sycl : fix the failed UT cases of conv_3d (#24900)

* hex-mm: new weight layout and fusion updates

* hvx-mm: unroll the new tiled vec_dots to optimize hvx register util

* hex-mm: optimize dyn.quant format for q8_0 and q8_1 to reduce overhead in vec_dots.

* hvx-mm: parallel quantizer per block for large rows

* hvx-mm: simplify and futher optimize dyn.quant and vec_dots

* hvx-mm: keep intermediate per tile accumulators in fp16

* hmx-mm: optimize weight dequant by aligning the repacked tiles with the DMA

* hmx-mm: remove qweight scratch and just use vtcm_weight

* hmx-mm: remove all unused and obsolete code

* hmx-mm: the new tiled repack format is here to stay -- rename all x4x2 to _tiled

* hmx-mm: improve activation processing with dma prefetch

* hex-mm: fix hmx/hvx fallback logic and MUL_MAT_ID allocation (unbreaks OLMoE)

* hex-mm: align the weight tiles with dma just like we did in hmx-mm

* hex-mm: factor out common mm bits into htp/matmul-ops.h

* hex-mm: start moving mm kernel selection to the host

* hex-mm: move all of the matmul param compute into the host

* hmx-mm: restore pipelined mode

* hmx-mm: unroll the dequant functions to optimize register usage

* hmx-mm: further improve activation process

* hex-mm: use vtcm_seq_alloc for all vtcm allocations and define more common functions

* hex-mm: improve mm optimizer to acount for number of activation threads

* hex-mm: fix matmul-id kernel params selection (unbreaks OLMoE and LFM)

* hexagon: remove support for arch < v73 since HMX is now required for most use-cases

* hex-mm: cleanup naming for consistency

* hex-mm: make sure matmul fusion accounts for vtcm allocation

* hex-mm: minor cleanup for kernel_params definition

* hex-mm: replace hardcoded limits with proper checks for vtcm requirements

* hex-mm: add support for non-tiled mm as a fallback option and factor out hvx kernels into separate header

* hex-mm: remove unused functions

* hex-mm: add shorthand for MM_SELECT in run-tool script

* hvx-mm: factor out hvx/hmx microkernels and unify matmul entry and dispatch

* hex-mm: further cleanup matmul fallback path

* hex-mm: refactor matmul entry point and dispatch a bit further

* hexagon: update cmake build to enable hmx for everything

* hex-ops: optimize kernel_param updates and include summary in the logs

* hex-mm: add support for GGML_HEXAGON_MM_SELECT

* hex-mm: add hex-common header

* hex-mm: pass correct number of tasks to workpool

* hex-mm: add proper checks for no-work in dyn.quant tasks

* hex-mm: convert all quantizers into a macro

* hex-mm: fix hvx-flat fallback to pass all MUL_MAT tests

* hex-mm: vectorize q8_1 quantizer

* hex-mm: improve fused ffn mm stride handling

* hex-mm: consistent use of n_threads and pipeline in kernel_params

* hexagon: minor formatting

* hex-mm: update MUL_MAT_ID kernel_param handling to make sure host/npu are in sync

* hvx-mm: go back to accumulating in fp32 in tiled hvx kernels, more accurate and same perf

* hvx-mm: unroll the loops and remove masking that is not needed for tiled accums

* hmx-mm: optimize activation processing (slit loops, some unrolling, etc)

* hmx-mm: minor optimization for output processing

* hex-mm: consistent use of uint32_t and size_t in mm kernels

* hex-mm: remove legacy restrictions for rows to be multiple of 256

* hexagon: replace sprintf with snprintf

* hex-mm: relax hardcoded nrows checks and rely on VTCM size requirements

* hexagon: minor alignment fix

* hexagon: fix trailing spaces

* hex-mm: relax padding from 256 to 128 (leftovers)

* hex-mm: remove redundant checks for weight align to 128

we always use 2D dma for the weights and align them properly

* hmx-mm: MUL_MAT_ID better work distribution between hvx threads and hmx tracing

* hex-mm: specialize per-token mmid activation handling

* hex-profile: update python scripts to handle kernel-params section in the logging output

* hex-mm: move n_prefetch (aka dma_depth) into kernel params and remove unused fields

* hex-trace: use easier to parse format, simply and fix post-proc scripts

* hmx-mm: relax 32 row limit for output processing which helps utilization

* hmx-mm: use start-chunk idx for tracing info

* hmx-mm: parameterize activation dma pipeline

* hexagon: add support for simple graph caching to avoid recomputing kernel-params

* hex-mm: remove left-over repack functions

* hex-mm: tighten n_prefetch asserts

* hex-mm: remove duplicate round/align_up helper

* hexagon: cleanup common header used in host/npu

* hexagon: update early wakeup threshold

* hmx-mm: define cost constants and update solver to assume that repacked ne[1] is padded to 32

* hmx-mm: make precompute_matmul a bit more readable (split into smaller functions, etc)

* hex-mm: remove n_threads constraint

* hex-mm: minor formatting updates

* hex-mm: remove obsolete profiling logs

* hex-mm: restore hardcode gate to refuse lm-head to avoid repacking that tensor

README.md

@@ -142,7 +142,9 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
-- [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
+- [x] [Liquid LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2)
+- [x] [Liquid LFM2.5 models](https://huggingface.co/collections/LiquidAI/lfm25)
+- [x] [Liquid Nanos](https://huggingface.co/collections/LiquidAI/liquid-nanos)
 - [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
 - [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
 - [x] [Mellum models](https://huggingface.co/JetBrains/models?search=mellum)

common/CMakeLists.txt

@@ -80,8 +80,6 @@ add_library(${TARGET}
     http.h
     imatrix-loader.cpp
     imatrix-loader.h
-    json-partial.cpp
-    json-partial.h
     json-schema-to-grammar.cpp
     llguidance.cpp
     log.cpp

common/chat.cpp

@@ -2758,5 +2758,9 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates) {
     GGML_ASSERT(chat_templates != nullptr);
     GGML_ASSERT(chat_templates->template_default != nullptr);
+    if (chat_templates->template_tool_use != nullptr) {
+        // take the more expressive template when available
+        return chat_templates->template_tool_use->caps.to_map();
+    }
     return chat_templates->template_default->caps.to_map();
 }

common/json-partial.cpp

@@ -1,324 +0,0 @@
-#include "json-partial.h"
-
-#include "log.h"
-
-#include <nlohmann/json.hpp>
-
-#include <string>
-#include <regex>
-
-using json = nlohmann::ordered_json;
-
-enum common_json_stack_element_type {
-    COMMON_JSON_STACK_ELEMENT_OBJECT,
-    COMMON_JSON_STACK_ELEMENT_KEY,
-    COMMON_JSON_STACK_ELEMENT_ARRAY,
-};
-
-struct common_json_stack_element {
-    common_json_stack_element_type type;
-    std::string key;
-};
-
-bool common_json_parse(
-    const std::string & input,
-    const std::string & healing_marker,
-    common_json & out)
-{
-    std::string::const_iterator it = input.begin();
-    const auto end = input.end();
-    return common_json_parse(it, end, healing_marker, out);
-}
-
-bool common_json_parse(
-    std::string::const_iterator & it,
-    const std::string::const_iterator & end,
-    const std::string & healing_marker,
-    common_json & out)
-{
-    // // https://json.nlohmann.me/features/parsing/sax_interface/
-    struct json_error_locator : public nlohmann::json_sax<json> {
-        std::size_t position;
-        bool found_error;
-        std::string last_token;
-        std::string exception_message;
-        std::vector<common_json_stack_element> stack;
-
-        json_error_locator() : position(0), found_error(false) {}
-
-        bool parse_error(std::size_t position, const std::string & last_token, const json::exception & ex) override { // NOLINT
-            this->position = position - 1;
-            this->found_error = true;
-            this->last_token = last_token;
-            this->exception_message = ex.what();
-            return false;
-        }
-        void close_value() {
-            if (!stack.empty() && (stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY)) {
-                stack.pop_back();
-            }
-        }
-        bool null() override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool boolean(bool) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool number_integer(number_integer_t) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool number_unsigned(number_unsigned_t) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool number_float(number_float_t, const string_t &) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool string(string_t &) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool binary(binary_t &) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool start_object(std::size_t) override { // NOLINT
-            stack.push_back({COMMON_JSON_STACK_ELEMENT_OBJECT, ""});
-            return true;
-        }
-        bool end_object() override {
-            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT);
-            stack.pop_back();
-            close_value();
-            return true;
-        }
-        bool key(string_t & key) override { // NOLINT
-            stack.push_back({COMMON_JSON_STACK_ELEMENT_KEY, key});
-            return true;
-        }
-        bool start_array(std::size_t) override { // NOLINT
-            stack.push_back({COMMON_JSON_STACK_ELEMENT_ARRAY, ""});
-            return true;
-        }
-        bool end_array() override {
-            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY);
-            stack.pop_back();
-            close_value();
-            return true;
-        }
-    };
-    json_error_locator err_loc;
-    auto start = it;
-    json::sax_parse(it, end, &err_loc);
-
-    if (err_loc.found_error) {
-        it = start;
-        auto temptative_end = it + err_loc.position;
-        // LOG_DBG("Error at position %zu (is_end = %s): %s\n", err_loc.position, temptative_end == end ? "true" : "false", err_loc.exception_message.c_str());
-
-        auto input = std::string(it, temptative_end);
-        try {
-            out.json = json::parse(input);
-            // out.json = json::parse(it, temptative_end);
-            it = temptative_end;
-            return true;
-        } catch (const std::exception & ex) {
-            // No, needs healing.
-            LOG_DBG("Failed to parse up to error: %s: <<<%s>>>\n", ex.what(), std::string(it, temptative_end).c_str());
-        }
-        auto can_parse = [](const std::string & str) {
-            try {
-                auto _ = json::parse(str); // NOLINT
-                return true;
-            } catch (const std::exception &) {
-                return false;
-            }
-        };
-        if (!healing_marker.empty() && !err_loc.stack.empty()) {
-            std::string str(it, temptative_end);
-            auto last_non_sp_pos = str.find_last_not_of(" \n\r\t");
-            if (last_non_sp_pos == std::string::npos) {
-                throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
-            }
-            auto last_non_sp_char = str[last_non_sp_pos];
-            // Used to detect stops on a number, which may not be complete.
-            auto was_maybe_number = [&]() {
-                if (!str.empty() && std::isspace(str.back())) {
-                    return false;
-                }
-                return std::isdigit(last_non_sp_char) ||
-                    last_non_sp_char == '.' ||
-                    last_non_sp_char == 'e' ||
-                    last_non_sp_char == 'E' ||
-                    last_non_sp_char == '-';
-            };
-
-            std::string closing;
-            for (size_t i = err_loc.stack.size(); i > 0; i--) {
-                auto & el = err_loc.stack[i - 1];
-                if (el.type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
-                    closing += "}";
-                } else if (el.type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
-                    closing += "]";
-                } else if (el.type != COMMON_JSON_STACK_ELEMENT_KEY) {
-                    throw std::runtime_error("Unexpected stack element type");
-                }
-            }
-
-            // Matches a potentially partial unicode escape sequence, e.g. \u, \uX, \uXX, \uXXX, \uXXXX
-            static const std::regex partial_unicode_regex(R"(\\u(?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F])?)?)?)?$)");
-
-            auto is_high_surrogate = [&](const std::string & s) {
-                // Check if a partial of a high surrogate (U+D800-U+DBFF)
-                return s.length() >= 4 &&
-                    s[0] == '\\' && s[1] == 'u' &&
-                    std::tolower(s[2]) == 'd' &&
-                    (s[3] == '8' || s[3] == '9' || std::tolower(s[3]) == 'a' || std::tolower(s[3]) == 'b');
-            };
-
-            // Initialize the unicode marker to a low surrogate to handle the edge case
-            // where a high surrogate (U+D800-U+DBFF) is immediately followed by a
-            // backslash (\)
-            std::string unicode_marker_padding = "udc00";
-            std::smatch last_unicode_seq;
-
-            if (std::regex_search(str, last_unicode_seq, partial_unicode_regex)) {
-                std::smatch second_last_seq;
-                std::string prelude = str.substr(0, last_unicode_seq.position());
-
-                // Pad the escape sequence with 0s until it forms a complete sequence of 6 characters
-                unicode_marker_padding = std::string(6 - last_unicode_seq.length(), '0');
-
-                if (is_high_surrogate(last_unicode_seq.str())) {
-                    // If the sequence is a partial match for a high surrogate, add a low surrogate (U+DC00-U+UDFF)
-                    unicode_marker_padding += "\\udc00";
-                } else if (std::regex_search(prelude, second_last_seq, partial_unicode_regex)) {
-                    if (is_high_surrogate(second_last_seq.str())) {
-                        // If this follows a high surrogate, pad it to be a low surrogate
-                        if (last_unicode_seq.length() == 2) {
-                            unicode_marker_padding = "dc00";
-                        } else if (last_unicode_seq.length() == 3) {
-                            unicode_marker_padding = "c00";
-                        } else {
-                            // The original unicode_marker_padding is already padded with 0s
-                        }
-                    }
-                }
-            }
-
-            const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
-
-            if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
-                // We're inside an object value
-                if (last_non_sp_char == ':' && can_parse(str + "1" + closing)) {
-                    // Was about to create an object value
-                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + ": 1" + closing)) {
-                    str += (out.healing_marker.json_dump_marker = ":\"" + magic_seed) + "\"" + closing;
-                } else if (last_non_sp_char == '{' && can_parse(str + closing)) {
-                    // Was about to create an object
-                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
-                } else if (can_parse(str + "\"" + closing)) {
-                    // Was inside an object value string
-                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
-                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
-                    // Was inside an object value string after an escape
-                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
-                    // Was inside an object value string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
-                } else {
-                    // find last :
-                    auto last_pos = str.find_last_of(':');
-                    if (last_pos == std::string::npos) {
-                        throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
-                    }
-                    // Cutting back to opening : for object value
-                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                }
-            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
-                if ((last_non_sp_char == ',' || last_non_sp_char == '[') && can_parse(str + "1" + closing)) {
-                    // Was about to create an array value
-                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + "\"" + closing)) {
-                    // Was inside an array value string
-                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
-                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
-                    // Was inside an array value string after an escape
-                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
-                    // Was inside an array value string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
-                } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
-                    // Had just finished a value
-                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
-                } else {
-                    auto last_pos = str.find_last_of("[,");
-                    if (last_pos == std::string::npos) {
-                        throw std::runtime_error("Cannot heal a truncated JSON array stopped in an unknown location");
-                    }
-                    // Cutting back to last [ or , for array value
-                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                }
-            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
-                if ((last_non_sp_char == '{' && can_parse(str + closing)) ||
-                        (last_non_sp_char == ',' && can_parse(str + "\"\": 1" + closing))) {
-                    // Was about to create an object key+value
-                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
-                } else if (!was_maybe_number() && can_parse(str + ",\"\": 1" + closing)) {
-                    // Was about to create an object key+value
-                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\": 1" + closing;
-                } else if (can_parse(str + "\": 1" + closing)) {
-                    // Was inside an object key string
-                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\": 1" + closing;
-                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
-                    // Was inside an object key string after an escape
-                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\": 1" + closing)) {
-                    // Was inside an object key string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\": 1" + closing;
-                } else {
-                    auto last_pos = str.find_last_of(':');
-                    if (last_pos == std::string::npos) {
-                        throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
-                    }
-                    // fprintf(stderr, "Cutting back to last : for object key+value\n");
-                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                }
-            } else {
-                throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
-            }
-            // fprintf(stderr, "HEALED:\nSTRING <<<\n%s\n>>>\n\nmagic_cut: <<<\n%s\n>>>\n\n", str.c_str(), out.healing_marker.json_dump_marker.c_str());
-            out.json = json::parse(str);
-            it = temptative_end;
-            return true;
-        }
-        // handle unclosed top-level primitive
-        if (err_loc.position != 0 && !healing_marker.empty() && err_loc.stack.empty()) {
-            std::string str(it, temptative_end);
-            const auto & magic_seed = out.healing_marker.marker = healing_marker;
-            if (can_parse(str + "\"")) {
-                // Was inside an string
-                str += (out.healing_marker.json_dump_marker = magic_seed) + "\"";
-            } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"")) {
-                // Was inside an string after an escape
-                str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"";
-            } else {
-                // TODO: handle more unclosed top-level primitive if the stack was empty but we got an error (e.g. "tru", "\"", etc...)
-                // fprintf(stderr, "Closing: TODO\n");
-                return false;
-            }
-            out.json = json::parse(str);
-            it = temptative_end;
-            return true;
-        }
-        return false;
-    }
-    out.json = json::parse(it, end);
-    it = end;
-    return true;
-}

common/json-partial.h

@@ -1,39 +0,0 @@
-#pragma once
-
-// TODO: use json_fwd.hpp when possible
-#include <nlohmann/json.hpp>
-
-// Healing marker (empty if the JSON was fully parsed / wasn't healed).
-struct common_healing_marker {
-    // Raw marker.
-    std::string marker;
-
-    // Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
-    std::string json_dump_marker;
-};
-
-// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
-struct common_json {
-    nlohmann::ordered_json json;
-
-    common_healing_marker healing_marker;
-};
-
-// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
-//
-// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
-// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
-// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
-//
-// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
-bool common_json_parse(
-    const std::string & input,
-    const std::string & healing_marker,
-    common_json & out);
-
-// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
-bool common_json_parse(
-    std::string::const_iterator & it,
-    const std::string::const_iterator & end,
-    const std::string & healing_marker,
-    common_json & out);

conversion/__init__.py

@@ -46,6 +46,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "DbrxForCausalLM": "dbrx",
     "DeciLMForCausalLM": "deci",
     "DeepseekForCausalLM": "deepseek",
+    "DeepseekOCRForCausalLM": "deepseek",
     "DeepseekV2ForCausalLM": "deepseek",
     "DeepseekV3ForCausalLM": "deepseek",
     "DeepseekV32ForCausalLM": "deepseek",
@@ -124,6 +125,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "LLaDAModelLM": "llada",
     "LLaMAForCausalLM": "llama",
     "Lfm25AudioTokenizer": "lfm2",
+    "Lfm2BidirectionalModel": "lfm2",
     "Lfm2ForCausalLM": "lfm2",
     "Lfm2Model": "lfm2",
     "Lfm2MoeForCausalLM": "lfm2",
@@ -232,6 +234,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "UMT5ForConditionalGeneration": "t5",
     "UMT5Model": "t5",
     "UltravoxModel": "ultravox",
+    "UnlimitedOCRForCausalLM": "deepseek",
     "VLlama3ForCausalLM": "llama",
     "VoxtralForConditionalGeneration": "llama",
     "WavTokenizerDec": "wavtokenizer",
@@ -298,6 +301,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "StepVLForConditionalGeneration": "step3",
     "Step3p7ForConditionalGeneration": "step3",
     "UltravoxModel": "ultravox",
+    "UnlimitedOCRForCausalLM": "deepseek",
     "VoxtralForConditionalGeneration": "ultravox",
     "YoutuVLForConditionalGeneration": "youtuvl",
 }

conversion/deepseek.py

@@ -14,7 +14,7 @@ from .base import MmprojModel, ModelBase, TextModel, gguf, logger
 from .qwen import QwenModel
 
 
-@ModelBase.register("DeepseekOCRForCausalLM")
+@ModelBase.register("DeepseekOCRForCausalLM", "UnlimitedOCRForCausalLM")
 class DeepseekOCRVisionModel(MmprojModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -205,6 +205,8 @@ class DeepseekModel(TextModel):
 @ModelBase.register(
     "DeepseekV2ForCausalLM",
     "DeepseekV3ForCausalLM",
+    "DeepseekOCRForCausalLM",
+    "UnlimitedOCRForCausalLM",
     "KimiVLForConditionalGeneration",
     "KimiK25ForConditionalGeneration",
     "YoutuForCausalLM",
@@ -224,7 +226,7 @@ class DeepseekV2Model(TextModel):
         self.origin_hf_arch = hparams.get('architectures', [None])[0]
 
         # special handling for Deepseek OCR
-        if self.origin_hf_arch in ("DeepseekOCRForCausalLM", "DeepseekOCR2ForCausalLM"):
+        if self.origin_hf_arch in ("DeepseekOCRForCausalLM", "DeepseekOCR2ForCausalLM", "UnlimitedOCRForCausalLM"):
             self.model_arch = gguf.MODEL_ARCH.DEEPSEEK2OCR
             self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
             self.gguf_writer.add_architecture()
@@ -350,6 +352,12 @@ class DeepseekV2Model(TextModel):
 
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
+        # Unlimited-OCR sliding window; written for metadata, the decoder ignores it (full MHA)
+        if is_ocr:
+            sliding_window = hparams.get("sliding_window_size") or hparams.get("sliding_window")
+            if sliding_window:
+                self.gguf_writer.add_sliding_window(sliding_window)
+
         if (rope_mscale_all := self.rope_parameters.get("mscale_all_dim")) is not None:
             # [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
             # note: for legacy reasons, this is not consistent with the other usages of self.gguf_writer.add_rope_scaling_yarn_log_mul

conversion/lfm2.py

@@ -64,11 +64,17 @@ class LFM2Model(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Lfm2Model")
+@ModelBase.register("Lfm2Model", "Lfm2BidirectionalModel")
 class LFM2ColBertModel(LFM2Model):
     model_arch = gguf.MODEL_ARCH.LFM2
     dense_tensor_name = "dense_2"
 
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.hf_arch == "Lfm2BidirectionalModel":
+            self.gguf_writer.add_causal_attention(False)
+        self._try_set_pooling_type()
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if not name.startswith(self.dense_tensor_name):
             name = "model." + name
@@ -76,10 +82,11 @@ class LFM2ColBertModel(LFM2Model):
         yield from super().modify_tensors(data_torch, name, bid)
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
-        # dense tensor is stored in a separate safetensors file
+        # optional dense tensor is stored in a separate safetensors file
         from safetensors.torch import load_file
         tensors_file = self.dir_model / "1_Dense" / "model.safetensors"
-        assert tensors_file.is_file()
+        if not tensors_file.is_file():
+            return
         tensor = load_file(tensors_file)["linear.weight"]
         self.gguf_writer.add_embedding_length_out(tensor.shape[0])
         yield f"{self.dense_tensor_name}.weight", tensor.clone()

docs/backend/snapdragon/CMakeUserPresets.json

@@ -24,7 +24,6 @@
             "GGML_LLAMAFILE":   "OFF",
             "GGML_OPENCL":      "ON",
             "GGML_HEXAGON":     "ON",
-            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
             "LLAMA_OPENSSL":    "OFF"
         }
     },
@@ -47,7 +46,6 @@
             "GGML_LLAMAFILE":   "OFF",
             "GGML_OPENCL":      "ON",
             "GGML_HEXAGON":     "ON",
-            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
             "LLAMA_OPENSSL":    "OFF"
         }
     },
@@ -73,7 +71,6 @@
             "GGML_LLAMAFILE":   "OFF",
             "GGML_OPENCL":      "OFF",
             "GGML_HEXAGON":     "ON",
-            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
             "LLAMA_OPENSSL":    "OFF"
         }
     },

ggml/CMakeLists.txt

@@ -266,7 +266,6 @@ set   (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING
                                             "ggml: OpenCL API version to target")
 
 option(GGML_HEXAGON                         "ggml: enable Hexagon backend"                    OFF)
-set(GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE 128 CACHE STRING "ggml: quantize group size (32, 64, or 128)")
 
 # toolchain for vulkan-shaders-gen
 set   (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen")

ggml/src/ggml-cpu/ops.cpp

@@ -3688,8 +3688,6 @@ static void ggml_compute_forward_norm_f32(
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -3703,25 +3701,49 @@ static void ggml_compute_forward_norm_f32(
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
 
-                float sum = 0.0;
-                ggml_vec_sum_f32(ne00, &sum, x);
-                float mean = sum/ne00;
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    const float * xf = (const float *) x;
 
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-                float variance = 0;
+                    float sum = 0.0;
+                    ggml_vec_sum_f32(ne00, &sum, xf);
+                    float mean = sum/ne00;
+
+                    float * yf = (float *) y;
+                    float variance = 0;
 
 #ifdef GGML_USE_ACCELERATE
-                mean = -mean;
-                vDSP_vsadd(x, 1, &mean, y, 1, ne00);
-                vDSP_measqv(y, 1, &variance, ne00);
+                    mean = -mean;
+                    vDSP_vsadd(xf, 1, &mean, yf, 1, ne00);
+                    vDSP_measqv(yf, 1, &variance, ne00);
 #else
-                variance = ggml_vec_cvar_f32(ne00, y, x, mean);
+                    variance = ggml_vec_cvar_f32(ne00, yf, xf, mean);
 #endif //GGML_USE_ACCELERATE
 
-                const float scale = 1.0f/sqrtf(variance + eps);
-                ggml_vec_scale_f32(ne00, y, scale);
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    ggml_vec_scale_f32(ne00, yf, scale);
+                } else {
+                    float sum = 0.0;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        sum += *(const float *) (x + i00*nb00);
+                    }
+                    const float mean = sum/ne00;
+
+                    float variance = 0.0f;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float v = *(const float *) (x + i00*nb00) - mean;
+                        *(float *) (y + i00*nb0) = v;
+                        variance += v * v;
+                    }
+                    variance /= ne00;
+
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        *(float *) (y + i00*nb0) *= scale;
+                    }
+                }
             }
         }
     }
@@ -4142,8 +4164,6 @@ static void ggml_compute_forward_l2_norm_f32(
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -4158,20 +4178,27 @@ static void ggml_compute_forward_l2_norm_f32(
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
 
                 ggml_float sum = 0.0;
                 for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    sum += (ggml_float)(x[i00] * x[i00]);
+                    const float xi = *(const float *) (x + i00*nb00);
+                    sum += (ggml_float)(xi * xi);
                 }
 
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-
-                memcpy(y, x, ne00 * sizeof(float));
-
                 const float scale = 1.0f/fmaxf(sqrtf(sum), eps);
 
-                ggml_vec_scale_f32(ne00, y, scale);
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
+
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    memcpy(y, x, ne00 * sizeof(float));
+                    ggml_vec_scale_f32(ne00, (float *) y, scale);
+                } else {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float xi = *(const float *) (x + i00*nb00);
+                        *(float *) (y + i00*nb0) = xi * scale;
+                    }
+                }
             }
         }
     }

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -5334,7 +5334,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_L2_NORM:
-            return true;
+            return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_RMS_NORM_BACK:
             return ggml_is_contiguous(op->src[0]);
             break;

ggml/src/ggml-hexagon/CMakeLists.txt

@@ -25,7 +25,6 @@ include(ExternalProject)
 option(GGML_HEXAGON_HTP_DEBUG  "ggml-hexagon: enable HTP debug output" OFF)
 option(GGML_HEXAGON_FA_EXP2_HF "ggml-hexagon: use FP16 exp2 polynomial in FA softmax instead of F32 exp round-trip" OFF)
 set(GGML_HEXAGON_HTP_CERT  "$ENV{HEXAGON_HTP_CERT}" CACHE PATH "ggml-hexagon: enable HTP library signing using certificate")
-set(GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE 128 CACHE STRING "ggml-hexagon: quantize group size (32, 64, or 128)")
 
 add_library(htp_iface OBJECT
     ${CMAKE_CURRENT_BINARY_DIR}/htp_iface_stub.c)
@@ -72,15 +71,12 @@ function(build_htp_skel V)
             -DHEXAGON_SDK_ROOT=${HEXAGON_SDK_ROOT}
             -DHEXAGON_TOOLS_ROOT=${HEXAGON_TOOLS_ROOT}
             -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG}
-            -DGGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE}
             -DDSP_VERSION=${V}
             -DPREBUILT_LIB_DIR="toolv19_${V}")
     list(APPEND HTP_SKELS ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-${V}.so)
     set(HTP_SKELS ${HTP_SKELS} PARENT_SCOPE)
 endfunction()
 
-build_htp_skel(v68)
-build_htp_skel(v69)
 build_htp_skel(v73)
 build_htp_skel(v75)
 build_htp_skel(v79)

ggml/src/ggml-hexagon/htp-opnode.h

@@ -5,10 +5,12 @@
 #include "ggml-backend-impl.h"
 #include "ggml-common.h"
 
+#include <algorithm>
 #include <string>
 #include <vector>
 #include <stdio.h>
 #include "htp-ops.h"
+#include "htp/matmul-ops.h"
 
 struct htp_opnode {
     ggml_tensor * node = nullptr;
@@ -17,6 +19,13 @@ struct htp_opnode {
 
     htp_op_code opcode = HTP_OP_INVALID;
 
+    std::vector<ggml_tensor *> extra_dsts;
+
+    int32_t kernel_params[HTP_OP_MAX_KERN_PARAMS] = {0};
+
+    htp_opnode(ggml_tensor * node = nullptr, std::vector<ggml_tensor *> fused = {}, htp_op_code opcode = HTP_OP_INVALID, std::vector<ggml_tensor *> extra_dsts = {})
+        : node(node), fused(std::move(fused)), opcode(opcode), extra_dsts(std::move(extra_dsts)) {}
+
     ggml_op op() const {
         return node->op;
     }
@@ -25,6 +34,26 @@ struct htp_opnode {
         return fused.empty() ? node : fused.back();
     }
 
+    void add_fused(ggml_tensor * t, bool extra_dst = false) {
+        fused.push_back(t);
+        if (extra_dst) {
+            extra_dsts.push_back(t);
+        }
+    }
+
+    std::vector<const ggml_tensor *> get_outputs() const {
+        std::vector<const ggml_tensor *> res;
+        if (extra_dsts.empty()) {
+            res.push_back(dst());
+        } else {
+            res.push_back(node);
+            for (const auto * x : extra_dsts) {
+                res.push_back(x);
+            }
+        }
+        return res;
+    }
+
     const ggml_tensor * src0() const {
         return node->src[0];
     }
@@ -37,10 +66,6 @@ struct htp_opnode {
         return ggml_op_is_empty(node->op);
     }
 
-    void add_fused(ggml_tensor * t) {
-        fused.push_back(t);
-    }
-
     bool stackable() const {
         switch (this->op()) {
             case GGML_OP_MUL_MAT:
@@ -131,87 +156,117 @@ struct htp_opformat {
     char types[16 * GGML_MAX_SRC];
     char buffs[64 * GGML_MAX_SRC];
     char names[64 * GGML_MAX_SRC];
+    char kparams[128];
 
-    int format_tensor_dims(char * str, const struct ggml_tensor * t) {
+    int format_tensor_dims(char * str, size_t max_size, const struct ggml_tensor * t) {
         if (!t) {
-            return sprintf(str, "NONE");
+            return snprintf(str, max_size, "NONE");
         }
         if (t->ne[2] == 1 && t->ne[3] == 1) {
-            return sprintf(str, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
+            return snprintf(str, max_size, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
         } else {
-            return sprintf(str, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
+            return snprintf(str, max_size, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
         }
     }
 
-    void format_op_dims(char * str, const htp_opnode & node) {
+    void format_op_dims(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += format_tensor_dims(p, inputs[0]);
+            p += std::min((size_t)format_tensor_dims(p, p_end - p, inputs[0]), (size_t)(p_end - p));
 
             for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += format_tensor_dims(p, inputs[i]);
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)format_tensor_dims(p, p_end - p, inputs[i]), (size_t)(p_end - p));
+                }
             }
 
-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
         char self[64];
-        format_tensor_dims(self, node.dst());
-        p += sprintf(p, "%s", self);
+        format_tensor_dims(self, sizeof(self), node.dst());
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", self), (size_t)(p_end - p));
+        }
     }
 
-    int format_tensor_strides(char * str, const struct ggml_tensor * t) {
+    int format_tensor_strides(char * str, size_t max_size, const struct ggml_tensor * t) {
         if (!t) {
-            return sprintf(str, "NONE");
+            return snprintf(str, max_size, "NONE");
         }
         const char * c = ggml_is_contiguous(t) ? "" : "!";
 
         if (t->ne[2] == 1 && t->ne[3] == 1) {
-            return sprintf(str, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
+            return snprintf(str, max_size, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
         } else {
-            return sprintf(str, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
+            return snprintf(str, max_size, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
         }
     }
 
-    void format_op_strides(char * str, const htp_opnode & node) {
+    void format_op_strides(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += format_tensor_strides(p, inputs[0]);
+            p += std::min((size_t)format_tensor_strides(p, p_end - p, inputs[0]), (size_t)(p_end - p));
 
             for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += format_tensor_strides(p, inputs[i]);
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)format_tensor_strides(p, p_end - p, inputs[i]), (size_t)(p_end - p));
+                }
             }
 
-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
         char self[64];
-        format_tensor_strides(self, node.dst());
-        p += sprintf(p, "%s", self);
+        format_tensor_strides(self, sizeof(self), node.dst());
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", self), (size_t)(p_end - p));
+        }
     }
 
-    void format_op_types(char * str, const htp_opnode & node) {
+    void format_op_types(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += sprintf(p, "%s", inputs[0] ? ggml_type_name(inputs[0]->type) : "NONE");
-
-            for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", inputs[i] ? ggml_type_name(inputs[i]->type) : "NONE");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[0] ? ggml_type_name(inputs[0]->type) : "NONE"), (size_t)(p_end - p));
             }
 
-            p += sprintf(p, " -> ");
+            for (size_t i = 1; i < inputs.size(); i++) {
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[i] ? ggml_type_name(inputs[i]->type) : "NONE"), (size_t)(p_end - p));
+                }
+            }
+
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
-        p += sprintf(p, "%s", ggml_type_name(node.dst()->type));
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", ggml_type_name(node.dst()->type)), (size_t)(p_end - p));
+        }
     }
 
     const char * tensor_buff_name(const struct ggml_tensor * t) {
@@ -221,51 +276,102 @@ struct htp_opformat {
         return "NONE";
     }
 
-    void format_op_buffs(char * str, const htp_opnode & node) {
+    void format_op_buffs(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += sprintf(p, "%s", tensor_buff_name(inputs[0]));
-
-            for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", tensor_buff_name(inputs[i]));
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(inputs[0])), (size_t)(p_end - p));
             }
 
-            p += sprintf(p, " -> ");
+            for (size_t i = 1; i < inputs.size(); i++) {
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(inputs[i])), (size_t)(p_end - p));
+                }
+            }
+
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
-        p += sprintf(p, "%s", tensor_buff_name(node.dst()));
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(node.dst())), (size_t)(p_end - p));
+        }
     }
 
-    void format_op_names(char * str, const htp_opnode & node) {
+    void format_op_names(char * str, size_t max_size, const htp_opnode & node) {
         char * p = str;
+        char * p_end = str + max_size;
         auto inputs = node.get_inputs();
 
         if (!inputs.empty()) {
-            p += sprintf(p, "%s", inputs[0] ? inputs[0]->name : "NONE");
-
-            for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", inputs[i] ? inputs[i]->name : "NONE");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[0] ? inputs[0]->name : "NONE"), (size_t)(p_end - p));
             }
 
-            p += sprintf(p, " -> ");
+            for (size_t i = 1; i < inputs.size(); i++) {
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[i] ? inputs[i]->name : "NONE"), (size_t)(p_end - p));
+                }
+            }
+
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
         }
 
-        p += sprintf(p, "%s", node.dst()->name);
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", node.dst()->name), (size_t)(p_end - p));
+        }
+    }
+    void format_kernel_params(char * str, size_t max_size, const htp_opnode & node) {
+        if (node.opcode == HTP_OP_MUL_MAT || node.opcode == HTP_OP_MUL_MAT_ID ||
+            node.opcode == HTP_OP_MUL_MAT_QKV || node.opcode == HTP_OP_MUL_MAT_FFN) {
+            const auto * kparams = (const struct htp_mm_kernel_params *) node.kernel_params;
+            const char * path = "unknown";
+            int32_t type = kparams->kernel_type;
+            if (type == HTP_MM_KERNEL_HMX_2D || type == HTP_MM_KERNEL_HMX_F16_BATCHED) {
+                path = "hmx-tiled";
+            } else if (type == HTP_MM_KERNEL_HVX_F16_F16_VTCM || type == HTP_MM_KERNEL_HVX_F32_F32_VTCM ||
+                       type == HTP_MM_KERNEL_HVX_QUANT_ROW    || type == HTP_MM_KERNEL_HVX_QUANT_BLOCK) {
+                path = "hvx-tiled";
+            } else if (type == HTP_MM_KERNEL_HVX_F16_F16_DDR  || type == HTP_MM_KERNEL_HVX_F16_F32_DDR ||
+                       type == HTP_MM_KERNEL_HVX_F32_F32_DDR  || type == HTP_MM_KERNEL_HVX_F32_F16_DDR ||
+                       type == HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT) {
+                path = "hvx-flat";
+            }
+            snprintf(str, max_size, "%s vtcm %d", path, (int) kparams->vtcm_size);
+        } else {
+            snprintf(str, max_size, "----");
+        }
     }
 
     void format(const htp_opnode & node) {
-        format_op_dims(dims, node);
-        format_op_strides(strides, node);
-        format_op_types(types, node);
-        format_op_buffs(buffs, node);
-        format_op_names(names, node);
+        format_op_dims(dims, sizeof(dims), node);
+        format_op_strides(strides, sizeof(strides), node);
+        format_op_types(types, sizeof(types), node);
+        format_op_buffs(buffs, sizeof(buffs), node);
+        format_op_names(names, sizeof(names), node);
+        format_kernel_params(kparams, sizeof(kparams), node);
     }
 
-    htp_opformat() {}
+    htp_opformat() {
+        strides[0] = '\0';
+        dims[0]    = '\0';
+        types[0]   = '\0';
+        buffs[0]   = '\0';
+        names[0]   = '\0';
+        kparams[0] = '\0';
+    }
     htp_opformat(const htp_opnode & node) { format(node); }
 };
 

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -19,43 +19,9 @@ add_library(${HTP_LIB} SHARED
     htp_iface_skel.c
     worker-pool.c
     hex-dma.c
-)
-
-target_compile_definitions(${HTP_LIB} PRIVATE
-    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,HTP_DEBUG=1,NDEBUG=1>
-    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,FARF_HIGH=1,>
-    FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE})
-
-if (GGML_HEXAGON_FA_EXP2_HF)
-    message(STATUS "ggml-htp: HMX_FA_USE_EXP2_HF=1 (use FP16 exp2 polynomial in FA softmax)")
-    target_compile_definitions(${HTP_LIB} PRIVATE HMX_FA_USE_EXP2_HF=1)
-endif()
-
-# HMX acceleration: available on v73+ architectures
-set(HTP_HMX_VERSIONS v73 v75 v79 v81)
-list(FIND HTP_HMX_VERSIONS ${DSP_VERSION} _hmx_idx)
-
-if (_hmx_idx GREATER_EQUAL 0)
-    target_sources(${HTP_LIB} PRIVATE
-        hmx-flash-attn-ops.c
-        hmx-matmul-ops.c
-        hmx-queue.c
-    )
-
-    # -mhmx enables HMX instruction set (needed by files that include hmx-utils.h)
-    set_source_files_properties(
-        hmx-flash-attn-ops.c
-        hmx-matmul-ops.c
-        hmx-queue.c
-        PROPERTIES COMPILE_OPTIONS "-mhmx"
-    )
-
-    target_compile_definitions(${HTP_LIB} PRIVATE HTP_HAS_HMX=1)
-endif()
-
-build_idl(htp_iface.idl ${HTP_LIB})
-
-target_sources(${HTP_LIB} PRIVATE
+    hmx-queue.c
+    flash-attn-ops.c
+    hmx-flash-attn-ops.c
     matmul-ops.c
     binary-ops.c
     unary-ops.c
@@ -63,7 +29,6 @@ target_sources(${HTP_LIB} PRIVATE
     softmax-ops.c
     act-ops.c
     rope-ops.c
-    flash-attn-ops.c
     set-rows-ops.c
     get-rows-ops.c
     cpy-ops.c
@@ -79,6 +44,17 @@ target_sources(${HTP_LIB} PRIVATE
     pad-ops.c
 )
 
+target_compile_definitions(${HTP_LIB} PRIVATE
+    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,HTP_DEBUG=1,NDEBUG=1>
+    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,FARF_HIGH=1,>)
+
+if (GGML_HEXAGON_FA_EXP2_HF)
+    message(STATUS "ggml-htp: HMX_FA_USE_EXP2_HF=1 (use FP16 exp2 polynomial in FA softmax)")
+    target_compile_definitions(${HTP_LIB} PRIVATE HMX_FA_USE_EXP2_HF=1)
+endif()
+
+build_idl(htp_iface.idl ${HTP_LIB})
+
 set_target_properties(${HTP_LIB} PROPERTIES EXPORT_COMPILE_COMMANDS ON)
 
 install(TARGETS ${HTP_LIB})

ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake

@@ -3,7 +3,7 @@ if (HEXAGON_TOOLCHAIN_INCLUDED)
 endif()
 set(HEXAGON_TOOLCHAIN_INCLUDED true)
 
-#Cross Compiling for Hexagon
+# Cross Compiling for Hexagon
 set(HEXAGON TRUE)
 set(CMAKE_SYSTEM_NAME QURT)
 set(CMAKE_SYSTEM_PROCESSOR Hexagon)
@@ -14,7 +14,6 @@ set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
 set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
 set(CUSTOM_RUNELF_PATH "")
 
-#To fix backward compatibility with EAI addon.
 if (NOT HEXAGON_SDK_ROOT)
     set(HEXAGON_SDK_ROOT $ENV{HEXAGON_SDK_ROOT})
 endif()
@@ -31,7 +30,6 @@ endif()
 file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
 file(TO_CMAKE_PATH "${HEXAGON_SDK_ROOT}"   HEXAGON_SDK_ROOT)
 
-#Get the Binary extension of the Hexagon Toolchain
 if(CMAKE_HOST_SYSTEM_NAME STREQUAL Windows)
     set(HEXAGON_TOOLCHAIN_SUFFIX .exe)
 endif()
@@ -48,12 +46,12 @@ set(CMAKE_TRY_COMPILE_PLATFORM_VARIABLES
     HEXAGON_TOOLS_ROOT
 )
 
-#QURT Related includes and linker flags
+# QURT Related includes and linker flags
 set(V_ARCH ${HEXAGON_ARCH})
 set(_QURT_INSTALL_DIR "${HEXAGON_SDK_ROOT}/rtos/qurt/ADSP${V_ARCH}MP${V_ARCH_EXTN}")
 set(_QURT_INSTALL_DIR "${HEXAGON_SDK_ROOT}/rtos/qurt/compute${V_ARCH}${V_ARCH_EXTN}")
 
-if( ${TREE} MATCHES PAKMAN )
+if (${TREE} MATCHES PAKMAN)
     set(_QURT_INSTALL_DIR "${QURT_IMAGE_DIR}/compute${V_ARCH}${V_ARCH_EXTN}")
 endif()
 message(DEBUG "_QURT_INSTALL_DIR:${_QURT_INSTALL_DIR}")
@@ -83,11 +81,9 @@ set(QURT_START_LINK_LIBS
     )
 STRING(REPLACE ";" " " QURT_START_LINK_LIBS "${QURT_START_LINK_LIBS}")
 
-set(QURT_END_LINK_LIBS
-    ${TARGET_DIR}/fini.o
-    )
+set(QURT_END_LINK_LIBS ${TARGET_DIR}/fini.o)
 
-#Non QURT related includes and linker flags
+# Non QURT related includes and linker flags
 
 set(TARGET_DIR_NOOS "${HEXAGON_TOOLCHAIN}/Tools/target/hexagon/lib/${HEXAGON_ARCH}")
 
@@ -99,8 +95,10 @@ if (NOT NO_WRAP_MEM_API)
     set(WRAP_MEMALIGN -Wl,--wrap=memalign)
 endif()
 
+set(ARCH_FLAGS "-mcpu=${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH} -mhmx")
+
 set(PIC_SHARED_LD_FLAGS
-    -mcpu=${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH}
+    ${ARCH_FLAGS}
     -G0
     -fpic
     -Wl,-Bsymbolic
@@ -120,13 +118,13 @@ STRING(REPLACE ";" " " PIC_SHARED_LD_FLAGS "${PIC_SHARED_LD_FLAGS}")
 
 set(HEXAGON_PIC_SHARED_LINK_OPTIONS "${PIC_SHARED_LD_FLAGS}")
 
-#System include paths
+# System include paths
 include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/incs)
 include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/incs/stddef)
 include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/ipc/fastrpc/incs)
 
-#LLVM toolchain setup
-#Compiler paths, options and architecture
+# LLVM toolchain setup
+# Compiler paths, options and architecture
 set(CMAKE_C_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang${HEXAGON_TOOLCHAIN_SUFFIX})
 set(CMAKE_CXX_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang++${HEXAGON_TOOLCHAIN_SUFFIX})
 set(CMAKE_AR ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-ar${HEXAGON_TOOLCHAIN_SUFFIX})
@@ -137,8 +135,8 @@ set(CMAKE_PREFIX_PATH ${HEXAGON_TOOLCHAIN}/Tools/target/hexagon)
 set(CMAKE_SHARED_LIBRARY_SONAME_C_FLAG   "-Wl,-soname,")
 set(CMAKE_SHARED_LIBRARY_SONAME_CXX_FLAG "-Wl,-soname,")
 
-#Compiler Options
-set(COMMON_FLAGS "-mcpu=hexagon${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH} -fvectorize -flto -Wall -Werror -fno-zero-initialized-in-bss -G0 -fdata-sections -fpic ${XQF_ARGS}")
+# Compiler Options
+set(COMMON_FLAGS "${ARCH_FLAGS} -fvectorize -flto -Wall -Werror -fno-zero-initialized-in-bss -G0 -fdata-sections -fpic ${XQF_ARGS}")
 
 set(CMAKE_CXX_FLAGS_DEBUG          "${COMMON_FLAGS} -O0 -D_DEBUG -g")
 set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS} -O2 -g")

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -18,7 +18,8 @@
 #include "htp-ctx.h"
 #include "htp-ops.h"
 #include "htp-ops.h"
-#include "hmx-ops.h"
+
+int hmx_flash_attn_ext(struct htp_ops_context * octx);
 
 // Must be multiple of 32
 #define FLASH_ATTN_BLOCK_SIZE (32 * 2)
@@ -633,7 +634,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-#ifdef HTP_HAS_HMX
     // HMX path: head_dim multiple of 64, F16 KV, and no sinks
     if (k->type == HTP_TYPE_F16 && v->type == HTP_TYPE_F16 && k->ne[0] % 64 == 0 && v->ne[0] % 64 == 0 && octx->src[4] == NULL) {
         int ret = hmx_flash_attn_ext(octx);
@@ -642,7 +642,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
         }
         // VTCM too small or other failure -> fall through to HVX path
     }
-#endif
 
     struct htp_fa_context factx;
     factx.octx = octx;

ggml/src/ggml-hexagon/htp/hex-common.h

@@ -0,0 +1,80 @@
+#ifndef HEX_COMMON_H
+#define HEX_COMMON_H
+
+#include <stdint.h>
+#include <stddef.h>
+#include <stdbool.h>
+
+#ifndef SIZE_MAX
+#define SIZE_MAX ((size_t)-1)
+#endif
+
+#ifndef MAX
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#endif
+
+#ifndef MIN
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#endif
+
+static inline uint32_t hex_ceil_pow2(uint32_t x) {
+    if (x <= 1) { return 1; }
+    int p = 2;
+    x--;
+    while (x >>= 1) { p <<= 1; }
+    return p;
+}
+
+static inline size_t hmx_ceil_div(size_t num, size_t den) {
+    return (num + den - 1) / den;
+}
+
+static inline int32_t hex_is_aligned(const void * addr, uint32_t align) {
+    return ((size_t) addr & (align - 1)) == 0;
+}
+
+static inline size_t hex_align_up(size_t v, size_t align) {
+    return hmx_ceil_div(v, align) * align;
+}
+
+static inline size_t hex_align_down(size_t v, size_t align) {
+    return (v / align) * align;
+}
+
+static inline int32_t hex_is_one_chunk(void * addr, uint32_t n, uint32_t chunk_size) {
+    uint32_t left_off  = (size_t) addr & (chunk_size - 1);
+    uint32_t right_off = left_off + n;
+    return right_off <= chunk_size;
+}
+
+static inline uint32_t hex_round_up(uint32_t n, uint32_t m) {
+    return m * ((n + m - 1) / m);
+}
+
+static inline size_t hex_smin(size_t a, size_t b) {
+    return a < b ? a : b;
+}
+
+static inline size_t hex_smax(size_t a, size_t b) {
+    return a > b ? a : b;
+}
+
+static inline void hex_swap_ptr(void ** p1, void ** p2) {
+    void * t = *p1;
+    *p1      = *p2;
+    *p2      = t;
+}
+
+static inline bool hex_mul_overflow(size_t a, size_t b, size_t *out) {
+    if (a != 0 && b > SIZE_MAX / a) return true;
+    *out = a * b;
+    return false;
+}
+
+static inline bool hex_add_overflow(size_t a, size_t b, size_t *out) {
+    if (a > SIZE_MAX - b) return true;
+    *out = a + b;
+    return false;
+}
+
+#endif // HEX_COMMON_H

ggml/src/ggml-hexagon/htp/hex-dma.h

@@ -5,6 +5,7 @@
 #include <hexagon_types.h>
 #include <stdbool.h>
 #include <stdint.h>
+#include "hex-utils.h"
 
 #include "hex-profile.h"
 
@@ -127,13 +128,8 @@ static inline dma_ptr dma_make_ptr(void *dst, const void *src)
     return p;
 }
 
-#if __HVX_ARCH__ < 73
-static const uint32_t dma_src_l2_bypass_on = 1;
-static const uint32_t dma_dst_l2_bypass_on = 0;
-#else
 static const uint32_t dma_src_l2_bypass_on = 1;
 static const uint32_t dma_dst_l2_bypass_on = 1;
-#endif
 
 static inline bool dma_queue_push_single_1d(dma_queue * q, dma_ptr dptr, size_t size) {
     if (((q->push_idx + 1) & q->idx_mask) == q->pop_idx) {

ggml/src/ggml-hexagon/htp/hex-utils.h

@@ -11,14 +11,7 @@
 
 #include "hex-fastdiv.h"
 #include "hex-dump.h"
-
-#ifndef MAX
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-#endif
-
-#ifndef MIN
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#endif
+#include "hex-common.h"
 
 static inline uint64_t hex_get_cycles() {
     uint64_t cycles = 0;
@@ -32,54 +25,6 @@ static inline uint64_t hex_get_pktcnt() {
     return pktcnt;
 }
 
-static inline uint32_t hex_ceil_pow2(uint32_t x) {
-    if (x <= 1) { return 1; }
-    int p = 2;
-    x--;
-    while (x >>= 1) { p <<= 1; }
-    return p;
-}
-
-static inline size_t hmx_ceil_div(size_t num, size_t den) {
-    return (num + den - 1) / den;
-}
-
-static inline int32_t hex_is_aligned(const void * addr, uint32_t align) {
-    return ((size_t) addr & (align - 1)) == 0;
-}
-
-static inline size_t hex_align_up(size_t v, size_t align) {
-    return hmx_ceil_div(v, align) * align;
-}
-
-static inline size_t hex_align_down(size_t v, size_t align) {
-    return (v / align) * align;
-}
-
-static inline int32_t hex_is_one_chunk(void * addr, uint32_t n, uint32_t chunk_size) {
-    uint32_t left_off  = (size_t) addr & (chunk_size - 1);
-    uint32_t right_off = left_off + n;
-    return right_off <= chunk_size;
-}
-
-static inline uint32_t hex_round_up(uint32_t n, uint32_t m) {
-    return m * ((n + m - 1) / m);
-}
-
-static inline size_t hex_smin(size_t a, size_t b) {
-    return a < b ? a : b;
-}
-
-static inline size_t hex_smax(size_t a, size_t b) {
-    return a > b ? a : b;
-}
-
-static inline void hex_swap_ptr(void ** p1, void ** p2) {
-    void * t = *p1;
-    *p1      = *p2;
-    *p2      = t;
-}
-
 static inline void hex_l2fetch(const void * p, uint32_t width, uint32_t stride, uint32_t height) {
     const uint64_t control = Q6_P_combine_RR(stride, Q6_R_combine_RlRl(width, height));
     Q6_l2fetch_AP((void *) p, control);

ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c

@@ -49,7 +49,7 @@
 // g_br = hex_align_up(gqa_factor * Br, 32) replaces Br for all Q/O/S/P/D dimensions.
 // Layout: Q + O_ping + O_pong + K_dma*2 + V_dma*2 + K_tile + V_tile + S + P + D + vectors + scales
 // Mask is DMA'd into a VTCM buffer (Br rows per KV block) to avoid DDR reads in softmax.
-static size_t hmx_fa_compute_vtcm_usage(size_t gqa_factor, size_t DK, size_t DV, size_t Br, size_t Bc, size_t n_threads, bool use_pipeline) {
+static size_t hmx_fa_compute_vtcm_usage(size_t gqa_factor, size_t DK, size_t DV, size_t Br, size_t Bc, size_t n_threads, bool pipeline) {
     const size_t g_br         = hex_align_up(gqa_factor * Br, HMX_FP16_TILE_N_ROWS);
     const size_t q_tile_size  = hex_align_up(g_br * DK * sizeof(__fp16), 4096);    // Q:  [g_br, DK]
     const size_t o_tile_size  = hex_align_up(g_br * DV * sizeof(__fp16), 4096);    // O:  [g_br, DV] x2 ping-pong
@@ -70,7 +70,7 @@ static size_t hmx_fa_compute_vtcm_usage(size_t gqa_factor, size_t DK, size_t DV,
            + k_dma_size  * 2               // K DMA x2
            + v_dma_size  * 2               // V DMA x2
            + k_tile_size * 1               // K tiles
-           + v_tile_size * (use_pipeline ? 2 : 1) // V tiles (double-buffered if pipelining)
+           + v_tile_size * (pipeline ? 2 : 1) // V tiles (double-buffered if pipelining)
            + s_tile_size * 2               // S + P
            + d_tile_size * 1               // D (diagonal matrix)
            + col_vec_size * 4              // m_vec, l_vec, s_rowmax, p_rowsum
@@ -290,7 +290,7 @@ static const int16_t d_tile_scatter_offsets[64] __attribute__((aligned(128))) =
 
 struct hmx_fa_context {
     const struct htp_ops_context * octx;
-    bool         use_pipeline;  // true when n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads >= 2
+    bool         pipeline;  // true when n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads >= 2
     uint32_t     n_threads;
 
     // Op parameters
@@ -409,7 +409,7 @@ static void fa_v_interleave_thread(unsigned int n, unsigned int i, void * data)
         return;
     }
 
-    __fp16 * v_tiles_dest = factx->use_pipeline ? factx->vtcm_v_tiles[args->buf_idx] : factx->vtcm_v_tiles[0];
+    __fp16 * v_tiles_dest = factx->pipeline ? factx->vtcm_v_tiles[args->buf_idx] : factx->vtcm_v_tiles[0];
 
     struct htp_thread_trace * tr = factx->octx->ctx ? &factx->octx->ctx->trace[i] : NULL;
     htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, start);
@@ -1312,13 +1312,13 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     const size_t g_br = hex_align_up(G * Br, HMX_FP16_TILE_N_ROWS);
 
     const uint32_t n_kv_blocks  = (nek1 + Bc - 1) / Bc;
-    const bool     use_pipeline = (n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads_init >= 2);
+    const bool     pipeline = (n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads_init >= 2);
 
     // Bypass thread pool dispatch for small prompts/non-pipelined prefill by setting n_threads = 1
-    const uint32_t n_threads = use_pipeline ? n_threads_init : 1;
+    const uint32_t n_threads = pipeline ? n_threads_init : 1;
 
     FARF(HIGH, "hmx-fa: neq1=%u nek1=%u DK=%u DV=%u G=%u Br=%zu Bc=%zu g_br=%zu n_kv_blocks=%u pipeline=%d vtcm=%zu",
-         neq1, nek1, DK, DV, G, Br, Bc, g_br, n_kv_blocks, use_pipeline, vtcm_budget);
+         neq1, nek1, DK, DV, G, Br, Bc, g_br, n_kv_blocks, pipeline, vtcm_budget);
 
     // ======== Build context ========
     struct hmx_fa_context factx;
@@ -1339,7 +1339,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     factx.n_kv_blocks    = n_kv_blocks;
     factx.is_q_fp32      = (q->type == HTP_TYPE_F32);
     factx.is_dst_fp32    = (dst->type == HTP_TYPE_F32);
-    factx.use_pipeline   = use_pipeline;
+    factx.pipeline   = pipeline;
     factx.mask_broadcast = (mask != NULL && mask->ne[2] == 1);
 
     // Extract op parameters (mutable during softcap adjustment, then stored as const in factx)
@@ -1405,7 +1405,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     factx.vtcm_v_fp16[1]      = (__fp16 *) vtcm_seq_alloc(&vtcm_cur, v_dma_bytes);
     factx.vtcm_k_tiles        = (__fp16 *) vtcm_seq_alloc(&vtcm_cur, k_tile_bytes);
     factx.vtcm_v_tiles[0]     = (__fp16 *) vtcm_seq_alloc(&vtcm_cur, v_tile_bytes);
-    if (use_pipeline) {
+    if (pipeline) {
         factx.vtcm_v_tiles[1] = (__fp16 *) vtcm_seq_alloc(&vtcm_cur, v_tile_bytes);
     } else {
         factx.vtcm_v_tiles[1] = NULL;
@@ -1456,7 +1456,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     // ======== HMX lock strategy ========
     // Pipeline: queue thread auto-acquires HMX lock on first push; released by suspend.
     // Fallback: main thread holds the lock (original behavior).
-    if (!factx.use_pipeline) {
+    if (!factx.pipeline) {
         HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
     }
 
@@ -1550,7 +1550,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                 const size_t k_src_stride = size_k_row_padded / sizeof(__fp16);
                 const size_t v_src_stride = size_v_row_padded / sizeof(__fp16);
 
-                if (factx.use_pipeline) {
+                if (factx.pipeline) {
                     // ==================================================================
                     // Pipeline path: HVX phases ‖ HMX queue worker
                     // ==================================================================
@@ -1780,7 +1780,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                     fa_build_d_diag_inv_l(&factx, n_row_tiles, n_row_tiles_g_br);
 
                     // HMX: O_final = diag(1/l) @ O_prev
-                    if (factx.use_pipeline) {
+                    if (factx.pipeline) {
                         on_job.o_curr           = o_tile_curr;
                         on_job.o_prev           = o_tile_prev;
                         on_job.d_tiles          = factx.vtcm_d_tiles;
@@ -1826,7 +1826,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
         }  // end KV head loop
     }  // end batch loop
 
-    if (factx.use_pipeline) {
+    if (factx.pipeline) {
         hmx_queue_suspend(ctx->hmx_queue);
     } else {
         HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);

ggml/src/ggml-hexagon/htp/hmx-ops.c

@@ -1,6 +0,0 @@
-// HMX operations compiled as a single translation unit.
-// This allows interprocedural optimizations within HMX ops without requiring global HTP LTO.
-
-#include "hmx-queue.c"
-#include "hmx-matmul-ops.c"
-#include "hmx-flash-attn-ops.c"

ggml/src/ggml-hexagon/htp/hmx-ops.h

@@ -1,88 +0,0 @@
-// HMX operation entry-point declarations.
-// Ported from htp-ops-lib/include/dsp/ops.h (renamed, benchmark kernels removed). (https://github.com/haozixu/htp-ops-lib)
-
-#ifndef HMX_OPS_H
-#define HMX_OPS_H
-
-#include <stddef.h>
-#include <stdint.h>
-
-#include "htp-ops.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-typedef struct {
-    float        *dst;
-    const float  *activation;
-    const __fp16 *permuted_weight;
-    int           m;
-    int           k;
-    int           n;
-    int           act_stride;
-    int           weight_stride;
-    int           dst_stride;
-    int           ne02;
-    int           ne03;
-    int           ne12;
-    int           ne13;
-    size_t        src0_nb2;
-    size_t        src0_nb3;
-    size_t        src1_nb2;
-    size_t        src1_nb3;
-    size_t        dst_nb2;
-    size_t        dst_nb3;
-} hmx_matmul_f16_f32_batched_params_t;
-
-// HMX matrix multiplication — tile-permuted FP16 weights, FP32 activation/output
-// act_stride: activation row stride in elements (= k for contiguous, or
-//             nb[1]/sizeof(float) for permuted tensors like attention Q).
-// weight_stride: weight row stride in elements (= k for compact weights, or
-//                nb[1]/sizeof(__fp16) for permuted KV-cache views used by QK).
-int hmx_matmul_f16_f32(struct htp_context *ctx,
-                                float *restrict dst,
-                                const float *activation,
-                                const __fp16 *permuted_weight,
-                                int m, int k, int n,
-                                int act_stride,
-                                int weight_stride);
-
-// Batched F16 wrapper over hmx_mat_mul_f16_f32.
-// Batch semantics match ggml_mul_mat(): src0 broadcasts to src1 in dims 2/3.
-int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32_batched_params_t *params);
-
-// HMX matrix multiplication — all supported weight types (F16/F32/Q4_0/Q4_1/Q8_0/IQ4_NL/MXFP4)
-int hmx_matmul_2d_f32(struct htp_context *ctx,
-                                      float *restrict dst,
-                                      const float *activation,
-                                      const uint8_t *permuted_weight,
-                                      int m, int k, int n,
-                                      int act_stride,
-                                      int weight_stride,
-                                      int weight_type);
-
-struct mmid_row_mapping;
-
-int hmx_matmul_id_2d_f32(struct htp_context *ctx,
-                                         float *restrict dst,
-                                         const float *activation,
-                                         const uint8_t *permuted_weight,
-                                         int m, int k, int n,
-                                         int ne11,
-                                         size_t act_nb1, size_t act_nb2,
-                                         size_t dst_nb1, size_t dst_nb2,
-                                         int weight_stride,
-                                         int weight_type,
-                                         const struct mmid_row_mapping *matrix_rows,
-                                         int cur_a,
-                                         int mapping_stride);
-
-// HMX flash attention
-int hmx_flash_attn_ext(struct htp_ops_context * octx);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // HMX_OPS_H

ggml/src/ggml-hexagon/htp/htp-ctx.h

@@ -13,7 +13,9 @@
 #include <stdint.h>
 #include <stdbool.h>
 
+#ifndef HTP_MAX_NTHREADS
 #define HTP_MAX_NTHREADS 10
+#endif
 #define HTP_MAX_MMAPS    16
 
 // Memory mapping
@@ -42,9 +44,13 @@ struct htp_ops_context {
 
     enum htp_op_code    op; // FIXME: rename to opcode
     int32_t             op_params[HTP_OP_MAX_PARAMS];
+    int32_t             kernel_params[HTP_OP_MAX_KERN_PARAMS];
 
     const struct htp_tensor * src[HTP_OP_MAX_INPUTS];
-    const struct htp_tensor * dst;
+    union {
+        const struct htp_tensor * dst;
+        const struct htp_tensor * dsts[HTP_OP_MAX_OUTPUTS];
+    };
 
     // TODO convert these to an array
     struct htp_spad src0_spad;
@@ -87,13 +93,13 @@ struct htp_context {
 
     struct htp_ops_context octx;
 
-#ifdef HTP_HAS_HMX
     struct hmx_queue *     hmx_queue; // Async HMX queue for pipeline overlap
-#endif
 };
 
 int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(struct htp_ops_context * octx);
+int op_matmul_qkv(struct htp_ops_context * octx);
+int op_matmul_ffn(struct htp_ops_context * octx);
 int op_binary(struct htp_ops_context * octx);
 int op_unary(struct htp_ops_context * octx);
 int op_sum_rows(struct htp_ops_context * octx);

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -28,18 +28,19 @@ enum htp_data_type {
     HTP_TYPE_MXFP4  = 39,
 
     // types used internally for repack, dyn.quant, etc
-    HTP_TYPE_Q4_0x4x2 = 200,
-    HTP_TYPE_Q4_1x4x2,
-    HTP_TYPE_Q8_0x4x2,
-    HTP_TYPE_MXFP4x4x2,
+    HTP_TYPE_Q4_0_TILED = 200,
+    HTP_TYPE_Q4_1_TILED,
+    HTP_TYPE_Q8_0_TILED,
+    HTP_TYPE_MXFP4_TILED,
 
     HTP_TYPE_INVALID
 };
 
 // Constats for internal types
-#define QK_Q4_0x4x2  256  // 4x Q4_0  blocks packed with next 4x Q4_0 blocks (size in bytes 128)
-#define QK_Q8_0x4x2  256  // 4x Q8_0  blocks concat with next 4x Q8_0 blocks
-#define QK_MXFP4x4x2 256  // 4x MXFP4 blocks concat with next 4x MXFP4 blocks
+#define QK_Q4_0_TILED  256  // 32x32 Q4_0 tiled layout
+#define QK_Q8_0_TILED  128  // 32x32 Q8_0 tiled layout
+#define QK_MXFP4_TILED 256  // 32x32 MXFP4 tiled layout
+
 
 
 // Mask to enable various stages of the Ops.
@@ -57,6 +58,8 @@ enum htp_op_code {
     HTP_OP_DIV = 3,
     HTP_OP_MUL_MAT,
     HTP_OP_MUL_MAT_ID,
+    HTP_OP_MUL_MAT_QKV,
+    HTP_OP_MUL_MAT_FFN,
     HTP_OP_RMS_NORM,
     HTP_OP_RMS_NORM_MUL,
     HTP_OP_UNARY_SILU,
@@ -99,7 +102,9 @@ enum htp_op_code {
 
 #define HTP_OP_MAX_DIMS    4    // aka GGML_MAX_DIMS
 #define HTP_OP_MAX_INPUTS  6    // aka GGML_MAX_SRCS
+#define HTP_OP_MAX_OUTPUTS 4
 #define HTP_OP_MAX_PARAMS  16   // aka GGML_MAX_OP_PARAMS
+#define HTP_OP_MAX_KERN_PARAMS 32
 
 #define HTP_OP_MAX_BUFS    16
 #define HTP_OP_MAX_REQS    256
@@ -142,8 +147,10 @@ struct htp_op_desc {
     uint32_t opcode;                    // GGML/HTP Op
     uint32_t flags;                     // Op flags
     int32_t  params[HTP_OP_MAX_PARAMS]; // Params for the op, e.g. epsilon of RMS norm
+    int32_t  kernel_params[HTP_OP_MAX_KERN_PARAMS]; // generic blob for host-precomputed parameters
     uint16_t src[HTP_OP_MAX_INPUTS];    // Input tensors indices
-    uint16_t dst;                       // Output tensor index
+    uint16_t dst[HTP_OP_MAX_OUTPUTS];   // Output tensor indices
+    uint16_t pad[2];                    // padding to align to 64 bits
 };
 
 #ifndef HTP_MAX_NTHREADS

ggml/src/ggml-hexagon/htp/htp_iface.idl

@@ -11,12 +11,13 @@ struct htp_iface_pmu_conf {
 };
 
 interface htp_iface : remote_handle64 {
-    AEEResult start(in uint32 sess_id, in uint64 dsp_queue_id, in uint32 n_hvx, in uint32 use_hmx, in uint64 max_vmem);
+    AEEResult start(in uint32 sess_id, in uint64 dsp_queue_id, in uint32 n_hvx, in uint32 n_hmx, in uint64 max_vmem);
     AEEResult stop();
     AEEResult mmap(in uint32 fd, in uint32 size);
     AEEResult munmap(in uint32 fd);
     AEEResult profiler(in uint32 mode, in htp_iface_pmu_conf pmu);
     AEEResult etm(in uint32 enable);
+    AEEResult hwinfo(rout uint32 n_threads, rout uint32 n_hvx, rout uint32 n_hmx, rout uint64 vtcm_size);
 };
 
 #endif /* HTP_IDL */

ggml/src/ggml-hexagon/htp/hvx-base.h

@@ -170,25 +170,7 @@ static inline HVX_VectorPair hvx_vec_f16_to_f32(HVX_Vector v) {
 }
 #endif
 
-/* Q6_Vsf_equals_Vw is only available on v73+.*/
-#if __HVX_ARCH__ < 73
-static inline HVX_Vector hvx_vec_i32_to_qf32(HVX_Vector const in)
-{
-    HVX_Vector const vzero = Q6_V_vzero();
-    HVX_VectorPred is_zero = Q6_Q_vcmp_eq_VwVw(in, vzero);
-    HVX_Vector lshift = Q6_Vw_vnormamt_Vw(in);
-    HVX_Vector normalized = Q6_Vw_vasl_VwVw(in, lshift);
-    HVX_Vector vexp = Q6_Vw_vsub_VwVw(Q6_V_vsplat_R(0x7f + 30), lshift);
-    HVX_Vector mant = Q6_V_vand_VV(Q6_V_vsplat_R(0xFFFFFF00), normalized);
-    HVX_Vector ret = Q6_V_vmux_QVV(is_zero, vzero, Q6_Vw_vadd_VwVw(mant, vexp));
-    return ret;
-}
 
-static inline HVX_Vector Q6_Vsf_equals_Vw(HVX_Vector const in)
-{
-    return Q6_Vsf_equals_Vqf32(hvx_vec_i32_to_qf32(in));
-}
-#endif
 
 static inline HVX_Vector hvx_vec_i16_from_hf_rnd_sat(HVX_Vector vin) {
     // This looks complicated.
@@ -305,4 +287,17 @@ static inline HVX_Vector hvx_vec_mul_f32_f32(HVX_Vector a, HVX_Vector b) {
 
 #endif // __HVX_ARCH__ < 79
 
+static inline HVX_Vector hvx_vec_load_act_tile(const uint8_t * y_q, uint32_t kt, HVX_Vector * v_act_all) {
+    if (kt % 4 == 0) {
+        *v_act_all = hvx_vmem(y_q + kt * 32);
+        return *v_act_all;
+    } else if (kt % 4 == 1) {
+        return Q6_V_vror_VR(*v_act_all, 32);
+    } else if (kt % 4 == 2) {
+        return Q6_V_vror_VR(*v_act_all, 64);
+    } else {
+        return Q6_V_vror_VR(*v_act_all, 96);
+    }
+}
+
 #endif /* HVX_BASE_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -361,7 +361,7 @@ static void vtcm_free(struct htp_context * ctx) {
 static void htp_packet_callback(dspqueue_t queue, int error, void * context);
 static void htp_error_callback(dspqueue_t queue, int error, void * context);
 
-AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_queue_id, uint32 n_hvx, uint32 use_hmx, uint64_t max_vmem) {
+AEEResult htp_iface_start(remote_handle64 handle, uint32_t sess_id, uint64_t dsp_queue_id, uint32_t n_hvx, uint32_t n_hmx, uint64_t max_vmem) {
     struct htp_context * ctx = (struct htp_context *) handle;
 
     if (!ctx) {
@@ -395,10 +395,9 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
         return AEE_ENOMEMORY;
     }
 
-#ifdef HTP_HAS_HMX
-    ctx->hmx_enabled = use_hmx;
+    ctx->hmx_enabled = n_hmx;
     ctx->hmx_queue   = NULL;
-    if (use_hmx) {
+    if (n_hmx) {
         ctx->hmx_queue = hmx_queue_create(16, ctx->vtcm_rctx);
         if (ctx->hmx_queue) {
             ctx->hmx_queue->trace = &ctx->trace[HTP_MAX_NTHREADS];
@@ -407,8 +406,7 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
             ctx->hmx_enabled = false;
         }
     }
-    FARF(HIGH, "HMX %s (use_hmx=%d)", ctx->hmx_enabled ? "enabled" : "disabled", use_hmx);
-#endif
+    FARF(HIGH, "HMX %s (n_hmx=%d)", ctx->hmx_enabled ? "enabled" : "disabled", n_hmx);
 
     qurt_sysenv_max_hthreads_t hw_threads;
     qurt_sysenv_get_max_hw_threads(&hw_threads);
@@ -481,13 +479,11 @@ AEEResult htp_iface_stop(remote_handle64 handle) {
         dma_queue_delete(ctx->dma[i]);
     }
 
-#ifdef HTP_HAS_HMX
     if (ctx->hmx_queue) {
         hmx_queue_delete(ctx->hmx_queue);
         ctx->hmx_queue = NULL;
     }
     ctx->hmx_enabled = false;
-#endif
 
     vtcm_free(ctx);
 
@@ -500,6 +496,36 @@ AEEResult htp_iface_stop(remote_handle64 handle) {
     return AEE_SUCCESS;
 }
 
+AEEResult htp_iface_hwinfo(remote_handle64 handle, uint32_t * n_threads, uint32_t * n_hvx, uint32_t * n_hmx, uint64_t * vtcm_size) {
+    (void)handle;
+    if (!n_threads || !n_hvx || !n_hmx || !vtcm_size) {
+        return AEE_EBADPARM;
+    }
+
+    qurt_sysenv_max_hthreads_t hw_threads;
+    qurt_sysenv_get_max_hw_threads(&hw_threads);
+    uint32_t hw_nhvx = (qurt_hvx_get_units() >> 8) & 0xFF;
+
+    uint32_t n_hvx_val = hw_nhvx;
+    if (n_hvx_val > hw_threads.max_hthreads) {
+        n_hvx_val = hw_threads.max_hthreads;
+    }
+    if (n_hvx_val > HTP_MAX_NTHREADS) {
+        n_hvx_val = HTP_MAX_NTHREADS;
+    }
+
+    // for now we force n_threads == n_hvx
+    *n_threads = n_hvx_val;
+    *n_hvx     = n_hvx_val;
+    *n_hmx     = 1;
+
+    uint32_t vtcm_sz = 8 * 1024 * 1024; // 8MB default fallback
+    HAP_compute_res_query_VTCM(0, (unsigned int *)&vtcm_sz, NULL, NULL, NULL);
+    *vtcm_size = vtcm_sz;
+
+    return AEE_SUCCESS;
+}
+
 static void htp_error_callback(dspqueue_t queue, int error, void * context) {
     // No errors expected on the DSP.
     FARF(ERROR, "Error callback: 0x%08x", (unsigned) error);
@@ -554,6 +580,12 @@ static int execute_op(struct htp_ops_context * octx) {
         case HTP_OP_MUL_MAT_ID:
             return op_matmul_id(octx);
 
+        case HTP_OP_MUL_MAT_QKV:
+            return op_matmul_qkv(octx);
+
+        case HTP_OP_MUL_MAT_FFN:
+            return op_matmul_ffn(octx);
+
         case HTP_OP_MUL:
         case HTP_OP_ADD:
         case HTP_OP_SUB:
@@ -762,8 +794,9 @@ static void prep_tensors(struct htp_context *ctx, struct htp_buf_desc *bufs, str
     }
 }
 
-static void proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens, uint32_t idx, struct htp_op_desc * op) {
+static int proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens, uint32_t idx, struct htp_op_desc * op) {
     memcpy(octx->op_params, op->params, sizeof(octx->op_params));
+    memcpy(octx->kernel_params, op->kernel_params, sizeof(octx->kernel_params));
     octx->flags = op->flags;
     octx->op    = op->opcode;
 
@@ -785,22 +818,41 @@ static void proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens,
             src->ne[0], src->ne[1], src->ne[3], src->ne[3]);
     }
 
-    // Prep output tensor
-    struct htp_tensor *dst = tens + op->dst;
+    // Prep output tensors
+    for (uint32_t i = 0; i < HTP_OP_MAX_OUTPUTS; i++) {
+        uint16_t dst_idx = op->dst[i];
+        if (dst_idx == 0xffff) {
+            octx->dsts[i] = NULL;
+            continue;
+        }
+        struct htp_tensor *dst = tens + dst_idx;
+        octx->dsts[i] = dst;
 
-    octx->dst = dst;
+        FARF(HIGH, "prep-dst[%u] #%u: data %p size %u : %u:%u:%u:%u", i, dst_idx, (void*) dst->data, dst->size,
+            dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
+    }
 
-    FARF(HIGH, "prep-dst #%u: data %p size %u : %u:%u:%u:%u", op->dst, (void*) dst->data, dst->size,
-        dst->ne[0], dst->ne[1], dst->ne[3], dst->ne[3]);
+    int status = execute_op(octx);
 
-    (void) execute_op(octx);
+    octx->src0_spad.src = NULL;
+    octx->src1_spad.src = NULL;
+    octx->src2_spad.src = NULL;
+    octx->src3_spad.src = NULL;
+    octx->dst_spad.src  = NULL;
 
     // flush buffers on output
-    hex_l2flush((void *) dst->data, dst->size);
-    dst->flags |= HTP_TENSOR_FLUSHED;
+    for (uint32_t i = 0; i < HTP_OP_MAX_OUTPUTS; i++) {
+        if (octx->dsts[i]) {
+            struct htp_tensor *dst = (struct htp_tensor *)octx->dsts[i];
+            hex_l2flush((void *) dst->data, dst->size);
+            dst->flags |= HTP_TENSOR_FLUSHED;
 
-    FARF(HIGH, "post-dst #%u: data %p size %u : %u:%u:%u:%u", op->dst, (void*) dst->data, dst->size,
-        dst->ne[0], dst->ne[1], dst->ne[3], dst->ne[3]);
+            FARF(HIGH, "post-dst[%u] #%u: data %p size %u : %u:%u:%u:%u", i, op->dst[i], (void*) dst->data, dst->size,
+                dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
+        }
+    }
+
+    return status;
 }
 
 #define DSPQUEUE_POLL_TIMEOUT_USEC 100
@@ -892,20 +944,26 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
             }
         }
 
+        int      op_status = HTP_STATUS_OK;
+        uint32_t op_wakeup = n_ops / 2; // half-way throgh the batch
+
         for (uint32_t i=0; i < n_ops; i++) {
             struct profile_data prof;
 
-            if (i == (n_ops-1)) {
-                // wake up the host before starting the last op
+            if (i == op_wakeup) {
                 dspqueue_write_early_wakeup_noblock(queue, 0, 0);
             }
 
             profile_start(ctx->profiler, &prof);
 
-            proc_op_req(octx, tens, i, &ops[i]);
+            op_status = proc_op_req(octx, tens, i, &ops[i]);
 
             profile_stop(ctx->profiler, &prof);
 
+            if (op_status != HTP_STATUS_OK) {
+                break;
+            }
+
             if (ctx->profiler) {
                 pds[i].opcode = ops[i].opcode;
                 pds[i].usecs  = prof.usecs;
@@ -919,7 +977,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
 
         struct htp_opbatch_rsp rsp;
         rsp.id        = req.id;
-        rsp.status    = HTP_STATUS_OK;
+        rsp.status    = op_status;
         rsp.n_bufs    = n_bufs;
         rsp.n_tensors = n_tens;
         rsp.n_ops     = n_ops;

ggml/src/ggml-hexagon/htp/matmul-ops.h

@@ -0,0 +1,508 @@
+#ifndef HTP_MATMUL_OPS_H
+#define HTP_MATMUL_OPS_H
+
+#include <stdint.h>
+#include <stddef.h>
+#include "htp-ops.h"
+#include "hex-fastdiv.h"
+#include "hex-common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// --- HMX Tile Constraints ---
+#define HTP_MM_HMX_TILE_N_COLS 32
+#define HTP_MM_HMX_TILE_N_ROWS 32
+#define HTP_MM_HMX_TILE_SIZE   (32 * 32 * sizeof(__fp16)) // 2048 bytes
+#define HTP_MM_HMX_TILE_N_ELMS 1024
+#define HTP_MM_HMX_MIN_NROWS   4
+
+// --- Weight Repacked Tile Sizes ---
+#define HTP_MM_WEIGHT_TILE_SIZE_Q4_0   576
+#define HTP_MM_WEIGHT_TILE_SIZE_Q4_1   640
+#define HTP_MM_WEIGHT_TILE_SIZE_Q8_0   1088
+#define HTP_MM_WEIGHT_TILE_SIZE_IQ4_NL 576
+#define HTP_MM_WEIGHT_TILE_SIZE_MXFP4  544
+
+// --- Weight Repacked Aligned Tile Sizes ---
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q4_0   640
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q4_1   640
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q8_0   1152
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_IQ4_NL 640
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_MXFP4  640
+
+// --- Activation Tiled Block Sizes (including padding) ---
+#define HTP_MM_ACT_TILE_SIZE_Q8_0      1152
+#define HTP_MM_ACT_TILE_SIZE_Q8_1      1280
+
+#define HTP_MM_MAX_PREFETCH 16
+
+// --- Solver Cost Model Penalty Weights (HMX-specific) ---
+#define HTP_MM_HMX_COST_W_DEQUANT 3 // cost penalty for quantized weight loading/dequantization
+#define HTP_MM_HMX_COST_A_CONVERT 2 // cost penalty for activation loading/conversion
+
+// --- DMA Activation Transfer Configuration ---
+#define HTP_MM_DMA_ACT_ROWS_PER_STEP 2
+#define HTP_MM_DMA_ACT_MULTIPLIER    4
+
+enum htp_mm_kernel_type {
+    HTP_MM_KERNEL_UNSUPPORTED = 0,
+
+    // HMX paths
+    HTP_MM_KERNEL_HMX_2D,
+    HTP_MM_KERNEL_HMX_F16_BATCHED,
+
+    // HVX floating-point paths
+    HTP_MM_KERNEL_HVX_F16_F16_VTCM,
+    HTP_MM_KERNEL_HVX_F16_F16_DDR,
+    HTP_MM_KERNEL_HVX_F16_F32_DDR,
+
+    HTP_MM_KERNEL_HVX_F32_F32_VTCM,
+    HTP_MM_KERNEL_HVX_F32_F32_DDR,
+    HTP_MM_KERNEL_HVX_F32_F16_DDR,
+
+    // HVX quantized paths
+    HTP_MM_KERNEL_HVX_QUANT_ROW,      // standard row-wise parallel quantization
+    HTP_MM_KERNEL_HVX_QUANT_BLOCK,    // parallel block-wise quantization
+    HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT, // row-wise fallback flat quantization
+};
+
+// Op-specific struct for precomputed matmul params
+struct htp_mm_kernel_params {
+    int32_t  kernel_type;        // enum htp_mm_kernel_type
+    int32_t  pipeline;           // 1 = pipelined execution, 0 = standard
+    int32_t  m_chunk;            // Row chunk size (M chunk)
+    int32_t  n_chunk;            // Col chunk size (N chunk)
+    int32_t  n_threads;          // Number of threads to spawn
+    int32_t  n_act_threads;      // Number of threads for activation preparation
+    int32_t  n_hmx;              // 1 = use HMX, 0 = use HVX
+    int32_t  n_prefetch;         // Prefetch lookahead buffers/rows in VTCM
+    int32_t  tile_size;          // Weight tile size
+    int32_t  aligned_tile_size;  // Aligned weight tile size (padded to 128)
+    int32_t  src1_row_size;      // Row size for quantized activation
+    int32_t  vtcm_size;          // Total required scratchpad size in VTCM
+    int32_t  vtcm_src0_size;     // src0 scratchpad size in VTCM
+    int32_t  vtcm_src1_size;     // src1 scratchpad size in VTCM
+    int32_t  vtcm_src2_size;     // src2 scratchpad size in VTCM (fused only)
+    int32_t  vtcm_src3_size;     // src3 scratchpad size in VTCM (fused only)
+    int32_t  vtcm_dst_size;      // dst scratchpad size in VTCM
+
+    // Precomputed division values
+    struct fastdiv_values div_ne12_ne1;
+    struct fastdiv_values div_ne1;
+    struct fastdiv_values div_r2;
+    struct fastdiv_values div_r3;
+    struct fastdiv_values div_ne11;
+};
+
+#if defined(__cplusplus)
+static_assert(sizeof(struct htp_mm_kernel_params) <= 128, "htp_matmul_kernel_params is too large for kernel_params blob");
+#else
+_Static_assert(sizeof(struct htp_mm_kernel_params) <= 128, "htp_matmul_kernel_params is too large for kernel_params blob");
+#endif
+
+struct mmid_row_mapping {
+    uint32_t i1;
+    uint32_t i2;
+};
+
+// Search for optimal (mc, nc) chunk sizes within VTCM budget.
+static inline int htp_mm_hmx_compute_chunks(size_t   vtcm_total,
+                              size_t   overhead,
+                              size_t   per_n_cost,
+                              size_t   per_m_cost,
+                              size_t   per_mn_cost,
+                              size_t   m,
+                              size_t   n,
+                              size_t   m_block_cost,
+                              size_t   n_block_cost,
+                              size_t * m_chunk_out,
+                              size_t * n_chunk_out,
+                              size_t * total_out) {
+    if (m == 0 || n == 0) return -1;
+    if (vtcm_total <= overhead) return -1;
+    if (per_n_cost == 0 || per_m_cost == 0 || per_mn_cost == 0) return -1;
+
+    const size_t usable = vtcm_total - overhead;
+
+    size_t best_cost = SIZE_MAX;
+    size_t best_mn   = 0;
+    size_t best_m = 0, best_n = 0;
+
+    const size_t n_max = hex_align_down((size_t)n, HTP_MM_HMX_TILE_N_COLS);
+    for (size_t nc = n_max; nc >= HTP_MM_HMX_TILE_N_COLS; nc -= HTP_MM_HMX_TILE_N_COLS) {
+        size_t n_fixed = 0, ncmn = 0, mc_denom = 0;
+        if (hex_mul_overflow(nc, per_n_cost, &n_fixed)) continue;
+        if (n_fixed >= usable) goto next_nc;
+
+        if (hex_mul_overflow(nc, per_mn_cost, &ncmn)) goto next_nc;
+        if (hex_add_overflow(per_m_cost, ncmn, &mc_denom) || mc_denom == 0) goto next_nc;
+
+        {
+            size_t remain = usable - n_fixed;
+            size_t mc = remain / mc_denom;
+            mc = hex_align_down(mc, HTP_MM_HMX_TILE_N_ROWS);
+            mc = hex_smin(mc, m);
+
+            if (mc == 0) {
+                goto next_nc;
+            }
+
+            size_t mblocks = ((size_t) m + mc - 1) / mc;
+            size_t nblocks = ((size_t) n + nc - 1) / nc;
+            size_t cost    = mblocks * m_block_cost + nblocks * n_block_cost;
+            size_t mn      = mc * nc;
+            if (cost < best_cost || (cost == best_cost && mn > best_mn)) {
+                best_cost = cost;
+                best_mn   = mn;
+                best_m    = mc;
+                best_n    = nc;
+            }
+        }
+
+next_nc:
+        if (nc == HTP_MM_HMX_TILE_N_COLS) break;  // avoid size_t underflow
+    }
+
+    if (best_m == 0 || best_n == 0) return -1;
+
+    // Compute exact total (with overflow checks)
+    size_t t0 = 0, t1 = 0, t2 = 0, mn = 0, total = 0;
+    if (hex_mul_overflow(best_n, per_n_cost, &t0)) return -1;
+    if (hex_mul_overflow(best_m, per_m_cost, &t1)) return -1;
+    if (hex_mul_overflow(best_m, best_n, &mn))     return -1;
+    if (hex_mul_overflow(mn, per_mn_cost, &t2))    return -1;
+    if (hex_add_overflow(t0, t1, &total))          return -1;
+    if (hex_add_overflow(total, t2, &total))       return -1;
+    if (hex_add_overflow(total, overhead, &total)) return -1;
+
+    *m_chunk_out = best_m;
+    *n_chunk_out = best_n;
+    *total_out   = total;
+    return 0;
+}
+
+// --- Tile Size Helpers ---
+static inline uint32_t htp_mm_get_weight_tile_size(int weight_type) {
+    switch (weight_type) {
+        case HTP_TYPE_Q4_0:
+        case HTP_TYPE_IQ4_NL:
+            return HTP_MM_WEIGHT_TILE_SIZE_Q4_0;
+        case HTP_TYPE_Q4_1:
+            return HTP_MM_WEIGHT_TILE_SIZE_Q4_1;
+        case HTP_TYPE_Q8_0:
+            return HTP_MM_WEIGHT_TILE_SIZE_Q8_0;
+        case HTP_TYPE_MXFP4:
+            return HTP_MM_WEIGHT_TILE_SIZE_MXFP4;
+        default:
+            return 0;
+    }
+}
+
+static inline uint32_t htp_mm_get_weight_aligned_tile_size(int weight_type) {
+    switch (weight_type) {
+        case HTP_TYPE_Q4_0:
+        case HTP_TYPE_IQ4_NL:
+            return HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q4_0;
+        case HTP_TYPE_Q4_1:
+            return HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q4_1;
+        case HTP_TYPE_Q8_0:
+            return HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q8_0;
+        case HTP_TYPE_MXFP4:
+            return HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_MXFP4;
+        default:
+            return 0;
+    }
+}
+
+// --- Activation/Row Size Helpers ---
+static inline size_t htp_mm_q8_0_tiled_row_size(uint32_t ne) {
+    const uint32_t ne_padded = ((ne + 127) / 128) * 128;
+    const uint32_t nb_32 = ne_padded / 32;
+    return nb_32 * HTP_MM_ACT_TILE_SIZE_Q8_0;
+}
+
+static inline size_t htp_mm_q8_1_tiled_row_size(uint32_t ne) {
+    const uint32_t ne_padded = ((ne + 127) / 128) * 128;
+    const uint32_t nb_32 = ne_padded / 32;
+    return nb_32 * HTP_MM_ACT_TILE_SIZE_Q8_1;
+}
+
+static inline size_t htp_mm_q8_0_flat_row_size(uint32_t ne) {
+    const uint32_t quants_size = hex_align_up(ne, 128);
+    const uint32_t num_scales = (ne + 31) / 32;
+    const uint32_t scales_size = hex_align_up(num_scales * 2, 128);
+    return quants_size + scales_size;
+}
+
+static inline size_t htp_mm_q8_1_flat_row_size(uint32_t ne) {
+    const uint32_t quants_size = hex_align_up(ne, 128);
+    const uint32_t num_scales = (ne + 31) / 32;
+    const uint32_t scales_size = hex_align_up(num_scales * 4, 128);
+    return quants_size + scales_size;
+}
+
+static inline size_t htp_mm_get_tiled_row_stride(int weight_type, uint32_t k) {
+    uint32_t nb = (k + QK_Q4_0_TILED - 1) / QK_Q4_0_TILED;
+    switch (weight_type) {
+        case HTP_TYPE_Q4_0:
+        case HTP_TYPE_IQ4_NL:
+        case HTP_TYPE_Q4_1:
+        case HTP_TYPE_Q8_0:
+        case HTP_TYPE_MXFP4:
+            return (size_t) nb * htp_mm_get_weight_tile_size(weight_type);
+        case HTP_TYPE_F16:
+            return (size_t) k * sizeof(__fp16);
+        case HTP_TYPE_F32:
+            return (size_t) k * sizeof(float);
+        default:
+            return 0;
+    }
+}
+
+static inline size_t htp_mm_round_up(size_t n, size_t m) {
+    return ((n + m - 1) / m) * m;
+}
+
+static inline bool htp_mm_hmx_pipeline(uint32_t m) {
+    return m > 32;
+}
+
+static inline void htp_mm_hmx_get_2d_chunk_costs(
+    int wtype, uint32_t k, bool pipeline, uint32_t aligned_tile_size,
+    size_t * size_per_n_out, size_t * size_per_m_out, size_t * size_per_mn_out
+) {
+    const bool is_quant = (wtype != HTP_TYPE_F16 && wtype != HTP_TYPE_F32);
+    const size_t row_stride = htp_mm_get_tiled_row_stride(wtype, k);
+    const size_t vec_dot_size = k * sizeof(uint16_t);
+    const uint32_t n_k_tiles = k / HTP_MM_HMX_TILE_N_COLS;
+    const size_t qweight_row_stride = is_quant ? (size_t)(n_k_tiles * aligned_tile_size) / 32 : 0;
+
+    *size_per_n_out = (pipeline ? 2 : 1) * (is_quant ? qweight_row_stride : row_stride) +
+                      (pipeline ? 2 * vec_dot_size : vec_dot_size);
+    *size_per_m_out = vec_dot_size;
+    *size_per_mn_out = (pipeline ? 2 : 1) * sizeof(uint16_t);
+}
+
+static inline void htp_mm_hmx_get_batched_chunk_costs(
+    uint32_t k, uint32_t group_size,
+    size_t * size_per_n_out, size_t * size_per_m_out, size_t * size_per_mn_out
+) {
+    const size_t vec_dot_size = k * sizeof(uint16_t);
+    *size_per_n_out = 3 * vec_dot_size;
+    *size_per_m_out = group_size * vec_dot_size;
+    *size_per_mn_out = sizeof(uint16_t);
+}
+
+static inline size_t htp_mm_hmx_get_2d_vtcm_size(
+    int wtype, uint32_t k, size_t mc, size_t nc, bool pipeline, uint32_t act_threads, uint32_t aligned_tile_size
+) {
+    const uint32_t n_k_tiles = k / HTP_MM_HMX_TILE_N_COLS;
+    const bool is_quant = (wtype != HTP_TYPE_F16 && wtype != HTP_TYPE_F32);
+    const size_t row_stride = htp_mm_get_tiled_row_stride(wtype, k);
+    const size_t vec_dot_size = k * sizeof(uint16_t);
+
+    const size_t act_f32_size = htp_mm_round_up(act_threads * 4 * k * sizeof(float), HTP_MM_HMX_TILE_SIZE);
+    size_t weight_area_size = is_quant
+        ? htp_mm_round_up((nc / 32) * n_k_tiles * aligned_tile_size, HTP_MM_HMX_TILE_SIZE)
+        : htp_mm_round_up(nc * row_stride, HTP_MM_HMX_TILE_SIZE);
+    if (pipeline) {
+        weight_area_size *= 2;
+    }
+    const size_t act_area_size    = htp_mm_round_up(mc * vec_dot_size, HTP_MM_HMX_TILE_SIZE);
+    const size_t output_area_size = htp_mm_round_up(mc * nc * sizeof(uint16_t), HTP_MM_HMX_TILE_SIZE);
+
+    size_t scratch0_size = htp_mm_round_up(nc * vec_dot_size, HTP_MM_HMX_TILE_SIZE);
+    size_t scratch1_size = pipeline ? scratch0_size : 0;
+    size_t scratch2_size = pipeline ? output_area_size : 0;
+
+    return weight_area_size + act_area_size + act_f32_size + output_area_size +
+           scratch0_size + scratch1_size + scratch2_size + 256;
+}
+
+static inline size_t htp_mm_hmx_get_batched_vtcm_size(
+    int wtype, uint32_t k, size_t mc, size_t nc, uint32_t group_size, bool use_dma_activation, bool pipeline, uint32_t act_threads) {
+    (void)wtype;
+    (void)pipeline;
+    const size_t vec_dot_size     = k * sizeof(uint16_t);
+    const size_t f32_scratch_size = use_dma_activation
+        ? htp_mm_round_up(act_threads * 4 * k * sizeof(float), HTP_MM_HMX_TILE_SIZE) : 0;
+
+    const size_t act_head_stride   = mc * k;
+    const size_t weight_area_size  = htp_mm_round_up(nc * vec_dot_size, HTP_MM_HMX_TILE_SIZE);
+    const size_t act_area_size     = htp_mm_round_up(group_size * act_head_stride * sizeof(uint16_t), HTP_MM_HMX_TILE_SIZE);
+    const size_t output_area_size  = htp_mm_round_up(group_size * mc * nc * sizeof(uint16_t), HTP_MM_HMX_TILE_SIZE);
+    const size_t scratch_area_size = htp_mm_round_up(nc * vec_dot_size, HTP_MM_HMX_TILE_SIZE);
+
+    return weight_area_size + act_area_size + output_area_size +
+           2 * scratch_area_size + 256 + f32_scratch_size;
+}
+
+static inline size_t htp_mm_hvx_get_vtcm_sizes(
+    int kernel_type,
+    int wtype,
+    uint32_t ne10,       // k
+    uint32_t src1_nrows, // m_total (or act_nrows)
+    uint32_t n_threads,
+    size_t dst_row_size,
+    size_t src0_row_size,
+    size_t src1_row_size,
+    uint32_t n_prefetch,
+    size_t * vtcm_src0_size_out,
+    size_t * vtcm_src1_size_out,
+    size_t * vtcm_dst_size_out
+) {
+    size_t vtcm_src0_size = 0;
+    size_t vtcm_src1_size = 0;
+    size_t vtcm_dst_size  = 0;
+
+    const bool is_repack = (wtype == HTP_TYPE_Q4_0 || wtype == HTP_TYPE_Q4_1 ||
+                            wtype == HTP_TYPE_Q8_0 || wtype == HTP_TYPE_IQ4_NL ||
+                            wtype == HTP_TYPE_MXFP4);
+
+    const size_t src0_row_size_padded = htp_mm_round_up(src0_row_size, 128);
+    const size_t dst_nrows = (src1_nrows > 1) ? 0 : 1;
+
+    switch (kernel_type) {
+        case HTP_MM_KERNEL_HVX_F16_F16_VTCM: {
+            size_t f16_src1_row_size = htp_mm_round_up(ne10 * 2, 128);
+            vtcm_src1_size = htp_mm_round_up(f16_src1_row_size * src1_nrows, 256);
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256) * n_threads;
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) * n_threads : 0;
+            break;
+        }
+        case HTP_MM_KERNEL_HVX_F16_F32_DDR:
+        case HTP_MM_KERNEL_HVX_F16_F16_DDR:
+        case HTP_MM_KERNEL_HVX_F32_F32_DDR:
+        case HTP_MM_KERNEL_HVX_F32_F16_DDR: {
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size, 256) * n_threads;
+            vtcm_src1_size = htp_mm_round_up(n_prefetch * src1_row_size, 256) * n_threads;
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) * n_threads : 0;
+            break;
+        }
+        case HTP_MM_KERNEL_HVX_F32_F32_VTCM: {
+            size_t f32_src1_row_size = htp_mm_round_up(ne10 * 4, 128);
+            vtcm_src1_size = htp_mm_round_up(f32_src1_row_size * src1_nrows, 256);
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256) * n_threads;
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) * n_threads : 0;
+            break;
+        }
+        case HTP_MM_KERNEL_HVX_QUANT_BLOCK:
+        case HTP_MM_KERNEL_HVX_QUANT_ROW: {
+            size_t q_src1_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_tiled_row_size(ne10) : htp_mm_q8_0_tiled_row_size(ne10);
+
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) : 0;
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256);
+            vtcm_src1_size = htp_mm_round_up(q_src1_row_size * src1_nrows, 256);
+
+            // src0 spad is also used in dynamic quantizer to store padded src1 rows
+            size_t src1_row_size_padded = htp_mm_round_up(q_src1_row_size, QK_Q8_0_TILED * sizeof(float));
+            if (vtcm_src0_size < src1_row_size_padded) {
+                vtcm_src0_size = src1_row_size_padded;
+            }
+
+            vtcm_src0_size = vtcm_src0_size * n_threads;
+            vtcm_dst_size  = vtcm_dst_size * n_threads;
+
+            if (is_repack) {
+                uint32_t aligned_tile_size = htp_mm_get_weight_aligned_tile_size(wtype);
+                uint32_t n_k_tiles = ne10 / 32;
+                uint32_t tile_row_size = n_k_tiles * aligned_tile_size;
+                size_t repacked_vtcm_size = htp_mm_round_up(n_prefetch * tile_row_size, 256);
+                if (repacked_vtcm_size < src1_row_size_padded) {
+                    repacked_vtcm_size = src1_row_size_padded;
+                }
+                vtcm_src0_size = repacked_vtcm_size * n_threads;
+            }
+            break;
+        }
+        case HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT: {
+            size_t q_src1_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_flat_row_size(ne10) : htp_mm_q8_0_flat_row_size(ne10);
+
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) : 0;
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256);
+            vtcm_src1_size = htp_mm_round_up(q_src1_row_size * src1_nrows, 256);
+
+            size_t src1_row_size_padded = htp_mm_round_up(q_src1_row_size, 256);
+            if (vtcm_src0_size < src1_row_size_padded) {
+                vtcm_src0_size = src1_row_size_padded;
+            }
+
+            vtcm_src0_size = vtcm_src0_size * n_threads;
+            vtcm_dst_size  = vtcm_dst_size * n_threads;
+
+            if (is_repack) {
+                uint32_t aligned_tile_size = htp_mm_get_weight_aligned_tile_size(wtype);
+                uint32_t n_k_tiles = ne10 / 32;
+                uint32_t tile_row_size = n_k_tiles * aligned_tile_size;
+                size_t repacked_vtcm_size = htp_mm_round_up(n_prefetch * tile_row_size, 256);
+                if (repacked_vtcm_size < src1_row_size_padded) {
+                    repacked_vtcm_size = src1_row_size_padded;
+                }
+                vtcm_src0_size = repacked_vtcm_size * n_threads;
+            }
+            break;
+        }
+        default:
+            break;
+    }
+
+    *vtcm_src0_size_out = vtcm_src0_size;
+    *vtcm_src1_size_out = vtcm_src1_size;
+    *vtcm_dst_size_out  = vtcm_dst_size;
+
+    return vtcm_src0_size + vtcm_src1_size + vtcm_dst_size;
+}
+
+static inline size_t htp_mm_hvx_id_get_vtcm_sizes(
+    int wtype,
+    uint32_t ne10,                // k
+    uint32_t src1_nrows,
+    uint32_t n_threads,
+    size_t src0_row_size,    // nb01
+    uint32_t n_prefetch,
+    size_t * vtcm_src0_size_out,
+    size_t * vtcm_src1_size_out
+) {
+    const bool is_repack = (wtype == HTP_TYPE_Q4_0 || wtype == HTP_TYPE_Q4_1 ||
+                            wtype == HTP_TYPE_Q8_0 || wtype == HTP_TYPE_IQ4_NL ||
+                            wtype == HTP_TYPE_MXFP4);
+
+    const size_t src0_row_size_padded = htp_mm_round_up(src0_row_size, 128);
+    const size_t src1_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_tiled_row_size(ne10)
+                                                          : htp_mm_q8_0_tiled_row_size(ne10);
+
+    size_t src0_sz_per_thread = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256);
+    size_t src1_sz            = htp_mm_round_up(src1_row_size * src1_nrows, 256);
+
+    // src0 spad also holds temporary transposed src1 columns during dynamic quantization.
+    const size_t src1_row_size_padded = htp_mm_round_up(src1_row_size, QK_Q8_0_TILED * sizeof(float));
+    if (src0_sz_per_thread < src1_row_size_padded) {
+        src0_sz_per_thread = src1_row_size_padded;
+    }
+
+    if (is_repack) {
+        const uint32_t aligned_tile_size = htp_mm_get_weight_aligned_tile_size(wtype);
+        const uint32_t n_k_tiles    = ne10 / 32;
+        const uint32_t tile_row_size = n_k_tiles * aligned_tile_size;
+        size_t repacked_vtcm_size = htp_mm_round_up(n_prefetch * tile_row_size, 256);
+        if (repacked_vtcm_size < src1_row_size_padded) {
+            repacked_vtcm_size = src1_row_size_padded;
+        }
+        src0_sz_per_thread = repacked_vtcm_size;
+    }
+
+    const size_t vtcm_src0_size = src0_sz_per_thread * n_threads;
+
+    *vtcm_src0_size_out = vtcm_src0_size;
+    *vtcm_src1_size_out = src1_sz;
+
+    return vtcm_src0_size + src1_sz;
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // HTP_MATMUL_OPS_H

ggml/src/ggml-hexagon/libggml-htp.inf

@@ -14,8 +14,6 @@ Drivers_Dir = 13
 1 = %DiskId%
 
 [SourceDisksFiles]
-libggml-htp-v68.so = 1
-libggml-htp-v69.so = 1
 libggml-htp-v73.so = 1
 libggml-htp-v75.so = 1
 libggml-htp-v79.so = 1
@@ -28,8 +26,6 @@ ExcludeFromSelect = *
 CopyFiles=Drivers_Dir
 
 [Drivers_Dir]
-libggml-htp-v68.so,,,0x10 ;COPYFLG_NO_OVERWRITE
-libggml-htp-v69.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v73.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v75.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v79.so,,,0x10 ;COPYFLG_NO_OVERWRITE

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

@@ -10152,14 +10152,8 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
 
-    const int ne00 = src0 ? src0->ne[0] : 0;
-    const int ne01 = src0 ? src0->ne[1] : 0;
-    const int ne02 = src0 ? src0->ne[2] : 0;
-    const int ne03 = src0 ? src0->ne[3] : 0;
-
-    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
-    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
-    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);
 
     const int nth = MIN(64, ne00);
 
@@ -10173,11 +10167,12 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
     CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),       &ne01));
     CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),       &ne02));
     CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),       &ne03));
-    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),  &nb01));
-    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),  &nb02));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),  &nb03));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float),     &eps));
-    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth, NULL));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),  &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),  &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),     &eps));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float)*nth, NULL));
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};

ggml/src/ggml-opencl/kernels/norm.cl

@@ -24,6 +24,7 @@ kernel void kernel_norm(
         int ne01,
         int ne02,
         int ne03,
+        ulong nb00,
         ulong nb01,
         ulong nb02,
         ulong nb03,
@@ -43,7 +44,8 @@ kernel void kernel_norm(
     // parallel sum
     sum[get_local_id(0)] = 0.0f;
     for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
-        sum[get_local_id(0)] += x[i00];
+        // this kernel handles float, nb00/4 translates byte offset to element offset
+        sum[get_local_id(0)] += x[i00*nb00/4];
     }
     // reduce
     barrier(CLK_LOCAL_MEM_FENCE);
@@ -60,7 +62,8 @@ kernel void kernel_norm(
     global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
     sum[get_local_id(0)] = 0.0f;
     for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
-        y[i00] = x[i00] - mean;
+        // this kernel handles float, nb00/4 translates byte offset to element offset
+        y[i00] = x[i00*nb00/4] - mean;
         sum[get_local_id(0)] += y[i00] * y[i00];
     }
 

ggml/src/ggml-sycl/conv3d.cpp

@@ -103,8 +103,8 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     // allocate packed arrays: A_packed (k x m), B_packed (k x n)
     ggml_sycl_pool_alloc<float> A_packed_alloc(ctx.pool());
     ggml_sycl_pool_alloc<float> B_packed_alloc(ctx.pool());
-    A_packed_alloc.alloc((size_t) knl_n_total * patch_total * sizeof(float));
-    B_packed_alloc.alloc((size_t) knl_n_total * oc * sizeof(float));
+    A_packed_alloc.alloc((size_t) knl_n_total * patch_total);
+    B_packed_alloc.alloc((size_t) knl_n_total * oc);
 
     float * A_packed = A_packed_alloc.get();
     float * B_packed = B_packed_alloc.get();
@@ -115,10 +115,16 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
     // Combined kernel: im2col -> pack A, and pack B simultaneously
     const char * src1_base = (const char *) src1->data;
+    const char * src0_base = (const char *) src0->data;
     const int64_t src1_nb0 = src1->nb[0];
     const int64_t src1_nb1 = src1->nb[1];
     const int64_t src1_nb2 = src1->nb[2];
     const int64_t src1_nb3 = src1->nb[3];
+    const int64_t src1_w = src1->ne[0];
+    const int64_t src1_h = src1->ne[1];
+    const int64_t src1_d = src1->ne[2];
+
+    const bool src0_is_f32 = (src0->type == GGML_TYPE_F32);
 
     // Compute correct strides for src0 as (knl_n_total, oc) matrix
     const int64_t src0_packed_nb0 = kernel_type_size;
@@ -165,7 +171,7 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
         const int64_t sz = dst_z * s2 + kz * d2 - p2;
 
         float val = 0.0f;
-        if (sx >= 0 && sx < src1->ne[0] && sy >= 0 && sy < src1->ne[1] && sz >= 0 && sz < src1->ne[2]) {
+        if (sx >= 0 && sx < src1_w && sy >= 0 && sy < src1_h && sz >= 0 && sz < src1_d) {
             const int64_t channel_idx = batch_idx * c + ic;
             const char * ptr = src1_base + sx * src1_nb0 + sy * src1_nb1 + sz * src1_nb2 + channel_idx * src1_nb3;
             val = *(const float *) ptr;
@@ -184,9 +190,9 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 
         const int64_t row = t % k;
         const int64_t col = t / k;
-        const char * src_ptr = (const char *) src0->data + row * src0_packed_nb0 + col * src0_packed_nb1;
+        const char * src_ptr = src0_base + row * src0_packed_nb0 + col * src0_packed_nb1;
         float v;
-        if (src0->type == GGML_TYPE_F32) {
+        if (src0_is_f32) {
             v = *(const float *) src_ptr;
         } else {
             v = sycl::vec<sycl::half, 1>(*(const sycl::half *) src_ptr).convert<float, sycl::rounding_mode::automatic>()[0];

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

@@ -493,6 +493,20 @@ struct vk_conv2d_pipeline_state {
     }
 };
 
+struct vk_conv3d_pipeline_state {
+    vk_conv3d_pipeline_state(uint32_t s0, uint32_t s1, uint32_t s2, uint32_t p0, uint32_t p1, uint32_t p2,
+                             uint32_t d0, uint32_t d1, uint32_t d2, uint32_t KW, uint32_t KH, uint32_t KD, uint32_t aligned)
+        : s0(s0), s1(s1), s2(s2), p0(p0), p1(p1), p2(p2), d0(d0), d1(d1), d2(d2), KW(KW), KH(KH), KD(KD), aligned(aligned) {}
+
+    uint32_t s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD;
+    uint32_t aligned;
+
+    bool operator<(const vk_conv3d_pipeline_state &b) const {
+        return std::tie(s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD, aligned) <
+               std::tie(b.s0, b.s1, b.s2, b.p0, b.p1, b.p2, b.d0, b.d1, b.d2, b.KW, b.KH, b.KD, b.aligned);
+    }
+};
+
 struct vk_solve_tri_pipeline_state {
     vk_solve_tri_pipeline_state(uint32_t N, uint32_t K)
         : N(N), K(K) {}
@@ -685,6 +699,7 @@ struct vk_device_struct {
 
     bool add_rms_fusion;
     uint32_t partials_binding_alignment;
+    uint32_t max_nodes_per_submit;
 
     bool shader_64b_indexing;
 
@@ -777,6 +792,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
     vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_get_rows_back_f32;
     vk_pipeline pipeline_acc_f32;
     vk_pipeline pipeline_set_f32;
 
@@ -801,14 +817,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_concat_i8, pipeline_concat_i16, pipeline_concat_i32, pipeline_concat_i64;
     vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32, pipeline_upscale_bilinear_antialias_f32;
     vk_pipeline pipeline_scale_f32;
-    vk_pipeline pipeline_sqr_f32;
-    vk_pipeline pipeline_sqrt_f32;
-    vk_pipeline pipeline_sin_f32;
-    vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_log[2];
     vk_pipeline pipeline_tri[2];
     vk_pipeline pipeline_diag[2];
-    vk_pipeline pipeline_clamp_f32;
+    vk_pipeline pipeline_clamp[2];
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
@@ -840,6 +852,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_gelu_quick[2];
     vk_pipeline pipeline_silu[2];
     vk_pipeline pipeline_relu[2];
+    vk_pipeline pipeline_sqr[2];
+    vk_pipeline pipeline_sqrt[2];
+    vk_pipeline pipeline_sin[2];
+    vk_pipeline pipeline_cos[2];
     vk_pipeline pipeline_xielu[2];
     vk_pipeline pipeline_neg[2];
     vk_pipeline pipeline_tanh[2];
@@ -871,7 +887,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_geglu_erf[2];
     vk_pipeline pipeline_geglu_quick[2];
 
-    vk_pipeline pipeline_leaky_relu_f32;
+    vk_pipeline pipeline_leaky_relu[2];
     vk_pipeline pipeline_silu_back_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
@@ -924,6 +940,8 @@ struct vk_device_struct {
     std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
     std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv_transpose_2d_f32[CONV_SHAPE_COUNT];
     std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv_transpose_2d_f16_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv3d_pipeline_state, vk_pipeline> pipeline_conv3d_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv3d_pipeline_state, vk_pipeline> pipeline_conv3d_f16_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_dw_whcn_f32, pipeline_conv2d_dw_whcn_f16_f32;
     vk_pipeline pipeline_conv2d_dw_cwhn_f32, pipeline_conv2d_dw_cwhn_f16_f32;
 
@@ -1669,6 +1687,41 @@ template <> void init_pushconst_fastdiv(vk_op_conv2d_push_constants &p) {
     init_fastdiv_values(p.OW*p.OH,  p.OWOHmp,  p.OWOHL);
 }
 
+struct vk_op_conv3d_push_constants {
+    uint32_t OC;
+    uint32_t IC;
+    uint32_t N;
+
+    uint32_t IW;
+    uint32_t IH;
+    uint32_t ID;
+    uint32_t OW;
+    uint32_t OH;
+    uint32_t OD;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+
+    uint32_t nb1;
+    uint32_t nb2;
+    uint32_t nb3;
+
+    uint32_t OWmp;     uint32_t OWL;
+    uint32_t OWOHmp;   uint32_t OWOHL;
+    uint32_t OWOHODmp; uint32_t OWOHODL;
+};
+
+template <> void init_pushconst_fastdiv(vk_op_conv3d_push_constants &p) {
+    init_fastdiv_values(p.OW,             p.OWmp,     p.OWL);
+    init_fastdiv_values(p.OW*p.OH,        p.OWOHmp,   p.OWOHL);
+    init_fastdiv_values(p.OW*p.OH*p.OD,   p.OWOHODmp, p.OWOHODL);
+}
+
 struct vk_op_conv2d_dw_push_constants {
     uint32_t ne;
     uint32_t batches;
@@ -4074,19 +4127,35 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     }
 #endif
 
+    auto const &ggml_vk_mul_mm_spec = [](std::vector<uint32_t> spec, bool aligned) {
+        spec.push_back(aligned ? 1u : 0u);
+        return spec;
+    };
+
     const int mul_mat_id_param_count = 5;
 
 #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
     if (device->coopmat2) {
+        auto const &ggml_vk_mul_mm_cm2_spec = [](std::vector<uint32_t> spec, bool aligned, bool mul_mat_id) {
+            if (mul_mat_id && spec.size() > 5) {
+                spec.insert(spec.begin() + 5, aligned ? 1u : 0u);
+            } else {
+                spec.push_back(aligned ? 1u : 0u);
+            }
+            if (mul_mat_id && spec.size() == 6) {
+                spec.push_back(32);
+            }
+            return spec;
+        };
 
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(l_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(m_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(s_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(l_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), l_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(m_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), m_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(s_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), s_align, true);   \
 
         // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
@@ -4161,17 +4230,17 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, true);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, true);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, true);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, true);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, true);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, true);   \
 
         // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
@@ -4284,32 +4353,32 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         // Selects dot2 SPIR-V variant at runtime when device->dot2_f16 is true
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
 
         // bf16 scalar path promotes to f32, no dot2 variant
 #define CREATE_MM_NODOT2(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
 
 #define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l_int[TYPE]) { \
@@ -4474,17 +4543,17 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
 
 #define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
         if (device->mul_mat ## ID ## _l_int[TYPE]) \
@@ -4879,6 +4948,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL],  "get_rows_iq4_nl_f32",  get_rows_iq4_nl_f32_len,  get_rows_iq4_nl_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4],   "get_rows_mxfp4_f32",   get_rows_mxfp4_f32_len,   get_rows_mxfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_NVFP4],   "get_rows_nvfp4_f32",   get_rows_nvfp4_f32_len,   get_rows_nvfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_back_f32, "get_rows_back_f32", get_rows_back_f32_len, get_rows_back_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {256, 1, 1}, {}, 1, true);
 
     ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, sizeof(vk_op_flash_attn_split_k_reduce_push_constants), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
@@ -4903,7 +4973,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     }
     ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_nc_push_constants), {1, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1, true);
@@ -5023,11 +5093,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
 
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_sqrt_f32, "sqrt_f32", sqrt_f32_len, sqrt_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
     ggml_vk_create_pipeline(device, device->pipeline_log[0], "log_f32", log_f32_len, log_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_log[1], "log_f16", log_f16_len, log_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -5037,8 +5102,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_diag[0], "diag_f32", diag_f32_len, diag_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -5058,6 +5121,12 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     CREATE_UNARY(gelu_quick)
     CREATE_UNARY(silu)
     CREATE_UNARY(relu)
+    CREATE_UNARY(sqr)
+    CREATE_UNARY(sqrt)
+    CREATE_UNARY(sin)
+    CREATE_UNARY(cos)
+    CREATE_UNARY(clamp)
+    CREATE_UNARY(leaky_relu)
     CREATE_UNARY(xielu)
     CREATE_UNARY(neg)
     CREATE_UNARY(tanh)
@@ -5097,7 +5166,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     CREATE_GLU(geglu_quick)
 #undef CREATE_GLU
 
-    ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true);
@@ -5314,7 +5382,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
 
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_sgd_f32, "opt_step_sgd_f32", opt_step_sgd_f32_len, opt_step_sgd_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
-    // conv2d, conv_transpose_2d
+    // conv2d, conv_transpose_2d, conv3d
     for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
         // smaller WG for the small-tile fallback gives more concurrent WGs per SM
         uint32_t conv2d_WG_SIZE  = (s == CONV_SHAPE_64x32) ? 128 : 256;
@@ -5377,8 +5445,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
             return (conv2d_BS.K * (conv2d_BS.CRS + pad) + conv2d_BS.CRS * (conv2d_BS.NPQ + pad) + csh_elems) * elem_size;
         };
 
-        // coopmat1 needs to store the output through shared memory, so check up front
-        // whether it'll fit and disable it before applying coopmat1 parameters.
+        // 2D, transpose-2D, and 3D conv use the same KxCRS @ CRSxNPQ shmem
+        // layout. cm1 needs Csh for output, so check before applying cm1 params.
         if (conv2d_use_cm1 && device->properties.limits.maxComputeSharedMemorySize < shmem_req(conv2d_cm1_shmem_pad, true, true)) {
             conv2d_use_cm1 = false;
         }
@@ -5470,6 +5538,53 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         }
 #undef CREATE_CONV
 #undef CREATE_CONVS
+
+        std::vector<uint32_t> conv3d_spec_constants = { conv2d_WG_SIZE, conv2d_BS.K, conv2d_BS.CRS, conv2d_BS.NPQ, conv2d_TS_K, conv2d_SHMEM_PAD };
+#define CREATE_CONV3D(type_suffix, spv_suffix) \
+        for (auto &c : device->pipeline_conv3d##type_suffix[s]) { \
+            const vk_conv3d_pipeline_state &state = c.first; \
+            std::vector<uint32_t> spec_constants_cpy = conv3d_spec_constants; \
+            spec_constants_cpy.push_back(state.s0); \
+            spec_constants_cpy.push_back(state.s1); \
+            spec_constants_cpy.push_back(state.s2); \
+            spec_constants_cpy.push_back(state.p0); \
+            spec_constants_cpy.push_back(state.p1); \
+            spec_constants_cpy.push_back(state.p2); \
+            spec_constants_cpy.push_back(state.d0); \
+            spec_constants_cpy.push_back(state.d1); \
+            spec_constants_cpy.push_back(state.d2); \
+            spec_constants_cpy.push_back(state.KW); \
+            spec_constants_cpy.push_back(state.KH); \
+            spec_constants_cpy.push_back(state.KD); \
+            spec_constants_cpy.push_back(state.aligned); \
+            spec_constants_cpy.push_back(conv2d_csh_store); \
+            spec_constants_cpy.push_back(conv2d_WM); \
+            spec_constants_cpy.push_back(conv2d_WN); \
+            ggml_vk_create_pipeline( \
+                device, c.second, "conv3d" #type_suffix, \
+                conv3d##type_suffix##spv_suffix##_len, conv3d##type_suffix##spv_suffix##_data, "main", 3, \
+                sizeof(vk_op_conv3d_push_constants), wg_denoms, spec_constants_cpy, 1, true, conv2d_required_subgroup_size != 0, conv2d_required_subgroup_size); \
+        }
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        if (device->coopmat2) {
+            CREATE_CONV3D(_f32, _cm2)
+            CREATE_CONV3D(_f16_f32, _cm2)
+        } else
+#endif
+#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+        if (conv2d_use_cm1) {
+            CREATE_CONV3D(_f32, _cm1)
+            CREATE_CONV3D(_f16_f32, _cm1)
+        } else
+#endif
+        if (conv2d_UNROLL) {
+            CREATE_CONV3D(_f32, _unroll)
+            CREATE_CONV3D(_f16_f32, _unroll)
+        } else {
+            CREATE_CONV3D(_f32, )
+            CREATE_CONV3D(_f16_f32, )
+        }
+#undef CREATE_CONV3D
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
@@ -5764,6 +5879,14 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->subgroup_vote = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                 (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eVote);
 
+        // Submit at least every 100 nodes, in case there are workloads without as much matmul.
+        device->max_nodes_per_submit = 100;
+        const char* GGML_VK_MAX_NODES_PER_SUBMIT = getenv("GGML_VK_MAX_NODES_PER_SUBMIT");
+        if (GGML_VK_MAX_NODES_PER_SUBMIT != nullptr) {
+            uint32_t max_nodes_per_submit = std::stoul(GGML_VK_MAX_NODES_PER_SUBMIT);
+            device->max_nodes_per_submit = std::max(max_nodes_per_submit, 1u);
+        }
+
         const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;
 
         device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
@@ -10294,6 +10417,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_get_rows_f32[src0->type];
         }
         return nullptr;
+    case GGML_OP_GET_ROWS_BACK:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_get_rows_back_f32;
+        }
+        return nullptr;
     case GGML_OP_ACC:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_acc_f32;
@@ -10400,23 +10528,27 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_SQR:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sqr_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sqr[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_SQRT:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sqrt_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sqrt[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_SIN:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sin_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sin[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_COS:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_cos_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_cos[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_LOG:
@@ -10438,8 +10570,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_CLAMP:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_clamp_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_clamp[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_PAD:
@@ -10807,8 +10940,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
         return nullptr;
     case GGML_OP_LEAKY_RELU:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_leaky_relu_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_leaky_relu[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
     case GGML_OP_CONV_2D:
@@ -10885,6 +11019,61 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             }
         }
         return nullptr;
+    case GGML_OP_CONV_3D:
+        if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            const uint32_t OC = (uint32_t)ggml_get_op_params_i32(dst, 11);
+            const uint32_t IC = (uint32_t)ggml_get_op_params_i32(dst, 9);
+            const uint32_t N  = (uint32_t)ggml_get_op_params_i32(dst, 10);
+            const uint32_t NPQ = N * dst->ne[2] * dst->ne[1] * dst->ne[0];
+            const vk_conv_shapes shape = ggml_vk_conv_select_shape(ctx, OC, NPQ);
+
+            const uint32_t KW = (uint32_t)src0->ne[0];
+            const uint32_t KH = (uint32_t)src0->ne[1];
+            const uint32_t KD = (uint32_t)src0->ne[2];
+            const uint32_t s0 = (uint32_t)ggml_get_op_params_i32(dst, 0);
+            const uint32_t s1 = (uint32_t)ggml_get_op_params_i32(dst, 1);
+            const uint32_t s2 = (uint32_t)ggml_get_op_params_i32(dst, 2);
+            const uint32_t p0 = (uint32_t)ggml_get_op_params_i32(dst, 3);
+            const uint32_t p1 = (uint32_t)ggml_get_op_params_i32(dst, 4);
+            const uint32_t p2 = (uint32_t)ggml_get_op_params_i32(dst, 5);
+            const uint32_t d0 = (uint32_t)ggml_get_op_params_i32(dst, 6);
+            const uint32_t d1 = (uint32_t)ggml_get_op_params_i32(dst, 7);
+            const uint32_t d2 = (uint32_t)ggml_get_op_params_i32(dst, 8);
+
+            const uint32_t CRS = IC * KW * KH * KD;
+            const uint32_t BS_K   = vk_conv_block_sizes[shape].K;
+            const uint32_t BS_CRS = vk_conv_block_sizes[shape].CRS;
+            const uint32_t BS_NPQ = vk_conv_block_sizes[shape].NPQ;
+            const uint32_t aligned = ((OC  % BS_K   == 0) &&
+                                      (CRS % BS_CRS == 0) &&
+                                      (NPQ % BS_NPQ == 0)) ? 1u : 0u;
+
+            vk_conv3d_pipeline_state conv3d_pipeline_state(s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD, aligned);
+
+            std::map<vk_conv3d_pipeline_state, vk_pipeline> *pipelines = nullptr;
+            if (src0->type == GGML_TYPE_F32) {
+                pipelines = &ctx->device->pipeline_conv3d_f32[shape];
+            } else if (src0->type == GGML_TYPE_F16) {
+                pipelines = &ctx->device->pipeline_conv3d_f16_f32[shape];
+            } else {
+                return nullptr;
+            }
+
+            vk_pipeline pipeline = nullptr;
+
+            {
+                std::lock_guard<std::mutex> guard(ctx->device->compile_mutex);
+                auto it = pipelines->find(conv3d_pipeline_state);
+                if (it != pipelines->end()) {
+                    pipeline = it->second;
+                } else {
+                    (*pipelines)[conv3d_pipeline_state] = pipeline = std::make_shared<vk_pipeline_struct>();
+                }
+            }
+
+            return pipeline;
+        }
+        return nullptr;
     case GGML_OP_ADD1:
         if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
             return ctx->device->pipeline_add1_f16_f16;
@@ -11135,6 +11324,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
         elements[2] = std::min(elements[2], ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
         break;
+    case GGML_OP_GET_ROWS_BACK:
+        elements = { (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], 1 };
+        elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+        break;
     case GGML_OP_ARGSORT:
         GGML_ASSERT(0);
         break;
@@ -11220,6 +11413,21 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             GGML_ABORT("invalid push constant type for CONV_2D");
         }
         break;
+    case GGML_OP_CONV_3D:
+        if constexpr (std::is_same_v<PC, vk_op_conv3d_push_constants>) {
+            const uint32_t NPQ = pc.N * pc.OD * pc.OH * pc.OW;
+            const vk_conv_shapes shape = ggml_vk_conv_select_shape(ctx, pc.OC, NPQ);
+            const uint32_t NPQ_blocks = CEIL_DIV(NPQ, vk_conv_block_sizes[shape].NPQ);
+
+            elements = { pc.OC, NPQ_blocks, 1 };
+            if (elements[1] > 512) {
+                elements[2] = CEIL_DIV(elements[1], 512);
+                elements[1] = 512;
+            }
+        } else {
+            GGML_ABORT("invalid push constant type for CONV_3D");
+        }
+        break;
     case GGML_OP_ADD:
     case GGML_OP_SUB:
     case GGML_OP_DIV:
@@ -11236,6 +11444,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_TRI:
     case GGML_OP_DIAG:
     case GGML_OP_CLAMP:
+    case GGML_OP_LEAKY_RELU:
     case GGML_OP_PAD:
     case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
@@ -11380,6 +11589,21 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
     });
 }
 
+static void ggml_vk_get_rows_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_GET_ROWS_BACK, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2], (uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    });
+}
+
 static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
@@ -12087,8 +12311,10 @@ static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx,
 
 static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f, 0.0f, 0.0f });
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, std::move(p));
 }
 
 static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -13118,6 +13344,51 @@ static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx,
     ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, dst->op, std::move(p));
 }
 
+static void ggml_vk_conv_3d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
+                            const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_ASSERT(nb00 == sizeof(float) || nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb10 == sizeof(float));
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    vk_op_conv3d_push_constants p{};
+    p.IC = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 9));
+    p.N  = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 10));
+    p.OC = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 11));
+    GGML_ASSERT(src0->ne[3] == (int64_t)p.IC * p.OC);
+    GGML_ASSERT(src1->ne[3] == (int64_t)p.IC * p.N);
+    GGML_ASSERT(dst->ne[3] == (int64_t)p.OC * p.N);
+
+    p.IW = static_cast<uint32_t>(ne10);
+    p.IH = static_cast<uint32_t>(ne11);
+    p.ID = static_cast<uint32_t>(ne12);
+    p.OW = static_cast<uint32_t>(ne0);
+    p.OH = static_cast<uint32_t>(ne1);
+    p.OD = static_cast<uint32_t>(ne2);
+
+    // the shader clamps src addresses to p.IC * p.N * p.IW * p.IH * p.ID - 1 in uint32, so the
+    // total input element count must fit in a uint32.
+    GGML_ASSERT((uint64_t)p.IC * p.N * p.IW * p.IH * p.ID <= 0xFFFFFFFFull);
+
+    p.nb01 = static_cast<uint32_t>(nb01 / nb00);
+    p.nb02 = static_cast<uint32_t>(nb02 / nb00);
+    p.nb03 = static_cast<uint32_t>(nb03 / nb00);
+
+    p.nb11 = static_cast<uint32_t>(nb11 / nb10);
+    p.nb12 = static_cast<uint32_t>(nb12 / nb10);
+    p.nb13 = static_cast<uint32_t>(nb13 / nb10);
+
+    p.nb1 = static_cast<uint32_t>(nb1 / nb0);
+    p.nb2 = static_cast<uint32_t>(nb2 / nb0);
+    p.nb3 = static_cast<uint32_t>(nb3 / nb0);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_3D, std::move(p));
+}
+
 static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     vk_op_conv2d_dw_push_constants p{};
     p.ne = ggml_nelements(dst);
@@ -13144,7 +13415,10 @@ static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx
 
 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const float * op_params = (const float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f, 0.0f, 0.0f });
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, std::move(p));
 }
 
 #ifdef GGML_VULKAN_RUN_TESTS
@@ -14247,6 +14521,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_GET_ROWS:
         ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node);
 
+        break;
+    case GGML_OP_GET_ROWS_BACK:
+        ggml_vk_get_rows_back(ctx, compute_ctx, src0, src1, node);
+
         break;
     case GGML_OP_ADD:
         if (ctx->num_additional_fused_ops) {
@@ -14515,6 +14793,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_CONV_TRANSPOSE_2D:
         ggml_vk_conv_2d(ctx, compute_ctx, src0, src1, node);
 
+        break;
+    case GGML_OP_CONV_3D:
+        ggml_vk_conv_3d(ctx, compute_ctx, src0, src1, node);
+
         break;
     case GGML_OP_CONV_2D_DW:
         ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node);
@@ -15900,8 +16182,6 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
     // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB
     // (and scaled down based on model size, so smaller models submit earlier).
-    // Also submit at least every 100 nodes, in case there are workloads without as much matmul.
-    int nodes_per_submit = 100;
     int submitted_nodes = 0;
     int submit_count = 0;
     uint64_t mul_mat_bytes = 0;
@@ -16127,7 +16407,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
         bool almost_ready = (cgraph->n_nodes - i) < cgraph->n_nodes / 5;
-        bool submit = (submitted_nodes >= nodes_per_submit) ||
+        bool submit = ((uint32_t)submitted_nodes >= ctx->device->max_nodes_per_submit) ||
                       (mul_mat_bytes_per_submit != 0 && mul_mat_bytes >= mul_mat_bytes_per_submit) ||
                       (i + ctx->num_additional_fused_ops >= last_node) ||
                       (almost_ready && !ctx->almost_ready_fence_pending);
@@ -16964,6 +17244,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                         return false;
                 }
             }
+        case GGML_OP_GET_ROWS_BACK:
+            return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SET_ROWS:
             {
                 switch (op->type) {
@@ -17060,12 +17342,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_TRANSPOSE:
         case GGML_OP_RMS_NORM:
             return true;
-        case GGML_OP_NORM:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_NORM:
         case GGML_OP_L2_NORM:
-            return ggml_is_contiguous_rows(op->src[0]) &&
-                   op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_ADD:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
@@ -17084,8 +17365,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
-            return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_LEAKY_RELU:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                   op->type == op->src[0]->type;
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
@@ -17285,6 +17567,13 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     ggml_is_contiguous(op->src[1]) &&
                     ggml_is_contiguous(op));
             }
+        case GGML_OP_CONV_3D:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                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);
         default:
             return false;
     }
@@ -18128,6 +18417,20 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             const int32_t d0 = tensor->op_params[4];
             const int32_t d1 = tensor->op_params[5];
             tensor_clone = ggml_conv_2d(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1);
+        } else if (tensor->op == GGML_OP_CONV_3D) {
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t s1 = tensor->op_params[1];
+            const int32_t s2 = tensor->op_params[2];
+            const int32_t p0 = tensor->op_params[3];
+            const int32_t p1 = tensor->op_params[4];
+            const int32_t p2 = tensor->op_params[5];
+            const int32_t d0 = tensor->op_params[6];
+            const int32_t d1 = tensor->op_params[7];
+            const int32_t d2 = tensor->op_params[8];
+            const int32_t IC = tensor->op_params[9];
+            const int32_t N  = tensor->op_params[10];
+            const int32_t OC = tensor->op_params[11];
+            tensor_clone = ggml_conv_3d_direct(ggml_ctx, src_clone[0], src_clone[1], s0, s1, s2, p0, p1, p2, d0, d1, d2, IC, N, OC);
         } else if (tensor->op == GGML_OP_CONV_2D_DW) {
             const int32_t s0 = tensor->op_params[0];
             const int32_t s1 = tensor->op_params[1];

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
-}

ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp

@@ -0,0 +1,431 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#ifdef COOPMAT2
+#extension GL_NV_cooperative_matrix2 : enable
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_KHR_memory_scope_semantics : enable
+#endif
+
+#ifdef COOPMAT
+#extension GL_KHR_cooperative_matrix : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_KHR_memory_scope_semantics : enable
+#endif
+
+#include "types.glsl"
+
+// shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
+layout(binding = 0) readonly buffer A {
+    A_TYPE knl_data[];
+};  // src0 - kernel:   [KW, KH, KD, IC*OC]
+
+layout(binding = 1) readonly buffer B {
+    B_TYPE src_data[];
+};  // src1 - input:    [IW, IH, ID, IC*N] -- channel_first format
+
+layout(binding = 2) writeonly buffer D {
+    D_TYPE dst_data[];
+};  // dst - result:    [OW, OH, OD, OC*N]
+
+layout(push_constant) uniform parameter {
+    // I/O channels, batch size
+    uint32_t OC;
+    uint32_t IC;
+    uint32_t N;
+
+    // Tensor spatial sizes: input, output
+    uint32_t IW;
+    uint32_t IH;
+    uint32_t ID;
+    uint32_t OW;
+    uint32_t OH;
+    uint32_t OD;
+
+    // Strides in elements
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+
+    uint32_t nb1;
+    uint32_t nb2;
+    uint32_t nb3;
+
+    // fastdiv helper values
+    uint32_t OWmp;   uint32_t OWL;
+    uint32_t OWOHmp; uint32_t OWOHL;
+    uint32_t OWOHODmp; uint32_t OWOHODL;
+}
+
+p;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+// Blocktile sizes
+layout(constant_id = 1) const uint BS_K            = 128;
+layout(constant_id = 2) const uint BS_CRS          = 16;
+layout(constant_id = 3) const uint BS_NPQ          = 128;
+// Thread-tile sizes
+layout(constant_id = 4) const uint TS_K            = 8;
+layout(constant_id = 5) const uint SHMEM_PAD       = 4;
+// Stride, padding, dilation
+layout(constant_id = 6)  const uint s0             = 1;
+layout(constant_id = 7)  const uint s1             = 1;
+layout(constant_id = 8)  const uint s2             = 1;
+layout(constant_id = 9)  const uint p0             = 0;
+layout(constant_id = 10) const uint p1             = 0;
+layout(constant_id = 11) const uint p2             = 0;
+layout(constant_id = 12) const uint d0             = 1;
+layout(constant_id = 13) const uint d1             = 1;
+layout(constant_id = 14) const uint d2             = 1;
+// Kernel spatial sizes
+layout(constant_id = 15) const uint KW             = 1;
+layout(constant_id = 16) const uint KH             = 1;
+layout(constant_id = 17) const uint KD             = 1;
+// when set, skip bounds checks and address clamps (K/CRS/NPQ are tile-aligned)
+layout(constant_id = 18) const uint aligned        = 0;
+// stage cm2 result through shmem (Csh) for coalesced stores. cm1 always does this.
+layout(constant_id = 19) const uint csh_store      = 0;
+
+#ifdef COOPMAT
+// cm1 subgroup tile: each subgroup computes a WM x WN region as a grid of
+// TM x TN x TK fragments. Requires WM%TM == WN%TN == BS_K%WM == BS_NPQ%WN ==
+// BS_CRS%TK == 0, and WG_SIZE == (BS_K/WM) * (BS_NPQ/WN) * subgroup_size.
+layout(constant_id = 20) const uint WM             = 32;
+layout(constant_id = 21) const uint WN             = 32;
+const uint TM = 16;
+const uint TN = 16;
+const uint TK = 16;
+const uint cms_per_row = WM / TM;
+const uint cms_per_col = WN / TN;
+const uint warps_M     = BS_K / WM;
+const uint warps_N     = BS_NPQ / WN;
+#endif
+
+// without padding, ID_idx/IH_idx/IW_idx are in bounds by construction
+const bool dhw_in_bounds = (p0 == 0) && (p1 == 0) && (p2 == 0);
+
+uint32_t       tid     = gl_LocalInvocationID.x;
+const uint32_t WG_SIZE = gl_WorkGroupSize.x;
+
+uint splitWork(uint work_size, uint block_size) {
+    return (block_size + work_size - 1) / block_size;
+}
+
+uint32_t K   = p.OC;
+uint32_t CRS = p.IC * KD * KH * KW;
+uint32_t NPQ = p.N * p.OD * p.OH * p.OW;
+
+// Number of blocktiles per input
+uint32_t NB_CRS = splitWork(CRS, BS_CRS);
+
+#if defined(COOPMAT2) || defined(COOPMAT)
+#define SHMEM_TYPE float16_t
+#else
+#define SHMEM_TYPE float
+#endif
+
+const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
+const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
+
+const uint32_t Ash_len = BS_K * Ash_stride;
+const uint32_t Bsh_len = BS_CRS * Bsh_stride;
+
+shared SHMEM_TYPE Ash[Ash_len];  // K x CRS
+shared SHMEM_TYPE Bsh[Bsh_len];  // CRS x NPQ
+
+#if defined(COOPMAT2) || defined(COOPMAT)
+// stage matC through shmem so global stores are row-major (NPQ-contiguous)
+const uint32_t Csh_stride = BS_NPQ;
+#ifdef COOPMAT
+const uint32_t Csh_len    = BS_K * Csh_stride;
+#else
+const uint32_t Csh_len    = csh_store != 0 ? BS_K * Csh_stride : 1;
+#endif
+shared SHMEM_TYPE Csh[Csh_len];  // K x NPQ
+#endif
+
+// Threadtile sizes
+const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
+
+// Number of threadtiles per blocktile
+const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
+
+/*
+Compute
+KxCRS @ CRSxNPQ = K x NPQ
+K=OC
+C=IC
+D,R,S=KD,KH,KW
+Z,P,Q=OD,OH,OW
+*/
+
+uint32_t B_idx_K   = gl_WorkGroupID.x;
+uint32_t B_idx_NPQ = gl_WorkGroupID.y + gl_WorkGroupID.z * 512;
+
+uint32_t T_y = tid / NT_NPQ;
+uint32_t T_x = tid % NT_NPQ;
+
+uint32_t       Ar    = tid / BS_CRS;
+uint32_t       Ac    = tid % BS_CRS;
+const uint32_t ArpWg = WG_SIZE / BS_CRS;
+
+uint32_t       Br    = tid / BS_NPQ;
+uint32_t       Bc    = tid % BS_NPQ;
+const uint32_t BrpWg = WG_SIZE / BS_NPQ;
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
+void split_crs(uint32_t crs_idx, out uint32_t ic, out uint32_t kd, out uint32_t kh, out uint32_t kw) {
+    const uint32_t KHKW = KH * KW;
+    const uint32_t KDKHKW = KD * KHKW;
+    ic = crs_idx / KDKHKW;
+    uint32_t rem = crs_idx - ic * KDKHKW;
+    kd = rem / KHKW;
+    rem = rem - kd * KHKW;
+    kh = rem / KW;
+    kw = rem - kh * KW;
+}
+
+void split_npq(uint32_t npq_idx, out uint32_t n, out uint32_t od, out uint32_t oh, out uint32_t ow) {
+    const uint32_t OWOH = p.OW * p.OH;
+    n = fastdiv(npq_idx, p.OWOHODmp, p.OWOHODL);
+    uint32_t rem = npq_idx - n * p.OD * OWOH;
+    od = fastdiv(rem, p.OWOHmp, p.OWOHL);
+    rem = rem - od * OWOH;
+    oh = fastdiv(rem, p.OWmp, p.OWL);
+    ow = rem - oh * p.OW;
+}
+
+#ifdef COOPMAT2
+#define ACC_TYPE float16_t
+
+ACC_TYPE perElemOpStore(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem)
+{
+    uint32_t K_idx   = B_idx_K * BS_K + r;
+    uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + c;
+    uint32_t N_idx;
+    uint32_t OD_idx;
+    uint32_t OH_idx;
+    uint32_t OW_idx;
+    split_npq(NPQ_idx, N_idx, OD_idx, OH_idx, OW_idx);
+    uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + OD_idx * p.nb2 + (N_idx * p.OC + K_idx) * p.nb3;
+    if (aligned != 0 || (K_idx < K && NPQ_idx < NPQ)) {
+        dst_data[dst_idx] = D_TYPE(elem);
+    }
+    return elem;
+}
+#endif
+
+void main() {
+    if (B_idx_NPQ * BS_NPQ >= NPQ) {
+        return;
+    }
+
+#ifdef COOPMAT2
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator> matC;
+    matC = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator>(0.0);
+#elif defined(COOPMAT)
+    coopmat<float16_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
+    [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
+        sums[i] = coopmat<float16_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0);
+    }
+    const uint warp_r = gl_SubgroupID / warps_N;
+    const uint warp_c = gl_SubgroupID % warps_N;
+#else
+    float regC[TS_K][TS_NPQ];
+    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+            regC[T_ly][T_lx] = 0.0;
+        }
+    }
+#endif
+    /* Advance block in CRS dim */
+    [[dont_unroll]] for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
+        uint32_t CRS_idx_a = B_idx_CRS * BS_CRS + Ac;
+        uint32_t IC_idx_a;
+        uint32_t KD_idx_a;
+        uint32_t KH_idx_a;
+        uint32_t KW_idx_a;
+        split_crs(CRS_idx_a, IC_idx_a, KD_idx_a, KH_idx_a, KW_idx_a);
+
+        /* Load kernel to A_block: (BS_K x BS_CRS)*/
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
+            uint32_t B_ly    = r_offset + Ar;
+            uint32_t B_lx    = Ac;
+            uint32_t K_idx   = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
+            uint32_t knl_idx = KW_idx_a + KH_idx_a * p.nb01 + KD_idx_a * p.nb02 + (K_idx * p.IC + IC_idx_a) * p.nb03;
+            if (aligned == 0) {
+                knl_idx = min(knl_idx, K * CRS - 1);
+            }
+            float    val     = knl_data[knl_idx];
+            if (aligned == 0 && (K_idx >= K || CRS_idx_a >= CRS)) {
+                val = 0.0;
+            }
+            Ash[B_ly * Ash_stride + B_lx] = SHMEM_TYPE(val);
+        }
+        /* Load input to B_block: (BS_CRS x BS_NPQ) */
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
+            uint32_t B_ly          = r_offset + Br;             /* Row index of B block */
+            uint32_t B_lx          = Bc;
+            uint32_t NPQ_idx       = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
+            uint32_t N_idx;
+            uint32_t OD_idx;
+            uint32_t OH_idx;
+            uint32_t OW_idx;
+            split_npq(NPQ_idx, N_idx, OD_idx, OH_idx, OW_idx);
+
+            uint32_t CRS_idx_b = B_idx_CRS * BS_CRS + B_ly;
+            uint32_t IC_idx_b;
+            uint32_t KD_idx_b;
+            uint32_t KH_idx_b;
+            uint32_t KW_idx_b;
+            split_crs(CRS_idx_b, IC_idx_b, KD_idx_b, KH_idx_b, KW_idx_b);
+
+            uint32_t ID_idx = OD_idx * s2 + KD_idx_b * d2 - p2;
+            uint32_t IH_idx = OH_idx * s1 + KH_idx_b * d1 - p1;
+            uint32_t IW_idx = OW_idx * s0 + KW_idx_b * d0 - p0;
+
+            uint32_t src_idx = IW_idx + IH_idx * p.nb11 + ID_idx * p.nb12 + (N_idx * p.IC + IC_idx_b) * p.nb13;
+            // skip clamp when address can't go OOB
+            if (aligned == 0 || !dhw_in_bounds) {
+                src_idx = min(src_idx, p.IC * p.N * p.IW * p.IH * p.ID - 1);
+            }
+            float val = src_data[src_idx];
+            bool oob = false;
+            if (aligned == 0 && (CRS_idx_b >= CRS || NPQ_idx >= NPQ)) {
+                oob = true;
+            }
+            // also catches lower-bound underflow (idx wraps to 0x80000000+)
+            if (!dhw_in_bounds && (ID_idx >= p.ID || IH_idx >= p.IH || IW_idx >= p.IW)) {
+                oob = true;
+            }
+            if (oob) {
+                val = 0.0;
+            }
+            Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);
+        }
+        barrier();
+#ifdef COOPMAT2
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_K, BS_CRS, gl_MatrixUseA> matA;
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_CRS, BS_NPQ, gl_MatrixUseB> matB;
+
+        coopMatLoad(matA, Ash, 0, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
+        coopMatLoad(matB, Bsh, 0, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+        matC = coopMatMulAdd(matA, matB, matC);
+#elif defined(COOPMAT)
+        // each subgroup multiplies its grid of fragments per TK-sized CRS chunk
+        [[unroll]] for (uint k_step = 0; k_step < BS_CRS / TK; k_step++) {
+            coopmat<float16_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a[cms_per_row];
+            [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+                const uint a_off = (warp_r * WM + cm_row * TM) * Ash_stride + k_step * TK;
+                coopMatLoad(cache_a[cm_row], Ash, a_off, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
+            }
+            [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+                coopmat<float16_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
+                const uint b_off = k_step * TK * Bsh_stride + warp_c * WN + cm_col * TN;
+                coopMatLoad(cache_b, Bsh, b_off, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+                    sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a[cm_row], cache_b, sums[cm_col * cms_per_row + cm_row]);
+                }
+            }
+        }
+#else
+        if (T_y * TS_K < K) {
+            UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+                float regA[TS_K];
+                float regB[TS_NPQ];
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
+                }
+                for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                    regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
+                }
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                        regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                    }
+                }
+            }
+        }
+#endif
+        barrier();
+    }
+    /* Save C* */
+#if defined(COOPMAT2) || defined(COOPMAT)
+    // stage matC into Csh, then write to dst with coalesced NPQ-contiguous stores
+#ifdef COOPMAT
+    const bool use_staged_store = true;
+#else
+    const bool use_staged_store = (csh_store != 0);
+#endif
+    if (use_staged_store) {
+#ifdef COOPMAT
+        // cm1: each subgroup stores its fragment grid into its Csh slot
+        [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+            [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+                const uint csh_off = (warp_r * WM + cm_row * TM) * Csh_stride + warp_c * WN + cm_col * TN;
+                coopMatStore(sums[cm_col * cms_per_row + cm_row], Csh, csh_off, Csh_stride, gl_CooperativeMatrixLayoutRowMajor);
+            }
+        }
+#else
+        coopMatStore(matC, Csh, 0, Csh_stride, gl_CooperativeMatrixLayoutRowMajor);
+#endif
+        barrier();
+
+        // cooperative shmem->global: WG threads spread across BS_NPQ (the
+        // contiguous direction of dst), each iter covers store_rows_per_iter K-rows
+        const uint32_t store_rows_per_iter = WG_SIZE / BS_NPQ;
+        const uint32_t store_iters         = BS_K / store_rows_per_iter;
+        const uint32_t k_thread_offset     = tid / BS_NPQ;
+        const uint32_t npq_thread          = tid % BS_NPQ;
+        [[unroll]] for (uint32_t i = 0; i < store_iters; i++) {
+            uint32_t k_local = i * store_rows_per_iter + k_thread_offset;
+            uint32_t K_idx   = B_idx_K * BS_K + k_local;
+            uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + npq_thread;
+            uint32_t N_idx;
+            uint32_t OD_idx;
+            uint32_t OH_idx;
+            uint32_t OW_idx;
+            split_npq(NPQ_idx, N_idx, OD_idx, OH_idx, OW_idx);
+            uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + OD_idx * p.nb2 + (N_idx * p.OC + K_idx) * p.nb3;
+            if (aligned != 0 || (K_idx < K && NPQ_idx < NPQ)) {
+                dst_data[dst_idx] = D_TYPE(Csh[k_local * Csh_stride + npq_thread]);
+            }
+        }
+    }
+#ifdef COOPMAT2
+    else {
+        coopMatPerElementNV(matC, matC, perElemOpStore);
+    }
+#endif
+#else
+    if (T_y * TS_K < K) {
+        for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
+                uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
+                uint32_t N_idx;
+                uint32_t OD_idx;
+                uint32_t OH_idx;
+                uint32_t OW_idx;
+                split_npq(NPQ_idx, N_idx, OD_idx, OH_idx, OW_idx);
+                uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + OD_idx * p.nb2 + (N_idx * p.OC + K_idx) * p.nb3;
+                if (aligned != 0 || (K_idx < K && NPQ_idx < NPQ)) {
+                    dst_data[dst_idx] = D_TYPE(regC[T_ly][T_lx]);
+                }
+            }
+        }
+    }
+#endif
+}

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(cos(val));
-}

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

@@ -463,6 +463,7 @@ void main() {
                 }
                 rowmaxf = max(rowmaxf, float(Sf[r][c]));
             }
+            rowmaxf += FATTN_KQ_MAX_OFFSET;
             float Moldf = Mf[r];
 
             // M = max(rowmax, Mold)

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -352,6 +352,7 @@ void main() {
                 }
                 rowmaxf = max(rowmaxf, float(sfsh[r_vec + (c * cols_per_iter + col_tid) * sfshstride][r_comp]));
             }
+            rowmaxf += FATTN_KQ_MAX_OFFSET;
             float Moldf = Mf[r];
 
             // Compute max across the row

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

@@ -0,0 +1,25 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_binary_head.glsl"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint col = gl_GlobalInvocationID.x;
+
+    if (col >= p.ne20) {
+        return;
+    }
+
+    for (uint row = gl_GlobalInvocationID.y; row < p.ne21; row += gl_WorkGroupSize.y * gl_NumWorkGroups.y) {
+        float sum = 0.0f;
+        for (uint i = 0; i < p.ne10; ++i) {
+            if (data_b[get_boffset() + i*p.nb10] == int(row)) {
+                sum += data_a[get_aoffset() + i*p.nb01 + col*p.nb00];
+            }
+        }
+
+        data_d[get_doffset() + row*p.nb21 + col*p.nb20] = sum;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp

@@ -14,16 +14,13 @@ void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
-    const uint i3 = row / (p.ne11 * p.ne12);
-    const uint i3_offset = i3 * p.ne12 * p.ne11;
-    const uint i2 = (row - i3_offset) / p.ne11;
-    const uint i2_offset = i2 * p.ne11;
-    const uint i1 = row - i3_offset - i2_offset;
+    const uint a_base = get_aoffset() + src0_idx(row * p.ne00);
+    const uint d_base = get_doffset() + dst_idx(row * p.ne10);
 
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
     [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]);
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_base + i0*p.nb00]);
         sum[tid] += xi * xi;
     }
 
@@ -39,6 +36,6 @@ void main() {
     const FLOAT_TYPE scale = 1.0f / max(sqrt(sum[0]), FLOAT_TYPE(p.param1));
 
     [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
-        data_d[i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0] = D_TYPE(scale * FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]));
+        data_d[d_base + i0*p.nb10] = D_TYPE(scale * FLOAT_TYPE(data_a[a_base + i0*p.nb00]));
     }
 }

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

@@ -1,22 +0,0 @@
-#version 450
-
-#include "generic_head.glsl"
-#include "types.glsl"
-
-#extension GL_EXT_control_flow_attributes : enable
-
-layout(local_size_x = 512, 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 i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
-
-    if (i >= p.KX) {
-        return;
-    }
-
-    const float val = float(data_a[i]);
-    data_d[i] = D_TYPE(max(val, 0.0f) + min(val, 0.0f) * p.param1);
-}

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

@@ -38,17 +38,7 @@
 #define LOAD_VEC_B 1
 #endif
 
-// Load 2 values at once without affecting index calculations through LOAD_VEC
-#if (defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)) && !defined(ALIGNED)
-#define LOAD_VEC_BATCH_A 2
-#else
-#define LOAD_VEC_BATCH_A 1
-#endif
-#if !defined(ALIGNED)
-#define LOAD_VEC_BATCH_B 2
-#else
-#define LOAD_VEC_BATCH_B 1
-#endif
+layout (constant_id = 11) const uint ALIGNED = 0;
 
 #if !defined(TO_FLOAT_TYPE)
 #define TO_FLOAT_TYPE FLOAT_TYPE
@@ -57,6 +47,13 @@
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(DATA_A_F32)
+layout (binding = 0) readonly buffer A_SCALAR {float data_a_scalar[];};
+#elif defined(DATA_A_F16)
+layout (binding = 0) readonly buffer A_SCALAR {float16_t data_a_scalar[];};
+#elif defined(DATA_A_BF16)
+layout (binding = 0) readonly buffer A_SCALAR {uint16_t data_a_scalar[];};
+#endif
 #if defined(A_TYPE_PACKED16)
 layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
 #endif
@@ -65,6 +62,7 @@ layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32
 #endif
 
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 1) readonly buffer B_SCALAR {B_TYPE_SCALAR data_b_scalar[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 #ifdef MUL_MAT_ID
@@ -194,13 +192,23 @@ void main() {
     const uint warp_r = warp_i % (BM / WM);
     const uint warp_c = warp_i / (BM / WM);
 
-    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
-    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
-    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
-    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
+#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
+    const uint LOAD_VEC_A_EFF = (ALIGNED != 0) ? LOAD_VEC_A : 1;
+    const uint LOAD_VEC_BATCH_A = (ALIGNED != 0) ? 1 : 2;
+#else
+    const uint LOAD_VEC_A_EFF = LOAD_VEC_A;
+    const uint LOAD_VEC_BATCH_A = 1;
+#endif
+    const uint LOAD_VEC_B_EFF = (ALIGNED != 0) ? LOAD_VEC_B : 1;
+    const uint LOAD_VEC_BATCH_B = (ALIGNED != 0) ? 1 : 2;
 
-    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A * LOAD_VEC_BATCH_A / BK;
-    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B * LOAD_VEC_BATCH_B / BK;
+    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A_EFF / LOAD_VEC_BATCH_A);
+    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A_EFF / LOAD_VEC_BATCH_A);
+    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B_EFF / LOAD_VEC_BATCH_B);
+    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B_EFF / LOAD_VEC_BATCH_B);
+
+    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A_EFF * LOAD_VEC_BATCH_A / BK;
+    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B_EFF * LOAD_VEC_BATCH_B / BK;
 
 #ifdef MUL_MAT_ID
 #ifdef MUL_MAT_ID_USE_SUBGROUPS
@@ -239,15 +247,15 @@ void main() {
 
     uint pos_a =
 #ifdef MUL_MAT_ID
-        expert_idx * (p.batch_stride_a / LOAD_VEC_A) +
+        expert_idx * (p.batch_stride_a / LOAD_VEC_A_EFF) +
 #else
-        batch_idx_a * (p.batch_stride_a / LOAD_VEC_A) +
+        batch_idx_a * (p.batch_stride_a / LOAD_VEC_A_EFF) +
 #endif
-        (ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
+        (ir * BM * p.stride_a + start_k) / LOAD_VEC_A_EFF;
 #ifdef MUL_MAT_ID
     uint pos_b = 0;
 #else
-    uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B;
+    uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B_EFF;
 #endif
 
 #ifdef COOPMAT
@@ -287,8 +295,8 @@ void main() {
 
         barrier();
 
-        pos_a += BK / LOAD_VEC_A;
-        pos_b += BK / LOAD_VEC_B;
+        pos_a += BK / LOAD_VEC_A_EFF;
+        pos_b += BK / LOAD_VEC_B_EFF;
 
 #ifdef COOPMAT
         [[unroll]] for (uint i = 0; i < BK; i += TK) {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -36,6 +36,7 @@ layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working wit
 layout (constant_id = 4) const bool enable_smaller_matrices = false;
 const uint BNover2 = enable_smaller_matrices ? (BN / 2) : BN;
 const uint BNover4 = enable_smaller_matrices ? (BN / 4) : BN;
+layout (constant_id = 5) const uint ALIGNED = 0;
 
 layout (push_constant) uniform parameter
 {
@@ -111,7 +112,7 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB {
 };
 
 uint _ne1;
-layout (constant_id = 5) const uint subgroup_size = 32;
+layout (constant_id = 6) const uint subgroup_size = 32;
 shared uvec4 ballots_sh[BLOCK_SIZE / subgroup_size];
 
 B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const in uint coordInBlock[2])
@@ -297,12 +298,12 @@ void main() {
 
     // Hint to the compiler that values are aligned (want 16B alignment).
     // Quants are always block-aligned, no alignment needed.
-#if ALIGNED
+    if (ALIGNED != 0) {
 #if QUANT_K == 1
-    stride_a &= ~7;
-#endif
-    stride_b &= ~7;
+        stride_a &= ~7;
 #endif
+        stride_b &= ~7;
+    }
 
     // Create layouts for both clamped and unclamped accesses
     tensorLayoutNV<2> tensorLayoutA = createTensorLayoutNV(2);

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl

@@ -1,50 +1,57 @@
 void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uint idx_m, const uint block, const uint end_k) {
 #if defined(DATA_A_F32) || defined(DATA_A_F16)
 #if LOAD_VEC_A == 8
-            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
-            FLOAT_TYPEV8 aa = FLOAT_TYPEV8(data_a[idx]);
-            buf_a[buf_idx    ] = aa[0].xy;
-            buf_a[buf_idx + 1] = aa[0].zw;
-            buf_a[buf_idx + 2] = aa[1].xy;
-            buf_a[buf_idx + 3] = aa[1].zw;
+            if (ALIGNED != 0) {
+                const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
+                FLOAT_TYPEV8 aa = FLOAT_TYPEV8(data_a[idx]);
+                buf_a[buf_idx    ] = aa[0].xy;
+                buf_a[buf_idx + 1] = aa[0].zw;
+                buf_a[buf_idx + 2] = aa[1].xy;
+                buf_a[buf_idx + 3] = aa[1].zw;
+                return;
+            }
 #elif LOAD_VEC_A == 4
-            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
-            FLOAT_TYPEV4 aa = FLOAT_TYPEV4(data_a[idx]);
-            buf_a[buf_idx    ] = aa.xy;
-            buf_a[buf_idx + 1] = aa.zw;
-#else // LOAD_VEC_BATCH_A == 2
+            if (ALIGNED != 0) {
+                const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
+                FLOAT_TYPEV4 aa = FLOAT_TYPEV4(data_a[idx]);
+                buf_a[buf_idx    ] = aa.xy;
+                buf_a[buf_idx + 1] = aa.zw;
+                return;
+            }
+#endif
             const uint idx = pos_a + col * p.stride_a + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_m < p.M && block + row * 2 + 1 < end_k) {
-                buf_a[buf_idx] = FLOAT_TYPEV2(data_a[idx],
-                                              data_a[idx + 1]);
+                buf_a[buf_idx] = FLOAT_TYPEV2(data_a_scalar[idx],
+                                              data_a_scalar[idx + 1]);
             } else if (idx_m < p.M && block + row * 2 < end_k) {
-                buf_a[buf_idx] = FLOAT_TYPEV2(data_a[idx], 0.0f);
+                buf_a[buf_idx] = FLOAT_TYPEV2(data_a_scalar[idx], 0.0f);
             } else {
                 buf_a[buf_idx] = FLOAT_TYPEV2(0.0f);
             }
-#endif
 #elif defined(DATA_A_BF16)
 #if LOAD_VEC_A == 4
-            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
-            FLOAT_TYPEV4 aa = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_a[idx]));
-            buf_a[buf_idx    ] = aa.xy;
-            buf_a[buf_idx + 1] = aa.zw;
-#else // LOAD_VEC_BATCH_A == 2
+            if (ALIGNED != 0) {
+                const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
+                FLOAT_TYPEV4 aa = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_a[idx]));
+                buf_a[buf_idx    ] = aa.xy;
+                buf_a[buf_idx + 1] = aa.zw;
+                return;
+            }
+#endif
             const uint idx = pos_a + col * p.stride_a + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_m < p.M && block + row * 2 + 1 < end_k) {
-                buf_a[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_a[idx]),
-                                              TO_FLOAT_TYPE(data_a[idx + 1]));
+                buf_a[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_a_scalar[idx]),
+                                              TO_FLOAT_TYPE(data_a_scalar[idx + 1]));
             } else if (idx_m < p.M && block + row * 2 < end_k) {
-                buf_a[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_a[idx]), 0.0f);
+                buf_a[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_a_scalar[idx]), 0.0f);
             } else {
                 buf_a[buf_idx] = FLOAT_TYPEV2(0.0f);
             }
-#endif
 #elif defined(DATA_A_Q4_0)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 4;
@@ -526,75 +533,85 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
 #if !defined(MUL_MAT_ID)
 void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uint idx_n, const uint block, const uint end_k) {
 #if LOAD_VEC_B == 8
-            // Not supported for b_type bf16 because bf16mat2x4 does not exist
-            const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
-            FLOAT_TYPEV8 bb = FLOAT_TYPEV8(data_b[idx]);
-            buf_b[buf_idx + 0] = bb[0].xy;
-            buf_b[buf_idx + 1] = bb[0].zw;
-            buf_b[buf_idx + 2] = bb[1].xy;
-            buf_b[buf_idx + 3] = bb[1].zw;
+            if (ALIGNED != 0) {
+                // Not supported for b_type bf16 because bf16mat2x4 does not exist
+                const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
+                FLOAT_TYPEV8 bb = FLOAT_TYPEV8(data_b[idx]);
+                buf_b[buf_idx + 0] = bb[0].xy;
+                buf_b[buf_idx + 1] = bb[0].zw;
+                buf_b[buf_idx + 2] = bb[1].xy;
+                buf_b[buf_idx + 3] = bb[1].zw;
+                return;
+            }
 #elif LOAD_VEC_B == 4
-            const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
+            if (ALIGNED != 0) {
+                const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
 #if defined(DATA_B_BF16)
-            FLOAT_TYPEV4 bb = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_b[idx]));
+                FLOAT_TYPEV4 bb = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_b[idx]));
 #else
-            FLOAT_TYPEV4 bb = FLOAT_TYPEV4(data_b[idx]);
+                FLOAT_TYPEV4 bb = FLOAT_TYPEV4(data_b[idx]);
+#endif
+                buf_b[buf_idx + 0] = bb.xy;
+                buf_b[buf_idx + 1] = bb.zw;
+                return;
+            }
 #endif
-            buf_b[buf_idx + 0] = bb.xy;
-            buf_b[buf_idx + 1] = bb.zw;
-#else // LOAD_VEC_BATCH_B == 2
             const uint idx = pos_b + col * p.stride_b + row * 2;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (idx_n < p.N && block + row * 2 + 1 < end_k) {
-                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b[idx]),
-                                              TO_FLOAT_TYPE(data_b[idx + 1]));
+                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b_scalar[idx]),
+                                              TO_FLOAT_TYPE(data_b_scalar[idx + 1]));
             } else if (idx_n < p.N && block + row * 2 < end_k) {
-                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b[idx]), 0.0f);
+                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b_scalar[idx]), 0.0f);
             } else {
                 buf_b[buf_idx] = FLOAT_TYPEV2(0.0f);
             }
-#endif
 }
 #else
 void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uint ic, const uint _ne1, const uint block, const uint end_k) {
 #if LOAD_VEC_B == 8
-            // Not supported for b_type bf16 because bf16mat2x4 does not exist
-            const u16vec2 row_idx = row_ids[col];
-            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
-            FLOAT_TYPEV8 bb = FLOAT_TYPEV8(data_b[idx]);
-            buf_b[buf_idx + 0] = bb[0].xy;
-            buf_b[buf_idx + 1] = bb[0].zw;
-            buf_b[buf_idx + 2] = bb[1].xy;
-            buf_b[buf_idx + 3] = bb[1].zw;
+            if (ALIGNED != 0) {
+                // Not supported for b_type bf16 because bf16mat2x4 does not exist
+                const u16vec2 row_idx = row_ids[col];
+                const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
+                FLOAT_TYPEV8 bb = FLOAT_TYPEV8(data_b[idx]);
+                buf_b[buf_idx + 0] = bb[0].xy;
+                buf_b[buf_idx + 1] = bb[0].zw;
+                buf_b[buf_idx + 2] = bb[1].xy;
+                buf_b[buf_idx + 3] = bb[1].zw;
+                return;
+            }
 #elif LOAD_VEC_B == 4
-            const u16vec2 row_idx = row_ids[col];
-            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
+            if (ALIGNED != 0) {
+                const u16vec2 row_idx = row_ids[col];
+                const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
+                const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
 #if defined(DATA_B_BF16)
-            FLOAT_TYPEV4 bb = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_b[idx]));
+                FLOAT_TYPEV4 bb = FLOAT_TYPEV4(TO_FLOAT_TYPE(data_b[idx]));
 #else
-            FLOAT_TYPEV4 bb = FLOAT_TYPEV4(data_b[idx]);
+                FLOAT_TYPEV4 bb = FLOAT_TYPEV4(data_b[idx]);
+#endif
+                buf_b[buf_idx + 0] = bb.xy;
+                buf_b[buf_idx + 1] = bb.zw;
+                return;
+            }
 #endif
-            buf_b[buf_idx + 0] = bb.xy;
-            buf_b[buf_idx + 1] = bb.zw;
-#else // LOAD_VEC_BATCH_B == 2
             const uint row_i = ic * BN + col;
             const uint buf_idx = col * SHMEM_STRIDE + row;
             if (row_i < _ne1 && block + row * 2 + 1 < end_k) {
                 const u16vec2 row_idx = row_ids[col];
                 const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
-                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b[idx]),
-                                              TO_FLOAT_TYPE(data_b[idx + 1]));
+                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b_scalar[idx]),
+                                              TO_FLOAT_TYPE(data_b_scalar[idx + 1]));
             } else if (row_i < _ne1 && block + row * 2 < end_k) {
                 const u16vec2 row_idx = row_ids[col];
                 const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
-                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b[idx]), 0.0f);
+                buf_b[buf_idx] = FLOAT_TYPEV2(TO_FLOAT_TYPE(data_b_scalar[idx]), 0.0f);
             } else {
                 buf_b[buf_idx] = FLOAT_TYPEV2(0.0f);
             }
-#endif
 }
 #endif

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

@@ -1,26 +1,26 @@
 #version 450
 
-#include "generic_head.glsl"
 #include "types.glsl"
+#include "generic_unary_head.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512
 
 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[];};
-
 shared vec2 sum[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
 
+    const uint a_base = get_aoffset() + src0_idx(row * p.ne00);
+    const uint d_base = get_doffset() + dst_idx(row * p.ne10);
+
     sum[tid] = vec2(0.0f, 0.0f);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        const float xi = float(data_a[row*p.KX + col]);
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        const float xi = float(data_a[a_base + i0*p.nb00]);
         sum[tid].x += xi;
         sum[tid].y += xi * xi;
     }
@@ -34,11 +34,11 @@ void main() {
         barrier();
     }
 
-    const float mean = sum[0].x / p.KX;
-    const float var = sum[0].y / p.KX - mean * mean;
+    const float mean = sum[0].x / p.ne00;
+    const float var = sum[0].y / p.ne00 - mean * mean;
     const float inv_std = inversesqrt(var + p.param1);
 
-    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
-        data_d[row*p.KX + col] = D_TYPE((float(data_a[row*p.KX + col]) - mean) * inv_std);
+    [[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
+        data_d[d_base + i0*p.nb10] = D_TYPE((float(data_a[a_base + i0*p.nb00]) - mean) * inv_std);
     }
 }

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sin(val));
-}

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sqrt(val));
-}

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

@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val * val);
-}

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

@@ -17,6 +17,30 @@ float op_neg(float x) {
     return -x;
 }
 
+float op_sqr(float x) {
+    return x * x;
+}
+
+float op_sqrt(float x) {
+    return sqrt(x);
+}
+
+float op_sin(float x) {
+    return sin(x);
+}
+
+float op_cos(float x) {
+    return cos(x);
+}
+
+float op_clamp(float x) {
+    return clamp(x, p.param1, p.param2);
+}
+
+float op_leaky_relu(float x) {
+    return max(x, 0.0f) + min(x, 0.0f) * p.param1;
+}
+
 float op_step(float x) {
     return x >= 0.0f ? 1.0f : 0.0f;
 }

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

@@ -11,6 +11,7 @@
 #include <future>
 #include <queue>
 #include <condition_variable>
+#include <atomic>
 #include <cstdio>
 #include <cstring>
 #include <cstdlib>
@@ -34,6 +35,9 @@
 
 std::mutex lock;
 std::vector<std::pair<std::string, std::string>> shader_fnames;
+// Set when any shader subprocess fails (non-zero exit / stderr / launch failure) so the
+// build is stopped instead of silently producing a broken libggml-vulkan. (issue #24393)
+static std::atomic<bool> compile_failed{false};
 std::locale c_locale("C");
 
 std::string GLSLC = "glslc";
@@ -78,7 +82,7 @@ enum MatMulIdType {
 
 namespace {
 
-void execute_command(std::vector<std::string>& command, std::string& stdout_str, std::string& stderr_str) {
+int execute_command(std::vector<std::string>& command, std::string& stdout_str, std::string& stderr_str) {
 #ifdef _WIN32
     HANDLE stdout_read, stdout_write;
     HANDLE stderr_read, stderr_write;
@@ -127,8 +131,11 @@ void execute_command(std::vector<std::string>& command, std::string& stdout_str,
     CloseHandle(stdout_read);
     CloseHandle(stderr_read);
     WaitForSingleObject(pi.hProcess, INFINITE);
+    DWORD exit_code = 1;
+    GetExitCodeProcess(pi.hProcess, &exit_code);
     CloseHandle(pi.hProcess);
     CloseHandle(pi.hThread);
+    return (int)exit_code;
 #else
     int stdout_pipe[2];
     int stderr_pipe[2];
@@ -175,7 +182,9 @@ void execute_command(std::vector<std::string>& command, std::string& stdout_str,
 
         close(stdout_pipe[0]);
         close(stderr_pipe[0]);
-        waitpid(pid, nullptr, 0);
+        int status = 0;
+        waitpid(pid, &status, 0);
+        return WIFEXITED(status) ? WEXITSTATUS(status) : -1;
     }
 #endif
 }
@@ -372,13 +381,14 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
         // }
         // std::cout << std::endl;
 
-        execute_command(cmd, stdout_str, stderr_str);
-        if (!stderr_str.empty()) {
-            std::cerr << "cannot compile " << name << "\n\n";
+        int exit_code = execute_command(cmd, stdout_str, stderr_str);
+        if (exit_code != 0 || !stderr_str.empty()) {
+            std::cerr << "cannot compile " << name << " (exit code " << exit_code << ")\n\n";
             for (const auto& part : cmd) {
                 std::cerr << part << " ";
             }
             std::cerr << "\n\n" << stderr_str << std::endl;
+            compile_failed = true;
             return;
         }
 
@@ -398,6 +408,7 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
         shader_fnames.push_back(std::make_pair(name, out_path));
     } catch (const std::exception& e) {
         std::cerr << "Error executing command for " << name << ": " << e.what() << std::endl;
+        compile_failed = true;
     }
 }
 
@@ -539,11 +550,9 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
     };
 
     // Shaders with f16 B_TYPE
-    string_to_spv(shader_name + "_f32_f16" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"},                                                     {"B_TYPE", "float16_t"},        {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f32_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f32_f16" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
 
-    string_to_spv(shader_name + "_f16" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"},                                                     {"B_TYPE", "float16_t"},            {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f16" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16},     {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f16" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict_f16), {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
 
     // bf16
     {
@@ -565,8 +574,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 #endif
         {
             if (!dot2) {
-                string_to_spv(shader_name + "_bf16",         source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"},                             {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
-                string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},  {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+                string_to_spv(shader_name + "_bf16", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"B_TYPE_SCALAR", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"B_TYPEV4", "bf16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}), fp16, coopmat, coopmat2, f16acc);
             }
         }
     }
@@ -583,8 +591,6 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
         }
 
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-        // For unaligned, load one at a time for f32/f16, or two at a time for quants
-        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant;
         // For aligned matmul loads
         std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
 
@@ -597,13 +603,11 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 
         // don't generate f32 variants for coopmat2
         if (!coopmat2) {
-            string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx,              source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx + "_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32" + dot2_sfx, source_name, merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"B_TYPE_SCALAR", "float"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
         if (tname != "f16" && tname != "f32") {
-            string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx,              source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx + "_aligned", source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16" + dot2_sfx, source_name,  merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"B_TYPE_SCALAR", "float16_t"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
@@ -850,21 +854,12 @@ void process_shaders() {
 
     string_to_spv("repeat_i32", "repeat.comp", {{"A_TYPE", "int32_t"}, {"D_TYPE", "int32_t"}});
     string_to_spv("repeat_back_f32", "repeat_back.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    string_to_spv("get_rows_back_f32", "get_rows_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}});
 
     string_to_spv("repeat_i16", "repeat.comp", {{"A_TYPE", "int16_t"}, {"D_TYPE", "int16_t"}});
 
     string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
-    string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
-    string_to_spv("sqrt_f32", "sqrt.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
-    string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
-    string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
-    string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-
     string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("concat_i8", "concat.comp", {{"A_TYPE", "uint8_t"}, {"B_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}});
@@ -891,6 +886,18 @@ void process_shaders() {
     string_to_spv("silu_f32",       "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_silu"}});
     string_to_spv("relu_f16",       "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_relu"}});
     string_to_spv("relu_f32",       "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_relu"}});
+    string_to_spv("sqr_f16",        "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_sqr"}});
+    string_to_spv("sqr_f32",        "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_sqr"}});
+    string_to_spv("sqrt_f16",       "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_sqrt"}});
+    string_to_spv("sqrt_f32",       "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_sqrt"}});
+    string_to_spv("sin_f16",        "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_sin"}});
+    string_to_spv("sin_f32",        "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_sin"}});
+    string_to_spv("cos_f16",        "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_cos"}});
+    string_to_spv("cos_f32",        "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_cos"}});
+    string_to_spv("clamp_f16",      "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_clamp"}});
+    string_to_spv("clamp_f32",      "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_clamp"}});
+    string_to_spv("leaky_relu_f16", "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_leaky_relu"}});
+    string_to_spv("leaky_relu_f32", "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_leaky_relu"}});
     string_to_spv("neg_f16",        "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_neg"}});
     string_to_spv("neg_f32",        "unary.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},     {"OP", "op_neg"}});
     string_to_spv("tanh_f16",       "unary.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}, {"OP", "op_tanh"}});
@@ -948,7 +955,6 @@ void process_shaders() {
     string_to_spv("geglu_quick_f16","geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("geglu_quick_f32","geglu_quick.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
-    string_to_spv("leaky_relu_f32", "leaky_relu.comp",  {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("silu_back_f32",  "silu_back.comp",   {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -1060,6 +1066,31 @@ void process_shaders() {
         }
     }
 
+    for (auto unroll : {false, true}) {
+        for (auto a_f16 : {false, true}) {
+            std::map<std::string, std::string> defines = {
+                {"A_TYPE", a_f16 ? "float16_t" : "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"},
+                {"UNROLL", unroll ? "[[unroll]]" : ""},
+            };
+            std::string name = std::string("conv3d") + (a_f16 ? "_f16" : "") + "_f32";
+            string_to_spv(name + (unroll ? "_unroll" : ""), "conv3d_mm.comp", defines);
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+            if (unroll) {
+                auto cm2_defines = defines;
+                cm2_defines["COOPMAT2"] = "1";
+                string_to_spv(name, "conv3d_mm.comp", cm2_defines, true, false, true);
+            }
+#endif
+#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+            if (unroll) {
+                auto cm1_defines = defines;
+                cm1_defines["COOPMAT"] = "1";
+                string_to_spv(name, "conv3d_mm.comp", cm1_defines, true, true, false);
+            }
+#endif
+        }
+    }
+
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
     string_to_spv("conv2d_dw_whcn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
@@ -1251,6 +1282,11 @@ int main(int argc, char** argv) {
 
     process_shaders();
 
+    if (compile_failed) {
+        std::cerr << "vulkan-shaders-gen: one or more shaders failed to compile" << std::endl;
+        return EXIT_FAILURE;
+    }
+
     write_output_files();
 
     return EXIT_SUCCESS;

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

@@ -4270,7 +4270,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_RMS_NORM:
         case GGML_OP_NORM:
         case GGML_OP_L2_NORM:
-            supports_op = op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32;
+            supports_op = (op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32) && ggml_is_contiguous_rows(src0);
             break;
         case GGML_OP_ROPE:
             supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;

scripts/snapdragon/adb/run-completion.sh

@@ -57,19 +57,25 @@ oppoll=
 opflt=
 [ "$OF" != "" ] && opflt="GGML_HEXAGON_OPFILTER=$OF"
 
+opfuse=
+[ "$OC" != "" ] && opfuse="GGML_HEXAGON_OPFUSION=$OC"
+
 vmem=
 [ "$VM" != "" ] && vmem="GGML_HEXAGON_VMEM=$VM"
 
 mbuf=
 [ "$MB" != "" ] && mbuf="GGML_HEXAGON_MBUF=$MB"
 
+mmsel=
+[ "$MM" != "" ] && mmsel="GGML_HEXAGON_MM_SELECT=$MM"
+
 set -x
 
 adb $adbserial $adbhost shell " \
   cd $basedir; ulimit -c unlimited;        \
     LD_LIBRARY_PATH=$basedir/$branch/lib   \
     ADSP_LIBRARY_PATH=$basedir/$branch/lib \
-    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $oppoll $opflt $vmem $mbuf \
+    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $oppoll $opflt $opfuse $vmem $mbuf $mmsel \
       ./$branch/bin/llama-completion --no-mmap -m $basedir/../gguf/$model \
          --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1                  \
          --ctx-size 8192 --ubatch-size 1024 -fa on                        \

scripts/snapdragon/adb/run-tool.sh

@@ -51,6 +51,12 @@ opqueue=
 oppoll=
 [ "$OP" != "" ] && oppoll="GGML_HEXAGON_OPPOLL=$OP"
 
+opfuse=
+[ "$OC" != "" ] && opfuse="GGML_HEXAGON_OPFUSION=$OC"
+
+mmsel=
+[ "$MM" != "" ] && mmsel="GGML_HEXAGON_MM_SELECT=$MM"
+
 set -x
 
 tool=$1; shift
@@ -59,5 +65,5 @@ adb $adbserial $adbhost shell " \
   cd $basedir; ulimit -c unlimited;        \
     LD_LIBRARY_PATH=$basedir/$branch/lib   \
     ADSP_LIBRARY_PATH=$basedir/$branch/lib \
-    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $oppoll ./$branch/bin/$tool $@ \
+    $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $oppoll $opfuse $mmsel ./$branch/bin/$tool $@ \
 "

scripts/snapdragon/ggml-hexagon-profile.py

@@ -26,7 +26,7 @@ COL_MAP = {
 }
 
 op_pattern = re.compile(
-    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+.*?\s+(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?"
+    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+.*?\s+:\s+(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?"
 )
 
 trace_pattern = re.compile(
@@ -93,9 +93,40 @@ def parse_log(file_path, pmu_index=None):
                 + int(ts_match.group('us'))
             )
 
-        op_match = op_pattern.search(line)
+        if "|" in line and "profile-op" in line:
+            parts = [p.strip() for p in line.split("|")]
+            prefix = parts[0]
+            prefix_match = re.search(r"profile-op\s+(?P<op_name>[A-Z_0-9+]+)", prefix)
+            if not prefix_match:
+                continue
+
+            if len(parts) == 7:
+                dims, types, timings = parts[2], parts[3], parts[6]
+            elif len(parts) == 6:
+                dims, types, timings = parts[2], parts[3], parts[5]
+            else:
+                continue
+
+            timing_match = re.search(
+                r"(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?",
+                timings
+            )
+            if not timing_match:
+                continue
+
+            op_match = timing_match
+            op_name = prefix_match.group("op_name")
+        else:
+            op_match = op_pattern.search(line)
+            if op_match:
+                op_name = op_match.group('op_name')
+                dims = op_match.group('dims').strip()
+                types = op_match.group('types').strip()
+            else:
+                op_match = None
+
         if op_match:
-            pmu_raw = op_match.group('pmu')
+            pmu_raw = op_match.group('pmu') if 'pmu' in op_match.groupdict() else None
             pmu_val = None
             if pmu_raw and pmu_index is not None:
                 try:
@@ -105,7 +136,7 @@ def parse_log(file_path, pmu_index=None):
                 except (ValueError, IndexError):
                     pmu_val = None
 
-            evt_raw = op_match.group('evt')
+            evt_raw = op_match.group('evt') if 'evt' in op_match.groupdict() else None
             evt_val = None
             if evt_raw:
                 try:
@@ -122,9 +153,9 @@ def parse_log(file_path, pmu_index=None):
             op_text = line[idx + 11:].strip() if idx != -1 else line.strip()
 
             current_op = {
-                'name':         op_match.group('op_name'),
-                'dims':         op_match.group('dims').strip(),
-                'types':        op_match.group('types').strip(),
+                'name':         op_name,
+                'dims':         dims,
+                'types':        types,
                 'op_text':      op_text,
                 'usec':         int(op_match.group('usec')),
                 'cycles':       int(op_match.group('cycles')),

scripts/snapdragon/ggml-hexagon-trace.py

@@ -12,7 +12,7 @@ from collections import defaultdict
 logger = logging.getLogger("ggml-hexagon-trace")
 
 op_pattern = re.compile(
-    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+(?P<strides>[\d:x\s\->!]+)\s+:\s+(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?"
+    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+(?P<strides>[\d:x\s\->!]+?)\s+:\s+(?:(?P<params>.*?)\s+:\s+)?(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?"
 )
 
 trace_pattern = re.compile(
@@ -66,7 +66,40 @@ def parse_log(file_path):
 
     for line in f:
         line_idx += 1
-        op_match = op_pattern.search(line)
+        if "|" in line and "profile-op" in line:
+            parts = [p.strip() for p in line.split("|")]
+            prefix = parts[0]
+            prefix_match = re.search(r"profile-op\s+(?P<op_name>[A-Z_0-9+]+)", prefix)
+            if not prefix_match:
+                continue
+
+            if len(parts) == 7:
+                dims, types, strides, params, timings = parts[2], parts[3], parts[4], parts[5], parts[6]
+            elif len(parts) == 6:
+                dims, types, strides, params, timings = parts[2], parts[3], parts[4], "", parts[5]
+            else:
+                continue
+
+            timing_match = re.search(
+                r"(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?",
+                timings
+            )
+            if not timing_match:
+                continue
+
+            op_match = timing_match
+            op_name = prefix_match.group("op_name")
+        else:
+            op_match = op_pattern.search(line)
+            if op_match:
+                op_name = op_match.group('op_name')
+                dims = op_match.group('dims').strip() if op_match.group('dims') else ''
+                types = op_match.group('types').strip() if op_match.group('types') else ''
+                strides = op_match.group('strides').strip() if op_match.group('strides') else ''
+                params = op_match.group('params').strip() if ('params' in op_match.groupdict() and op_match.group('params')) else ''
+            else:
+                op_match = None
+
         if op_match:
             cycles_start_raw = op_match.group('start')
             unwrapped_cycles_start = None
@@ -77,10 +110,11 @@ def parse_log(file_path):
             op_text = line[idx + 11:].strip() if idx != -1 else line.strip()
 
             current_op = {
-                'name':         op_match.group('op_name'),
-                'dims':         op_match.group('dims').strip() if op_match.group('dims') else '',
-                'types':        op_match.group('types').strip() if op_match.group('types') else '',
-                'strides':      op_match.group('strides').strip() if op_match.group('strides') else '',
+                'name':         op_name,
+                'dims':         dims,
+                'types':        types,
+                'strides':      strides,
+                'params':       params,
                 'op_text':      op_text,
                 'usec':         int(op_match.group('usec')),
                 'cycles':       int(op_match.group('cycles')),
@@ -397,6 +431,8 @@ def generate_perfetto_trace(filtered_ops, output_path):
                 debug_annots.append(make_debug_annotation("line", int_val=op['line_num']))
             if 'strides' in op and op['strides']:
                 debug_annots.append(make_debug_annotation("strides", string_val=op['strides']))
+            if 'params' in op and op['params'] and op['params'] != '----':
+                debug_annots.append(make_debug_annotation("params", string_val=op['params']))
 
             # Slice Begin
             evt_begin = make_track_event(1, 2, name=f"{op['name']} ({op['dims']})", category="operator", debug_annotations=debug_annots)

src/models/lfm2.cpp

@@ -190,7 +190,15 @@ llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_
         auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
         auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
 
-        bx = ggml_concat(ctx0, conv, bx, 0);
+        // causal prepends the state, non-causal pads symmetrically for a centered window
+        if (hparams.causal_attn) {
+            bx = ggml_concat(ctx0, conv, bx, 0);
+        } else {
+            const int64_t pad = (hparams.n_shortconv_l_cache - 1) / 2;
+            auto * left = ggml_cont(ctx0,
+                ggml_view_3d(ctx0, conv, pad, hparams.n_embd, n_seqs, conv->nb[1], conv->nb[2], (d_conv - pad) * conv->nb[0]));
+            bx = ggml_pad_ext(ctx0, ggml_concat(ctx0, left, bx, 0), 0, pad, 0, 0, 0, 0, 0, 0);
+        }
         GGML_ASSERT(bx->ne[0] > conv->ne[0]);
 
         // last d_conv columns is a new conv state
@@ -266,10 +274,12 @@ llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
-    cur = build_lora_mm(model.output, cur, model.output_s);
-    cb(cur, "result_output", -1);
+    if (!cparams.embeddings) {
+        cur = build_lora_mm(model.output, cur, model.output_s);
+        cb(cur, "result_output", -1);
 
-    res->t_logits = cur;
+        res->t_logits = cur;
+    }
 
     ggml_build_forward_expand(gf, cur);
 }

tests/CMakeLists.txt

@@ -199,7 +199,6 @@ llama_build_and_test(test-jinja.cpp)
 llama_test(test-jinja NAME test-jinja-py ARGS -py LABEL python)
 llama_build_and_test(test-chat-auto-parser.cpp WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})
 llama_build_and_test(test-chat-template.cpp)
-llama_build_and_test(test-json-partial.cpp)
 llama_build_and_test(test-log.cpp)
 llama_build_and_test(
     test-peg-parser.cpp

tests/test-backend-ops.cpp

@@ -3298,21 +3298,29 @@ struct test_norm : public test_case {
     const std::array<int64_t, 4> ne;
     const bool v; // whether a is a non-contiguous view
     const float eps;
+    const bool noncontig_rows;
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, v, eps);
+        return VARS_TO_STR5(type, ne, v, eps, noncontig_rows);
     }
 
     test_norm(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 5, 4, 3},
             bool v = false,
-            float eps = 1e-6f)
-        : type(type), ne(ne), v(v), eps(eps) {}
+            float eps = 1e-6f,
+            bool noncontig_rows = false)
+        : type(type), ne(ne), v(v), eps(eps), noncontig_rows(noncontig_rows) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        const std::array<int64_t, 4> ne_a = noncontig_rows ?
+            std::array<int64_t, 4>{ ne[1], ne[0], ne[2], ne[3] } : ne;
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
         ggml_set_name(a, "a");
 
+        if (noncontig_rows) {
+            a = ggml_permute(ctx, a, 1, 0, 2, 3);
+            ggml_set_name(a, "permuted a");
+        }
         if (v) {
             a = ggml_view_4d(ctx, a, a->ne[0]/2, a->ne[1]/2, a->ne[2]/2, a->ne[3]/2, a->nb[1], a->nb[2], a->nb[3], 0);
             ggml_set_name(a, "view of a");
@@ -6193,21 +6201,29 @@ struct test_l2_norm : public test_case {
     const std::array<int64_t, 4> ne;
     const float eps;
     bool v;
+    bool noncontig_rows;
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, eps, v);
+        return VARS_TO_STR5(type, ne, eps, v, noncontig_rows);
     }
 
     test_l2_norm(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 64, 320, 1},
             float eps = 1e-12f,
-            bool v = false)
-        : type(type), ne(ne), eps(eps), v(v) {}
+            bool v = false,
+            bool noncontig_rows = false)
+        : type(type), ne(ne), eps(eps), v(v), noncontig_rows(noncontig_rows) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        const std::array<int64_t, 4> ne_a = noncontig_rows ?
+            std::array<int64_t, 4>{ ne[1], ne[0], ne[2], ne[3] } : ne;
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
         ggml_set_name(a, "a");
 
+        if (noncontig_rows) {
+            a = ggml_permute(ctx, a, 1, 0, 2, 3);
+            ggml_set_name(a, "permuted a");
+        }
         if (v) {
             a = ggml_view_4d(ctx, a, a->ne[0]/2, a->ne[1]/2, a->ne[2]/2, a->ne[3]/2, a->nb[1], a->nb[2], a->nb[3], 0);
             ggml_set_name(a, "view of a");
@@ -8282,9 +8298,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                 test_cases.emplace_back(new test_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, v, eps));
                 test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, v, eps));
             }
+            test_cases.emplace_back(new test_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, false, eps, true));
             test_cases.emplace_back(new test_rms_norm_back(GGML_TYPE_F32, { n, 5, 4, 3 }, eps));
             test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false));
             test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, true));
+            test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false, true));
         }
     }
 
@@ -8402,6 +8420,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 2880, 32, 2880, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 2880, 32, 2880, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_MXFP4, GGML_TYPE_F32, 2880, 32, 2880, {1, 1}, {1, 1}));
+
+
 #if 0
     {
         // Test paths in OpenCL
@@ -8576,6 +8599,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     // gpt-oss issue with Vulkan mmq_id
     test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_MXFP4, GGML_TYPE_F32, 32, 2, false, 2880, 32, 2880));
+    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_Q4_0, GGML_TYPE_F32, 32, 2, false, 2880, 32, 2880));
 
     for (ggml_type type_a : all_types) {
         test_cases.emplace_back(new test_mul_mat_id(type_a, GGML_TYPE_F32, 4, 2, false, 64, 16, 3*ggml_blck_size(type_a)));
@@ -9272,6 +9296,34 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
 
+    struct conv3d_perf_case {
+        int N, IC, ID, IH, IW, OC, KD, KH, KW, s0, s1, s2, p0, p1, p2, d0, d1, d2;
+    };
+
+    const std::vector<conv3d_perf_case> conv3d_cases = {
+        {1,  320, 8,  38,  26, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1, 1280, 8,  38,  26, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1,  320, 8,  76,  52, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1, 1280, 8,  76,  52, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1,  320, 8, 152, 104, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+#if 0
+        // too slow on some devices
+        {1, 1280, 8, 152, 104, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1,  320, 4, 304, 208, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+        {1,  640, 4, 304, 208, 1280, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1},
+#endif
+    };
+
+    for (ggml_type kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+        for (const conv3d_perf_case & c : conv3d_cases) {
+            test_cases.emplace_back(new test_conv_3d(
+                c.N, c.IC, c.ID, c.IH, c.IW,
+                c.OC, c.KD, c.KH, c.KW,
+                c.s0, c.s1, c.s2, c.p0, c.p1, c.p2, c.d0, c.d1, c.d2,
+                kernel_type));
+        }
+    }
+
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1,   1, 1, 1}));
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 512, 1, 1}));
 

tests/test-json-partial.cpp

@@ -1,287 +0,0 @@
-#include "common.h"
-#include "json-partial.h"
-#include <exception>
-#include <iostream>
-#include <stdexcept>
-
-template <class T> static void assert_equals(const T & expected, const T & actual) {
-  if (expected != actual) {
-      std::cerr << "Expected: " << expected << std::endl;
-      std::cerr << "Actual: " << actual << std::endl;
-      std::cerr << std::flush;
-      throw std::runtime_error("Test failed");
-  }
-}
-
-static void test_json_healing() {
-  auto parse = [](const std::string & str) {
-      std::cerr << "# Parsing: " << str << '\n';
-      std::string::const_iterator it = str.begin();
-      const auto end = str.end();
-      common_json out;
-      std::string healing_marker = "$llama.cpp.json$";
-      if (common_json_parse(it, end, healing_marker, out)) {
-          auto dump = out.json.dump();
-          std::cerr << "Parsed: " << dump << '\n';
-          std::cerr << "Magic: " << out.healing_marker.json_dump_marker << '\n';
-          std::string result;
-          if (!out.healing_marker.json_dump_marker.empty()) {
-              auto i = dump.find(out.healing_marker.json_dump_marker);
-              if (i == std::string::npos) {
-                  throw std::runtime_error("Failed to find magic in dump " + dump + " (magic: " + out.healing_marker.json_dump_marker + ")");
-              }
-              result = dump.substr(0, i);
-          } else {
-            result = dump;
-          }
-          std::cerr << "Result: " << result << '\n';
-          if (string_starts_with(str, result)) {
-            std::cerr << "Failure!\n";
-          }
-        //   return dump;
-      } else {
-        throw std::runtime_error("Failed to parse: " + str);
-      }
-
-  };
-  auto parse_all = [&](const std::string & str) {
-      for (size_t i = 1; i < str.size(); i++) {
-          parse(str.substr(0, i));
-      }
-  };
-  parse_all("{\"a\": \"b\"}");
-  parse_all("{\"hey\": 1, \"ho\\\"ha\": [1]}");
-
-  parse_all("[{\"a\": \"b\"}]");
-
-  auto test = [&](const std::vector<std::string> & inputs, const std::string & expected, const std::string & expected_marker) {
-      for (const auto & input : inputs) {
-        common_json out;
-        assert_equals(true, common_json_parse(input, "$foo", out));
-        assert_equals<std::string>(expected, out.json.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true));
-        assert_equals<std::string>(expected_marker, out.healing_marker.json_dump_marker);
-      }
-  };
-  // No healing needed:
-  test(
-    {
-      R"([{"a":"b"}, "y"])",
-    },
-    R"([{"a":"b"},"y"])",
-    ""
-  );
-  // Partial literals can't be healed:
-  test(
-    {
-      R"([1)",
-      R"([tru)",
-      R"([n)",
-      R"([nul)",
-      R"([23.2)",
-    },
-    R"(["$foo"])",
-    R"("$foo)"
-  );
-  test(
-    {
-      R"({"a": 1)",
-      R"({"a": tru)",
-      R"({"a": n)",
-      R"({"a": nul)",
-      R"({"a": 23.2)",
-    },
-    R"({"a":"$foo"})",
-    R"("$foo)"
-  );
-  test(
-    {
-      R"({)",
-    },
-    R"({"$foo":1})",
-    R"("$foo)"
-  );
-  test(
-    {
-      R"([)",
-    },
-    R"(["$foo"])",
-    R"("$foo)"
-  );
-  // Healing right after a full literal
-  test(
-    {
-      R"(1 )",
-    },
-    R"(1)",
-    ""
-  );
-  test(
-    {
-      R"(true)",
-      R"(true )",
-    },
-    R"(true)",
-    ""
-  );
-  test(
-    {
-      R"(null)",
-      R"(null )",
-    },
-    R"(null)",
-    ""
-  );
-  test(
-    {
-      R"([1 )",
-    },
-    R"([1,"$foo"])",
-    R"(,"$foo)"
-  );
-  test(
-    {
-      R"([{})",
-      R"([{} )",
-    },
-    R"([{},"$foo"])",
-    R"(,"$foo)"
-  );
-  test(
-    {
-      R"([true)",
-    },
-    // TODO: detect the true/false/null literal was complete
-    R"(["$foo"])",
-    R"("$foo)"
-  );
-  test(
-    {
-      R"([true )",
-    },
-    R"([true,"$foo"])",
-    R"(,"$foo)"
-  );
-  test(
-    {
-      R"([true,)",
-    },
-    R"([true,"$foo"])",
-    R"("$foo)"
-  );
-  // Test nesting
-  test(
-    {
-      R"([{"a": [{"b": [{)",
-    },
-    R"([{"a":[{"b":[{"$foo":1}]}]}])",
-    R"("$foo)"
-  );
-  test(
-    {
-      R"([{"a": [{"b": [)",
-    },
-    R"([{"a":[{"b":["$foo"]}]}])",
-    R"("$foo)"
-  );
-
-  test(
-    {
-      R"([{"a": "b"})",
-      R"([{"a": "b"} )",
-    },
-    R"([{"a":"b"},"$foo"])",
-    R"(,"$foo)"
-  );
-  test(
-    {
-      R"([{"a": "b"},)",
-      R"([{"a": "b"}, )",
-    },
-    R"([{"a":"b"},"$foo"])",
-    R"("$foo)"
-  );
-  test(
-    {
-      R"({ "code)",
-    },
-    R"({"code$foo":1})",
-    R"($foo)"
-  );
-  test(
-    {
-      R"({ "code\)",
-    },
-    R"({"code\\$foo":1})",
-    R"(\$foo)"
-  );
-  test(
-    {
-      R"({ "code")",
-    },
-    R"({"code":"$foo"})",
-    R"(:"$foo)"
-  );
-  test(
-    {
-      R"({ "key")",
-    },
-    R"({"key":"$foo"})",
-    R"(:"$foo)"
-  );
-  // Test unicode escape sequences
-  test(
-    {
-      R"({"a":"\u)",
-    },
-    R"({"a":"\u0000$foo"})",
-    R"(0000$foo)"
-  );
-  test(
-    {
-      R"({"a":"\u00)",
-    },
-    R"({"a":"\u0000$foo"})",
-    R"(00$foo)"
-  );
-  test(
-    {
-      R"({"a":"\ud300)",
-    },
-    R"({"a":"\ud300$foo"})",
-    R"($foo)"
-  );
-  test(
-    {
-      R"({"a":"\ud800)",
-    },
-    R"({"a":"\ud800\udc00$foo"})",
-    R"(\udc00$foo)"
-  );
-  test(
-    {
-      R"({"a":"\ud800\)",
-    },
-    R"({"a":"\ud800\udc00$foo"})",
-    R"(udc00$foo)"
-  );
-  test(
-    {
-      R"({"a":"\ud800\u)",
-    },
-    R"({"a":"\ud800\udc00$foo"})",
-    R"(dc00$foo)"
-  );
-  test(
-    {
-      R"({"a":"\ud800\udc00)",
-    },
-    R"({"a":"\ud800\udc00$foo"})",
-    R"($foo)"
-  );
-}
-
-int main() {
-    test_json_healing();
-    std::cerr << "All tests passed.\n";
-    return 0;
-}

tools/mtmd/tests/test-deepseek-ocr.py

@@ -9,6 +9,7 @@ its output, and holds them against the HF model's scores.
 
 import argparse
 import logging
+import re
 import subprocess
 import sys
 import unicodedata
@@ -28,6 +29,12 @@ class ModelSpec:
     mmproj_arg: str
     model_default: str
     mmproj_default: str
+    prompt: str = "Free OCR. "
+    n_predict: int = 512
+    n_ctx: int | None = None
+    # Unlimited-OCR's "document parsing" prompt emits <|det|> grounding markup that
+    # the HF reference strips in result.md; drop it before scoring to match.
+    strip_grounding: bool = False
 
 
 @dataclass
@@ -63,6 +70,20 @@ MODELS = {
         model_default="gguf_models/deepseek-ai/deepseek-ocr-2-bf16.gguf",
         mmproj_default="gguf_models/deepseek-ai/mmproj-deepseek-ocr-2-bf16.gguf",
     ),
+    "unlimited": ModelSpec(
+        key="unlimited", label="Unlimited-OCR",
+        model_arg="--llama-model-unlimited", mmproj_arg="--mmproj-unlimited",
+        model_default="gguf_models/baidu/unlimited-ocr-bf16.gguf",
+        mmproj_default="gguf_models/baidu/mmproj-unlimited-ocr-bf16.gguf",
+        # "Free OCR." immediately emits EOS on this checkpoint; the HF reference
+        # (demo/unlimited_ocr_scores.py) uses "document parsing.", which grounds.
+        prompt="document parsing.",
+        # Grounding emits ~3x the tokens of plain OCR, so it needs a larger budget
+        # and context to reach the article body the ground truth covers.
+        n_predict=4096,
+        n_ctx=16384,
+        strip_grounding=True,
+    ),
 }
 
 CASES = [
@@ -82,9 +103,26 @@ CASES = [
         # is one pixel off and lands at ~0.69 instead.
         hf_cer=0.7761, hf_chrf=28.70, cer_tol=0.12, chrf_tol=8.0,
     ),
+    TestCase(
+        model_key="unlimited", label="single-view scan",
+        image="tools/mtmd/test-1.jpeg",
+        ground_truth="tools/mtmd/tests/test-1-ground-truth.txt",
+        # HF reference: Unlimited-OCR scoring (gundam, bf16) on this image/ground-truth.
+        # Decoder runs full MHA, not R-SWA; the band absorbs that gap + bf16 variance.
+        hf_cer=0.1869, hf_chrf=75.23, cer_tol=0.06, chrf_tol=6.0,
+    ),
 ]
 
 
+GROUNDING_TAG_RE = re.compile(r"<\|(ref|det)\|>.*?<\|/\1\|>", re.DOTALL)
+
+
+def strip_grounding(text: str) -> str:
+    """Drop <|ref|>..<|/ref|> / <|det|>..<|/det|> grounding markup, matching the
+    cleaned result.md the HF reference scores against."""
+    return GROUNDING_TAG_RE.sub("", text)
+
+
 def arg_dest(flag: str) -> str:
     return flag.lstrip("-").replace("-", "_")
 
@@ -129,19 +167,19 @@ def compute_chrf(expected: str, ocr_out: str) -> float:
     return CHRF().sentence_score(ocr_out, [expected]).score
 
 
-def run_mtmd_cli(model_path, mmproj_path, image_path, bin_path) -> str:
+def run_mtmd_cli(spec: "ModelSpec", model_path, mmproj_path, image_path, bin_path) -> str:
     """Run mtmd-cli on the image and return its output."""
     cmd = [
         str(bin_path),
         "-m", str(model_path),
         "--mmproj", str(mmproj_path),
         "--image", str(image_path),
-        "-p", "Free OCR. ",
+        "-p", spec.prompt,
         "--chat-template", "deepseek-ocr",
         "--temp", "0",
         "--flash-attn", "off",  # match the HF "eager" attention reference
         "--no-warmup",
-        "-n", "512",  # cap loops on hard images (KV would otherwise fill)
+        "-n", str(spec.n_predict),  # cap loops on hard images (KV would otherwise fill)
         # HF decodes with no_repeat_ngram_size; llama.cpp's analog is DRY.
         # Default DRY breakers include "\n", so they are cleared below.
         "--dry-multiplier", "0.8",
@@ -150,6 +188,8 @@ def run_mtmd_cli(model_path, mmproj_path, image_path, bin_path) -> str:
         "--dry-penalty-last-n", "-1",
         "--dry-sequence-breaker", "none",
     ]
+    if spec.n_ctx is not None:
+        cmd += ["-c", str(spec.n_ctx)]
     logger.debug(f"  command: {' '.join(cmd)}")
 
     try:
@@ -164,6 +204,8 @@ def run_mtmd_cli(model_path, mmproj_path, image_path, bin_path) -> str:
         raise RuntimeError(f"llama-mtmd-cli failed with code {result.returncode}")
 
     output = result.stdout.decode("utf-8", errors="replace").strip()
+    if spec.strip_grounding:
+        output = strip_grounding(output)
     if not output:
         raise RuntimeError("llama-mtmd-cli produced no output on stdout")
     logger.info(f"  output: {len(output)} chars")
@@ -193,7 +235,7 @@ def evaluate(case: "TestCase", expected: str, ocr_out: str) -> bool:
 
     logger.info("")
     logger.info("=" * 60)
-    logger.info("Free OCR evaluation:")
+    logger.info("OCR evaluation:")
     logger.info("=" * 60)
     logger.info(f"  CER               {cer:>7.4f}    (HF {case.hf_cer:.4f}, <= {case.cer_max:>7.4f}  -> {verdict(cer_pass)})")
     logger.info(f"  chrF (0-100)      {chrf:>7.2f}    (HF {case.hf_chrf:.2f}, >= {case.chrf_min:>7.2f}  -> {verdict(chrf_pass)})")
@@ -269,9 +311,9 @@ def main() -> int:
         expected = read_expected_text(ground_truth)
         logger.info(f"  Image: {case.image}")
         logger.info(f"  Expected text: {len(expected)} chars")
-        logger.info("  Running llama.cpp 'Free OCR'")
+        logger.info(f"  Running llama.cpp prompt {model_spec.prompt!r}")
         try:
-            ocr_out = run_mtmd_cli(model, mmproj, image, binary)
+            ocr_out = run_mtmd_cli(model_spec, model, mmproj, image, binary)
         except RuntimeError as e:
             logger.error(f"  Error: {e}")
             results[title] = False

tools/parser/debug-template-parser.cpp

@@ -40,6 +40,7 @@ struct debug_options {
     bool               enable_reasoning  = true;
     bool               debug_jinja       = false;
     bool               force_tool_call   = false;
+    bool               parallel_tool_calls = true;
     output_mode        mode              = output_mode::BOTH;
     input_message_type input_message     = input_message_type::NONE;
 };
@@ -87,6 +88,7 @@ static void print_usage(const char * program_name) {
     LOG_ERR("\nOptions:\n");
     LOG_ERR("  --no-tools              Disable tool definitions\n");
     LOG_ERR("  --force-tool-call       Set tool calls to forced\n");
+    LOG_ERR("  --parallel-tool-calls=0|1 Set parallel_tool_calls (default: 1)\n");
     LOG_ERR("  --generation-prompt=0|1 Set add_generation_prompt (default: 1)\n");
     LOG_ERR("  --enable-reasoning=0|1  Enable reasoning parsing (default: 1)\n");
     LOG_ERR("  --output=MODE           Output mode: analysis, template, both (default: both)\n");
@@ -121,6 +123,8 @@ static bool parse_options(int argc, char ** argv, debug_options & opts) {
             opts.debug_jinja = true;
         } else if (arg == "--no-tools") {
             opts.with_tools = false;
+        } else if (arg.rfind("--parallel-tool-calls=", 0) == 0) {
+            opts.parallel_tool_calls = parse_bool_option(arg.substr(22));
         } else if (arg.rfind("--generation-prompt=", 0) == 0) {
             opts.generation_prompt = parse_bool_option(arg.substr(20));
         } else if (arg.rfind("--enable-reasoning=", 0) == 0) {
@@ -349,7 +353,7 @@ static autoparser::generation_params prepare_params(const debug_options & opts,
         params.tools       = json();
         params.tool_choice = COMMON_CHAT_TOOL_CHOICE_NONE;
     }
-    params.parallel_tool_calls = false;
+    params.parallel_tool_calls = opts.parallel_tool_calls;
     return params;
 }
 

tools/server/server-context.cpp

@@ -89,7 +89,9 @@ struct server_batch {
     }
 
     ~server_batch() {
-        llama_batch_free(batch);
+        if (batch.token != nullptr) {
+            llama_batch_free(batch);
+        }
     }
 
     void init(int32_t n_tokens_alloc) {
@@ -1215,6 +1217,10 @@ private:
             cparams.ctx_other = ctx_tgt;
 
             ctx_dft.reset(llama_init_from_model(model_dft.get(), cparams));
+            if (ctx_dft == nullptr) {
+                SRV_ERR("%s", "failed to create draft context\n");
+                return false;
+            }
 
             params_base.speculative.draft.ctx_tgt = ctx_tgt;
             params_base.speculative.draft.ctx_dft = ctx_dft.get();

tools/ui/.gitignore

@@ -28,10 +28,9 @@ vite.config.ts.timestamp-*
 # PWA Artifacts
 apple-splash-*.png
 apple-touch-icon-*.png
-favicon.ico
-favicon-dark.ico
 maskable-icon-*.png
 pwa-*.png
+static/favicon*
 
 # Storybook
 *storybook.log

tools/ui/package-lock.json

@@ -35,7 +35,7 @@
 				"bits-ui": "2.18.1",
 				"clsx": "2.1.1",
 				"dexie": "4.4.3",
-				"dompurify": "3.4.5",
+				"dompurify": "3.4.11",
 				"eslint": "9.39.4",
 				"eslint-config-prettier": "10.1.8",
 				"eslint-plugin-storybook": "10.4.2",
@@ -8653,9 +8653,9 @@
 			"peer": true
 		},
 		"node_modules/dompurify": {
-			"version": "3.4.5",
-			"resolved": "https://registry.npmjs.org/dompurify/-/dompurify-3.4.5.tgz",
-			"integrity": "sha512-OrwIBKsdNSVEeubdJ1HBv/wNENRM9ytAVCv7YXt//A3vPdVMNuACRqK9mXCGCBW2ln7BT/A4X0jXHo2Gu89miA==",
+			"version": "3.4.11",
+			"resolved": "https://registry.npmjs.org/dompurify/-/dompurify-3.4.11.tgz",
+			"integrity": "sha512-zhlUV12GsaRzMsf9q5M254YhA4+VuF0fG+QFqu6aYpoGlKtz+w8//jBcGVYBgQkR5GHjUomejY84AV+/uPbWdw==",
 			"dev": true,
 			"license": "(MPL-2.0 OR Apache-2.0)",
 			"optionalDependencies": {
@@ -10226,9 +10226,9 @@
 			}
 		},
 		"node_modules/hono": {
-			"version": "4.12.23",
-			"resolved": "https://registry.npmjs.org/hono/-/hono-4.12.23.tgz",
-			"integrity": "sha512-eIaZ9qDgu7XV0pxOCrg7/WhnQ6Ivm22UcxhXx/A3dcbqbbYgBEkc6e/J/s7j2tS96zoB0S9VBdLwQNCWwUo4LA==",
+			"version": "4.12.26",
+			"resolved": "https://registry.npmjs.org/hono/-/hono-4.12.26.tgz",
+			"integrity": "sha512-uyZtpnYxM9CmQ7QsQknM4zN8EftNqhON1qYeIKM0Se67CCEe2c44xyGURwB0axX2fBDu1dqHrHAc1hmNT8ITkw==",
 			"dev": true,
 			"license": "MIT",
 			"engines": {

tools/ui/package.json

@@ -54,7 +54,7 @@
 		"bits-ui": "2.18.1",
 		"clsx": "2.1.1",
 		"dexie": "4.4.3",
-		"dompurify": "3.4.5",
+		"dompurify": "3.4.11",
 		"eslint": "9.39.4",
 		"eslint-config-prettier": "10.1.8",
 		"eslint-plugin-storybook": "10.4.2",

tools/ui/pwa-assets-dark.config.ts

@@ -1,4 +1,10 @@
 import { defineConfig } from '@vite-pwa/assets-generator/config';
+import { FAVICON_COLORS, PWA_ASSET_GENERATOR } from './src/lib/constants/pwa';
+import { writeThemeFavicons } from './scripts/favicon-colorize';
+
+writeThemeFavicons(FAVICON_COLORS.LIGHT, FAVICON_COLORS.DARK, {
+	padding: PWA_ASSET_GENERATOR.FAVICON_PADDING
+});
 
 export default defineConfig({
 	headLinkOptions: {
@@ -7,7 +13,8 @@ export default defineConfig({
 	preset: {
 		transparent: {
 			sizes: [],
-			favicons: [[48, 'favicon-dark.ico']]
+			favicons: [[48, 'favicon-dark.ico']],
+			padding: PWA_ASSET_GENERATOR.FAVICON_PADDING
 		},
 		maskable: {
 			sizes: []

tools/ui/pwa-assets.config.ts

@@ -5,15 +5,32 @@ import {
 } from '@vite-pwa/assets-generator/config';
 import { readFileSync } from 'node:fs';
 import { resolve } from 'node:path';
-import { THEME_COLORS, PWA_GENERATOR_DEVICES, PWA_ASSET_GENERATOR } from './src/lib/constants/pwa';
+import {
+	THEME_COLORS,
+	PWA_GENERATOR_DEVICES,
+	PWA_ASSET_GENERATOR,
+	FAVICON_COLORS
+} from './src/lib/constants/pwa';
 import { SplashOrientation } from './src/lib/enums/splash.enums';
+import { writeThemeFavicons } from './scripts/favicon-colorize';
+
+writeThemeFavicons(FAVICON_COLORS.LIGHT, FAVICON_COLORS.DARK, {
+	padding: PWA_ASSET_GENERATOR.FAVICON_PADDING
+});
 
 export default defineConfig({
 	headLinkOptions: {
 		preset: PWA_ASSET_GENERATOR.LINK_PRESET
 	},
 	preset: combinePresetAndAppleSplashScreens(
-		minimal2023Preset,
+		{
+			...minimal2023Preset,
+			// tiny margin so favicon.ico / pwa-*.png breathe inside the canvas
+			transparent: {
+				...minimal2023Preset.transparent,
+				padding: PWA_ASSET_GENERATOR.FAVICON_PADDING
+			}
+		},
 		{
 			padding: PWA_ASSET_GENERATOR.SPLASH_PADDING,
 			resizeOptions: {

tools/ui/scripts/favicon-colorize.ts

@@ -0,0 +1,107 @@
+import { mkdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { dirname, resolve } from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const HERE = dirname(fileURLToPath(import.meta.url));
+const PROJECT_ROOT = resolve(HERE, '..');
+
+const DEFAULT_LOGO = resolve(PROJECT_ROOT, 'src/lib/assets/logo.svg');
+const DEFAULT_OUT_DIR = resolve(PROJECT_ROOT, 'static');
+const DEFAULT_OUT_LIGHT = resolve(DEFAULT_OUT_DIR, 'favicon.svg');
+const DEFAULT_OUT_DARK = resolve(DEFAULT_OUT_DIR, 'favicon-dark.svg');
+
+const CURRENT_COLOR = 'currentColor';
+
+export interface ColorizedFavicon {
+	light: string;
+	dark: string;
+}
+
+export interface WriteThemeFaviconsOptions {
+	sourcePath?: string;
+	lightOutPath?: string;
+	darkOutPath?: string;
+	/**
+	 * Fraction of the icon (0..1) to leave as an even margin on each side.
+	 * Applied by wrapping the inner content in a `<g transform="...">` so the
+	 * source `src/lib/assets/logo.svg` is not modified. Pass 0 to disable.
+	 */
+	padding?: number;
+}
+
+/**
+ * Replace every `currentColor` occurrence in the SVG with the given color.
+ * Pure: no filesystem access, so it is straightforward to unit-test.
+ */
+export function colorizeFaviconSvg(
+	svg: string,
+	lightColor: string,
+	darkColor: string
+): ColorizedFavicon {
+	return {
+		light: svg.replaceAll(CURRENT_COLOR, lightColor),
+		dark: svg.replaceAll(CURRENT_COLOR, darkColor)
+	};
+}
+
+/**
+ * Shrink the inner SVG content uniformly and re-center it so `padding` (a
+ * 0..1 fraction) is reserved as equal margin on each side. Returns the input
+ * unchanged for non-positive padding, missing/invalid `viewBox`, or unexpected
+ * markup so the caller always gets a renderable SVG.
+ */
+export function padFaviconSvg(svg: string, padding: number): string {
+	if (!(padding > 0) || padding >= 1) return svg;
+
+	const viewBoxMatch = svg.match(/viewBox\s*=\s*["']([^"']+)["']/i);
+	if (!viewBoxMatch) return svg;
+
+	const parts = viewBoxMatch[1]
+		.trim()
+		.split(/[\s,]+/)
+		.map(Number);
+	if (parts.length !== 4 || parts.some((n) => !Number.isFinite(n))) return svg;
+
+	const [, , width, height] = parts;
+	if (width <= 0 || height <= 0) return svg;
+
+	const scale = 1 - padding;
+	const translateX = (padding * width) / 2;
+	const translateY = (padding * height) / 2;
+
+	const openTagStart = svg.search(/<svg\b/i);
+	if (openTagStart === -1) return svg;
+	const openTagEnd = svg.indexOf('>', openTagStart);
+	if (openTagEnd === -1) return svg;
+	const closeStart = svg.lastIndexOf('</svg');
+	if (closeStart === -1 || closeStart <= openTagEnd) return svg;
+
+	const openTag = svg.slice(0, openTagEnd + 1);
+	const inner = svg.slice(openTagEnd + 1, closeStart);
+	const closeTag = svg.slice(closeStart);
+
+	const group = `<g transform="translate(${translateX} ${translateY}) scale(${scale})">`;
+	return `${openTag}${group}${inner}</g>${closeTag}`;
+}
+
+/**
+ * Read `src/lib/assets/logo.svg`, colorize it for both themes, and write
+ * the results to the static directory so the PWA asset generator can consume
+ * them. Paths can be overridden for tests.
+ */
+export function writeThemeFavicons(
+	lightColor: string,
+	darkColor: string,
+	{
+		sourcePath = DEFAULT_LOGO,
+		lightOutPath = DEFAULT_OUT_LIGHT,
+		darkOutPath = DEFAULT_OUT_DARK,
+		padding = 0
+	}: WriteThemeFaviconsOptions = {}
+): void {
+	const source = readFileSync(sourcePath, 'utf-8');
+	const { light, dark } = colorizeFaviconSvg(source, lightColor, darkColor);
+	mkdirSync(dirname(lightOutPath), { recursive: true });
+	writeFileSync(lightOutPath, padFaviconSvg(light, padding));
+	writeFileSync(darkOutPath, padFaviconSvg(dark, padding));
+}

tools/ui/src/app.css

@@ -48,6 +48,7 @@
 
 	--chat-form-area-height: 8rem;
 	--chat-form-area-offset: 2rem;
+	--chat-form-padding-top: 6rem;
 	--max-message-height: max(24rem, min(80dvh, calc(100dvh - var(--chat-form-area-height) - 12rem)));
 }
 
@@ -55,6 +56,7 @@
 	:root {
 		--chat-form-area-height: 24rem;
 		--chat-form-area-offset: 12rem;
+		--chat-form-padding-top: 6rem;
 	}
 }
 
@@ -141,7 +143,6 @@
 		@apply bg-background text-foreground;
 		scrollbar-width: thin;
 		scrollbar-gutter: stable;
-		overflow: hidden; /* Added due to Mermaid rendering somehow causing the double scrollbar */
 	}
 
 	/* Global scrollbar styling - visible only on hover */
@@ -193,3 +194,7 @@
 		scrollbar-width: none;
 	}
 }
+
+.mermaidTooltip {
+	display: none !important;
+}

tools/ui/src/lib/actions/fade-in-view.svelte.ts

@@ -10,9 +10,9 @@ import { isElementInViewport } from '$lib/utils/viewport';
  */
 export function fadeInView(
 	node: HTMLElement,
-	options: { duration?: number; y?: number; skipIfVisible?: boolean } = {}
+	options: { duration?: number; y?: number; delay?: number; skipIfVisible?: boolean } = {}
 ) {
-	const { duration = 300, y = 0, skipIfVisible = false } = options;
+	const { duration = 300, y = 0, delay = 0, skipIfVisible = false } = options;
 
 	if (skipIfVisible && isElementInViewport(node)) {
 		return;
@@ -27,10 +27,12 @@ export function fadeInView(
 			(entries) => {
 				for (const entry of entries) {
 					if (entry.isIntersecting) {
-						requestAnimationFrame(() => {
-							node.style.opacity = '1';
-							node.style.transform = 'translateY(0)';
-						});
+						setTimeout(() => {
+							requestAnimationFrame(() => {
+								node.style.opacity = '1';
+								node.style.transform = 'translateY(0)';
+							});
+						}, delay);
 						observer.disconnect();
 					}
 				}

tools/ui/src/lib/assets/logo.svg

@@ -0,0 +1,7 @@
+<svg width="512" height="512" viewBox="0 0 512 512" fill="none" xmlns="http://www.w3.org/2000/svg">
+    <path d="M244.95 8C215.233 8 187.774 23.8591 172.923 49.5999L95.6009 183.625C60.2162 244.959 104.481 321.6 175.29 321.6H208L316.977 132.708C348.959 77.2719 308.95 8 244.95 8ZM208 321.6H351.947C415.982 321.6 456.013 390.91 424.013 446.377C409.155 472.132 381.681 488 351.947 488H271.29C200.481 488 156.216 411.359 191.601 350.026L208 321.6Z" fill="currentColor"/>
+    <path d="M208 321.6H16L106.462 164.8L208 321.6Z" fill="currentColor"/>
+    <path d="M388.923 8L208 321.6L253.6 8H388.923Z" fill="currentColor"/>
+    <path d="M304 488H112L202.462 331.2L304 488Z" fill="currentColor"/>
+    <path d="M496 321.6H208L419.399 454.4L496 321.6Z" fill="currentColor"/>
+</svg>

tools/ui/src/lib/components/app/actions/ActionIcon.svelte

@@ -8,12 +8,13 @@
 		ariaLabel?: string;
 		class?: string;
 		disabled?: boolean;
+		href?: string;
 		icon: Component;
 		iconSize?: string;
-		onclick: (e?: MouseEvent) => void;
+		onclick?: (e?: MouseEvent) => void;
 		size?: ButtonSize;
 		stopPropagationOnClick?: boolean;
-		tooltip: string;
+		tooltip?: string;
 		variant?: ButtonVariant;
 		tooltipSide?: TooltipSide;
 	}
@@ -22,6 +23,7 @@
 		icon,
 		tooltip,
 		variant = 'ghost',
+		href = '',
 		size = 'sm',
 		class: className = '',
 		disabled = false,
@@ -31,34 +33,49 @@
 		onclick,
 		ariaLabel
 	}: Props = $props();
+
+	let innerWidth = $state(0);
+	const showTooltip = $derived(!!tooltip && innerWidth > 768);
 </script>
 
-<Tooltip.Root>
-	<Tooltip.Trigger>
-		<!-- prevent another nested button element -->
-		{#snippet child({ props })}
-			<Button
-				{...props}
-				{variant}
-				{size}
-				{disabled}
-				onclick={(e: MouseEvent) => {
-					if (stopPropagationOnClick) e.stopPropagation();
+{#snippet button(props = {})}
+	<Button
+		{...props}
+		{href}
+		{variant}
+		{size}
+		{disabled}
+		onclick={(e: MouseEvent) => {
+			if (stopPropagationOnClick) e.stopPropagation();
 
-					onclick?.(e);
-				}}
-				class="h-6 w-6 p-0 {className} flex hover:bg-transparent data-[state=open]:bg-transparent!"
-				aria-label={ariaLabel || tooltip}
-			>
-				{#if icon}
-					{@const IconComponent = icon}
-					<IconComponent class={iconSize} />
-				{/if}
-			</Button>
-		{/snippet}
-	</Tooltip.Trigger>
+			onclick?.(e);
+		}}
+		class="h-6 w-6 p-0 {className} flex hover:bg-transparent data-[state=open]:bg-transparent!"
+		aria-label={ariaLabel || tooltip}
+	>
+		{#if icon}
+			{@const IconComponent = icon}
 
-	<Tooltip.Content side={tooltipSide}>
-		<p>{tooltip}</p>
-	</Tooltip.Content>
-</Tooltip.Root>
+			<IconComponent class={iconSize} />
+		{/if}
+	</Button>
+{/snippet}
+
+{#if showTooltip}
+	<Tooltip.Root>
+		<Tooltip.Trigger>
+			<!-- prevent another nested button element -->
+			{#snippet child({ props })}
+				{@render button(props)}
+			{/snippet}
+		</Tooltip.Trigger>
+
+		<Tooltip.Content side={tooltipSide}>
+			<p>{tooltip}</p>
+		</Tooltip.Content>
+	</Tooltip.Root>
+{:else}
+	{@render button({ href })}
+{/if}
+
+<svelte:window bind:innerWidth />

tools/ui/src/lib/components/app/chat/ChatForm/ChatForm.svelte

@@ -494,7 +494,7 @@
 	/>
 
 	<div
-		class="{INPUT_CLASSES} overflow-hidden rounded-3xl backdrop-blur-md {disabled
+		class="{INPUT_CLASSES} overflow-hidden rounded-4xl md:rounded-3xl backdrop-blur-md {disabled
 			? 'cursor-not-allowed opacity-60'
 			: ''}"
 		data-slot="input-area"
@@ -510,7 +510,7 @@
 		/>
 
 		<div
-			class="flex-column relative min-h-[48px] items-center rounded-3xl py-2 pb-2.25 shadow-sm transition-all focus-within:shadow-md md:!py-3"
+			class="flex-column relative min-h-12 items-center rounded-4xl md:rounded-3xl py-2 pb-2.25 shadow-sm transition-all focus-within:shadow-md md:py-3!"
 			onpaste={handlePaste}
 		>
 			<ChatFormTextarea

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddButton.svelte

@@ -15,7 +15,7 @@
 <Tooltip.Root>
 	<Tooltip.Trigger class="w-full">
 		<Button
-			class="file-upload-button h-8 w-8 rounded-full p-0"
+			class="file-upload-button md:h-8 md:w-8 h-9 w-9 rounded-full p-0"
 			{disabled}
 			{onclick}
 			variant="secondary"

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddSheet.svelte

@@ -15,6 +15,7 @@
 	import { McpLogo } from '$lib/components/app';
 	import { PencilRuler, ChevronDown, ChevronRight } from '@lucide/svelte';
 	import { HealthCheckStatus } from '$lib/enums';
+	import { AttachmentAction } from '$lib/enums/attachment.enums';
 
 	interface Props {
 		class?: string;
@@ -270,14 +271,22 @@
 					</Collapsible.Root>
 				{/if}
 
-				<button type="button" class={sheetItemClass} onclick={onSystemPromptClick}>
+				<button
+					type="button"
+					class={sheetItemClass}
+					onclick={() => attachmentMenu.callbacks[AttachmentAction.SYSTEM_PROMPT_CLICK]()}
+				>
 					<MessageSquare class="h-4 w-4 shrink-0" />
 
 					<span>System Message</span>
 				</button>
 
 				{#if hasMcpPromptsSupport}
-					<button type="button" class={sheetItemClass} onclick={onMcpPromptClick}>
+					<button
+						type="button"
+						class={sheetItemClass}
+						onclick={() => attachmentMenu.callbacks[AttachmentAction.MCP_PROMPT_CLICK]()}
+					>
 						<Zap class="h-4 w-4 shrink-0" />
 
 						<span>MCP Prompt</span>
@@ -285,7 +294,11 @@
 				{/if}
 
 				{#if hasMcpResourcesSupport}
-					<button type="button" class={sheetItemClass} onclick={onMcpResourcesClick}>
+					<button
+						type="button"
+						class={sheetItemClass}
+						onclick={() => attachmentMenu.callbacks[AttachmentAction.MCP_RESOURCES_CLICK]()}
+					>
 						<FolderOpen class="h-4 w-4 shrink-0" />
 
 						<span>MCP Resources</span>

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionsAdd.svelte

@@ -42,6 +42,7 @@
 		{hasMcpPromptsSupport}
 		{hasMcpResourcesSupport}
 		{onFileUpload}
+		{onSystemPromptClick}
 		{onMcpPromptClick}
 		{onMcpResourcesClick}
 	>

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormActions/ChatFormActionSubmit.svelte

@@ -20,7 +20,7 @@
 		type="submit"
 		disabled={isDisabled}
 		class={[
-			'h-8 w-8 rounded-full p-0',
+			'md:h-8 md:w-8 h-9 w-9 rounded-full p-0',
 			showErrorState &&
 				'bg-red-400/10 text-red-400 hover:bg-red-400/20 hover:text-red-400 disabled:opacity-100'
 		]}

tools/ui/src/lib/components/app/chat/ChatForm/ChatFormTextarea.svelte

@@ -1,4 +1,5 @@
 <script lang="ts">
+	import { isMobile } from '$lib/stores/viewport.svelte';
 	import { autoResizeTextarea } from '$lib/utils';
 	import { onMount } from 'svelte';
 
@@ -37,7 +38,9 @@
 	}
 
 	export function focus() {
-		textareaElement?.focus();
+		if (isMobile.current) return;
+
+		textareaElement?.focus({ preventScroll: true });
 	}
 
 	export function resetHeight() {

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessage.svelte

@@ -231,7 +231,7 @@
 			editedContent = message.content;
 		}
 
-		textareaElement?.focus();
+		textareaElement?.focus({ preventScroll: true });
 		editedExtras = message.extra ? [...message.extra] : [];
 		editedUploadedFiles = [];
 
@@ -324,7 +324,7 @@
 	}
 </script>
 
-<div use:fadeInView>
+<div use:fadeInView class="chat-message">
 	{#if message.role === MessageRole.SYSTEM}
 		<ChatMessageSystem
 			bind:textareaElement

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageAssistant/ChatMessageAssistant.svelte

@@ -180,6 +180,9 @@
 
 	let displayedModel = $derived(message.model ?? null);
 
+	// model being switched to while it loads, so the selector bar tracks it
+	let pendingModel = $state<string | null>(null);
+
 	let isCurrentlyLoading = $derived(isLoading());
 	let isStreaming = $derived(isChatStreaming());
 	let hasNoContent = $derived(!message?.content?.trim());
@@ -207,6 +210,42 @@
 			isLastAssistantMessage
 	);
 
+	let assistantEl: HTMLDivElement | undefined = $state();
+	let lastUserMessageHeight = $state(0);
+	let assistantMarginTop = $state(0);
+
+	$effect(() => {
+		if (!assistantEl) return;
+
+		assistantMarginTop = Math.round(parseFloat(getComputedStyle(assistantEl).marginTop));
+
+		const chatMessageEl = assistantEl.closest('.chat-message');
+		const previousChatMessage = chatMessageEl?.previousElementSibling;
+		const userMessageEl = previousChatMessage?.querySelector(
+			'.chat-message-user'
+		) as HTMLElement | null;
+
+		if (!userMessageEl) {
+			lastUserMessageHeight = 0;
+			return;
+		}
+
+		const updateHeight = () => {
+			const rect = userMessageEl.getBoundingClientRect();
+			const marginTop = Math.round(parseFloat(getComputedStyle(userMessageEl).marginTop));
+			lastUserMessageHeight = Math.round(rect.height + marginTop);
+		};
+
+		updateHeight();
+
+		const resizeObserver = new ResizeObserver(updateHeight);
+		resizeObserver.observe(userMessageEl);
+
+		return () => {
+			resizeObserver.disconnect();
+		};
+	});
+
 	function handleCopyModel() {
 		void copyToClipboard(displayedModel ?? '');
 	}
@@ -219,12 +258,17 @@
 </script>
 
 <div
-	class="text-md group w-full leading-7.5 {className}"
+	bind:this={assistantEl}
+	class="chat-message-assistant text-md group w-full leading-7.5 {className}"
+	style:--last-user-message-height={lastUserMessageHeight > 0
+		? `${lastUserMessageHeight}px`
+		: undefined}
+	style:--assistant-margin-top={assistantMarginTop > 0 ? `${assistantMarginTop}px` : undefined}
 	role="group"
 	aria-label="Assistant message with actions"
 >
 	{#if showProcessingInfoTop}
-		<div class="mt-6 w-full max-w-[48rem]" in:fade>
+		<div class="mt-6 w-full max-w-3xl" in:fade>
 			<div class="processing-container">
 				<span class="processing-text">
 					{modelLoadingText ??
@@ -257,7 +301,7 @@
 	{/if}
 
 	{#if showProcessingInfoBottom}
-		<div class="mt-4 w-full max-w-[48rem]" in:fade>
+		<div class="mt-4 w-full max-w-3xl" in:fade>
 			<div class="processing-container">
 				<span class="processing-text">
 					{modelLoadingText ??
@@ -277,13 +321,19 @@
 			>
 				{#if isRouter}
 					<ModelsSelectorDropdown
-						currentModel={displayedModel}
+						currentModel={pendingModel ?? displayedModel}
 						disabled={isLoading()}
 						onModelChange={async (modelId: string, modelName: string) => {
 							const status = modelsStore.getModelStatus(modelId);
 
 							if (status !== ServerModelStatus.LOADED) {
-								await modelsStore.loadModel(modelId);
+								pendingModel = modelId;
+
+								try {
+									await modelsStore.loadModel(modelId);
+								} finally {
+									pendingModel = null;
+								}
 							}
 
 							onRegenerate(modelName);
@@ -351,6 +401,23 @@
 </div>
 
 <style>
+	:global(.chat-message):last-child .chat-message-assistant {
+		--assistant-min-height-offset: calc(
+			var(--last-user-message-height, 19rem) + var(--chat-form-height, 6rem) +
+				var(--chat-form-bottom-position, 0.5rem) + var(--chat-form-padding-top, 6rem) +
+				var(--assistant-margin-top, 3rem)
+		);
+		min-height: calc(100dvh - var(--assistant-min-height-offset));
+
+		@media (width > 768px) {
+			--assistant-min-height-offset: calc(
+				var(--last-user-message-height, 18rem) + var(--chat-form-height, 6rem) +
+					var(--chat-form-bottom-position, 1rem) + var(--chat-form-padding-top, 6rem) +
+					var(--assistant-margin-top, 3rem)
+			);
+		}
+	}
+
 	.processing-container {
 		display: flex;
 		flex-direction: column;

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUser.svelte

@@ -48,7 +48,7 @@
 
 <div
 	aria-label="User message with actions"
-	class="group flex flex-col items-end gap-3 md:gap-2 {className}"
+	class="chat-message-user group flex flex-col items-end gap-3 md:gap-2 {className}"
 	role="group"
 >
 	{#if editCtx.isEditing}

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUserBubble.svelte

@@ -19,7 +19,7 @@
 		renderMarkdown = false,
 		textColorClass = 'text-foreground',
 		cardBgClass = 'dark:bg-primary/15',
-		maxHeightStyle = 'max-height: var(--max-message-height);'
+		maxHeightStyle = ''
 	}: Props = $props();
 
 	let isMultiline = $state(false);
@@ -59,7 +59,7 @@
 
 {#if content.trim()}
 	<Card
-		class="max-w-[80%] overflow-y-auto rounded-[1.125rem] border-none bg-primary/5 px-3.75 py-1.5 {textColorClass} backdrop-blur-md data-[multiline]:py-2.5 {cardBgClass}"
+		class="chat-message-user-bubble max-w-[80%] overflow-y-auto rounded-[1.125rem] border-none bg-primary/5 px-3.75 py-1.5 {textColorClass} backdrop-blur-md data-multiline:py-2.5 {cardBgClass}"
 		data-multiline={isMultiline ? '' : undefined}
 		style="{maxHeightStyle} overflow-wrap: anywhere; word-break: break-word;"
 	>

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessages.svelte

@@ -37,6 +37,7 @@
 	let allConversationMessages = $state<DatabaseMessage[]>([]);
 	let isVisible = $state(false);
 	let previousConversationId = $state<string | null>(null);
+	let previousRouteId = $state<string | null>(null);
 
 	const currentConfig = config();
 
@@ -157,8 +158,9 @@
 		});
 	});
 
-	beforeNavigate(() => {
+	beforeNavigate((navigation) => {
 		isVisible = false;
+		previousRouteId = navigation.from?.route.id ?? null;
 	});
 
 	afterNavigate(() => {
@@ -249,12 +251,13 @@
 </script>
 
 <div
-	class="transition-opacity delay-300 duration-500 ease-out
-		{isVisible ? 'opacity-100' : 'opacity-0'}"
+	class="transition-opacity duration-500 ease-out
+		{isVisible ? 'opacity-100' : 'opacity-0'}
+		{previousRouteId === '/(chat)/chat/[id]' ? '' : 'delay-300'}"
 >
 	{#each displayMessages as { message, toolMessages, isLastAssistantMessage, siblingInfo } (message.id)}
 		<ChatMessage
-			class="mx-auto mt-12 w-full max-w-[48rem]"
+			class="mx-auto mt-12 w-full max-w-3xl"
 			{message}
 			{toolMessages}
 			{isLastAssistantMessage}

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte

@@ -1,31 +1,28 @@
 <script lang="ts">
-	import { Trash2 } from '@lucide/svelte';
-	import { afterNavigate } from '$app/navigation';
 	import { page } from '$app/state';
 	import {
 		ChatScreenForm,
 		ChatMessages,
 		ChatScreenDragOverlay,
 		ChatScreenProcessingInfo,
-		ChatScreenActionScrollDown,
-		DialogEmptyFileAlert,
-		DialogFileUploadError,
-		DialogChatError,
 		ServerLoadingSplash,
-		DialogConfirmation,
 		ChatScreenServerError
 	} from '$lib/components/app';
 	import { setProcessingInfoContext } from '$lib/contexts';
-	import { ErrorDialogType } from '$lib/enums';
 	import { createAutoScrollController } from '$lib/hooks/use-auto-scroll.svelte';
+	import { useChatScreenActiveModel } from '$lib/hooks/use-chat-screen-active-model.svelte';
+	import { useChatScreenDragAndDrop } from '$lib/hooks/use-chat-screen-drag-and-drop.svelte';
+	import { useChatScreenFileUpload } from '$lib/hooks/use-chat-screen-file-upload.svelte';
+	import { useChatScreenScroll } from '$lib/hooks/use-chat-screen-scroll.svelte';
 	import { useKeyboardShortcuts } from '$lib/hooks/use-keyboard-shortcuts.svelte';
+	import { device } from '$lib/stores/device.svelte';
+	import { isMobile } from '$lib/stores/viewport.svelte';
 	import {
 		chatStore,
 		errorDialog,
 		isLoading,
 		isChatStreaming,
 		isEditing,
-		getAddFilesHandler,
 		activeProcessingState
 	} from '$lib/stores/chat.svelte';
 	import {
@@ -34,138 +31,81 @@
 		activeConversation
 	} from '$lib/stores/conversations.svelte';
 	import { config } from '$lib/stores/settings.svelte';
-	import { serverLoading, serverError, isRouterMode } from '$lib/stores/server.svelte';
-	import { modelsStore, modelOptions, selectedModelId } from '$lib/stores/models.svelte';
-	import { isFileTypeSupported, filterFilesByModalities } from '$lib/utils';
-	import { parseFilesToMessageExtras, processFilesToChatUploaded } from '$lib/utils/browser-only';
-	import { onMount } from 'svelte';
+	import { serverLoading, serverError } from '$lib/stores/server.svelte';
+	import { parseFilesToMessageExtras } from '$lib/utils/browser-only';
+	import { onDestroy, onMount } from 'svelte';
 	import ChatScreenGreeting from './ChatScreenGreeting.svelte';
+	import ChatScreenActionScrollDown from './ChatScreenActionScrollDown.svelte';
+	import ChatScreenDialogsAndAlerts from './ChatScreenDialogsAndAlerts.svelte';
+	import { ROUTES } from '$lib/constants';
 
 	let { showCenteredEmpty = false } = $props();
 
-	const autoScroll = createAutoScrollController();
-
-	let disableAutoScroll = $derived(Boolean(config().disableAutoScroll));
-	let chatScrollContainer: HTMLDivElement | undefined = $state();
-	let dragCounter = $state(0);
-	let isDragOver = $state(false);
-	let showFileErrorDialog = $state(false);
-	let uploadedFiles = $state<ChatUploadedFile[]>([]);
-
-	let fileErrorData = $state<{
-		generallyUnsupported: File[];
-		modalityUnsupported: File[];
-		modalityReasons: Record<string, string>;
-		supportedTypes: string[];
-	}>({
-		generallyUnsupported: [],
-		modalityUnsupported: [],
-		modalityReasons: {},
-		supportedTypes: []
-	});
-
-	let showDeleteDialog = $state(false);
-
-	let showEmptyFileDialog = $state(false);
-
-	let processingInfoVisible = $state(false);
-
-	let emptyFileNames = $state<string[]>([]);
-
-	let initialMessage = $state('');
-
-	let isEmpty = $derived(
-		showCenteredEmpty && !activeConversation() && activeMessages().length === 0 && !isLoading()
-	);
-
-	let activeErrorDialog = $derived(errorDialog());
-	let isServerLoading = $derived(serverLoading());
-	let hasPropsError = $derived(!!serverError());
-
-	let isCurrentConversationLoading = $derived(isLoading() || isChatStreaming());
-
-	let showProcessingInfo = $derived(
-		isCurrentConversationLoading ||
-			(config().keepStatsVisible && !!page.params.id) ||
-			activeProcessingState() !== null
-	);
-
-	let isRouter = $derived(isRouterMode());
-
-	let conversationModel = $derived(
-		chatStore.getConversationModel(activeMessages() as DatabaseMessage[])
-	);
-
-	let activeModelId = $derived.by(() => {
-		const options = modelOptions();
-
-		if (!isRouter) {
-			return options.length > 0 ? options[0].model : null;
-		}
-
-		const selectedId = selectedModelId();
-		if (selectedId) {
-			const model = options.find((m) => m.id === selectedId);
-			if (model) return model.model;
-		}
-
-		if (conversationModel) {
-			const model = options.find((m) => m.model === conversationModel);
-			if (model) return model.model;
-		}
-
-		return null;
-	});
-
-	let modelPropsVersion = $state(0);
-
 	setProcessingInfoContext({
 		get showProcessingInfo() {
 			return showProcessingInfo;
 		}
 	});
 
-	$effect(() => {
-		if (activeModelId) {
-			const cached = modelsStore.getModelProps(activeModelId);
+	let disableAutoScroll = $derived(Boolean(config().disableAutoScroll) || isMobile.current);
+	let isMobileUserScrolledUp = $state(false);
+	let mobileScrollDownHint = $state(false);
+	let mobileScrollDownHintLockedUntil = $state(0);
+	let emptyFileNames = $state<string[]>([]);
+	let initialMessage = $state('');
+	let showDeleteDialog = $state(false);
+	let showEmptyFileDialog = $state(false);
+	let isEmpty = $derived(
+		showCenteredEmpty && !activeConversation() && activeMessages().length === 0 && !isLoading()
+	);
+	let activeErrorDialog = $derived(errorDialog());
+	let isServerLoading = $derived(serverLoading());
+	let hasPropsError = $derived(!!serverError());
+	let isCurrentConversationLoading = $derived(isLoading() || isChatStreaming());
+	let showProcessingInfo = $derived(
+		isCurrentConversationLoading ||
+			(config().keepStatsVisible && !!page.params.id) ||
+			activeProcessingState() !== null
+	);
+	let chatFormBottomPosition = $derived.by(() => {
+		if (!isMobile.current) return '1rem';
+		if (device.isStandalone) return '1.5rem';
+		if (device.isIOSSafari) return '0.25rem';
+		return '0.5rem';
+	});
 
-			if (!cached) {
-				modelsStore.fetchModelProps(activeModelId).then(() => {
-					modelPropsVersion++;
-				});
+	const autoScroll = createAutoScrollController();
+	const scroll = useChatScreenScroll(autoScroll);
+	const activeModel = useChatScreenActiveModel();
+	const fileUpload = useChatScreenFileUpload({
+		capabilities: () => ({
+			hasVision: activeModel.hasVisionModality,
+			hasAudio: activeModel.hasAudioModality,
+			hasVideo: activeModel.hasVideoModality
+		}),
+		activeModelId: () => activeModel.activeModelId
+	});
+	const dragAndDrop = useChatScreenDragAndDrop({
+		onDrop: fileUpload.handleFileUpload
+	});
+	const { handleKeydown } = useKeyboardShortcuts({
+		deleteActiveConversation: () => {
+			if (activeConversation()) {
+				showDeleteDialog = true;
 			}
 		}
 	});
 
-	let hasAudioModality = $derived.by(() => {
-		if (activeModelId) {
-			void modelPropsVersion;
+	function handleMobileScroll() {
+		if (!isMobile.current) return;
 
-			return modelsStore.modelSupportsAudio(activeModelId);
-		}
+		const container = scroll.chatScrollContainer;
+		if (!container) return;
 
-		return false;
-	});
-
-	let hasVideoModality = $derived.by(() => {
-		if (activeModelId) {
-			void modelPropsVersion;
-
-			return modelsStore.modelSupportsVideo(activeModelId);
-		}
-
-		return false;
-	});
-
-	let hasVisionModality = $derived.by(() => {
-		if (activeModelId) {
-			void modelPropsVersion;
-
-			return modelsStore.modelSupportsVision(activeModelId);
-		}
-
-		return false;
-	});
+		const distanceFromBottom =
+			container.scrollHeight - container.clientHeight - container.scrollTop;
+		isMobileUserScrolledUp = distanceFromBottom > 300;
+	}
 
 	async function handleDeleteConfirm() {
 		const conversation = activeConversation();
@@ -177,27 +117,69 @@
 		showDeleteDialog = false;
 	}
 
-	function handleProcessingInfoVisibility(visible: boolean) {
-		processingInfoVisible = visible;
-	}
+	async function handleSendMessage(message: string, files?: ChatUploadedFile[]): Promise<boolean> {
+		const plainFiles = files ? $state.snapshot(files) : undefined;
+		const result = plainFiles
+			? await parseFilesToMessageExtras(plainFiles, activeModel.activeModelId ?? undefined)
+			: undefined;
 
-	function handleDragEnter(event: DragEvent) {
-		event.preventDefault();
-
-		dragCounter++;
-
-		if (event.dataTransfer?.types.includes('Files')) {
-			isDragOver = true;
+		if (result?.emptyFiles && result.emptyFiles.length > 0) {
+			emptyFileNames = result.emptyFiles;
+			showEmptyFileDialog = true;
+			if (files) {
+				const emptyFileNamesSet = new Set(result.emptyFiles);
+				fileUpload.uploadedFiles = fileUpload.uploadedFiles.filter(
+					(file) => !emptyFileNamesSet.has(file.name)
+				);
+			}
+			return false;
 		}
+
+		handleSendLikeScroll();
+
+		await chatStore.sendMessage(message, result?.extras);
+		return true;
 	}
 
-	function handleDragLeave(event: DragEvent) {
-		event.preventDefault();
+	function handleSendLikeScroll() {
+		if (!isMobile.current) {
+			autoScroll.enable();
+		}
 
-		dragCounter--;
+		setTimeout(() => {
+			const container = scroll.chatScrollContainer;
+			if (!container) return;
 
-		if (dragCounter === 0) {
-			isDragOver = false;
+			const lastUserBubble = container.querySelector(
+				'.chat-message:nth-last-child(2) .chat-message-user .chat-message-user-bubble'
+			) as HTMLElement | null;
+
+			if (isMobile.current) {
+				// Keep the last user message bubble just above the input on mobile
+				const bubbleHeight = lastUserBubble?.scrollHeight ?? 0;
+				const baseHeight = container.scrollHeight - innerHeight;
+
+				container.scrollTo({
+					top: bubbleHeight > 0 ? baseHeight - bubbleHeight : baseHeight,
+					behavior: 'smooth'
+				});
+			} else if (lastUserBubble) {
+				// On desktop, place the last user message near the top of the viewport
+				const topPadding = 24;
+				const bubbleRect = lastUserBubble.getBoundingClientRect();
+				container.scrollTo({
+					top: Math.max(0, container.scrollTop + bubbleRect.top - topPadding),
+					behavior: 'smooth'
+				});
+			} else {
+				autoScroll.scrollToBottom();
+			}
+		}, 100);
+
+		if (isMobile.current) {
+			autoScroll.setDisabled(disableAutoScroll);
+			mobileScrollDownHint = true;
+			mobileScrollDownHintLockedUntil = Date.now() + 500;
 		}
 	}
 
@@ -207,273 +189,138 @@
 		}
 	}
 
-	function handleDragOver(event: DragEvent) {
-		event.preventDefault();
-	}
-
-	function handleDrop(event: DragEvent) {
-		event.preventDefault();
-
-		isDragOver = false;
-		dragCounter = 0;
-
-		if (event.dataTransfer?.files) {
-			const files = Array.from(event.dataTransfer.files);
-
-			if (isEditing()) {
-				const handler = getAddFilesHandler();
-
-				if (handler) {
-					handler(files);
-					return;
-				}
-			}
-
-			processFiles(files);
-		}
-	}
-
-	function handleFileRemove(fileId: string) {
-		uploadedFiles = uploadedFiles.filter((f) => f.id !== fileId);
-	}
-
-	function handleFileUpload(files: File[]) {
-		processFiles(files);
-	}
-
-	const { handleKeydown } = useKeyboardShortcuts({
-		deleteActiveConversation: () => {
-			if (activeConversation()) {
-				showDeleteDialog = true;
-			}
-		}
-	});
-
 	async function handleSystemPromptAdd(draft: { message: string; files: ChatUploadedFile[] }) {
 		if (draft.message || draft.files.length > 0) {
 			chatStore.savePendingDraft(draft.message, draft.files);
 		}
-
 		await chatStore.addSystemPrompt();
 	}
 
-	function handleScroll() {
-		autoScroll.handleScroll();
-	}
-
-	async function handleSendMessage(message: string, files?: ChatUploadedFile[]): Promise<boolean> {
-		const plainFiles = files ? $state.snapshot(files) : undefined;
-		const result = plainFiles
-			? await parseFilesToMessageExtras(plainFiles, activeModelId ?? undefined)
-			: undefined;
-
-		if (result?.emptyFiles && result.emptyFiles.length > 0) {
-			emptyFileNames = result.emptyFiles;
-			showEmptyFileDialog = true;
-
-			if (files) {
-				const emptyFileNamesSet = new Set(result.emptyFiles);
-				uploadedFiles = uploadedFiles.filter((file) => !emptyFileNamesSet.has(file.name));
-			}
-			return false;
-		}
-
-		const extras = result?.extras;
-
-		// Enable autoscroll for user-initiated message sending
-		autoScroll.enable();
-		await chatStore.sendMessage(message, extras);
-		autoScroll.scrollToBottom();
-
-		return true;
-	}
-
-	async function processFiles(files: File[]) {
-		const generallySupported: File[] = [];
-		const generallyUnsupported: File[] = [];
-
-		for (const file of files) {
-			if (isFileTypeSupported(file.name, file.type)) {
-				generallySupported.push(file);
-			} else {
-				generallyUnsupported.push(file);
-			}
-		}
-
-		// Use model-specific capabilities for file validation
-		const capabilities = {
-			hasVision: hasVisionModality,
-			hasAudio: hasAudioModality,
-			hasVideo: hasVideoModality
-		};
-		const { supportedFiles, unsupportedFiles, modalityReasons } = filterFilesByModalities(
-			generallySupported,
-			capabilities
-		);
-
-		const allUnsupportedFiles = [...generallyUnsupported, ...unsupportedFiles];
-
-		if (allUnsupportedFiles.length > 0) {
-			const supportedTypes: string[] = ['text files', 'PDFs'];
-
-			if (hasVisionModality) supportedTypes.push('images');
-			if (hasAudioModality) supportedTypes.push('audio files');
-			if (hasVideoModality) supportedTypes.push('video files');
-
-			fileErrorData = {
-				generallyUnsupported,
-				modalityUnsupported: unsupportedFiles,
-				modalityReasons,
-				supportedTypes
-			};
-			showFileErrorDialog = true;
-		}
-
-		if (supportedFiles.length > 0) {
-			const processed = await processFilesToChatUploaded(
-				supportedFiles,
-				activeModelId ?? undefined
-			);
-			uploadedFiles = [...uploadedFiles, ...processed];
-		}
-	}
-
-	afterNavigate(() => {
-		if (!disableAutoScroll) {
+	$effect(() => {
+		const shouldDisableAutoScroll =
+			config().disableAutoScroll || (isMobile.current && isCurrentConversationLoading);
+		autoScroll.setDisabled(shouldDisableAutoScroll);
+		if (!shouldDisableAutoScroll) {
 			autoScroll.enable();
 		}
 	});
 
-	function handleMessagesReady() {
-		if (disableAutoScroll) return;
-
-		if (!autoScroll.userScrolledUp) {
-			requestAnimationFrame(() => {
-				autoScroll.scrollToBottom('instant');
-			});
-		}
-	}
-
 	onMount(() => {
+		const pendingDraft = chatStore.consumePendingDraft();
+		if (pendingDraft) {
+			initialMessage = pendingDraft.message;
+			fileUpload.uploadedFiles = pendingDraft.files;
+		}
+
 		autoScroll.startObserving();
 
 		if (!disableAutoScroll) {
 			autoScroll.enable();
 		}
 
-		const pendingDraft = chatStore.consumePendingDraft();
-		if (pendingDraft) {
-			initialMessage = pendingDraft.message;
-			uploadedFiles = pendingDraft.files;
+		if (isMobile.current && isCurrentConversationLoading) {
+			mobileScrollDownHint = true;
+			mobileScrollDownHintLockedUntil = Date.now() + 500;
 		}
+
+		handleMobileScroll();
 	});
 
-	$effect(() => {
-		autoScroll.setContainer(chatScrollContainer);
-	});
-
-	$effect(() => {
-		autoScroll.setDisabled(disableAutoScroll);
-	});
+	onDestroy(() => autoScroll.destroy());
 </script>
 
-{#if isDragOver}
+{#if dragAndDrop.isDragOver}
 	<ChatScreenDragOverlay />
 {/if}
 
-<svelte:window onkeydown={handleKeydown} />
+<svelte:window
+	onkeydown={handleKeydown}
+	onscroll={(e) => {
+		scroll.handleScroll(e);
+		handleMobileScroll();
+		if (e.isTrusted && Date.now() > mobileScrollDownHintLockedUntil) {
+			mobileScrollDownHint = false;
+		}
+	}}
+/>
 
 {#if isServerLoading}
 	<ServerLoadingSplash />
 {:else}
 	<div
-		bind:this={chatScrollContainer}
-		aria-label="Chat interface with file drop zone"
-		class="flex h-full flex-col overflow-y-auto px-4 md:px-6"
-		ondragenter={handleDragEnter}
-		ondragleave={handleDragLeave}
-		ondragover={handleDragOver}
-		ondrop={handleDrop}
-		onscroll={handleScroll}
+		class="chat-screen flex grow flex-col min-h-[calc(100dvh-1rem)] md:min-h-full px-4 md:py-0 pt-12 pb-48 md:pb-4"
+		style:--chat-form-bottom-position={chatFormBottomPosition}
+		ondragenter={dragAndDrop.dragHandlers.dragenter}
+		ondragleave={dragAndDrop.dragHandlers.dragleave}
+		ondragover={dragAndDrop.dragHandlers.dragover}
+		ondrop={dragAndDrop.dragHandlers.drop}
 		role="main"
 	>
-		<div class="flex grow flex-col pt-14">
-			{#if !isEmpty}
-				<ChatMessages
-					messages={activeMessages()}
-					onMessagesReady={handleMessagesReady}
-					onUserAction={() => {
-						autoScroll.enable();
-						if (!autoScroll.userScrolledUp) {
-							autoScroll.scrollToBottom();
-						}
-					}}
-				/>
-			{/if}
+		{#if !isEmpty}
+			<ChatMessages
+				messages={activeMessages()}
+				onUserAction={() => {
+					handleSendLikeScroll();
+				}}
+			/>
+		{/if}
 
-			<div
-				class={[
-					'pointer-events-none sticky right-4 left-4 mt-auto transition-all duration-200',
-					isEmpty ? 'bottom-[calc(50dvh-7rem)]' : 'bottom-4 pt-24 md:pt-32'
-				]}
-			>
-				<ChatScreenGreeting {isEmpty} />
+		<div
+			class={[
+				'pointer-events-none md:sticky fixed  mt-auto transition-all duration-200',
+				device.isStandalone
+					? 'bottom-6 right-4 left-4'
+					: device.isIOSSafari
+						? 'bottom-1 left-2 right-2'
+						: 'bottom-2 right-2 left-2',
+				isEmpty ? 'md:bottom-[calc(50dvh-7rem)] 2xl:bottom-[calc(50dvh-4rem)]' : 'md:bottom-4'
+			]}
+			style:padding-top={!isEmpty ? 'var(--chat-form-padding-top)' : undefined}
+		>
+			<ChatScreenGreeting {isEmpty} />
 
-				<ChatScreenActionScrollDown
-					container={chatScrollContainer}
-					hasProcessingInfoVisible={processingInfoVisible}
-				/>
+			<ChatScreenServerError />
 
-				<ChatScreenProcessingInfo onVisibilityChange={handleProcessingInfoVisibility} />
-
-				<ChatScreenServerError />
-
-				<div class="conversation-chat-form pointer-events-auto rounded-t-3xl">
-					<ChatScreenForm
-						disabled={hasPropsError || isEditing()}
-						{initialMessage}
-						isLoading={isCurrentConversationLoading}
-						onFileRemove={handleFileRemove}
-						onFileUpload={handleFileUpload}
-						onSend={handleSendMessage}
-						onStop={() => chatStore.stopGeneration()}
-						onSystemPromptAdd={handleSystemPromptAdd}
-						bind:uploadedFiles
+			<div class="pointer-events-none flex flex-col gap-6 items-center w-full">
+				{#if (isMobile.current ? mobileScrollDownHint || isMobileUserScrolledUp : autoScroll.userScrolledUp) && page.url.hash.includes(ROUTES.CHAT) && page.params.id}
+					<ChatScreenActionScrollDown
+						onclick={() => {
+							mobileScrollDownHint = false;
+							scroll.chatScrollContainer?.scrollTo({
+								top: scroll.chatScrollContainer.scrollHeight,
+								behavior: 'smooth'
+							});
+						}}
 					/>
-				</div>
+				{/if}
+
+				{#if showProcessingInfo}
+					<ChatScreenProcessingInfo />
+				{/if}
 			</div>
+
+			<ChatScreenForm
+				class="pointer-events-auto conversation-chat-form"
+				disabled={hasPropsError || isEditing()}
+				{initialMessage}
+				isLoading={isCurrentConversationLoading}
+				onFileRemove={fileUpload.handleFileRemove}
+				onFileUpload={fileUpload.handleFileUpload}
+				onSend={handleSendMessage}
+				onStop={() => chatStore.stopGeneration()}
+				onSystemPromptAdd={handleSystemPromptAdd}
+				bind:uploadedFiles={fileUpload.uploadedFiles}
+			/>
 		</div>
 	</div>
 {/if}
 
-<DialogFileUploadError bind:open={showFileErrorDialog} {fileErrorData} />
-
-<DialogConfirmation
-	bind:open={showDeleteDialog}
-	title="Delete Conversation"
-	description="Are you sure you want to delete this conversation? This action cannot be undone and will permanently remove all messages in this conversation."
-	confirmText="Delete"
-	cancelText="Cancel"
-	variant="destructive"
-	icon={Trash2}
-	onConfirm={handleDeleteConfirm}
-	onCancel={() => (showDeleteDialog = false)}
-/>
-
-<DialogEmptyFileAlert
-	bind:open={showEmptyFileDialog}
-	emptyFiles={emptyFileNames}
-	onOpenChange={(open) => {
-		if (!open) {
-			emptyFileNames = [];
-		}
-	}}
-/>
-
-<DialogChatError
-	message={activeErrorDialog?.message ?? ''}
-	contextInfo={activeErrorDialog?.contextInfo}
-	onOpenChange={handleErrorDialogOpenChange}
-	open={Boolean(activeErrorDialog)}
-	type={activeErrorDialog?.type ?? ErrorDialogType.SERVER}
+<ChatScreenDialogsAndAlerts
+	{showDeleteDialog}
+	{handleDeleteConfirm}
+	{showEmptyFileDialog}
+	{emptyFileNames}
+	{activeErrorDialog}
+	{handleErrorDialogOpenChange}
+	{fileUpload}
 />

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenActionScrollDown.svelte

@@ -1,58 +1,18 @@
 <script lang="ts">
 	import { ArrowDown } from '@lucide/svelte';
-	import { Button } from '$lib/components/ui/button';
+	import ActionIcon from '$lib/components/app/actions/ActionIcon.svelte';
 
-	interface Props {
-		container: HTMLDivElement | undefined;
-		hasProcessingInfoVisible: boolean;
-	}
-
-	let { container, hasProcessingInfoVisible }: Props = $props();
-
-	let show = $state(false);
-
-	let buttonBottom = $derived(hasProcessingInfoVisible ? '2rem' : '0');
-
-	function checkVisibility() {
-		if (!container) return;
-		const { scrollTop, scrollHeight, clientHeight } = container;
-		const distanceFromBottom = scrollHeight - clientHeight - scrollTop;
-		show = distanceFromBottom > clientHeight * 0.5;
-	}
-
-	function scrollToBottom() {
-		if (container) {
-			container.scrollTo({
-				top: container.scrollHeight,
-				behavior: 'smooth'
-			});
-		}
-	}
-
-	$effect(() => {
-		const c = container;
-		if (c) {
-			c.addEventListener('scroll', checkVisibility);
-			checkVisibility();
-			return () => {
-				c.removeEventListener('scroll', checkVisibility);
-			};
-		}
-	});
+	let { onclick }: { onclick: (e?: MouseEvent) => void } = $props();
 </script>
 
-<div class="relative z-50 mx-auto mb-4 flex max-w-[48rem] justify-center">
-	<Button
-		onclick={scrollToBottom}
-		variant="secondary"
-		size="icon"
-		disabled={!show}
-		class="pointer-events-auto absolute h-10 w-10 rounded-full bg-background/80 shadow-lg backdrop-blur-sm transition-all duration-200 hover:bg-muted/80"
-		style="bottom: {buttonBottom}; transform: translateY({show ? '0' : '2rem'}); opacity: {show
-			? 1
-			: 0};"
-		aria-label="Scroll to bottom"
-	>
-		<ArrowDown class="h-4 w-4" />
-	</Button>
+<div class="pointer-events-auto flex justify-center relative h-0">
+	<ActionIcon
+		icon={ArrowDown}
+		{onclick}
+		ariaLabel="Scroll to bottom"
+		tooltip="Scroll to bottom"
+		size="lg"
+		iconSize="h-4 w-4"
+		class="h-9 w-9 rounded-full bg-accent text-accent-foreground absolute bottom-4 shadow-md"
+	/>
 </div>

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenDialogsAndAlerts.svelte

@@ -0,0 +1,55 @@
+<script lang="ts">
+	import { Trash2 } from '@lucide/svelte';
+	import { ErrorDialogType } from '$lib/enums';
+	import {
+		DialogChatError,
+		DialogConfirmation,
+		DialogEmptyFileAlert,
+		DialogFileUploadError
+	} from '$lib/components/app';
+
+	let {
+		showDeleteDialog,
+		handleDeleteConfirm,
+		showEmptyFileDialog,
+		emptyFileNames,
+		activeErrorDialog,
+		handleErrorDialogOpenChange,
+		fileUpload
+	} = $props();
+</script>
+
+<DialogFileUploadError
+	bind:open={fileUpload.showFileErrorDialog}
+	fileErrorData={fileUpload.fileErrorData}
+/>
+
+<DialogConfirmation
+	bind:open={showDeleteDialog}
+	title="Delete Conversation"
+	description="Are you sure you want to delete this conversation? This action cannot be undone and will permanently remove all messages in this conversation."
+	confirmText="Delete"
+	cancelText="Cancel"
+	variant="destructive"
+	icon={Trash2}
+	onConfirm={handleDeleteConfirm}
+	onCancel={() => (showDeleteDialog = false)}
+/>
+
+<DialogEmptyFileAlert
+	bind:open={showEmptyFileDialog}
+	emptyFiles={emptyFileNames}
+	onOpenChange={(open) => {
+		if (!open) {
+			emptyFileNames = [];
+		}
+	}}
+/>
+
+<DialogChatError
+	message={activeErrorDialog?.message ?? ''}
+	contextInfo={activeErrorDialog?.contextInfo}
+	onOpenChange={handleErrorDialogOpenChange}
+	open={Boolean(activeErrorDialog)}
+	type={activeErrorDialog?.type ?? ErrorDialogType.SERVER}
+/>

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenForm.svelte

@@ -2,6 +2,7 @@
 	import { afterNavigate } from '$app/navigation';
 	import { page } from '$app/state';
 	import { ChatForm } from '$lib/components/app';
+	import { isMobile } from '$lib/stores/viewport.svelte';
 	import { onMount } from 'svelte';
 	import { useDraftMessages } from '$lib/hooks/use-draft-messages.svelte';
 
@@ -32,7 +33,30 @@
 	}: Props = $props();
 
 	let chatFormRef: ChatForm | undefined = $state(undefined);
+	let formWrapperEl: HTMLDivElement | undefined = $state();
 	let chatId = $derived(page.params.id as string | undefined);
+
+	$effect(() => {
+		if (!formWrapperEl) return;
+
+		const formEl = formWrapperEl.querySelector('form') as HTMLElement | null;
+		if (!formEl) return;
+
+		const updateHeight = () => {
+			const height = Math.round(formEl.getBoundingClientRect().height);
+			document.documentElement.style.setProperty('--chat-form-height', `${height}px`);
+		};
+
+		updateHeight();
+
+		const resizeObserver = new ResizeObserver(updateHeight);
+		resizeObserver.observe(formEl);
+
+		return () => {
+			resizeObserver.disconnect();
+			document.documentElement.style.removeProperty('--chat-form-height');
+		};
+	});
 	let hasLoadingAttachments = $derived(uploadedFiles.some((f) => f.isLoading));
 	let message = $derived(initialMessage);
 	let previousIsLoading = $derived(isLoading);
@@ -83,12 +107,14 @@
 	}
 
 	onMount(() => {
-		setTimeout(() => chatFormRef?.focus(), 10);
+		if (!isMobile.current) {
+			setTimeout(() => chatFormRef?.focus(), 100);
+		}
 	});
 
 	afterNavigate((navigation) => {
-		if (navigation?.from != null) {
-			setTimeout(() => chatFormRef?.focus(), 10);
+		if (navigation?.from != null && !isMobile.current) {
+			setTimeout(() => chatFormRef?.focus(), 100);
 		}
 	});
 
@@ -108,12 +134,12 @@
 	});
 </script>
 
-<div class="relative mx-auto max-w-[48rem]">
+<div class="chat-screen-form-wrapper" bind:this={formWrapperEl}>
 	<ChatForm
+		class="mx-auto max-w-3xl {className}"
 		bind:this={chatFormRef}
 		bind:value={message}
 		bind:uploadedFiles
-		class={className}
 		{disabled}
 		{isLoading}
 		showMcpPromptButton

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenGreeting.svelte

@@ -1,5 +1,4 @@
 <script lang="ts">
-	import { fadeInView } from '$lib/actions/fade-in-view.svelte';
 	import { serverStore } from '$lib/stores/server.svelte';
 
 	interface Props {
@@ -11,10 +10,9 @@
 
 <div
 	class={[
-		'pointer-events-none mb-4 hidden px-4 text-center',
-		isEmpty && 'pointer-events-auto block!'
+		'pointer-events-none mb-4 hidden px-4 text-center text-balance',
+		isEmpty && 'mb-[calc(50dvh-8rem)] md:mb-6 pointer-events-auto block!'
 	]}
-	use:fadeInView={{ duration: 300 }}
 >
 	<h1 class="mb-2 text-2xl font-semibold tracking-tight md:text-3xl">Hello there</h1>
 

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenProcessingInfo.svelte

@@ -5,13 +5,8 @@
 	import { chatStore, isLoading, isChatStreaming } from '$lib/stores/chat.svelte';
 	import { activeMessages, activeConversation } from '$lib/stores/conversations.svelte';
 	import { config } from '$lib/stores/settings.svelte';
-	import { getProcessingInfoContext } from '$lib/contexts';
-	import { page } from '$app/state';
 
 	const processingState = useProcessingState();
-	const processingInfoCtx = getProcessingInfoContext();
-
-	let showProcessingInfo = $derived(processingInfoCtx.showProcessingInfo);
 
 	let isCurrentConversationLoading = $derived(isLoading());
 	let isStreaming = $derived(isChatStreaming());
@@ -70,8 +65,8 @@
 
 <div
 	class={[
-		'chat-processing-info-container pointer-events-none relative',
-		page.params.id && showProcessingInfo && 'visible'
+		'chat-processing-info-container pointer-events-none relative w-full hidden md:block',
+		processingVisible && 'visible'
 	]}
 >
 	<div class="chat-processing-info-content absolute bottom-4 left-1/2 -translate-x-1/2">

tools/ui/src/lib/components/app/chat/index.ts

@@ -677,13 +677,6 @@ export { default as ChatScreenForm } from './ChatScreen/ChatScreenForm.svelte';
  */
 export { default as ChatScreenProcessingInfo } from './ChatScreen/ChatScreenProcessingInfo.svelte';
 
-/**
- * Scroll-to-bottom action button. Displays a floating button when the user
- * has scrolled up more than half a viewport height from the bottom.
- * Takes the chat container element as a prop to manage scroll state internally.
- */
-export { default as ChatScreenActionScrollDown } from './ChatScreen/ChatScreenActionScrollDown.svelte';
-
 /**
  * Server error alert displayed when the server is unreachable.
  * Shows the error message with a retry button.

tools/ui/src/lib/components/app/forms/SearchInput.svelte

@@ -3,6 +3,7 @@
 	import { Search, X } from '@lucide/svelte';
 
 	interface Props {
+		autofocus?: boolean;
 		value?: string;
 		placeholder?: string;
 		onInput?: (value: string) => void;
@@ -15,6 +16,7 @@
 	}
 
 	let {
+		autofocus,
 		value = $bindable(''),
 		placeholder = 'Search...',
 		onInput,
@@ -39,7 +41,7 @@
 		if (value) {
 			value = '';
 			onInput?.('');
-			ref?.focus();
+			ref?.focus({ preventScroll: true });
 		} else {
 			onClose?.();
 		}
@@ -52,6 +54,7 @@
 	/>
 
 	<Input
+		{autofocus}
 		{id}
 		bind:value
 		bind:ref

tools/ui/src/lib/components/app/misc/Logo.svelte

@@ -0,0 +1,15 @@
+<script>
+	import logoMark from '$lib/assets/logo.svg?raw';
+	let { class: className = '', style = '' } = $props();
+</script>
+
+<div class={className} {style}>
+	{@html logoMark}
+</div>
+
+<style>
+	div :global(svg) {
+		width: var(--size, 1rem);
+		height: var(--size, 1rem);
+	}
+</style>

tools/ui/src/lib/components/app/misc/index.ts

@@ -51,3 +51,11 @@ export { default as KeyboardShortcutInfo } from './KeyboardShortcutInfo.svelte';
  * Preview button is shown only for HTML code blocks.
  */
 export { default as CodeBlockActions } from './CodeBlockActions.svelte';
+
+/**
+ * **Logo** - Application brand mark
+ *
+ * Inline SVG of the application logo. Accepts styling via the standard
+ * `class` and `style` props and inherits color via `currentColor`.
+ */
+export { default as Logo } from './Logo.svelte';