sycl: fix mul_mat selection (#15092)

* webui: fix markdown table

* webui: fix table display with themes

common/arg.cpp

@@ -24,6 +24,7 @@
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
+#include <list>
 #include <regex>
 #include <set>
 #include <string>
@@ -2375,20 +2376,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                     }
                     throw std::invalid_argument("unknown buffer type");
                 }
-                // FIXME: this leaks memory
-                params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)});
+                // keep strings alive and avoid leaking memory by storing them in a static vector
+                static std::list<std::string> buft_overrides;
+                buft_overrides.push_back(tensor_name);
+                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
             }
         }
     ));
     add_opt(common_arg(
-        {"--cpu-moe"},
-        "use CPU for Mixture of Experts (MoE) weights",
+        {"--cpu-moe", "-cmoe"},
+        "keep all Mixture of Experts (MoE) weights in the CPU",
         [](common_params & params) {
-            params.tensor_buft_overrides.push_back({"\\.ffn_up_exps\\.weight$",   ggml_backend_cpu_buffer_type()});
-            params.tensor_buft_overrides.push_back({"\\.ffn_down_exps\\.weight$", ggml_backend_cpu_buffer_type()});
-            params.tensor_buft_overrides.push_back({"\\.ffn_gate_exps\\.weight$", ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
         }
     ).set_env("LLAMA_ARG_CPU_MOE"));
+    add_opt(common_arg(
+        {"--n-cpu-moe", "-ncmoe"}, "N",
+        "keep the Mixture of Experts (MoE) weights of the first N layers in the CPU",
+        [](common_params & params, int value) {
+            if (value < 0) {
+                throw std::invalid_argument("invalid value");
+            }
+            for (int i = 0; i < value; ++i) {
+                // keep strings alive and avoid leaking memory by storing them in a static vector
+                static std::list<std::string> buft_overrides;
+                buft_overrides.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
+        }
+    ).set_env("LLAMA_ARG_N_CPU_MOE"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -2649,10 +2665,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--output-format"}, "{gguf,dat}",
-        string_format("output format for imatrix file (default: %s)", params.imat_dat ? "dat" : "gguf"),
+        string_format("output format for imatrix file (default: %s)", params.imat_dat > 0 ? "dat" : "gguf"),
         [](common_params & params, const std::string & value) {
-            /**/ if (value == "gguf") { params.imat_dat = false; }
-            else if (value == "dat")  { params.imat_dat = true;  }
+            /**/ if (value == "gguf") { params.imat_dat = -1; }
+            else if (value == "dat")  { params.imat_dat = 1;  }
             else { throw std::invalid_argument("invalid output format"); }
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));

common/common.h

@@ -439,7 +439,7 @@ struct common_params {
     int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
-    bool    imat_dat    = false; // whether the legacy imatrix.dat format should be output
+    int8_t  imat_dat    =  0; // whether the legacy imatrix.dat format should be output (gguf <= 0 < dat)
 
     bool process_output  = false; // collect data for the output tensor
     bool compute_ppl     = true;  // whether to compute perplexity

ggml/CMakeLists.txt

@@ -39,8 +39,9 @@ if (WIN32)
     set(CMAKE_SHARED_MODULE_PREFIX  "")
 endif()
 
-option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
-option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
+option(BUILD_SHARED_LIBS           "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+option(GGML_BACKEND_DL             "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
+set(GGML_BACKEND_DIR "" CACHE PATH "ggml: directory to load dynamic backends from (requires GGML_BACKEND_DL")
 
 #
 # option list

ggml/cmake/ggml-config.cmake.in

@@ -106,7 +106,7 @@ if(NOT TARGET ggml::ggml)
 
     find_library(GGML_LIBRARY ggml
         REQUIRED
-        HINTS ${GGML_LIB_DIR}
+        HINTS ${GGML_LIB_DIR} ${GGML_BACKEND_DIR}
         NO_CMAKE_FIND_ROOT_PATH)
 
     add_library(ggml::ggml UNKNOWN IMPORTED)

ggml/src/CMakeLists.txt

@@ -214,6 +214,13 @@ add_library(ggml
             ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)
 
+if (GGML_BACKEND_DIR)
+    if (NOT GGML_BACKEND_DL)
+        message(FATAL_ERROR "GGML_BACKEND_DIR requires GGML_BACKEND_DL")
+    endif()
+    target_compile_definitions(ggml PUBLIC GGML_BACKEND_DIR="${GGML_BACKEND_DIR}")
+endif()
+
 target_link_libraries(ggml PUBLIC ggml-base)
 
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
@@ -227,7 +234,11 @@ function(ggml_add_backend_library backend)
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
         add_dependencies(ggml ${backend})
-        install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
+        if (GGML_BACKEND_DIR)
+            install(TARGETS ${backend} LIBRARY DESTINATION ${GGML_BACKEND_DIR})
+        else()
+            install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
+        endif()
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})

ggml/src/ggml-backend-reg.cpp

@@ -498,6 +498,9 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 
     std::vector<fs::path> search_paths;
     if (user_search_path == nullptr) {
+#ifdef GGML_BACKEND_DIR
+        search_paths.push_back(fs::u8path(GGML_BACKEND_DIR));
+#endif
         // default search paths: executable directory, current directory
         search_paths.push_back(get_executable_path());
         search_paths.push_back(fs::current_path());

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

@@ -2609,6 +2609,8 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
     GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->ne[1] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -3196,7 +3198,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
             // The kernel from the if path is faster for that specific case, but does not support all mul mats.
             ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
         }
-    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1 && src1->ne[3] == 1) {
         // KQV single-batch
         ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2] * src1->ne[3] > 1) {

gguf-py/gguf/scripts/gguf_new_metadata.py

@@ -111,6 +111,7 @@ def main() -> None:
     parser.add_argument("--general-description",                       type=str,  help="The models general.description", metavar='"Description ..."')
     parser.add_argument("--chat-template",                             type=str,  help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."')
     parser.add_argument("--chat-template-config",                      type=Path, help="Config file containing chat template(s)", metavar='tokenizer_config.json')
+    parser.add_argument("--chat-template-file",                        type=Path, help="Jinja file containing chat template", metavar='chat_template.jinja')
     parser.add_argument("--pre-tokenizer",                             type=str,  help="The models tokenizer.ggml.pre", metavar='"pre tokenizer"')
     parser.add_argument("--remove-metadata",      action="append",     type=str,  help="Remove metadata (by key name) from output model", metavar='general.url')
     parser.add_argument("--special-token",        action="append",     type=str,  help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '"<token>"'))
@@ -134,12 +135,17 @@ def main() -> None:
         new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template)
 
     if args.chat_template_config:
-        with open(args.chat_template_config, 'r') as fp:
+        with open(args.chat_template_config, 'r', encoding='utf-8') as fp:
             config = json.load(fp)
             template = config.get('chat_template')
             if template:
                 new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
 
+    if args.chat_template_file:
+        with open(args.chat_template_file, 'r', encoding='utf-8') as fp:
+            template = fp.read()
+            new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
+
     if args.pre_tokenizer:
         new_metadata[gguf.Keys.Tokenizer.PRE] = MetadataDetails(gguf.GGUFValueType.STRING, args.pre_tokenizer)
 

src/llama-context.cpp

@@ -786,7 +786,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_embd  = hparams.n_embd;
-    const int32_t n_vocab = model.vocab.n_tokens();
+    const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
     if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
@@ -959,7 +959,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & vocab   = model.vocab;
     const auto & hparams = model.hparams;
 
-    const int32_t n_vocab = vocab.n_tokens();
+    const int64_t n_vocab = vocab.n_tokens();
     const int64_t n_embd  = hparams.n_embd;
 
     // when computing embeddings, all tokens are output
@@ -1328,21 +1328,21 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 }
 
 void llama_context::output_reorder() {
-    const uint32_t n_vocab = model.vocab.n_tokens();
+    const uint64_t n_vocab = model.vocab.n_tokens();
     const uint64_t n_embd  = model.hparams.n_embd;
 
-    for (uint32_t s = 0; s < output_swaps.size(); ++s) {
-        const uint32_t i0 = output_swaps[s].i0;
-        const uint32_t i1 = output_swaps[s].i1;
+    for (size_t s = 0; s < output_swaps.size(); ++s) {
+        const uint64_t i0 = output_swaps[s].i0;
+        const uint64_t i1 = output_swaps[s].i1;
 
         if (logits_size > 0) {
-            for (uint32_t k = 0; k < n_vocab; k++) {
+            for (uint64_t k = 0; k < n_vocab; k++) {
                 std::swap(logits[i0*n_vocab + k], logits[i1*n_vocab + k]);
             }
         }
 
         if (embd_size > 0) {
-            for (uint32_t k = 0; k < n_embd; k++) {
+            for (uint64_t k = 0; k < n_embd; k++) {
                 std::swap(embd[i0*n_embd + k], embd[i1*n_embd + k]);
             }
         }

src/llama-model.cpp

@@ -13800,10 +13800,6 @@ struct llm_build_glm4_moe : public llm_graph_context {
                         LLM_FFN_SILU, LLM_FFN_PAR, il);
                 cb(cur, "ffn_out", il);
             } else {
-                // MoE layer with shared experts
-                const int64_t n_expert      = hparams.n_expert;
-                const int64_t n_expert_used = hparams.n_expert_used;
-
                 // Process routed experts using existing MoE infrastructure
                 ggml_tensor * routed_out = build_moe_ffn(cur,
                         model.layers[il].ffn_gate_inp,

tools/imatrix/imatrix.cpp

@@ -506,13 +506,17 @@ void IMatrixCollector::save_imatrix_legacy(int32_t ncall) const {
 
 void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
     auto fname = m_params.out_file;
-    bool use_legacy_format = m_params.imat_dat;
+    int8_t use_legacy_format = m_params.imat_dat;
 
-    if (use_legacy_format) {
+    if (use_legacy_format > 0) {
         this->save_imatrix_legacy(n_chunk);
         return;
     }
-    // else, default to GGUF imatrix
+    // only warn when `--output-format gguf` is not specified
+    if (use_legacy_format == 0 && !string_ends_with(fname, ".gguf")) {
+        LOG_WRN("\n%s: saving imatrix using GGUF format with a different suffix than .gguf\n", __func__);
+        LOG_WRN("%s: if you want the previous imatrix format, use --output-format dat\n", __func__);
+    }
 
     if (n_chunk > 0) {
         fname += ".at_";

tools/server/webui/src/index.scss

@@ -31,7 +31,24 @@ html {
   hr {
     @apply my-4 border-base-content/20 border-1;
   }
-  /* TODO: fix markdown table */
+  table {
+    @apply w-full border-collapse text-sm font-sans my-4 text-base-content;
+  }
+  thead {
+    @apply bg-base-200 text-base-content;
+  }
+  th {
+    @apply border border-base-300 px-4 py-2 text-left font-semibold;
+  }
+  td {
+    @apply border border-base-300 px-4 py-2 align-top;
+  }
+  tbody tr:nth-child(even) {
+    @apply bg-base-100;
+  }
+  tbody tr:hover {
+    @apply bg-base-200;
+  }
 }
 
 .btn-mini {