s\n",
+                mkl_call_count, DKQ, DV, n_queries, n_kv,
+                n_q_heads, n_kv_heads, gqa_ratio, n_batch,
+                ggml_type_name(K->type), ggml_type_name(V->type),
+                chunk_size,
+                (double)((int64_t)n_query_rows * chunk_size * sizeof(float))
+                    / (1024.0 * 1024.0),
+                k_early_interleaved ? " K_ILV" : "",
+                v_early_interleaved ? " V_ILV" : "");
+        GGML_LOG_INFO("[MKL-FA] #%d Q-nb1=%lld Q-nb2=%lld "
+                "q_rs=%lld q_hs=%lld dst_rs=%lld dst_hs=%lld\n",
+                mkl_call_count,
+                (long long)Q->nb[1], (long long)Q->nb[2],
+                (long long)q_row_stride, (long long)q_head_stride,
+                (long long)(KQV->nb[1] / sizeof(float)),
+                (long long)(KQV->nb[2] / sizeof(float)));
+    }
+
+    // --- Stream and allocators ---
+    dpct::queue_ptr stream = ctx.stream();
+
+#define MKL_TAKE_TIME(t0)  auto t0 = std::chrono::steady_clock::now()
+#define MKL_ACCUM(acc, t0) do { if (do_print) { \
+    acc += (int64_t)std::chrono::duration_cast \
+    <std::chrono::microseconds>(std::chrono::steady_clock::now() - (t0)).count(); \
+} } while(0)
+
+    int64_t gemm_kq_time_us  = 0;
+    int64_t gemm_vkq_time_us = 0;
+    int64_t softmax_time_us  = 0;
+    int64_t dequant_time_us  = 0;
+
+    MKL_TAKE_TIME(t_deq);
+
+    // --- K/V dequant descriptors ---
+    // Dequant is done per-chunk inside the KV loop (footprint independent of
+    // context). Output is always dense row-major fp16 [this_chunk x D], lda=D.
+    // Interleaved detection: ne[1]*nb[1] != nb[2] means heads are interleaved.
+    const bool k_interleaved =
+        ((int64_t)K->ne[1] * K->nb[1] != K->nb[2]) && K->ne[2] > 1;
+    const bool v_interleaved =
+        ((int64_t)V->ne[1] * V->nb[1] != V->nb[2]) && V->ne[2] > 1;
+
+    const mkl_fa_kv_desc K_desc = mkl_fa_make_desc(K, k_interleaved, n_kv_heads);
+    const mkl_fa_kv_desc V_desc = V_is_K_view
+        ? K_desc : mkl_fa_make_desc(V, v_interleaved, n_kv_heads);
+
+    MKL_ACCUM(dequant_time_us, t_deq);
+
+    // --- Resolve mask pointers ---
+    const sycl::half * mask_data = nullptr;
+    int64_t mask_head_stride = 0;
+    int64_t mask_row_stride  = 0;
+    int     mask_n_heads     = 0;
+
+    if (mask) {
+        // Use actual fp16 device size (2 bytes), NOT sizeof(sycl::half)
+        // which may be 4 on the host in oneAPI.
+        mask_head_stride = mask->nb[2] / 2;
+        mask_row_stride  = mask->nb[1] / 2;
+        mask_n_heads     = (int)mask->ne[2];
+    }
+
+    // --- Allocate intermediates from pool ---
+    ggml_sycl_pool & pool = ctx.pool();
+
+    ggml_sycl_pool_alloc<float>      KQ_f32(pool);      // [q_tile_rows x chunk]
+    ggml_sycl_pool_alloc<sycl::half> S_f16(pool);       // [q_tile_rows x chunk]
+    ggml_sycl_pool_alloc<float>      VKQ_chunk(pool);   // [q_tile_rows x DV]
+    ggml_sycl_pool_alloc<float>      VKQ_accum(pool);   // [n_query_rows x DV] (full)
+    ggml_sycl_pool_alloc<float>      KQ_max(pool);      // [n_query_rows] (full)
+    ggml_sycl_pool_alloc<float>      KQ_sum(pool);      // [n_query_rows] (full)
+    ggml_sycl_pool_alloc<sycl::half> Q_head_f16(pool);  // [n_query_rows x DKQ] (full)
+    ggml_sycl_pool_alloc<sycl::half> K_chunk_f16(pool); // [chunk x DKQ] (per-chunk dequant)
+    ggml_sycl_pool_alloc<sycl::half> V_chunk_f16(pool); // [chunk x DV] (per-chunk dequant)
+
+    KQ_f32.alloc((size_t)q_tile_rows * chunk_size);
+    S_f16.alloc((size_t)q_tile_rows * chunk_size);
+    VKQ_chunk.alloc((size_t)q_tile_rows * DV);
+    VKQ_accum.alloc((size_t)n_query_rows * DV);
+    KQ_max.alloc(n_query_rows);
+    KQ_sum.alloc(n_query_rows);
+    Q_head_f16.alloc((size_t)n_query_rows * DKQ);
+    K_chunk_f16.alloc((size_t)chunk_size * DKQ);
+
+    sycl::half * V_chunk_f16_ptr;
+    if (V_is_K_view) {
+        V_chunk_f16_ptr = K_chunk_f16.ptr;   // V aliases K (DV == DKQ)
+    } else {
+        V_chunk_f16.alloc((size_t)chunk_size * DV);
+        V_chunk_f16_ptr = V_chunk_f16.ptr;
+    }
+
+    sycl::half * Q_head_f16_ptr  = Q_head_f16.ptr;
+    float      * KQ_f32_ptr      = KQ_f32.ptr;
+    sycl::half * S_f16_ptr       = S_f16.ptr;
+    float      * VKQ_chunk_ptr   = VKQ_chunk.ptr;
+    float      * VKQ_accum_ptr   = VKQ_accum.ptr;
+    float      * KQ_max_ptr      = KQ_max.ptr;
+    float      * KQ_sum_ptr      = KQ_sum.ptr;
+    sycl::half * K_chunk_f16_ptr = K_chunk_f16.ptr;
+
+    const float alpha = 1.0f;
+    const float beta  = 0.0f;
+
+    for (int ib = 0; ib < n_batch; ib++) {
+        const float * Q_batch = (const float *)Q->data
+            + ib * (Q->nb[3] / sizeof(float));
+        float * dst_batch = (float *)KQV->data
+            + ib * (KQV->nb[3] / sizeof(float));
+
+        const sycl::half * mask_batch = nullptr;
+        if (mask) {
+            int m_batch = (mask->ne[3] > 1) ? ib : 0;
+            mask_batch = (const sycl::half *)mask->data
+                + m_batch * (mask->nb[3] / 2);  // 2 = actual fp16 device size
+        }
+
+        for (int ikvh = 0; ikvh < n_kv_heads; ikvh++) {
+            int kvh_base_head = ikvh * gqa_ratio;
+
+            // 1. Pack all GQA Q heads into fp16 (full n_query_rows)
+            mkl_fa_pack_q_fp16(stream,
+                Q_head_f16_ptr, Q_batch,
+                n_queries, n_query_rows, DKQ,
+                gqa_ratio, kvh_base_head,
+                q_scale, q_row_stride, q_head_stride, wg_size);
+
+            // 2. Initialize softmax state (full n_query_rows)
+            mkl_fa_init_softmax_state(stream,
+                KQ_max_ptr, KQ_sum_ptr, VKQ_accum_ptr,
+                n_query_rows, DV, wg_size);
+
+            // Sync before MKL GEMM (MKL may use an internal queue)
+            stream->wait();
+
+            // 3. KV chunk loop (OUTER): dequant each chunk once, then tile queries.
+            for (int chunk_start = 0; chunk_start < n_kv; chunk_start += chunk_size) {
+                int this_chunk = std::min(chunk_size, n_kv - chunk_start);
+
+                // 3a. Dequant this KV chunk to dense fp16 (once per chunk)
+                {
+                    MKL_TAKE_TIME(t0);
+                    mkl_fa_dequant_chunk(stream, K_desc, KQV,
+                        K_chunk_f16_ptr, ikvh, chunk_start, this_chunk);
+                    if (!V_is_K_view) {
+                        mkl_fa_dequant_chunk(stream, V_desc, KQV,
+                            V_chunk_f16_ptr, ikvh, chunk_start, this_chunk);
+                    }
+                    stream->wait();  // dequant must be ready before MKL GEMM
+                    MKL_ACCUM(dequant_time_us, t0);
+                }
+
+                // 3b. Query tile loop (INNER) — bounds KQ_f32/S_f16 footprint.
+                for (int q0 = 0; q0 < n_query_rows; q0 += q_tile_rows) {
+                    int q_rows = std::min(q_tile_rows, n_query_rows - q0);
+
+                    // GEMM: KQ = Q_tile × K_chunk^T
+                    {
+                        MKL_TAKE_TIME(t0);
+                        sycl::event ev = gemm(*stream,
+                            transpose::trans, transpose::nontrans,
+                            this_chunk, q_rows, DKQ,
+                            alpha,
+                            K_chunk_f16_ptr, DKQ,
+                            Q_head_f16_ptr + (int64_t)q0 * DKQ, DKQ,
+                            beta,
+                            KQ_f32_ptr, this_chunk);
+                        try { ev.wait_and_throw(); } catch (sycl::exception & e) {
+                            GGML_LOG_INFO("[MKL-FA] GEMM KQ: %s\n", e.what());
+                            GGML_ABORT("MKL GEMM KQ failed");
+                        }
+                        MKL_ACCUM(gemm_kq_time_us, t0);
+                    }
+                    // Online softmax over this chunk for this query tile
+                    {
+                        MKL_TAKE_TIME(t0);
+                        mkl_fa_online_softmax_chunk(stream,
+                            KQ_f32_ptr, S_f16_ptr,
+                            KQ_max_ptr, KQ_sum_ptr, VKQ_accum_ptr,
+                            q0, q_rows, n_queries, DV,
+                            this_chunk, chunk_start,
+                            kvh_base_head, gqa_ratio,
+                            mask_batch, mask_head_stride,
+                            mask_row_stride, mask_n_heads,
+                            logit_softcap, wg_size);
+                        stream->wait();  // S_f16 must be ready for GEMM
+                        MKL_ACCUM(softmax_time_us, t0);
+                    }
+
+                    // GEMM: VKQ_chunk = S × V_chunk
+                    {
+                        MKL_TAKE_TIME(t0);
+                        sycl::event ev = gemm(*stream,
+                            transpose::nontrans, transpose::nontrans,
+                            DV, q_rows, this_chunk,
+                            alpha,
+                            V_chunk_f16_ptr, DV,
+                            S_f16_ptr, this_chunk,
+                            beta,
+                            VKQ_chunk_ptr, DV);
+                        try { ev.wait_and_throw(); } catch (sycl::exception & e) {
+                            GGML_LOG_INFO("[MKL-FA] GEMM VKQ: %s\n", e.what());
+                            GGML_ABORT("MKL GEMM VKQ failed");
+                        }
+                        MKL_ACCUM(gemm_vkq_time_us, t0);
+                    }
+                    // VKQ_accum[q0..] += VKQ_chunk
+                    {
+                        const int64_t n_total = (int64_t)q_rows * DV;
+                        const int64_t wg = ((n_total + wg_size - 1) / wg_size)
+                            * wg_size;
+                        float * accum = VKQ_accum_ptr + (int64_t)q0 * DV;
+                        stream->submit([&](sycl::handler & cgh) {
+                            cgh.parallel_for(sycl::nd_range<1>(wg, wg_size),
+                                [=](sycl::nd_item<1> item) {
+                                    int64_t i = item.get_global_id(0);
+                                    if (i < n_total) {
+                                        accum[i] += VKQ_chunk_ptr[i];
+                                    }
+                                });
+                        });
+                    }
+                }
+            }
+
+            // 4. Normalize and scatter each GQA head to dst
+            for (int iqg = 0; iqg < gqa_ratio; iqg++) {
+                int     iqh        = kvh_base_head + iqg;
+                int64_t src_offset = (int64_t)iqg * n_queries * DV;
+                mkl_fa_normalize_head(stream,
+                    dst_batch, VKQ_accum_ptr, KQ_sum_ptr,
+                    iqh, n_queries, DV, n_q_heads,
+                    src_offset, wg_size);
+            }
+        }
+    }
+
+#undef MKL_TAKE_TIME
+#undef MKL_ACCUM
+
+    if (do_print) {
+        const int64_t v_chunk_elems = V_is_K_view ? 0 : (int64_t)chunk_size * DV;
+        double total_mb = (double)(
+            (int64_t)q_tile_rows * chunk_size * sizeof(float)      // KQ_f32
+          + (int64_t)q_tile_rows * chunk_size * sizeof(sycl::half) // S_f16
+          + (int64_t)q_tile_rows * DV * sizeof(float)              // VKQ_chunk
+          + (int64_t)n_query_rows * DV * sizeof(float)             // VKQ_accum
+          + (int64_t)n_query_rows * sizeof(float)                  // KQ_max
+          + (int64_t)n_query_rows * sizeof(float)                  // KQ_sum
+          + (int64_t)n_query_rows * DKQ * sizeof(sycl::half)       // Q_head_f16
+          + (int64_t)chunk_size * DKQ * sizeof(sycl::half)         // K_chunk_f16
+          + v_chunk_elems * (int64_t)sizeof(sycl::half)            // V_chunk_f16
+        ) / (1024.0 * 1024.0);
+        GGML_LOG_INFO("[MKL-FA] #%d n_kv=%d n_q=%d q_tile=%d time_us: "
+                "dequant=%lld GEMM_KQ=%lld softmax=%lld GEMM_VKQ=%lld "
+                "buf_mb=%.1f\n",
+                mkl_call_count, n_kv, n_queries, q_tile_rows,
+                (long long)dequant_time_us,
+                (long long)gemm_kq_time_us,
+                (long long)softmax_time_us,
+                (long long)gemm_vkq_time_us,
+                total_mb);
+    }
+}

ggml/src/ggml-sycl/fattn-onednn.cpp

@@ -2,11 +2,13 @@
 #include <cstdio>
 #include <cstring>
 #include <string>
+#include <optional>
 #include <unordered_map>
 #include <vector>
 
 #include "fattn-onednn.hpp"
 #include "fattn-tile.hpp"
+#include "convert.hpp"
 
 // set minimum query length to treat as prefill (32)
 #define GGML_SYCL_FA_ONEDNN_MIN_Q 32
@@ -33,10 +35,30 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
     const ggml_tensor * mask  = dst->src[3];
     const ggml_tensor * sinks = dst->src[4];
 
-    // gate for f16 KV only for now
-    // need to implement quantized KV
+    // F16 KV: native SDPA at any KV length.
+    // Non-F16: dequant to F16 then SDPA at prefill lengths. Only the
+    // standard quantized KV cache types (Q4_0-Q8_0) and F32 are accepted
+    // because their to_fp16_sycl conversion is verified. BF16 and IQ*
+    // are excluded: BF16 needs a strided conversion kernel that does not
+    // exist yet; IQ types are model-weight-only quants with no dequant
+    // registration and are never used as KV caches.
     if (K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16) {
-        return false;
+        auto kt = K->type, vt = V->type;
+        bool k_ok = kt == GGML_TYPE_F32 || kt == GGML_TYPE_Q4_0 || kt == GGML_TYPE_Q4_1 ||
+                    kt == GGML_TYPE_Q5_0 || kt == GGML_TYPE_Q5_1 || kt == GGML_TYPE_Q8_0;
+        bool v_ok = vt == GGML_TYPE_F32 || vt == GGML_TYPE_Q4_0 || vt == GGML_TYPE_Q4_1 ||
+                    vt == GGML_TYPE_Q5_0 || vt == GGML_TYPE_Q5_1 || vt == GGML_TYPE_Q8_0;
+        if (!k_ok || !v_ok) {
+            return false;
+        }
+        if (Q->ne[1] < 32 || K->ne[1] < 1024) {
+            return false;
+        }
+        for (const ggml_tensor * t : {K, V}) {
+            if (t->type == GGML_TYPE_F16 && t->nb[1] % (t->ne[0] * 2) != 0) {
+                return false;
+            }
+        }
     }
     // Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
     // very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
@@ -205,13 +227,101 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
     dnnl::engine    eng    = ctx.engine_dnnl(stream);
     dnnl::stream    strm   = ctx.stream_dnnl(stream);
 
-    // cont/cast inputs to contiguous f16 (head-major) -- the layout the fast systolic path wants.
-    ggml_sycl_pool_alloc<sycl::half> Qf(ctx.pool(), (size_t) H   * q   * d);
-    ggml_sycl_pool_alloc<sycl::half> Kf(ctx.pool(), (size_t) Hkv * seq * d);
-    ggml_sycl_pool_alloc<sycl::half> Vf(ctx.pool(), (size_t) Hkv * seq * d);
-    cont_to_f16_sycl<float>     ((const char *) Q->data, Qf.get(), d, q,   H,   mb, Q->nb[1], Q->nb[2], Q->nb[3], stream);
-    cont_to_f16_sycl<sycl::half>((const char *) K->data, Kf.get(), d, seq, Hkv, mb, K->nb[1], K->nb[2], K->nb[3], stream);
-    cont_to_f16_sycl<sycl::half>((const char *) V->data, Vf.get(), d, seq, Hkv, mb, V->nb[1], V->nb[2], V->nb[3], stream);
+    // Q: always f32 -- copy to dense f16.
+    ggml_sycl_pool_alloc<sycl::half> Qf(ctx.pool(), (size_t) H * q * d);
+    cont_to_f16_sycl<float>((const char *) Q->data, Qf.get(), d, q, H, mb, Q->nb[1], Q->nb[2], Q->nb[3], stream);
+
+    // K/V: use pool-alloc for both F16 and dequant paths.
+    sycl::half * K_ptr = nullptr;
+    sycl::half * V_ptr = nullptr;
+    std::optional<ggml_sycl_pool_alloc<sycl::half>> Kf_pool;
+    std::optional<ggml_sycl_pool_alloc<sycl::half>> Vf_pool;
+
+    if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+        Kf_pool.emplace(ctx.pool(), (size_t) Hkv * seq * d);
+        Vf_pool.emplace(ctx.pool(), (size_t) Hkv * seq * d);
+        cont_to_f16_sycl<sycl::half>((const char *) K->data, Kf_pool->get(), d, seq, Hkv, mb, K->nb[1], K->nb[2], K->nb[3], stream);
+        cont_to_f16_sycl<sycl::half>((const char *) V->data, Vf_pool->get(), d, seq, Hkv, mb, V->nb[1], V->nb[2], V->nb[3], stream);
+        K_ptr = Kf_pool->get();
+        V_ptr = Vf_pool->get();
+    } else if (ggml_is_quantized(K->type)) {
+        // Quantized K/V: dequant to dense F16 using pool, same lifetime as F16 path.
+        Kf_pool.emplace(ctx.pool(), ggml_nelements(K));
+        K_ptr = Kf_pool->get();
+        {
+            const char * K_data = (const char *)K->data;
+            const bool k_non_dense = ((int64_t)K->ne[1] * K->nb[1] != K->nb[2]) && K->ne[2] > 1;
+            const bool k_gemma = k_non_dense &&
+                ((int64_t)K->nb[2] < (int64_t)K->ne[1] * (int64_t)K->nb[1]);
+            if (ggml_is_contiguously_allocated(K) && !k_non_dense) {
+                to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(K->type, dst);
+                to_fp16(K_data, K_ptr, ggml_nelements(K), stream);
+            } else {
+                const size_t bs = ggml_blck_size(K->type);
+                const size_t ts = ggml_type_size(K->type);
+                to_fp16_nc_sycl_t to_fp16 = ggml_get_to_fp16_nc_sycl(K->type);
+                int64_t s01, s02, s03;
+                if (k_gemma) {
+                    const int64_t blk_per_row = (int64_t)K->ne[0] / bs;
+                    s01 = (int64_t)Hkv * blk_per_row;
+                    s02 = blk_per_row;
+                    s03 = (int64_t)K->ne[1] * s01;
+                } else {
+                    s01 = (int64_t)K->nb[1] / ts;
+                    s02 = (int64_t)K->nb[2] / ts;
+                    s03 = (int64_t)K->nb[3] / ts;
+                }
+                to_fp16(K_data, K_ptr,
+                        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                        s01, s02, s03, stream);
+            }
+        }
+        // Quantized V: always dequant separately. Even when K and V share
+        // the same underlying allocation (V is a view of K with the same
+        // data pointer), their logical values differ because the quantized
+        // elements at different positions/offsets represent different K/V
+        // data. Master's F16 path also never aliases K and V.
+        Vf_pool.emplace(ctx.pool(), ggml_nelements(V));
+        V_ptr = Vf_pool->get();
+        {
+            const char * V_data = (const char *)V->data;
+            const bool v_non_dense = ((int64_t)V->ne[1] * V->nb[1] != V->nb[2]) && V->ne[2] > 1;
+            const bool v_gemma = v_non_dense &&
+                ((int64_t)V->nb[2] < (int64_t)V->ne[1] * (int64_t)V->nb[1]);
+            if (ggml_is_contiguously_allocated(V) && !v_non_dense) {
+                to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(V->type, dst);
+                to_fp16(V_data, V_ptr, ggml_nelements(V), stream);
+            } else {
+                const size_t bs = ggml_blck_size(V->type);
+                const size_t ts = ggml_type_size(V->type);
+                to_fp16_nc_sycl_t to_fp16 = ggml_get_to_fp16_nc_sycl(V->type);
+                int64_t s01, s02, s03;
+                if (v_gemma) {
+                    const int64_t blk_per_row = (int64_t)V->ne[0] / bs;
+                    s01 = (int64_t)V->ne[2] * blk_per_row;
+                    s02 = blk_per_row;
+                    s03 = (int64_t)V->ne[1] * s01;
+                } else {
+                    s01 = (int64_t)V->nb[1] / ts;
+                    s02 = (int64_t)V->nb[2] / ts;
+                    s03 = (int64_t)V->nb[3] / ts;
+                }
+                to_fp16(V_data, V_ptr,
+                        V->ne[0], V->ne[1], V->ne[2], V->ne[3],
+                        s01, s02, s03, stream);
+            }
+        }
+    } else {
+        // F32: strided copy to dense F16 via cont_to_f16_sycl<float>.
+        Kf_pool.emplace(ctx.pool(), ggml_nelements(K));
+        K_ptr = Kf_pool->get();
+        cont_to_f16_sycl<float>((const char *) K->data, K_ptr, K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                                K->nb[1], K->nb[2], K->nb[3], stream);
+        Vf_pool.emplace(ctx.pool(), ggml_nelements(V));
+        V_ptr = Vf_pool->get();
+        cont_to_f16_sycl<float>((const char *) V->data, V_ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3],
+                                V->nb[1], V->nb[2], V->nb[3], stream);
+    }
 
     // divide-by-(1/scale) reproduces ggml's score *= kq_scale on the proven probe graph.
     //
@@ -244,8 +354,8 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
 
     auto id2ptr = [&](size_t r) -> void * {
         if (r == E.id_q)     return Qf.get();
-        if (r == E.id_k)     return Kf.get();
-        if (r == E.id_v)     return Vf.get();
+        if (r == E.id_k)     return K_ptr;
+        if (r == E.id_v)     return V_ptr;
         if (r == E.id_scale) return scale_dev;
         if (r == E.id_mask)  return (void *) mask->data;
         return nullptr;

ggml/src/ggml-sycl/fattn.cpp

@@ -97,10 +97,12 @@ static void ggml_sycl_flash_attn_ext_vec(ggml_backend_sycl_context & ctx, ggml_t
 enum best_fattn_kernel {
     BEST_FATTN_KERNEL_NONE     =   0,
     BEST_FATTN_KERNEL_VEC      = 100,
-    BEST_FATTN_KERNEL_ONEDNN   = 150, // added enum for onednn==150
+    BEST_FATTN_KERNEL_ONEDNN   = 150, // oneDNN SDPA: native F16 (PR #25222)
     BEST_FATTN_KERNEL_TILE     = 200,
+    BEST_FATTN_KERNEL_MKL      = 300,
 };
 
+
 static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const ggml_tensor * dst) {
     GGML_UNUSED(device);
 #ifndef SYCL_FLASH_ATTN
@@ -115,6 +117,7 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
     const ggml_tensor * K     = dst->src[1];
     const ggml_tensor * V     = dst->src[2];
     const ggml_tensor * mask  = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
 
     const int gqa_ratio = Q->ne[2] / K->ne[2];
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
@@ -122,7 +125,56 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
     float max_bias = 0.0f;
     memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
 
+    float logit_softcap = 0.0f;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
     bool gqa_opt_applies = gqa_ratio >= 2 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+
+    // XMX-accelerated path: oneDNN SDPA (native F16 and dequant+non-F16).
+    // ONEDNN requires min 32 query tokens — short-circuit decode to avoid
+    // calling _supported() on every decode FA call.
+    if (Q->ne[1] >= 32
+        && ggml_sycl_flash_attn_ext_onednn_supported(dst)) {
+        return BEST_FATTN_KERNEL_ONEDNN;
+    }
+
+    // MKL path: XMX-accelerated GEMM for prompt processing (all KV cache types).
+    // The MKL kernel converts non-F16 K/V to F16 via to_fp16_sycl before GEMM,
+    // so quantized, F16, BF16, and F32 caches all benefit from XMX acceleration.
+    // Activates automatically when flash-attn is enabled (--flash-attn on or -fa)
+    // and n_kv >= 1024. Falls through to TILE/VEC for ALiBi, logit softcap,
+    // and mismatched batch dimensions (unsupported by the MKL kernel).
+    // Set GGML_SYCL_ENABLE_MKL_FA=0 to force TILE/VEC path for A/B testing.
+    // Example: GGML_SYCL_ENABLE_MKL_FA=0 llama-cli -m model.gguf -fa -ngl 99 ...
+    // Note: MKL GEMM calls are incompatible with SYCL graph capture replay.
+    static int mkl_enable = ggml_sycl_get_env("GGML_SYCL_ENABLE_MKL_FA", 1);
+    // MKL is validated for the mainstream GQA envelope: grouped-query
+    // (gqa_ratio >= 2), head_dim a multiple of 64 in [64,512] with matching
+    // K/V head size, mask, no sinks/ALiBi/softcap. Gemma's global layers use
+    // head_dim 512, so the cap must include it. Head sizes not a multiple of
+    // 64 (72/80/96), MHA (gqa_ratio == 1), and MLA (DKQ != DV, e.g. 576/512)
+    // fall through to TILE/VEC; see follow-up work.
+    if (mkl_enable == 1 && mask && !sinks && gqa_ratio >= 2 &&
+        Q->ne[0] >= 64 && Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 &&
+        Q->ne[0] == V->ne[0] &&
+        Q->ne[1] >= 32 && K->ne[1] >= 1024 &&
+        max_bias == 0.0f && logit_softcap == 0.0f &&
+        (Q->ne[3] == K->ne[3] || K->ne[3] == 1)) {
+        // F16 K/V strides must be a multiple of ne[0]*2 (the natural row size
+        // in bytes). This passes both dense (nb1 == ne0*2) and interleaved
+        // (nb1 == H * ne0*2). Only pathological test strides like nb1=32 or
+        // nb1=75 for ne0=40 fall through to TILE.
+        bool kv_strides_ok = true;
+        for (const ggml_tensor * t : {K, V}) {
+            if (t->type == GGML_TYPE_F16 && t->nb[1] % (t->ne[0] * 2) != 0) {
+                kv_strides_ok = false;
+                break;
+            }
+        }
+        if (kv_strides_ok) {
+            return BEST_FATTN_KERNEL_MKL;
+        }
+    }
     for (const ggml_tensor * t : {Q, K, V, mask}) {
         if (t == nullptr || ggml_is_quantized(t->type)) {
             continue;
@@ -170,6 +222,7 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
     switch (K->type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
             break;
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -188,8 +241,11 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
         return BEST_FATTN_KERNEL_NONE;
     }
 
-    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
-    const bool can_use_vector_kernel = Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes.
+    // BF16 is excluded: the VEC kernel has no BF16 template (it needs GGML_SYCL_FA_ALL_QUANTS for non-F16/Q4_0/Q8_0).
+    const bool has_bf16 = (K->type == GGML_TYPE_BF16 || V->type == GGML_TYPE_BF16);
+    const bool can_use_vector_kernel = Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0
+        && !has_bf16;
 
     // Fused-XMX path: oneDNN Graph SDPA (flash attention). Strictly
     // additive -- taken only when statically supported, otherwise falls through to VEC/TILE below.
@@ -216,7 +272,40 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
 
 void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_set_device(ctx.device);
-    switch (ggml_sycl_get_best_fattn_kernel(ggml_sycl_get_device(), dst)) {
+
+    // n_kv watchdog: log when n_kv differs from the last FA call with
+    // the same D — helps detect cache-truncation issues.
+    static int nkv_debug = ggml_sycl_get_env("GGML_SYCL_MKL_FA_DEBUG", 0);
+    if (nkv_debug == 1) {
+        const ggml_tensor * K_dbg = dst->src[1];
+        const ggml_tensor * V_dbg = dst->src[2];
+        static int64_t last_nkv_d256 = 0, last_nkv_d512 = 0;
+        static int fa_call_seq = 0;
+        fa_call_seq++;
+        int64_t cur_nkv = K_dbg->ne[1];
+        int Dk = (int)K_dbg->ne[0];
+        const char * kname = "TILE";
+        best_fattn_kernel k = ggml_sycl_get_best_fattn_kernel(ctx.device, dst);
+        if (k == BEST_FATTN_KERNEL_MKL)  kname = "MKL";
+        if (k == BEST_FATTN_KERNEL_ONEDNN)  kname = "ONEDNN";
+        if (k == BEST_FATTN_KERNEL_VEC)  kname = "VEC";
+        int64_t delta = 0;
+        if (Dk == 256) {
+            delta = cur_nkv - last_nkv_d256;
+            last_nkv_d256 = cur_nkv;
+        } else if (Dk == 512) {
+            delta = cur_nkv - last_nkv_d512;
+            last_nkv_d512 = cur_nkv;
+        }
+        GGML_LOG_INFO("[FA-DISP] #%d %s D=%d n_kv=%lld delta=%lld "
+                "V_ne1=%lld\n",
+                fa_call_seq, kname, Dk,
+                (long long)cur_nkv, (long long)delta,
+                (long long)V_dbg->ne[1]);
+    }
+
+    const best_fattn_kernel fk = ggml_sycl_get_best_fattn_kernel(ggml_sycl_get_device(), dst);
+    switch (fk) {
         case BEST_FATTN_KERNEL_NONE:
             GGML_ABORT("Not support Flash-Attention");
         case BEST_FATTN_KERNEL_ONEDNN:
@@ -232,7 +321,54 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
         case BEST_FATTN_KERNEL_VEC:
             ggml_sycl_flash_attn_ext_vec(ctx, dst);
             break;
+        case BEST_FATTN_KERNEL_MKL:
+            ggml_sycl_flash_attn_ext_mkl(ctx, dst);
+            break;
     }
+
+    // --- Output fingerprint (GGML_SYCL_MKL_FA_DIAG=1) ---
+    // Copy first 64 float output values to host for fingerprinting.
+    // Compare MKL vs TILE (GGML_SYCL_ENABLE_MKL_FA=0) to detect divergence.
+    // Only fingerprints the first 6 FA calls with n_kv >= 1024.
+    static int fa_diag = ggml_sycl_get_env("GGML_SYCL_MKL_FA_DIAG", 0);
+    static int fa_diag_count = 0;
+    if (fa_diag == 1 && fa_diag_count < 6) {
+        const ggml_tensor * K_diag = dst->src[1];
+        const ggml_tensor * V_diag = dst->src[2];
+        const ggml_tensor * Q_diag = dst->src[0];
+        if (K_diag->ne[1] >= 1024) {
+            fa_diag_count++;
+            float diag_buf[64];
+            dpct::queue_ptr q = ctx.stream();
+            q->memcpy(diag_buf, dst->data, 64 * sizeof(float));
+            q->wait();
+            const char * kname = "???";
+            best_fattn_kernel kb = ggml_sycl_get_best_fattn_kernel(ctx.device, dst);
+            if (kb == BEST_FATTN_KERNEL_ONEDNN) kname = "ONEDNN";
+            if (kb == BEST_FATTN_KERNEL_MKL) kname = "MKL";
+            if (kb == BEST_FATTN_KERNEL_TILE) kname = "TILE";
+            if (kb == BEST_FATTN_KERNEL_VEC) kname = "VEC";
+            GGML_LOG_INFO("[FA-DIAG] #%d %s D=%d n_kv=%lld n_q=%lld "
+                    "n_qh=%lld n_kvh=%lld K=%s V=%s "
+                    "nb1=%zu nb2=%zu first 64 floats:\n",
+                    fa_diag_count, kname,
+                    (int)K_diag->ne[0], (long long)K_diag->ne[1],
+                    (long long)Q_diag->ne[1],
+                    (long long)Q_diag->ne[2], (long long)K_diag->ne[2],
+                    ggml_type_name(K_diag->type),
+                    ggml_type_name(V_diag->type),
+                    K_diag->nb[1], K_diag->nb[2]);
+            for (int i = 0; i < 64; i += 8) {
+                GGML_LOG_INFO("  [%2d] %08x %08x %08x %08x %08x %08x %08x %08x\n",
+                        i,
+                        *(unsigned *)&diag_buf[i+0], *(unsigned *)&diag_buf[i+1],
+                        *(unsigned *)&diag_buf[i+2], *(unsigned *)&diag_buf[i+3],
+                        *(unsigned *)&diag_buf[i+4], *(unsigned *)&diag_buf[i+5],
+                        *(unsigned *)&diag_buf[i+6], *(unsigned *)&diag_buf[i+7]);
+            }
+        }
+    }
+
 }
 
 bool ggml_sycl_flash_attn_ext_supported(int device, const ggml_tensor * dst) {

ggml/src/ggml-sycl/fattn.hpp

@@ -19,4 +19,6 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
 
 bool ggml_sycl_flash_attn_ext_supported(int device, const ggml_tensor * dst);
 
+void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
 #endif // GGML_SYCL_FATTN_HPP

ggml/src/ggml-sycl/fusion.cpp

@@ -0,0 +1,44 @@
+#include "fusion.hpp"
+
+bool ggml_sycl_can_fuse(const ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+    if (!g_ggml_sycl_enable_fusion) {
+        return false;
+    }
+
+    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
+        return false;
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_RMS_NORM && ops.begin()[1] == GGML_OP_MUL) {
+        const ggml_tensor * rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor * mul      = cgraph->nodes[node_idx + 1];
+
+        GGML_ASSERT(rms_norm->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(rms_norm->type == GGML_TYPE_F32);
+
+        if (mul->src[0]->type != GGML_TYPE_F32 ||
+            mul->src[1]->type != GGML_TYPE_F32 ||
+            mul->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        // if rms norm is the B operand, then we don't handle broadcast
+        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm)) {
+            return false;
+        }
+
+        const ggml_tensor * mul_w = (mul->src[0] == rms_norm) ? mul->src[1] : mul->src[0];
+        // the fused kernel indexes the weight as mul[col], so it must span ncols contiguously
+        if (mul_w->ne[0] != rms_norm->ne[0] || mul_w->nb[0] != ggml_type_size(mul_w->type)) {
+            return false;
+        }
+
+        if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+            return false;
+        }
+
+        return true;
+    }
+
+    return false;
+}

ggml/src/ggml-sycl/fusion.hpp

@@ -0,0 +1,15 @@
+#ifndef GGML_SYCL_FUSION_HPP
+#define GGML_SYCL_FUSION_HPP
+
+#include <initializer_list>
+
+#include "common.hpp"
+
+// Backend-side fusability test. `ops` names a candidate op sequence starting at cgraph node
+// `node_idx`; the result is true only if ggml considers that subgraph fusable *and* the SYCL
+// kernel which would service it accepts the tensors involved (types, shapes, contiguity).
+//
+// Lives in its own translation unit because it grows a branch per supported op sequence.
+bool ggml_sycl_can_fuse(const ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops);
+
+#endif  // GGML_SYCL_FUSION_HPP

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -167,7 +167,10 @@ static ggml_sycl_device_info ggml_sycl_init() {
             ze_device_properties_t props = {};
             props.stype = ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES;
             ze_result_t r = zeDeviceGetProperties(ze_dev, &props);
-            info.devices[i].l0_discrete_gpu = r == ZE_RESULT_SUCCESS && !(props.flags & ZE_DEVICE_PROPERTY_FLAG_INTEGRATED);
+            if (r == ZE_RESULT_SUCCESS) {
+                info.devices[i].l0_device_type_valid = true;
+                info.devices[i].l0_discrete_gpu = !(props.flags & ZE_DEVICE_PROPERTY_FLAG_INTEGRATED);
+            }
         }
 #endif
     }
@@ -274,6 +277,8 @@ static const char* dev2dev_int2str(int dev2dev) {
         return "SYCL API";
     } else if (dev2dev == DEV2DEV_MEMCPY_L0) {
         return "Level Zero API";
+    } else if (dev2dev == DEV2DEV_MEMCPY_FORWARD) {
+        return "Host Forward";
     } else {
         return "Unknown";
     }
@@ -684,7 +689,11 @@ static void dev2dev_memcpy(int device_dst, sycl::queue &q_dst, int device_src, s
     }
 
     // Host-staged copy
-    GGML_SYCL_DEBUG("[SYCL] dev2dev memcpy by host forward\n");
+    if(g_ggml_sycl_dev2dev_memcpy == DEV2DEV_MEMCPY_FORWARD) {
+        GGML_SYCL_DEBUG("[SYCL] dev2dev memcpy by host forward for setting GGML_SYCL_DEV2DEV_MEMCPY=2\n");
+    } else {
+        GGML_SYCL_DEBUG("[SYCL] dev2dev memcpy by host forward for SYCL/L0 fallback\n");
+    }
     char *host_buf = (char *)malloc(size);
     q_src.memcpy(host_buf, (const char *)ptr_src, size).wait();
     q_dst.memcpy((char *)ptr_dst, host_buf, size).wait();
@@ -5398,6 +5407,13 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
             }
         }
 #endif
+        if (node->op == GGML_OP_RMS_NORM &&
+            ggml_sycl_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+            ggml_sycl_op_rms_norm_fused(*sycl_ctx, node, cgraph->nodes[i + 1]);
+            i++;
+            continue;
+        }
+
         bool ok = ggml_sycl_compute_forward(*sycl_ctx, node);
         if (!ok) {
             GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
@@ -5593,7 +5609,11 @@ static void ggml_backend_sycl_device_get_memory(ggml_backend_dev_t dev, size_t *
 }
 
 static enum ggml_backend_dev_type ggml_backend_sycl_device_get_type(ggml_backend_dev_t dev) {
-    GGML_UNUSED(dev);
+    ggml_backend_sycl_device_context * ctx = (ggml_backend_sycl_device_context *)dev->context;
+    const sycl_device_info & info = ggml_sycl_info().devices[ctx->device];
+    if (info.l0_device_type_valid && !info.l0_discrete_gpu) {
+        return GGML_BACKEND_DEVICE_TYPE_IGPU;
+    }
     return GGML_BACKEND_DEVICE_TYPE_GPU;
 }
 

ggml/src/ggml-sycl/mmvq.cpp

@@ -1254,6 +1254,66 @@ static void mul_mat_vec_q1_0_q8_1_sycl_switch_ncols(
     }
 }
 
+static void mul_mat_vec_q2_0_q8_1_sycl(const void * vx, const void * vy,
+                                       float * dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK2_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                mul_mat_vec_q<QK2_0, QI2_0, block_q2_0,
+                              VDR_Q2_0_Q8_1_MMVQ, vec_dot_q2_0_q8_1>(
+                    vx, vy, dst, ncols, nrows, item_ct1);
+            });
+    });
+}
+
+template <int ncols_dst>
+static void mul_mat_vec_q2_0_q8_1_sycl_ncols(
+        const void * vx, const void * vy, float * dst,
+        const int ncols, const int nrows,
+        const int stride_col_y, const int stride_col_dst,
+        dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK2_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                mul_mat_vec_q_ncols<QK2_0, QI2_0, block_q2_0,
+                                    VDR_Q2_0_Q8_1_MMVQ, vec_dot_q2_0_q8_1, ncols_dst>(
+                    vx, vy, dst, ncols, nrows, stride_col_y, stride_col_dst, item_ct1);
+            });
+    });
+}
+
+static void mul_mat_vec_q2_0_q8_1_sycl_switch_ncols(
+        const void * vx, const void * vy, float * dst,
+        const int ncols, const int nrows, const int ncols_dst,
+        const int stride_col_y, const int stride_col_dst,
+        dpct::queue_ptr stream) {
+    switch (ncols_dst) {
+        case 1: mul_mat_vec_q2_0_q8_1_sycl(vx, vy, dst, ncols, nrows, stream); break;
+        case 2: mul_mat_vec_q2_0_q8_1_sycl_ncols<2>(vx, vy, dst, ncols, nrows, stride_col_y, stride_col_dst, stream); break;
+        case 3: mul_mat_vec_q2_0_q8_1_sycl_ncols<3>(vx, vy, dst, ncols, nrows, stride_col_y, stride_col_dst, stream); break;
+        case 4: mul_mat_vec_q2_0_q8_1_sycl_ncols<4>(vx, vy, dst, ncols, nrows, stride_col_y, stride_col_dst, stream); break;
+        case 5: mul_mat_vec_q2_0_q8_1_sycl_ncols<5>(vx, vy, dst, ncols, nrows, stride_col_y, stride_col_dst, stream); break;
+        case 6: mul_mat_vec_q2_0_q8_1_sycl_ncols<6>(vx, vy, dst, ncols, nrows, stride_col_y, stride_col_dst, stream); break;
+        case 7: mul_mat_vec_q2_0_q8_1_sycl_ncols<7>(vx, vy, dst, ncols, nrows, stride_col_y, stride_col_dst, stream); break;
+        case 8: mul_mat_vec_q2_0_q8_1_sycl_ncols<8>(vx, vy, dst, ncols, nrows, stride_col_y, stride_col_dst, stream); break;
+        default: GGML_ABORT("unsupported ncols_dst=%d for Q2_0 multi-col MMVQ", ncols_dst);
+    }
+}
+
 static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
                                        float *dst, const int ncols,
                                        const int nrows,
@@ -2194,6 +2254,20 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                     mul_mat_vec_q1_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                 }
                 break;
+            case GGML_TYPE_Q2_0:
+                if (i == 0 && src1_ncols > 1 && src1_ncols <= 8) {
+                    const int stride_col_y   = src1_padded_col_size / QK8_1;
+                    const int stride_col_dst = dst->ne[0];
+                    GGML_SYCL_DEBUG("Calling mul_mat_vec_q2_0_q8_1_sycl_switch_ncols ncols=%d\n", (int)src1_ncols);
+                    mul_mat_vec_q2_0_q8_1_sycl_switch_ncols(
+                        src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff,
+                        src1_ncols, stride_col_y, stride_col_dst, stream);
+                    return;
+                } else if (i == 0 || src1_ncols == 1) {
+                    GGML_SYCL_DEBUG("Calling mul_mat_vec_q2_0_q8_1_sycl\n");
+                    mul_mat_vec_q2_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                }
+                break;
             case GGML_TYPE_Q2_K:
                 if (i == 0 && src1_ncols > 1 && src1_ncols <= 8) {
                     const int stride_col_y   = src1_padded_col_size / QK8_1;
@@ -2503,6 +2577,11 @@ bool ggml_sycl_mul_mat_vec_q_id(
                 vx_base, vy, ids_dev, dst_base, ncols, nrows, n_experts_used,
                 expert_weight_stride, dst_row_stride, src1_row_stride, stream);
             return true;
+        case GGML_TYPE_Q2_0:
+            launch_mul_mat_vec_q_moe<QK2_0, QI2_0, block_q2_0, VDR_Q2_0_Q8_1_MMVQ, vec_dot_q2_0_q8_1>(
+                vx_base, vy, ids_dev, dst_base, ncols, nrows, n_experts_used,
+                expert_weight_stride, dst_row_stride, src1_row_stride, stream);
+            return true;
         case GGML_TYPE_Q2_K:
             launch_mul_mat_vec_q_moe<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
                 vx_base, vy, ids_dev, dst_base, ncols, nrows, n_experts_used,

ggml/src/ggml-sycl/norm.cpp

@@ -147,10 +147,13 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
     }
 }
 
+template <bool do_multiply = false>
 static void rms_norm_f32(const float* x, float* dst, const int ncols,
     const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
     const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
-    const float eps, const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
+    const float eps, const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size,
+    const float* mul = nullptr, const int64_t mul_stride_row = 0, const int64_t mul_stride_channel = 0,
+    const int64_t mul_stride_sample = 0, const int mul_nrows = 0, const int mul_nchannels = 0, const int mul_nsamples = 0) {
 
     const int nrows = item_ct1.get_group_range(2);
     const int nchannels = item_ct1.get_group_range(1);
@@ -170,6 +173,12 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols,
     x   += src_offset;
     dst += dst_offset;
 
+    if constexpr (do_multiply) {
+        const int mul_row     = row     % mul_nrows;
+        const int mul_channel = channel % mul_nchannels;
+        const int mul_sample  = sample  % mul_nsamples;
+        mul += mul_sample * mul_stride_sample + mul_channel * mul_stride_channel + mul_row * mul_stride_row;
+    }
 
     float tmp = 0.0f; // partial sum for thread in warp
 
@@ -202,7 +211,11 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols,
     const float scale = sycl::rsqrt(mean + eps);
 
     for (int col = tid; col < ncols; col += block_size) {
-        dst[col * dst_stride_col] = scale * x[col * src_stride_col];
+        if constexpr (do_multiply) {
+            dst[col * dst_stride_col] = scale * x[col * src_stride_col] * mul[col];
+        } else {
+            dst[col * dst_stride_col] = scale * x[col * src_stride_col];
+        }
     }
 }
 
@@ -376,6 +389,49 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
     }
 }
 
+static void rms_norm_mul_f32_sycl(const float* x, const float* mul, float* dst, const int ncols, const int nrows,
+        const int nchannels, const int nsamples,
+        const int64_t src_stride_col, const int64_t src_stride_row, const int64_t src_stride_channel, const int64_t src_stride_sample,
+        const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
+        const int64_t mul_stride_row, const int64_t mul_stride_channel, const int64_t mul_stride_sample,
+        const int mul_nrows, const int mul_nchannels, const int mul_nsamples,
+        const float eps, queue_ptr stream, int device) {
+    const sycl::range<3> global_dims(nsamples, nchannels, nrows);
+    if (ncols < 1024) {
+        const sycl::range<3> block_dims(1, 1, WARP_SIZE);
+        stream->submit([&](sycl::handler& cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    rms_norm_f32<true>(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, nullptr, WARP_SIZE,
+                        mul, mul_stride_row, mul_stride_channel, mul_stride_sample, mul_nrows, mul_nchannels, mul_nsamples);
+                });
+            });
+    }
+    else {
+        const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
+        assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
+        const sycl::range<3> block_dims(1, 1, work_group_size);
+        stream->submit([&](sycl::handler& cgh) {
+            sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), cgh);
+            cgh.parallel_for(
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    rms_norm_f32<true>(x, dst, ncols,
+                        src_stride_col, src_stride_row, src_stride_channel, src_stride_sample,
+                        dst_stride_col, dst_stride_row, dst_stride_channel, dst_stride_sample,
+                        eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size,
+                        mul, mul_stride_row, mul_stride_channel, mul_stride_sample, mul_nrows, mul_nchannels, mul_nsamples);
+                });
+            });
+    }
+}
+
 template<int warp_size>
 static void l2_norm_f32_sycl(const float *   x,
                              float *         dst,
@@ -518,6 +574,66 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
         ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, main_stream, ctx.device);
 }
 
+void ggml_sycl_op_rms_norm_fused(ggml_backend_sycl_context & ctx, ggml_tensor * dst, ggml_tensor * mul_tensor) {
+    const ggml_tensor * rms_norm_src = dst->src[0];
+    float eps = 0.0f;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    const float *       src0_dd = static_cast<const float *>(rms_norm_src->data);
+    const float *       mul_dd  = nullptr;
+    const ggml_tensor * mul_src = nullptr;
+    if (mul_tensor->src[0] == dst) {
+        mul_dd  = static_cast<const float *>(mul_tensor->src[1]->data);
+        mul_src = mul_tensor->src[1];
+    } else if (mul_tensor->src[1] == dst) {
+        mul_dd  = static_cast<const float *>(mul_tensor->src[0]->data);
+        mul_src = mul_tensor->src[0];
+    } else {
+        GGML_ASSERT(false);
+    }
+    float * dst_dd = static_cast<float *>(mul_tensor->data);
+
+    dpct::queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    GGML_ASSERT(rms_norm_src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(mul_tensor->type == GGML_TYPE_F32);
+    GGML_ASSERT(eps >= 0.0f);
+
+    const int64_t ne00 = rms_norm_src->ne[0];
+    const int64_t ne01 = rms_norm_src->ne[1];
+    const int64_t ne02 = rms_norm_src->ne[2];
+    const int64_t ne03 = rms_norm_src->ne[3];
+
+    const size_t ts0 = ggml_type_size(rms_norm_src->type);
+    GGML_ASSERT(rms_norm_src->nb[0] == ts0);
+    const int64_t s00 = rms_norm_src->nb[0] / ts0;
+    const int64_t s01 = rms_norm_src->nb[1] / ts0;
+    const int64_t s02 = rms_norm_src->nb[2] / ts0;
+    const int64_t s03 = rms_norm_src->nb[3] / ts0;
+
+    const size_t tdst = ggml_type_size(mul_tensor->type);
+    GGML_ASSERT(mul_tensor->nb[0] == tdst);
+    const int64_t d00 = mul_tensor->nb[0] / tdst;
+    const int64_t d01 = mul_tensor->nb[1] / tdst;
+    const int64_t d02 = mul_tensor->nb[2] / tdst;
+    const int64_t d03 = mul_tensor->nb[3] / tdst;
+
+    const size_t ts_mul = ggml_type_size(mul_src->type);
+    GGML_ASSERT(mul_src->nb[0] == ts_mul);
+    const int64_t mul_s01 = mul_src->nb[1] / ts_mul;
+    const int64_t mul_s02 = mul_src->nb[2] / ts_mul;
+    const int64_t mul_s03 = mul_src->nb[3] / ts_mul;
+    const int mul_nrows     = mul_src->ne[1];
+    const int mul_nchannels = mul_src->ne[2];
+    const int mul_nsamples  = mul_src->ne[3];
+
+    rms_norm_mul_f32_sycl(src0_dd, mul_dd, dst_dd, ne00, ne01, ne02, ne03,
+        s00, s01, s02, s03, d00, d01, d02, d03,
+        mul_s01, mul_s02, mul_s03, mul_nrows, mul_nchannels, mul_nsamples, eps, main_stream, ctx.device);
+}
+
 void ggml_sycl_op_rms_norm_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
 

ggml/src/ggml-sycl/norm.hpp

@@ -19,6 +19,8 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
 void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
+void ggml_sycl_op_rms_norm_fused(ggml_backend_sycl_context& ctx, ggml_tensor* dst, ggml_tensor* mul);
+
 void ggml_sycl_op_rms_norm_back(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
 void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);

ggml/src/ggml-sycl/vecdotq.hpp

@@ -658,6 +658,40 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
+#define VDR_Q2_0_Q8_1_MMVQ 1
+
+template <int vdr>
+static __dpct_inline__ float vec_dot_q2_0_q8_1_impl(
+    const int * v,
+    const int * u,
+    const float & d2,
+    const sycl::half2 & ds8) {
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+#pragma unroll
+        for (int j = 0; j < 4; ++j) {
+            const uint8_t q = (uint8_t) ((uint32_t) v[i] >> (8 * j));
+
+            // unpack 2-bit values to byte lanes (0..3), then apply zero-point
+            // correction with ds8f.y() below, mirroring the q4_0 style.
+            int vi = 0;
+            vi |= (((q >> 0) & 0x3) & 0xFF) << 0;
+            vi |= (((q >> 2) & 0x3) & 0xFF) << 8;
+            vi |= (((q >> 4) & 0x3) & 0xFF) << 16;
+            vi |= (((q >> 6) & 0x3) & 0xFF) << 24;
+
+            sumi = dpct::dp4a(vi, u[4 * i + j], sumi);
+        }
+    }
+
+    const sycl::float2 ds8f = ds8.convert<float, sycl::rounding_mode::automatic>();
+    // q2_0 has zero-point 1. Scale ds8f.y() by processed-lane ratio,
+    // consistent with q4_0's explicit zero-point subtraction style.
+    return d2 * (sumi * ds8f.x() - ((float) vdr / (float) QI2_0) * ds8f.y());
+}
+
 template <int vdr>
 static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int * v, const int * u, const float & d4,
                                                     const sycl::half2 & ds8) {
@@ -882,6 +916,41 @@ vec_dot_q4_0_q8_1(const void *__restrict__ vbq,
     return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);
 }
 
+static __dpct_inline__ float
+vec_dot_q2_0_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q2_0 * bq2_0 = (const block_q2_0 *) vbq;
+
+    int v[2 * VDR_Q2_0_Q8_1_MMVQ];
+    int u[8 * VDR_Q2_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q2_0_Q8_1_MMVQ; ++i) {
+        const int base = 4 * (iqs + i);
+
+        // Q2_0 has QK2_0 = 64 and uses 2 x QK8_1 blocks on the RHS.
+        v[2 * i + 0] = get_int_from_uint8(bq2_0->qs, iqs + i);
+        v[2 * i + 1] = get_int_from_uint8(bq2_0->qs, iqs + i + QI2_0);
+
+        u[8 * i + 0] = get_int_from_int8_aligned(bq8_1[0].qs, base + 0);
+        u[8 * i + 1] = get_int_from_int8_aligned(bq8_1[0].qs, base + 1);
+        u[8 * i + 2] = get_int_from_int8_aligned(bq8_1[0].qs, base + 2);
+        u[8 * i + 3] = get_int_from_int8_aligned(bq8_1[0].qs, base + 3);
+
+        u[8 * i + 4] = get_int_from_int8_aligned(bq8_1[1].qs, base + 0);
+        u[8 * i + 5] = get_int_from_int8_aligned(bq8_1[1].qs, base + 1);
+        u[8 * i + 6] = get_int_from_int8_aligned(bq8_1[1].qs, base + 2);
+        u[8 * i + 7] = get_int_from_int8_aligned(bq8_1[1].qs, base + 3);
+    }
+
+    const float sum0 = vec_dot_q2_0_q8_1_impl<VDR_Q2_0_Q8_1_MMVQ>(
+        v + 0, u + 0, bq2_0->d, bq8_1[0].ds);
+    const float sum1 = vec_dot_q2_0_q8_1_impl<VDR_Q2_0_Q8_1_MMVQ>(
+        v + VDR_Q2_0_Q8_1_MMVQ, u + 4 * VDR_Q2_0_Q8_1_MMVQ, bq2_0->d, bq8_1[1].ds);
+    return sum0 + sum1;
+}
+
 static __dpct_inline__ float
 vec_dot_q4_1_q8_1(const void *__restrict__ vbq,
                   const block_q8_1 *__restrict__ bq8_1, const int &iqs) {

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -610,6 +610,13 @@ static constexpr std::initializer_list<ggml_op> topk_moe_sigmoid_norm_bias{ GGML
                                                                             GGML_OP_RESHAPE,  GGML_OP_SUM_ROWS, GGML_OP_CLAMP,
                                                                             GGML_OP_DIV,      GGML_OP_RESHAPE };
 
+static constexpr std::initializer_list<ggml_op> topk_moe_sqrt_softplus_norm_bias{ GGML_OP_UNARY,    GGML_OP_SQRT,
+                                                                                  GGML_OP_RESHAPE,  GGML_OP_ADD,
+                                                                                  GGML_OP_ARGSORT,  GGML_OP_VIEW,
+                                                                                  GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
+                                                                                  GGML_OP_SUM_ROWS, GGML_OP_CLAMP,
+                                                                                  GGML_OP_DIV,      GGML_OP_RESHAPE };
+
 static constexpr std::initializer_list<ggml_op> topk_moe_early_softmax     { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
                                                                              GGML_OP_VIEW,     GGML_OP_GET_ROWS };
 
@@ -673,6 +680,22 @@ static constexpr std::initializer_list<std::array<int, 3>> topk_moe_sigmoid_norm
     {10, 0, 9 }, // reshape->src[0]  == div
 };
 
+static constexpr std::initializer_list<std::array<int, 3>> topk_moe_sqrt_softplus_norm_bias_edges {
+    { 1, 0, 0 }, // sqrt->src[0]     == softplus
+    { 2, 0, 1 }, // reshape->src[0]  == sqrt
+    { 3, 0, 1 }, // add->src[0]      == sqrt
+    { 4, 0, 3 }, // argsort->src[0]  == add
+    { 5, 0, 4 }, // view->src[0]     == argsort
+    { 6, 0, 2 }, // get_rows->src[0] == reshape
+    { 6, 1, 5 }, // get_rows->src[1] == view
+    { 7, 0, 6 }, // reshape->src[0]  == get_rows
+    { 8, 0, 7 }, // sum_rows->src[0] == reshape
+    { 9, 0, 8 }, // clamp->src[0]    == sum_rows
+    {10, 0, 7 }, // div->src[0]      == reshape
+    {10, 1, 9 }, // div->src[1]      == clamp
+    {11, 0,10 }, // reshape->src[0]  == div
+};
+
 // same as early_softmax_norm but ending after the get_rows
 static constexpr std::initializer_list<std::array<int, 3>> topk_moe_early_softmax_edges {
     { 1, 0, 0 }, // reshape->src[0]  == softmax
@@ -701,6 +724,7 @@ enum topk_moe_mode {
     TOPK_MOE_EARLY_SOFTMAX_NORM,
     TOPK_MOE_LATE_SOFTMAX,
     TOPK_MOE_SIGMOID_NORM_BIAS,
+    TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS,
     TOPK_MOE_COUNT,
 };
 
@@ -998,6 +1022,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_snake_f32;
     vk_pipeline pipeline_snake_f16;
     vk_pipeline pipeline_snake_bf16;
+    vk_pipeline pipeline_pool1d_f32;
     vk_pipeline pipeline_pool2d_f32;
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
@@ -1687,6 +1712,17 @@ struct vk_op_snake_push_constants {
     uint32_t ne1;
 };
 
+struct vk_op_pool1d_push_constants {
+    uint32_t IL;
+    uint32_t OL;
+    uint32_t OC;
+    uint32_t pelements;
+    uint32_t op;
+    int32_t k0;
+    int32_t s0;
+    int32_t p0;
+};
+
 struct vk_op_pool2d_push_constants {
     uint32_t IW; uint32_t IH;
     uint32_t OW; uint32_t OH;
@@ -1934,6 +1970,7 @@ struct ggml_vk_garbage_collector {
 static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_context subctx);
 static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested = nullptr);
 static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx);
+static bool ggml_vk_intel_windows_driver_equals_or_newer_than(uint32_t driver_version, uint32_t threshold_major, uint32_t threshold_minor);
 
 static bool vk_memory_logger_enabled = false;
 
@@ -5552,8 +5589,11 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    // Intel Arc B390 was observed segfaulting with this shader.
-    if (device->subgroup_basic && device->subgroup_shuffle && device->vendor_id != VK_VENDOR_ID_INTEL) {
+    // Intel Windows driver older than 32.0.101.8860 will crash when using fwht kernels on Xe2+ GPUS so we gate that here
+    const bool can_use_fwht = device->driver_id != vk::DriverId::eIntelProprietaryWindows ||
+        device->architecture != vk_device_architecture::INTEL_XE2 ||
+        (device->architecture == vk_device_architecture::INTEL_XE2 && ggml_vk_intel_windows_driver_equals_or_newer_than(device->properties.driverVersion, 101, 8860));
+    if (can_use_fwht && device->subgroup_basic && device->subgroup_shuffle) {
         int idx = 0;
         for (uint32_t n : {64, 128, 256, 512}) {
             if (device->subgroup_size <= n) {
@@ -5561,8 +5601,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
             }
             ++idx;
         }
-    } else if (device->driver_id != vk::DriverId::eIntelProprietaryWindows) {
-        // Disabled on Intel Windows due to a driver bug: https://github.com/ggml-org/llama.cpp/pull/23964#issuecomment-4598226147
+    } else if (can_use_fwht) {
         int idx = 0;
         for (uint32_t n : {64, 128, 256, 512}) {
             const uint32_t block_size = std::min(device->subgroup_size, n);
@@ -5619,6 +5658,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_snake_f16,  "snake_f16",  snake_f16_len,  snake_f16_data,  "main", 4, sizeof(vk_op_snake_push_constants), {256, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_snake_bf16, "snake_bf16", snake_bf16_len, snake_bf16_data, "main", 4, sizeof(vk_op_snake_push_constants), {256, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_pool1d_f32, "pool1d_f32", pool1d_f32_len, pool1d_f32_data, "main", 2, sizeof(vk_op_pool1d_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
@@ -11332,6 +11372,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             case GGML_TYPE_BF16: return ctx->device->pipeline_col2im_1d_bf16;
             default:             return nullptr;
         }
+    case GGML_OP_POOL_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pool1d_f32;
+        }
+        return nullptr;
     case GGML_OP_POOL_2D:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_pool2d_f32;
@@ -11854,6 +11899,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         {
             elements = { uint32_t(dst->ne[0]), uint32_t(dst->ne[1]), 1 };
         } break;
+    case GGML_OP_POOL_1D:
+        {
+            const uint32_t N = dst->ne[3] * dst->ne[2];
+            const uint32_t OC = dst->ne[1];
+            const uint32_t OL = dst->ne[0];
+            elements = { N * OC * OL, 1, 1};
+        } break;
     case GGML_OP_POOL_2D:
         {
             const uint32_t N = dst->ne[3];
@@ -13175,12 +13227,16 @@ static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& sub
 
 static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx) {
     topk_moe_mode mode = ctx->fused_topk_moe_mode;
+    const bool has_bias = mode == TOPK_MOE_SIGMOID_NORM_BIAS || mode == TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS;
     ggml_tensor * logits = cgraph->nodes[node_idx + 0]->src[0];
-    ggml_tensor * bias = (mode == TOPK_MOE_SIGMOID_NORM_BIAS) ? cgraph->nodes[node_idx + 2]->src[1] : logits;
+    ggml_tensor * bias = mode == TOPK_MOE_SIGMOID_NORM_BIAS       ? cgraph->nodes[node_idx + 2]->src[1] :
+                         mode == TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS ? cgraph->nodes[node_idx + 3]->src[1] :
+                                                                   logits;
     ggml_tensor * weights = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
-    ggml_tensor * ids = (mode == TOPK_MOE_SIGMOID_NORM_BIAS) ? cgraph->nodes[node_idx + 4] :
-                        (mode == TOPK_MOE_LATE_SOFTMAX) ?      cgraph->nodes[node_idx + 1] :
-                                                               cgraph->nodes[node_idx + 3];
+    ggml_tensor * ids = mode == TOPK_MOE_SIGMOID_NORM_BIAS       ? cgraph->nodes[node_idx + 4] :
+                        mode == TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS ? cgraph->nodes[node_idx + 5] :
+                        mode == TOPK_MOE_LATE_SOFTMAX             ? cgraph->nodes[node_idx + 1] :
+                                                                   cgraph->nodes[node_idx + 3];
 
     GGML_ASSERT(logits->type == GGML_TYPE_F32);
     GGML_ASSERT(bias->type == GGML_TYPE_F32);
@@ -13220,16 +13276,24 @@ static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx,
         pc.clamp_min = ggml_get_op_params_f32(clamp, 0);
         pc.clamp_max = ggml_get_op_params_f32(clamp, 1);
     }
+    if (mode == TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS) {
+        ggml_tensor * clamp = cgraph->nodes[node_idx + 9];
+        GGML_ASSERT(clamp->op == GGML_OP_CLAMP);
+        pc.clamp_min = ggml_get_op_params_f32(clamp, 0);
+        pc.clamp_max = ggml_get_op_params_f32(clamp, 1);
+    }
 
 #define GATING_FUNC_SOFTMAX 0
 #define GATING_FUNC_SIGMOID 1
 #define GATING_FUNC_SOFTMAX_WEIGHT 2
+#define GATING_FUNC_SQRT_SOFTPLUS 3
 
-    pc.gating_func = mode == TOPK_MOE_SIGMOID_NORM_BIAS ? GATING_FUNC_SIGMOID :
-                     mode == TOPK_MOE_LATE_SOFTMAX ?      GATING_FUNC_SOFTMAX_WEIGHT :
-                                                          GATING_FUNC_SOFTMAX;
-    pc.has_bias = mode == TOPK_MOE_SIGMOID_NORM_BIAS;
-    pc.with_norm = mode == TOPK_MOE_EARLY_SOFTMAX_NORM || mode == TOPK_MOE_SIGMOID_NORM_BIAS;
+    pc.gating_func = mode == TOPK_MOE_SIGMOID_NORM_BIAS       ? GATING_FUNC_SIGMOID :
+                     mode == TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS ? GATING_FUNC_SQRT_SOFTPLUS :
+                     mode == TOPK_MOE_LATE_SOFTMAX             ? GATING_FUNC_SOFTMAX_WEIGHT :
+                                                                 GATING_FUNC_SOFTMAX;
+    pc.has_bias = has_bias;
+    pc.with_norm = mode == TOPK_MOE_EARLY_SOFTMAX_NORM || has_bias;
     if (ctx->fused_topk_moe_scale) {
         GGML_ASSERT(weights->op == GGML_OP_SCALE);
         pc.output_scale = ggml_get_op_params_f32(weights, 0);
@@ -13773,6 +13837,29 @@ static void ggml_vk_snake_dispatch_fused(ggml_backend_vk_context * ctx, vk_conte
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { x_buf, a_buf, inv_b_buf, dst_buf }, pc, elements);
 }
 
+static void ggml_vk_pool_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
+    const int32_t k0 = dst->op_params[1];
+    const int32_t s0 = dst->op_params[2];
+    const int32_t p0 = dst->op_params[3];
+
+    const uint32_t IL = src0->ne[0];
+
+    const uint32_t N = dst->ne[3] * dst->ne[2];
+
+    const uint32_t OC = dst->ne[1];
+    const uint32_t OL = dst->ne[0];
+
+    const uint32_t parallel_elements = N * OC * OL;
+
+    ggml_vk_op_f32<vk_op_pool1d_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_POOL_1D, {
+        IL, OL, OC,
+        parallel_elements,
+        op,
+        k0, s0, p0,
+    });
+}
+
 static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
     const int32_t k1 = dst->op_params[1];
@@ -15284,6 +15371,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_CONV_TRANSPOSE_1D:
         ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node);
 
+        break;
+    case GGML_OP_POOL_1D:
+        ggml_vk_pool_1d(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_POOL_2D:
         ggml_vk_pool_2d(ctx, compute_ctx, src0, node);
@@ -16311,6 +16402,20 @@ static bool ggml_vk_can_fuse_topk_moe(ggml_backend_vk_context * ctx, const struc
             return false;
         }
         break;
+    case TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS:
+        softmax = cgraph->nodes[node_idx + 0]; // really softplus
+        weights = cgraph->nodes[node_idx + 11];
+        get_rows = cgraph->nodes[node_idx + 6];
+        argsort = cgraph->nodes[node_idx + 4];
+        if (ggml_get_unary_op(softmax) != GGML_UNARY_OP_SOFTPLUS) {
+            return false;
+        }
+        // bias is expected to be 1D
+        if (ggml_nrows(cgraph->nodes[node_idx + 3]->src[1]) != 1 ||
+            !ggml_is_contiguous(cgraph->nodes[node_idx + 3]->src[1])) {
+            return false;
+        }
+        break;
     case TOPK_MOE_EARLY_SOFTMAX:
         softmax = cgraph->nodes[node_idx + 0];
         weights = cgraph->nodes[node_idx + 4];
@@ -16334,7 +16439,9 @@ static bool ggml_vk_can_fuse_topk_moe(ggml_backend_vk_context * ctx, const struc
     probs = probs->src[0];
     ggml_tensor * selection_probs = argsort->src[0];
 
-    if (probs != selection_probs && mode != TOPK_MOE_SIGMOID_NORM_BIAS) {
+    if (probs != selection_probs &&
+        mode != TOPK_MOE_SIGMOID_NORM_BIAS &&
+        mode != TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS) {
         return false;
     }
 
@@ -16702,7 +16809,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         // the fused result in an elementwise-way. This affects whether the memory for
         // the src is allowed to overlap the memory for the destination.
         // The array is sized to handle the largest fusion (asserted later).
-        bool op_srcs_fused_elementwise[12];
+        bool op_srcs_fused_elementwise[13];
 
         ctx->fused_topk_moe_mode = TOPK_MOE_COUNT;
         ctx->fused_topk_moe_scale = false;
@@ -16813,6 +16920,15 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
                 ctx->fused_topk_moe_mode = TOPK_MOE_SIGMOID_NORM_BIAS;
                 fusion_string = "TOPK_MOE_SIGMOID_NORM_BIAS";
                 std::fill_n(op_srcs_fused_elementwise, ctx->num_additional_fused_ops + 1, false);
+            } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_sqrt_softplus_norm_bias, { i + 5, i + 11 }) &&
+                       ggml_check_edges(cgraph, i, topk_moe_sqrt_softplus_norm_bias_edges) &&
+                       ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS)) {
+                ctx->num_additional_fused_ops = topk_moe_sqrt_softplus_norm_bias.size() - 1;
+                // view of argsort writes to memory
+                ctx->fused_ops_write_mask |= 1 << 5;
+                ctx->fused_topk_moe_mode = TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS;
+                fusion_string = "TOPK_MOE_SQRT_SOFTPLUS_NORM_BIAS";
+                std::fill_n(op_srcs_fused_elementwise, ctx->num_additional_fused_ops + 1, false);
             } else if (ggml_can_fuse_subgraph(cgraph, i, topk_moe_early_softmax, { i + 3, i + 4 }) &&
                        ggml_check_edges(cgraph, i, topk_moe_early_softmax_edges) &&
                        ggml_vk_can_fuse_topk_moe(ctx, cgraph, i, TOPK_MOE_EARLY_SOFTMAX)) {
@@ -17079,6 +17195,9 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
         if (keep_pattern(topk_moe_sigmoid_norm_bias)) {
             continue;
         }
+        if (keep_pattern(topk_moe_sqrt_softplus_norm_bias)) {
+            continue;
+        }
         if (keep_pattern(topk_moe_early_softmax)) {
             continue;
         }
@@ -17109,6 +17228,7 @@ static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph *
             // Don't pull forward nodes from fusion patterns
             if (match_pattern(topk_moe_early_softmax_norm, j) ||
                 match_pattern(topk_moe_sigmoid_norm_bias, j) ||
+                match_pattern(topk_moe_sqrt_softplus_norm_bias, j) ||
                 match_pattern(topk_moe_early_softmax, j) ||
                 match_pattern(topk_moe_late_softmax, j) ||
                 match_pattern(snake_pattern, j)) {
@@ -18001,6 +18121,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_CONV_2D_DW:
             return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16)
                 && op->src[1]->type == GGML_TYPE_F32;
+        case GGML_OP_POOL_1D:
+            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_POOL_2D:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_RWKV_WKV6:
@@ -18466,6 +18588,22 @@ static uint32_t ggml_vk_intel_shader_core_count(const vk::PhysicalDevice& vkdev)
     }
 }
 
+static bool ggml_vk_intel_windows_driver_equals_or_newer_than(uint32_t driver_version, uint32_t threshold_major, uint32_t threshold_minor) {
+#if defined(_WIN32)
+    // Intel Windows encodes xxx.yyyy as [31:14].[13:0].
+    const uint32_t major = driver_version >> 14;
+    const uint32_t minor = driver_version & 0x3fff;
+
+    return major > threshold_major || (major == threshold_major && minor >= threshold_minor);
+#else
+    GGML_UNUSED(driver_version);
+    GGML_UNUSED(threshold_major);
+    GGML_UNUSED(threshold_minor);
+    return true;
+#endif
+}
+
+
 // checks
 
 #ifdef GGML_VULKAN_CHECK_RESULTS
@@ -18920,6 +19058,13 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             const int32_t oc     = tensor->op_params[1];
             const int32_t p0     = tensor->op_params[2];
             tensor_clone = ggml_col2im_1d(ggml_ctx, src_clone[0], stride, oc, p0);
+        } else if (tensor->op == GGML_OP_POOL_1D) {
+            enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
+            const int32_t k0 = tensor->op_params[1];
+            const int32_t s0 = tensor->op_params[2];
+            const int32_t p0 = tensor->op_params[3];
+
+            tensor_clone = ggml_pool_1d(ggml_ctx, src_clone[0], op, k0, s0, p0);
         } else if (tensor->op == GGML_OP_POOL_2D) {
             enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
             const int32_t k0 = tensor->op_params[1];

ggml/src/ggml-vulkan/vulkan-shaders/pool1d.comp

@@ -0,0 +1,65 @@
+#version 450
+
+#include "types.glsl"
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout(push_constant) uniform parameter {
+    uint IL;
+    uint OL;
+    uint OC;
+    uint pelements;
+    uint op;
+    int k0;
+    int s0;
+    int p0;
+} p;
+
+#define BLOCK_SIZE 512
+#define FLT_MAX 3.402823466e+38F
+#define OP_POOL_MAX 0u
+#define OP_POOL_AVG 1u
+
+layout (local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout(binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout(binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.x;
+    if (idx >= p.pelements) {
+        return;
+    }
+
+    const uint nc = idx / p.OL;
+    const uint cur_ol = idx % p.OL;
+
+    const int start = int(cur_ol) * p.s0 - p.p0;
+    const int bl = max(start, 0);
+    const int el = min(max(start + p.k0, 0), int(p.IL));
+
+    const int window_size = el - bl;
+    const float scale = window_size > 0 ? 1.0 / float(window_size) : 0.0;
+    float res;
+
+    if (p.op == OP_POOL_AVG) {
+        res = 0.0;
+    } else if (p.op == OP_POOL_MAX) {
+        res = -FLT_MAX;
+    } else {
+        return;
+    }
+
+    #pragma unroll
+    for (uint i = bl; i < el; i++) {
+        const float cur = D_TYPE(data_a[nc * p.IL + i]);
+
+        if (p.op == OP_POOL_AVG) {
+            res += cur * scale;
+        } else if (p.op == OP_POOL_MAX) {
+            res = max(res, cur);
+        }
+    }
+
+    data_d[nc * p.OL + cur_ol] = res;
+}

ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp

@@ -10,6 +10,7 @@
 #define GATING_FUNC_SOFTMAX 0
 #define GATING_FUNC_SIGMOID 1
 #define GATING_FUNC_SOFTMAX_WEIGHT 2
+#define GATING_FUNC_SQRT_SOFTPLUS 3
 
 layout (push_constant) uniform parameter
 {
@@ -120,6 +121,13 @@ void main() {
             const uint expert = i + lane;
             probs[i / WARP_SIZE] = (n_experts % WARP_SIZE == 0 || expert < n_experts) ? 1.f / (1.f + exp(-probs[i / WARP_SIZE])) : -INFINITY;
         }
+    } else if (gating_func == GATING_FUNC_SQRT_SOFTPLUS) {
+        [[unroll]]
+        for (uint i = 0; i < n_experts; i += WARP_SIZE) {
+            const uint expert = i + lane;
+            const float val = probs[i / WARP_SIZE];
+            probs[i / WARP_SIZE] = (n_experts % WARP_SIZE == 0 || expert < n_experts) ? sqrt(val > 20.0f ? val : log(1.0f + exp(val))) : -INFINITY;
+        }
     }
 
     float selection_probs[experts_per_thread];

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -1052,6 +1052,7 @@ void process_shaders() {
     string_to_spv("snake_f16",  "snake.comp", {{"DATA_A_F16", "1"},  {"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
     string_to_spv("snake_bf16", "snake.comp", {{"DATA_A_BF16", "1"}, {"DATA_D_BF16", "1"}, {"A_TYPE", "uint16_t"},  {"D_TYPE", "uint16_t"}});
 
+    string_to_spv("pool1d_f32", "pool1d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -2774,6 +2774,10 @@ class ggml_webgpu_shader_lib {
                 defines.push_back("TYPE_F32");
                 variant += "_f32";
                 break;
+            case GGML_TYPE_F16:
+                defines.push_back("TYPE_F16");
+                variant += "_f16";
+                break;
             case GGML_TYPE_I32:
                 defines.push_back("TYPE_I32");
                 variant += "_i32";

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -4290,7 +4290,8 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
             supports_op = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32);
             break;
         case GGML_OP_REPEAT:
-            supports_op = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32 || src0->type == GGML_TYPE_I16);
+            supports_op = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_I32 ||
+                           src0->type == GGML_TYPE_I16);
             break;
         case GGML_OP_CPY:
         case GGML_OP_CONT:

ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl

@@ -27,6 +27,9 @@ struct Params {
 #ifdef TYPE_I32
 #define DataType i32
 #endif
+#ifdef TYPE_F16
+#define DataType f16
+#endif
 #ifdef TYPE_I16
 // same size (16-bit) is sufficient for repeat
 #define DataType f16

ggml/src/ggml-zendnn/CMakeLists.txt

@@ -28,7 +28,7 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
     ExternalProject_Add(
         zendnn
         GIT_REPOSITORY https://github.com/amd/ZenDNN.git
-        GIT_TAG 253b94ce0d7e9284c265fefb485714944caff9d3    # ZenDNN-2026-WW19
+        GIT_TAG 1f399a75cc0993778374a51bea49b64a57879595    # ZenDNN-2026-WW28
         PREFIX      ${ZENDNN_PREFIX}
         SOURCE_DIR  ${ZENDNN_SOURCE_DIR}
         BINARY_DIR  ${ZENDNN_BUILD_DIR}

ggml/src/ggml-zendnn/ggml-zendnn.cpp

@@ -30,6 +30,29 @@ zendnnl::common::data_type_t ggml_to_zendnn_type() {
     }
 }
 
+/**
+ * Builds the matmul_params shared by ggml_zendnn_matmul() and ggml_zendnn_group_matmul():
+ * dtype selection plus, for Q8_0 weights, dynamic-quant setup. Callers still need to set
+ * quant_params.src_scale.dims themselves, since that depends on the batch size(s) in use.
+ */
+template <typename TA, typename TB, typename TC>
+static zendnnl::lowoha::matmul::matmul_params ggml_zendnn_make_matmul_params(ggml_backend_zendnn_context * ctx) {
+    zendnnl::lowoha::matmul::matmul_params params;
+    params.dtypes.src = ggml_to_zendnn_type<TB>();
+    params.dtypes.wei = ggml_to_zendnn_type<TA>();
+    params.dtypes.dst = ggml_to_zendnn_type<TC>();
+    params.num_threads = ctx->n_threads;
+
+    if constexpr (std::is_same_v<TA, block_q8_0>) {
+        params.dtypes.compute = zendnnl::common::data_type_t::s8;
+        params.dynamic_quant = true;
+        params.quant_params.src_scale.buff = nullptr;
+        params.quant_params.src_scale.dt   = zendnnl::common::data_type_t::bf16;
+        params.packing.pack_format_b = 1;
+    }
+    return params;
+}
+
 /**
  * ZenDNN matmul: computes C = B * A.
  *
@@ -47,22 +70,12 @@ static bool ggml_zendnn_matmul(ggml_backend_zendnn_context * ctx, int64_t m, int
                                const TA * A, int64_t lda, const TB * B, int64_t ldb, TC * C,
                                int64_t ldc) {
 
-    zendnnl::lowoha::matmul::matmul_params params;
-    params.dtypes.src = ggml_to_zendnn_type<TB>();
-    params.dtypes.wei = ggml_to_zendnn_type<TA>();
-    params.dtypes.dst = ggml_to_zendnn_type<TC>();
-    params.num_threads = ctx->n_threads;
+    zendnnl::lowoha::matmul::matmul_params params = ggml_zendnn_make_matmul_params<TA, TB, TC>(ctx);
 
     zendnnl::lowoha::matmul::matmul_batch_params_t batch_params;
 
     if constexpr (std::is_same_v<TA, block_q8_0>) {
-        params.dtypes.compute = zendnnl::common::data_type_t::s8;
-        const int64_t num_groups = k / QK8_0;
-        params.dynamic_quant = true;
-        params.quant_params.src_scale.buff = nullptr;
-        params.quant_params.src_scale.dt   = zendnnl::common::data_type_t::bf16;
-        params.quant_params.src_scale.dims = {n, num_groups};
-        params.packing.pack_format_b = 1;
+        params.quant_params.src_scale.dims = {n, k / QK8_0};
     }
 
     zendnnl::error_handling::status_t status = zendnnl::lowoha::matmul::matmul_direct(
@@ -223,6 +236,99 @@ struct mmid_row_mapping {
     int32_t i2;
 };
 
+/**
+ * ZenDNN batched matmul: computes C[i] = B[i] * A[i] for every active expert i via a single
+ * group_matmul_direct() call. Batched analogue of ggml_zendnn_matmul() - see its docs for the
+ * per-expert A/B/C shape convention. m and k are shared by every expert; n (batch size) varies
+ * per expert, hence the vector.
+ */
+template <typename TA, typename TB, typename TC>
+static bool ggml_zendnn_group_matmul(ggml_backend_zendnn_context * ctx, int64_t m, int64_t k,
+                                     const std::vector<int64_t> & n,
+                                     const std::vector<const void *> & A, int64_t lda,
+                                     const std::vector<const void *> & B, int64_t ldb,
+                                     const std::vector<void *> & C, int64_t ldc) {
+
+    const int n_experts = n.size();
+
+    zendnnl::lowoha::matmul::matmul_params base_params = ggml_zendnn_make_matmul_params<TA, TB, TC>(ctx);
+
+    std::vector<char>         layout(n_experts, 'r');
+    std::vector<bool>         trans_a(n_experts, false);
+    std::vector<bool>         trans_b(n_experts, true);
+    std::vector<int>          batch_m(n_experts);
+    std::vector<int>          batch_n(n_experts, m);
+    std::vector<int>          batch_k(n_experts, k);
+    std::vector<float>        alpha(n_experts, 1.0f);
+    std::vector<float>        beta(n_experts, 0.0f);
+    std::vector<const void *> bias(n_experts, nullptr);
+    std::vector<int>          lda_v(n_experts, lda);
+    std::vector<int>          ldb_v(n_experts, ldb);
+    std::vector<int>          ldc_v(n_experts, ldc);
+    std::vector<bool>         is_wei_const(n_experts, true);
+    std::vector<zendnnl::lowoha::matmul::matmul_params> params(n_experts, base_params);
+
+    for (int i = 0; i < n_experts; i++) {
+        batch_m[i] = n[i];
+
+        // src_scale.dims depends on this expert's row count, unlike the rest of base_params
+        if constexpr (std::is_same_v<TA, block_q8_0>) {
+            params[i].quant_params.src_scale.dims = {n[i], k / QK8_0};
+        }
+    }
+
+    zendnnl::error_handling::status_t status = zendnnl::lowoha::matmul::group_matmul_direct(
+        layout, trans_a, trans_b, batch_m, batch_n, batch_k, alpha,
+        B, ldb_v, A, lda_v, bias, beta,
+        C, ldc_v, is_wei_const, params);
+
+    if (status != zendnnl::error_handling::status_t::success) {
+        GGML_LOG_ERROR("%s, ZenDNN group matmul failed: status=%d\n", __func__, static_cast<int>(status));
+        return false;
+    }
+    return true;
+}
+
+static bool ggml_zendnn_group_gemm(ggml_backend_zendnn_context * ctx, int64_t m, int64_t k,
+                                    const std::vector<int64_t> & n,
+                                    const std::vector<const void *> & A, int64_t lda,
+                                    const std::vector<const void *> & B, int64_t ldb,
+                                    const std::vector<void *> & C, int64_t ldc,
+                                    int Atype, int Btype, int Ctype) {
+
+    assert(m >= 0);
+    for (size_t i = 0; i < n.size(); i++) {
+        assert(n[i] >= 0);
+    }
+    assert(k >= 0);
+    assert(lda >= k);
+    assert(ldb >= k);
+    assert(ldc >= m);
+
+    // categorize types
+    switch (Atype) {
+        case GGML_TYPE_F32:
+            if (Btype != GGML_TYPE_F32 || Ctype != GGML_TYPE_F32)
+                return false;
+            return ggml_zendnn_group_matmul<float, float, float>(ctx, m, k, n, A, lda, B, ldb, C, ldc);
+        case GGML_TYPE_BF16:
+            if (Btype != GGML_TYPE_BF16)
+                return false;
+            if (Ctype == GGML_TYPE_BF16)
+                return ggml_zendnn_group_matmul<ggml_bf16_t, ggml_bf16_t, ggml_bf16_t>(
+                    ctx, m, k, n, A, lda, B, ldb, C, ldc);
+            if (Ctype == GGML_TYPE_F32)
+                return ggml_zendnn_group_matmul<ggml_bf16_t, ggml_bf16_t, float>(ctx, m, k, n, A, lda, B, ldb, C, ldc);
+            return false;
+        case GGML_TYPE_Q8_0:
+            if (Btype != GGML_TYPE_F32 || Ctype != GGML_TYPE_F32)
+                return false;
+            return ggml_zendnn_group_matmul<block_q8_0, float, float>(ctx, m, k, n, A, lda, B, ldb, C, ldc);
+        default:
+            return false; // unsupported type
+    }
+}
+
 static void ggml_zendnn_compute_forward_mul_mat_id(
     ggml_backend_zendnn_context * ctx,
     ggml_tensor * dst) {
@@ -262,7 +368,8 @@ static void ggml_zendnn_compute_forward_mul_mat_id(
     std::vector<int64_t> matrix_row_counts(n_as, 0);
     std::vector<std::vector<mmid_row_mapping>> matrix_rows(n_as);
 
-    int64_t max_rows = 0;
+    int64_t total_rows = 0;
+    int     n_active_experts = 0;
     // group rows by expert (preprocessing step)
     for (int64_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
         for (int id = 0; id < n_ids; ++id) {
@@ -270,66 +377,74 @@ static void ggml_zendnn_compute_forward_mul_mat_id(
 
             GGML_ASSERT(i02 >= 0 && i02 < n_as);
 
+            if (matrix_row_counts[i02] == 0) {
+                n_active_experts++;
+            }
             matrix_rows[i02].push_back({id, iid1});
             matrix_row_counts[i02]++;
-            if (matrix_row_counts[i02] > max_rows) {
-                max_rows = matrix_row_counts[i02];
-            }
+            total_rows++;
         }
     }
 
-    if (max_rows == 0) {
+    if (total_rows == 0) {
         return; // no rows to process
     }
 
     const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-    // size for converting src1 rows to vec_dot_type if needed
-    const size_t nbw1 = row_size;
-    const size_t nbw2 = nbw1 * ne11;
-    const size_t nbw3 = nbw2 * ne12;
-    const size_t src1_conv_size = (src1->type != vec_dot_type && src0->type != GGML_TYPE_Q8_0) ? ne13 * nbw3 : 0;
-
     // For Q8_0, src1 is always F32; the gather buffer must hold F32 rows (ne10*4 bytes),
     // not Q8_0-encoded rows (row_size ≈ ne10/32*34 bytes) — they differ by ~4x.
     const size_t f32_row_size = (size_t)ne10 * sizeof(float);
     const size_t gather_row_size = (src0->type == GGML_TYPE_Q8_0) ? f32_row_size : row_size;
 
+    if (src1->type != vec_dot_type && src0->type != GGML_TYPE_Q8_0) {
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    }
+
     // size for MoE gather/scatter buffers
-    const size_t wdata_cur_size = max_rows * gather_row_size;
-    const size_t dst_cur_size = max_rows * ggml_row_size(dst->type, ne01);
+    const size_t wdata_cur_size = total_rows * gather_row_size;
+    const size_t dst_cur_size = total_rows * ggml_row_size(dst->type, ne01);
 
     // allocate single buffer for all needs
-    const size_t total_size = src1_conv_size + wdata_cur_size + dst_cur_size;
+    const size_t total_size = wdata_cur_size + dst_cur_size;
     if (ctx->work_size < total_size) {
         ctx->work_data.reset(new char[total_size]);
         ctx->work_size = total_size;
     }
 
     // partition the buffer
-    char * work_data = ctx->work_data.get();
-    char * wdata_cur = work_data + src1_conv_size;
+    char * wdata_cur = ctx->work_data.get();
     char * dst_cur = wdata_cur + wdata_cur_size;
 
-    // ZenDNN requires FP32 for dynamic quantization, so conversion is skipped
-    if (src1->type != vec_dot_type && src0->type != GGML_TYPE_Q8_0) {
-        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    // per-expert data collected during gather, handed to ggml_zendnn_group_gemm() as one batch
+    std::vector<int64_t>      expert_row_count(n_active_experts);
+    std::vector<const void *> batch_src(n_active_experts);
+    std::vector<const void *> batch_wei(n_active_experts);
+    std::vector<void *>       batch_dst(n_active_experts);
 
-        #pragma omp parallel for collapse(3) num_threads(ctx->n_threads) schedule(static)
-        for (int64_t i13 = 0; i13 < ne13; ++i13) {
-            for (int64_t i12 = 0; i12 < ne12; ++i12) {
-                for (int64_t i11 = 0; i11 < ne11; ++i11) {
-                    const float * src1_f32 = (float *)((char *)src1->data + i11*nb11 + i12*nb12 + i13*nb13);
-                    void * src1_conv = (char *)work_data + i11*nbw1 + i12*nbw2 + i13*nbw3;
-                    from_float(src1_f32, src1_conv, ne10);
-                }
+    // precompute per-expert buffer offsets and batch indices for the parallel loop below
+    std::vector<int64_t> expert_wdata_off(n_as, 0);
+    std::vector<int64_t> expert_dst_off(n_as, 0);
+    std::vector<int>     expert_batch_idx(n_as, -1);
+    {
+        int64_t w_off = 0;
+        int64_t d_off = 0;
+        int     batch_idx = 0;
+        for (int64_t cur_a = 0; cur_a < n_as; ++cur_a) {
+            if (matrix_row_counts[cur_a] == 0) {
+                continue;
             }
+            expert_wdata_off[cur_a] = w_off;
+            expert_dst_off[cur_a]   = d_off;
+            expert_batch_idx[cur_a] = batch_idx;
+            w_off += matrix_row_counts[cur_a] * gather_row_size;
+            d_off += matrix_row_counts[cur_a] * ggml_row_size(dst->type, ne01);
+            batch_idx++;
         }
     }
 
-    const void * wdata = (src1->type == vec_dot_type || src0->type == GGML_TYPE_Q8_0) ? src1->data : work_data;
-
-    // process each expert with gather -> gemm -> scatter pattern
+    // gather + inline-convert input rows into each expert's batch slot
+    #pragma omp parallel for num_threads(ctx->n_threads) schedule(static)
     for (int64_t cur_a = 0; cur_a < n_as; ++cur_a) {
         const int64_t cne1 = matrix_row_counts[cur_a];
 
@@ -337,42 +452,57 @@ static void ggml_zendnn_compute_forward_mul_mat_id(
             continue;
         }
 
-        const char * src0_cur = (const char *) src0->data + cur_a*nb02;
+        const int64_t w_off     = expert_wdata_off[cur_a];
+        const int64_t d_off     = expert_dst_off[cur_a];
+        const int     batch_idx = expert_batch_idx[cur_a];
 
-        // gather input rows for this expert
-        #pragma omp parallel for num_threads(ctx->n_threads) schedule(static)
         for (int64_t ir1 = 0; ir1 < cne1; ++ir1) {
             const mmid_row_mapping & row_mapping = matrix_rows[cur_a][ir1];
-            const int64_t id = row_mapping.i1;
+            const int64_t id  = row_mapping.i1;
             const int64_t i11 = id % ne11;
             const int64_t i12 = row_mapping.i2;
 
-            std::memcpy(
-                wdata_cur + ir1 * gather_row_size,
-                (const char *) wdata + (i11 + i12*ne11) * gather_row_size,
-                gather_row_size
-            );
+            const char * src_row = (const char *) src1->data + i11*nb11 + i12*nb12;
+            void *       dst_row = wdata_cur + w_off + ir1 * gather_row_size;
+
+            if (src1->type != vec_dot_type && src0->type != GGML_TYPE_Q8_0) {
+                from_float((const float *) src_row, dst_row, ne10);
+            } else {
+                // no conversion: src1 already matches vec_dot_type, or src0 is Q8_0, whose
+                // ZenDNN dynamic quantization requires the row to stay in F32
+                std::memcpy(dst_row, src_row, gather_row_size);
+            }
         }
 
-        // batched gemm for all tokens in this expert
-        if (!ggml_zendnn_gemm(ctx,
-                              ne01,       // m
-                              cne1,       // n
-                              ne10,       // k
-                              src0_cur,
-                              ne00,       // lda
-                              wdata_cur,
-                              ne10,       // ldb
-                              dst_cur,
-                              ne01,       // ldc
-                              src0->type,
-                              src0->type == GGML_TYPE_Q8_0 ? GGML_TYPE_F32 : vec_dot_type,
-                              dst->type)) {
-            GGML_ABORT("%s: ZenDNN gemm failed\n", __func__);
+        expert_row_count[batch_idx] = cne1;
+        batch_src[batch_idx]        = wdata_cur + w_off;
+        batch_wei[batch_idx]        = (const char *) src0->data + cur_a * nb02;
+        batch_dst[batch_idx]        = dst_cur + d_off;
+    }
+
+    if (!ggml_zendnn_group_gemm(ctx,
+                                ne01,             // m
+                                ne10,             // k
+                                expert_row_count, // n (per expert)
+                                batch_wei, ne00,  // A: weights, lda
+                                batch_src, ne10,  // B: input, ldb
+                                batch_dst, ne01,  // C: output, ldc
+                                src0->type,
+                                src0->type == GGML_TYPE_Q8_0 ? GGML_TYPE_F32 : vec_dot_type,
+                                dst->type))
+        GGML_ABORT("%s: ZenDNN group gemm failed\n", __func__);
+
+    // scatter output rows to destination
+    #pragma omp parallel for num_threads(ctx->n_threads) schedule(static)
+    for (int64_t cur_a = 0; cur_a < n_as; ++cur_a) {
+        const int64_t cne1 = matrix_row_counts[cur_a];
+
+        if (cne1 == 0) {
+            continue;
         }
 
-        // scatter output rows to destination
-        #pragma omp parallel for num_threads(ctx->n_threads) schedule(static)
+        const int64_t d_off = expert_dst_off[cur_a];
+
         for (int64_t ir1 = 0; ir1 < cne1; ++ir1) {
             const mmid_row_mapping & row_mapping = matrix_rows[cur_a][ir1];
             const int64_t id = row_mapping.i1;
@@ -381,7 +511,7 @@ static void ggml_zendnn_compute_forward_mul_mat_id(
 
             std::memcpy(
                 (char *) dst->data + i1*nb1 + i2*nb2,
-                dst_cur + ir1 * ggml_row_size(dst->type, ne01),
+                dst_cur + d_off + ir1 * ggml_row_size(dst->type, ne01),
                 ggml_row_size(dst->type, ne01)
             );
         }
@@ -591,22 +721,26 @@ static bool ggml_backend_zendnn_device_supports_op(ggml_backend_dev_t dev, const
                 if(K <= 256 || N <= 128 || M <= 96) {
                     return false;
                 }
+
+                // MUL_MAT_ID's gather+matmul+scatter approach favors a moderate expert count
+                if (op->op == GGML_OP_MUL_MAT_ID) {
+                    const int64_t n_experts = weights->ne[2];
+                    const int64_t max_experts = 32;
+                    if (n_experts > max_experts) {
+                        return false;
+                    }
+
+                    // fall back once the average rows per expert (N / n_experts) is too thin
+                    // to amortize each per-expert GEMM's overhead
+                    if (N / n_experts <= 32) {
+                        return false;
+                    }
+                }
             }
             else if (ne0 < min_batch || ne1 < min_batch || ne10 < min_batch) {
                 return false;
             }
 
-            // MUL_MAT_ID performs best with a moderate number of experts due to its
-            // gather + batched matmul + scatter approach. Future versions will leverage
-            // ZenDNN's grouped_gemm for better scalability with larger expert counts:
-            // https://github.com/amd/ZenDNN/blob/main/docs/operator/lowoha_group_gemm_operator.md
-            if (op->op == GGML_OP_MUL_MAT_ID) {
-                const int64_t n_experts = weights->ne[2];
-                const int64_t max_experts = 32;
-                if (n_experts > max_experts) {
-                    return false;
-                }
-            }
             switch (weights->type) {
                 case GGML_TYPE_F32:
                 case GGML_TYPE_BF16:

gguf-py/gguf/constants.py

@@ -353,6 +353,7 @@ class Keys:
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
             HEAD_COUNT_KV   = "clip.vision.attention.head_count_kv" # used by mimovl (GQA)
+            HEAD_DIM        = "clip.vision.attention.head_dim" # set when qkv width != n_embd
             LAYERNORM_EPS   = "clip.vision.attention.layer_norm_epsilon"
 
         class Projector:
@@ -2328,7 +2329,13 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.SSM_NORM,
         MODEL_TENSOR.SSM_IN,
         MODEL_TENSOR.SSM_BETA_ALPHA,
-        MODEL_TENSOR.SSM_OUT
+        MODEL_TENSOR.SSM_OUT,
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
     ],
     MODEL_ARCH.QWEN3VL: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -3214,6 +3221,13 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
         MODEL_TENSOR.FFN_EXP_PROBS_B,
+        # NextN/MTP tensors
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
     ],
     MODEL_ARCH.DEEPSEEK2OCR: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -3330,6 +3344,12 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE_SHEXP,
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
     ],
     MODEL_ARCH.ERNIE4_5_MOE: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -4376,10 +4396,35 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_OUT,
         MODEL_TENSOR.ATTN_Q_NORM,
         MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_SINKS,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV,
+        MODEL_TENSOR.ATTN_KV_NORM,
+        MODEL_TENSOR.ATTN_OUT_A,
+        MODEL_TENSOR.ATTN_OUT_B,
+        MODEL_TENSOR.HC_ATTN_FN,
+        MODEL_TENSOR.HC_ATTN_BASE,
+        MODEL_TENSOR.HC_ATTN_SCALE,
+        MODEL_TENSOR.HC_FFN_FN,
+        MODEL_TENSOR.HC_FFN_BASE,
+        MODEL_TENSOR.HC_FFN_SCALE,
+        MODEL_TENSOR.HC_HEAD_FN,
+        MODEL_TENSOR.HC_HEAD_BASE,
+        MODEL_TENSOR.HC_HEAD_SCALE,
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_GATE,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
         MODEL_TENSOR.FC,
         MODEL_TENSOR.ENC_OUTPUT_NORM,
         # optional DSpark heads

gguf-py/gguf/gguf_writer.py

@@ -1226,6 +1226,9 @@ class GGUFWriter:
     def add_vision_head_count_kv(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT_KV, value)
 
+    def add_vision_head_dim(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Attention.HEAD_DIM, value)
+
     def add_vision_attention_layernorm_eps(self, value: float) -> None:
         self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value)
 

include/llama.h

@@ -337,6 +337,7 @@ extern "C" {
         bool use_extra_bufts; // use extra buffer types (used for weight repacking)
         bool no_host;         // bypass host buffer allowing extra buffers to be used
         bool no_alloc;        // only load metadata and simulate memory allocations
+        bool load_mtp;        // whether to load MTP layers
     };
 
     struct llama_sampler_seq_config {
@@ -1423,10 +1424,11 @@ extern "C" {
 
     /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
     LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
+                             int32_t   n_vocab,
                              int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty, -1 = context size)
-                               float   penalty_repeat,   // 1.0 = disabled
-                               float   penalty_freq,     // 0.0 = disabled
-                               float   penalty_present); // 0.0 = disabled
+                               float   penalty_repeat,   // must be > 0.0, 1.0 = disabled
+                               float   penalty_freq,     // must be finite, 0.0 = disabled
+                               float   penalty_present); // must be finite, 0.0 = disabled
 
     ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
     LLAMA_API struct llama_sampler * llama_sampler_init_dry(

models/templates/deepseek-ai-DeepSeek-V4-Flash-0731.jinja

@@ -0,0 +1,140 @@
+{%- if not add_generation_prompt is defined -%}
+  {%- set add_generation_prompt = false -%}
+{%- endif -%}
+{%- if not thinking is defined -%}
+  {%- if enable_thinking is defined -%}
+    {%- set thinking = enable_thinking -%}
+  {%- else -%}
+    {%- set thinking = false -%}
+  {%- endif -%}
+{%- endif -%}
+{%- if not drop_thinking is defined -%}
+  {%- set drop_thinking = true -%}
+{%- endif -%}
+{%- set dsml_token = '｜DSML｜' -%}
+{%- set thinking_start_token = '<think>' -%}
+{%- set thinking_end_token = '</think>' -%}
+{%- set reasoning_effort_high = 'Reasoning Effort: Absolute maximum with no shortcuts permitted.\nYou MUST be very thorough in your thinking and comprehensively decompose the problem to resolve the root cause, rigorously stress-testing your logic against all potential paths, edge cases, and adversarial scenarios.\nExplicitly write out your entire deliberation process, documenting every intermediate step, considered alternative, and rejected hypothesis to ensure absolutely no assumption is left unchecked.\n\n' -%}
+{%- set reasoning_effort_max = 'Reasoning Effort: Beyond maximum — exhaustive, relentless, and uncompromising.\nYou MUST reason with the utmost depth and rigor, leaving absolutely nothing to chance: exhaustively decompose the problem into its most fundamental components, trace every causal chain to its root, and resolve the underlying cause rather than any surface symptom.\nDo not stop reasoning until you have independently verified the solution from multiple angles and are certain that no assumption remains unchecked and no error remains undiscovered.\n\n' -%}
+{%- set response_format_template = '## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n' -%}
+{%- set has_tools = false -%}
+{%- set tools_header = '## Tools\n\nYou have access to a set of tools to help answer the user\'s question. You can invoke tools by writing a "<' + dsml_token + 'tool_calls>" block like the following:\n\n<' + dsml_token + 'tool_calls>\n<' + dsml_token + 'invoke name="$TOOL_NAME">\n<' + dsml_token + 'parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</' + dsml_token + 'parameter>\n...\n</' + dsml_token + 'invoke>\n<' + dsml_token + 'invoke name="$TOOL_NAME2">\n...\n</' + dsml_token + 'invoke>\n</' + dsml_token + 'tool_calls>\n\nString parameters should be specified as is and set `string="true"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string="false"`.\n\nIf thinking_mode is enabled (triggered by ' + thinking_start_token + '), you MUST output your complete reasoning inside ' + thinking_start_token + '...' + thinking_end_token + ' BEFORE any tool calls or final response.\n\nOtherwise, output directly after ' + thinking_end_token + ' with tool calls or final response.\n\n### Available Tool Schemas\n\n' -%}
+{%- set tools_footer = '\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n' -%}
+{%- set ns = namespace(system_prompt='', is_first_sp=true, has_tool_calls=false) -%}
+{%- for message in messages -%}
+  {%- if message['role'] == 'system' -%}
+    {%- if ns.is_first_sp -%}
+      {%- set ns.system_prompt = ns.system_prompt + (message['content'] or '') -%}
+      {%- set ns.is_first_sp = false -%}
+    {%- else -%}
+      {%- set ns.system_prompt = ns.system_prompt + '\n\n' + (message['content'] or '') -%}
+    {%- endif -%}
+  {%- endif -%}
+{%- endfor -%}
+{%- if tools is defined and tools -%}
+  {%- set has_tools = true -%}
+  {%- set ts = namespace(schemas='') -%}
+  {%- for tool in tools -%}
+    {%- if tool['type'] == 'function' -%}
+      {%- set ts.schemas = ts.schemas + (tool['function'] | tojson) + '\n' -%}
+    {%- endif -%}
+  {%- endfor -%}
+  {%- if ns.system_prompt -%}
+    {%- set ns.system_prompt = ns.system_prompt + '\n\n' + tools_header + ts.schemas + tools_footer -%}
+  {%- else -%}
+    {%- set ns.system_prompt = tools_header + ts.schemas + tools_footer -%}
+  {%- endif -%}
+{%- endif -%}
+{%- if response_format is defined -%}
+  {%- if ns.system_prompt -%}
+    {%- set ns.system_prompt = ns.system_prompt + '\n\n' -%}
+  {%- endif -%}
+  {%- set ns.system_prompt = ns.system_prompt + response_format_template + (response_format | tojson) -%}
+{%- endif -%}
+{{- bos_token -}}
+{%- if messages and thinking and reasoning_effort is defined and reasoning_effort == 'high' -%}
+  {{- reasoning_effort_high -}}
+{%- elif messages and thinking and reasoning_effort is defined and reasoning_effort == 'max' -%}
+  {{- reasoning_effort_max -}}
+{%- endif -%}
+{{- ns.system_prompt -}}
+{%- set last_user_idx = namespace(value=-1) -%}
+{%- for message in messages -%}
+  {%- if message['role'] == 'user' or message['role'] == 'developer' or message['role'] == 'tool' -%}
+    {%- set last_user_idx.value = loop.index0 -%}
+  {%- endif -%}
+{%- endfor -%}
+{%- set state = namespace(in_user=false) -%}
+{%- for message in messages -%}
+  {%- if message['role'] == 'tool' -%}
+    {%- set ns.has_tool_calls = true -%}
+  {%- endif -%}
+{%- endfor -%}
+{%- for message in messages -%}
+  {%- if message['role'] == 'user' or message['role'] == 'developer' -%}
+    {%- if state.in_user -%}
+      {{- '\n\n' -}}
+    {%- else -%}
+      {{- '<｜User｜>' -}}
+      {%- set state.in_user = true -%}
+    {%- endif -%}
+    {{- message['content'] or '' -}}
+  {%- elif message['role'] == 'tool' -%}
+    {%- if state.in_user -%}
+      {{- '\n\n' -}}
+    {%- else -%}
+      {{- '<｜User｜>' -}}
+      {%- set state.in_user = true -%}
+    {%- endif -%}
+    {{- '<tool_result>' + (message['content'] or '') + '</tool_result>' -}}
+  {%- elif message['role'] == 'assistant' -%}
+    {%- set state.in_user = false -%}
+    {{- '<｜Assistant｜>' -}}
+    {%- set is_after_last_user = loop.index0 > last_user_idx.value -%}
+    {%- set keep_reasoning = thinking and ((not drop_thinking) or has_tools or is_after_last_user or ns.has_tool_calls) -%}
+    {%- if keep_reasoning -%}
+      {{- thinking_start_token -}}
+      {%- if message['reasoning_content'] is defined and message['reasoning_content'] -%}
+        {{- message['reasoning_content'] -}}
+      {%- endif -%}
+      {{- thinking_end_token -}}
+    {%- else -%}
+      {{- thinking_end_token -}}
+    {%- endif -%}
+    {%- if message['content'] is defined and message['content'] -%}
+      {{- message['content'] -}}
+    {%- endif -%}
+    {%- if message['tool_calls'] -%}
+      {{- '\n\n<' + dsml_token + 'tool_calls>\n' -}}
+      {%- for tool in message['tool_calls'] -%}
+        {%- set func = tool['function'] -%}
+        {{- '<' + dsml_token + 'invoke name="' + func['name'] + '">\n' -}}
+        {%- set args = func['arguments'] -%}
+        {%- if args is string -%}
+          {%- set args = args | from_json -%}
+        {%- endif -%}
+        {%- for key, val in args.items() -%}
+          {%- if val is string -%}
+            {{- '<' + dsml_token + 'parameter name="' + key + '" string="true">' + val + '</' + dsml_token + 'parameter>\n' -}}
+          {%- else -%}
+            {{- '<' + dsml_token + 'parameter name="' + key + '" string="false">' + (val | tojson) + '</' + dsml_token + 'parameter>\n' -}}
+          {%- endif -%}
+        {%- endfor -%}
+        {%- if not args -%}
+          {{- '\n' -}}
+        {%- endif -%}
+        {{- '</' + dsml_token + 'invoke>\n' -}}
+      {%- endfor -%}
+      {{- '</' + dsml_token + 'tool_calls>' -}}
+    {%- endif -%}
+    {{- '<｜end▁of▁sentence｜>' -}}
+  {%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+  {{- '<｜Assistant｜>' -}}
+  {%- if thinking -%}
+    {{- thinking_start_token -}}
+  {%- else -%}
+    {{- thinking_end_token -}}
+  {%- endif -%}
+{%- endif -%}

models/templates/deepseek-ai-DeepSeek-V4.jinja

@@ -9,11 +9,14 @@
   {%- endif -%}
 {%- endif -%}
 {%- if not drop_thinking is defined -%}
-  {%- set drop_thinking = false -%}
+  {%- set drop_thinking = true -%}
 {%- endif -%}
 {%- set dsml_token = '｜DSML｜' -%}
 {%- set thinking_start_token = '<think>' -%}
 {%- set thinking_end_token = '</think>' -%}
+{%- set reasoning_effort_max = 'Reasoning Effort: Absolute maximum with no shortcuts permitted.\nYou MUST be very thorough in your thinking and comprehensively decompose the problem to resolve the root cause, rigorously stress-testing your logic against all potential paths, edge cases, and adversarial scenarios.\nExplicitly write out your entire deliberation process, documenting every intermediate step, considered alternative, and rejected hypothesis to ensure absolutely no assumption is left unchecked.\n\n' -%}
+{%- set response_format_template = '## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n' -%}
+{%- set has_tools = false -%}
 {%- set tools_header = '## Tools\n\nYou have access to a set of tools to help answer the user\'s question. You can invoke tools by writing a "<' + dsml_token + 'tool_calls>" block like the following:\n\n<' + dsml_token + 'tool_calls>\n<' + dsml_token + 'invoke name="$TOOL_NAME">\n<' + dsml_token + 'parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</' + dsml_token + 'parameter>\n...\n</' + dsml_token + 'invoke>\n<' + dsml_token + 'invoke name="$TOOL_NAME2">\n...\n</' + dsml_token + 'invoke>\n</' + dsml_token + 'tool_calls>\n\nString parameters should be specified as is and set `string="true"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string="false"`.\n\nIf thinking_mode is enabled (triggered by ' + thinking_start_token + '), you MUST output your complete reasoning inside ' + thinking_start_token + '...' + thinking_end_token + ' BEFORE any tool calls or final response.\n\nOtherwise, output directly after ' + thinking_end_token + ' with tool calls or final response.\n\n### Available Tool Schemas\n\n' -%}
 {%- set tools_footer = '\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n' -%}
 {%- set ns = namespace(system_prompt='', is_first_sp=true, has_tool_calls=false) -%}
@@ -28,6 +31,7 @@
   {%- endif -%}
 {%- endfor -%}
 {%- if tools is defined and tools -%}
+  {%- set has_tools = true -%}
   {%- set ts = namespace(schemas='') -%}
   {%- for tool in tools -%}
     {%- if tool['type'] == 'function' -%}
@@ -40,7 +44,16 @@
     {%- set ns.system_prompt = tools_header + ts.schemas + tools_footer -%}
   {%- endif -%}
 {%- endif -%}
+{%- if response_format is defined -%}
+  {%- if ns.system_prompt -%}
+    {%- set ns.system_prompt = ns.system_prompt + '\n\n' -%}
+  {%- endif -%}
+  {%- set ns.system_prompt = ns.system_prompt + response_format_template + (response_format | tojson) -%}
+{%- endif -%}
 {{- bos_token -}}
+{%- if messages and thinking and reasoning_effort is defined and reasoning_effort == 'max' -%}
+  {{- reasoning_effort_max -}}
+{%- endif -%}
 {{- ns.system_prompt -}}
 {%- set last_user_idx = namespace(value=-1) -%}
 {%- for message in messages -%}
@@ -75,8 +88,8 @@
     {%- set state.in_user = false -%}
     {{- '<｜Assistant｜>' -}}
     {%- set is_after_last_user = loop.index0 > last_user_idx.value -%}
-    {%- set retain_reasoning = (not drop_thinking) or (is_after_last_user or ns.has_tool_calls) -%}
-    {%- if retain_reasoning and thinking -%}
+    {%- set keep_reasoning = thinking and ((not drop_thinking) or has_tools or is_after_last_user or ns.has_tool_calls) -%}
+    {%- if keep_reasoning -%}
       {{- thinking_start_token -}}
       {%- if message['reasoning_content'] is defined and message['reasoning_content'] -%}
         {{- message['reasoning_content'] -}}
@@ -104,6 +117,9 @@
             {{- '<' + dsml_token + 'parameter name="' + key + '" string="false">' + (val | tojson) + '</' + dsml_token + 'parameter>\n' -}}
           {%- endif -%}
         {%- endfor -%}
+        {%- if not args -%}
+          {{- '\n' -}}
+        {%- endif -%}
         {{- '</' + dsml_token + 'invoke>\n' -}}
       {%- endfor -%}
       {{- '</' + dsml_token + 'tool_calls>' -}}
@@ -118,4 +134,4 @@
   {%- else -%}
     {{- thinking_end_token -}}
   {%- endif -%}
-{%- endif -%}
+{%- endif -%}

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "refs/tags/v0.51.0"
+HTTPLIB_VERSION = "refs/tags/v0.52.0"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",

skills/code-review/SKILL.md

@@ -47,6 +47,7 @@ Mandatory on every review; any finding here is **blocking**. Rule of thumb: GGUF
 - **Sizes/counts from tensor dims:** validate before allocating. Products like `ne[i]*nb[i]`/nbytes can overflow on crafted dims into an undersized alloc then heap overflow. Overflow checks must run BEFORE the arithmetic they guard - padding/alignment macros wrap to 0 near `SIZE_MAX`, so a guard after the pad passes.
 - **GGUF strings/arrays:** cap declared lengths and element counts before using them to size a loop or buffer; validate element type and length before casting an array to a pointer or reading fixed indices (`[i+1]`, `[0..2]`).
 - **File-supplied counts indexing fixed arrays:** bound any count (e.g. layer/block count into a `LLAMA_MAX_*` array) before indexing; watch checks that only fire when an optional key is present.
+- **Declared vs actual array length:** check the declared length of a GGUF array against the count actually read, not just against a buffer size.
 - **Bounds comparisons:** flag narrowing casts (`size_t`->`int32_t`) and signed/unsigned mixing that can bypass a length check and copy past a buffer.
 - **Parsed/derived indices:** range-check `stoi`/`atoi` results and catch parse throws; never use a default or derived token id (EOS/BOS/...) as an index without a bounds check.
 - **Reused/reserved buffers:** recheck bounds after a buffer is shrunk or reused; watch `reserve()` then index-by-assumed-size, and header fields read before their length is checked.
@@ -131,6 +132,8 @@ Enforce the `AGENTS.md` / `CONTRIBUTING.md` coding and naming guidelines on ever
 - Reuse existing infrastructure over introducing new components; no new third-party dependencies, extra headers, or files unless clearly justified.
 - Keep it simple: a simpler change doing 90% is often preferable to a complex one doing 100%. Flag unnecessary templates/fancy STL; basic `for` loops are fine here.
 - Every added line should be something the contributor can explain and defend to a reviewer without AI help - flag anything that looks copied-in without understanding.
+- `Co-authored-by:` must be reserved for human co-authors; AI contributions (claude, cursor, codex, etc.) must use `Assisted-by:`; if this point is violated, it's a blocking finding.
+- Any mentions of Minja must be treated as blocking; see `AGENTS.md` for why.
 
 ## Reporting
 

src/CMakeLists.txt

@@ -25,6 +25,7 @@ add_library(llama
             llama-kv-cache.cpp
             llama-kv-cache-iswa.cpp
             llama-kv-cache-dsa.cpp
+            llama-kv-cache-msa.cpp
             llama-kv-cache-dsv4.cpp
             llama-memory.cpp
             llama-memory-hybrid.cpp

src/llama-arch.cpp

@@ -968,6 +968,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_KIMI_LINEAR:
         case LLM_ARCH_QWEN35:
         case LLM_ARCH_QWEN35MOE:
+        case LLM_ARCH_DEEPSEEK4:
             return true;
         default:
             return false;
@@ -990,6 +991,7 @@ bool llm_arch_supports_rs_rollback(const llm_arch & arch) {
     switch (arch) {
         case LLM_ARCH_QWEN35:
         case LLM_ARCH_QWEN35MOE:
+        case LLM_ARCH_DEEPSEEK4:
             return true;
         default:
             return false;

src/llama-context.cpp

@@ -120,8 +120,9 @@ llama_context::llama_context(
     cparams.no_perf                 = params.no_perf;
     cparams.warmup                  = false;
 
-    cparams.embeddings_layer_inp.resize(hparams.n_layer(), false);
-    embd_layer_inp.resize(hparams.n_layer());
+    // +1: id n_layer() taps the output of the last layer ("input" of the head)
+    cparams.embeddings_layer_inp.resize(hparams.n_layer() + 1, false);
+    embd_layer_inp.resize(hparams.n_layer() + 1);
 
     cparams.ctx_type     = params.ctx_type;
     cparams.pooling_type = params.pooling_type;
@@ -1164,7 +1165,7 @@ void llama_context::set_embeddings_nextn(bool value, bool masked) {
 void llama_context::set_embeddings_layer_inp(uint32_t lid, bool enable) {
     LLAMA_LOG_DEBUG("%s: lid = %d, enable = %d\n", __func__, lid, enable);
 
-    GGML_ASSERT(lid < model.hparams.n_layer());
+    GGML_ASSERT(lid <= model.hparams.n_layer());
 
     cparams.embeddings_layer_inp[lid] = enable;
 
@@ -1716,7 +1717,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_vocab = vocab.n_tokens();
-    const int64_t n_embd  = hparams.n_embd_inp();
+    const bool    mtp_embd = cparams.ctx_type == LLAMA_CONTEXT_TYPE_MTP && batch_inp.embd;
+    const int64_t n_embd  = mtp_embd ? hparams.n_embd_out() : hparams.n_embd_inp();
 
     // when computing embeddings, all tokens are output
     const bool output_all   = cparams.embeddings;
@@ -2275,8 +2277,9 @@ void llama_context::extract_layer_inputs(const llm_graph_result * res, size_t to
 }
 
 void llama_context::output_reorder() {
-    const uint64_t n_vocab = model.vocab.n_tokens();
-    const uint64_t n_embd  = model.hparams.n_embd;
+    const uint64_t n_vocab     = model.vocab.n_tokens();
+    const uint64_t n_embd      = model.hparams.n_embd;
+    const uint64_t n_embd_out  = model.hparams.n_embd_out();
 
     for (size_t s = 0; s < output_swaps.size(); ++s) {
         const uint64_t i0 = output_swaps[s].i0;
@@ -2289,14 +2292,14 @@ void llama_context::output_reorder() {
         }
 
         if (embd.size > 0) {
-            for (uint64_t k = 0; k < n_embd; k++) {
-                std::swap(embd.data[i0*n_embd + k], embd.data[i1*n_embd + k]);
+            for (uint64_t k = 0; k < n_embd_out; k++) {
+                std::swap(embd.data[i0*n_embd_out + k], embd.data[i1*n_embd_out + k]);
             }
         }
 
         if (embd_nextn.size > 0) {
-            for (uint64_t k = 0; k < n_embd; k++) {
-                std::swap(embd_nextn.data[i0*n_embd + k], embd_nextn.data[i1*n_embd + k]);
+            for (uint64_t k = 0; k < n_embd_out; k++) {
+                std::swap(embd_nextn.data[i0*n_embd_out + k], embd_nextn.data[i1*n_embd_out + k]);
             }
         }
 
@@ -2351,6 +2354,7 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
         model.arch == LLM_ARCH_QWEN35 ||
         model.arch == LLM_ARCH_QWEN35MOE ||
         model.arch == LLM_ARCH_DEEPSEEK4 ||
+        (model.arch == LLM_ARCH_DFLASH && model.hparams.dsv4_hc_mult > 0) ||
         model.arch == LLM_ARCH_NANBEIGE ||
         model.arch == LLM_ARCH_MINIMAX_M3) {
         return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
@@ -3553,6 +3557,22 @@ llama_context * llama_init_from_model(
         }
     }
 
+    if ((model->hparams.is_mla() || model->arch == LLM_ARCH_DEEPSEEK4) && params.type_k != params.type_v) {
+        LLAMA_LOG_ERROR("%s: model does not support different K (%s) and V (%s) cache types\n", __func__, ggml_type_name(params.type_k), ggml_type_name(params.type_v));
+        return nullptr;
+    }
+
+    if (ggml_is_quantized(params.type_v) && params.flash_attn_type != LLAMA_FLASH_ATTN_TYPE_ENABLED) {
+        if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO) {
+            LLAMA_LOG_INFO("%s: enabling flash_attn since it is required for quantized V cache\n", __func__);
+            params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
+        }
+        if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_DISABLED) {
+            LLAMA_LOG_ERROR("%s: quantized V cache requires flash_attn to be enabled\n", __func__);
+            return nullptr;
+        }
+    }
+
     if (params.flash_attn_type != LLAMA_FLASH_ATTN_TYPE_DISABLED && ggml_is_quantized(params.type_k)) {
         const uint32_t blck_size = ggml_blck_size(params.type_k);
         for (uint32_t il = 0; il < model->hparams.n_layer(); ++il) {
@@ -3575,11 +3595,6 @@ llama_context * llama_init_from_model(
         }
     }
 
-    if (ggml_is_quantized(params.type_v) && params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_DISABLED) {
-        LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
-        return nullptr;
-    }
-
     if (params.pooling_type != LLAMA_POOLING_TYPE_UNSPECIFIED &&
         params.pooling_type != model->hparams.pooling_type) {
         //user-specified pooling-type is different from the model default