cont : fix comments

ggml-ci

.devops/intel.Dockerfile

@@ -49,23 +49,19 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
-RUN apt-get update && \
-    apt-get install -y \
-        git \
-        python3 \
-        python3-pip \
-        python3-venv && \
-    python3 -m venv /opt/venv && \
-    . /opt/venv/bin/activate && \
-    pip install --upgrade pip setuptools wheel && \
-    pip install -r requirements.txt && \
-    apt autoremove -y && \
-    apt clean -y && \
-    rm -rf /tmp/* /var/tmp/* && \
-    find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
-    find /var/cache -type f -delete
+RUN apt-get update \
+    && apt-get install -y \
+    git \
+    python3 \
+    python3-pip \
+    && pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt \
+    && apt autoremove -y \
+    && apt clean -y \
+    && rm -rf /tmp/* /var/tmp/* \
+    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
+    && find /var/cache -type f -delete
 
-ENV PATH="/opt/venv/bin:$PATH"
 
 ENTRYPOINT ["/app/tools.sh"]
 

.github/workflows/build.yml

@@ -693,7 +693,7 @@ jobs:
           - build: 'openblas-x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
-            defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'llvm-arm64-opencl-adreno'
@@ -778,7 +778,6 @@ jobs:
           cmake -S . -B build ${{ matrix.defines }} `
             -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-          cp $env:CURL_PATH/bin/libcurl-*.dll build/bin/Release
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll

CMakeLists.txt

@@ -89,14 +89,6 @@ option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
-if (NOT DEFINED LLAMA_BUILD_NUMBER)
-    set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
-endif()
-if (NOT DEFINED LLAMA_BUILD_COMMIT)
-    set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
-endif()
-set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
-
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
 set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
@@ -163,8 +155,6 @@ if (LLAMA_USE_SYSTEM_GGML)
 endif()
 
 if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
-    set(GGML_BUILD_NUMBER ${LLAMA_BUILD_NUMBER})
-    set(GGML_BUILD_COMMIT ${LLAMA_BUILD_COMMIT})
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
@@ -214,6 +204,10 @@ endif()
 include(GNUInstallDirs)
 include(CMakePackageConfigHelpers)
 
+set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
+set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+
 set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")

README.md

@@ -6,7 +6,7 @@
 [![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
@@ -18,6 +18,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 ## Hot topics
 
 - 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
+- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
 - A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639

ci/run.sh

@@ -39,7 +39,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
@@ -779,7 +779,7 @@ function gg_run_rerank_tiny {
     model_f16="${path_models}/ggml-model-f16.gguf"
 
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

common/CMakeLists.txt

@@ -23,21 +23,31 @@ if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
     endif()
 
     if(EXISTS "${GIT_DIR}/index")
-        # For build-info.cpp below
-        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${GIT_DIR}/index")
+        set(GIT_INDEX "${GIT_DIR}/index")
     else()
         message(WARNING "Git index not found in git repository.")
+        set(GIT_INDEX "")
     endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
+    set(GIT_INDEX "")
 endif()
 
-set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
-set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
-configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-
+# Add a custom command to rebuild build-info.cpp when .git/index changes
+add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
+    COMMENT "Generating build details from Git"
+    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
+            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
+            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+    WORKING_DIRECTORY "${PROJECT_SOURCE_DIR}"
+    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
+    VERBATIM
+)
 set(TARGET build_info)
-add_library(${TARGET} OBJECT ${OUTPUT_FILE})
+add_library(${TARGET} OBJECT build-info.cpp)
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

common/arg.cpp

@@ -988,6 +988,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
+    if (params.reranking && params.embedding) {
+        throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
+    }
+
     if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
         throw std::runtime_error(string_format(
             "error: the supplied chat template is not supported: %s%s\n",
@@ -2706,13 +2710,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.embd_sep = value;
         }
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
-    add_opt(common_arg(
-        {"--cls-separator"}, "STRING",
-        "separator of classification sequences (default \\t) for example \"<#seq#>\"",
-        [](common_params & params, const std::string & value) {
-            params.cls_sep = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
         string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
@@ -2750,10 +2747,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
     add_opt(common_arg(
         {"--reranking", "--rerank"},
-        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", params.reranking ? "enabled" : "disabled"),
         [](common_params & params) {
-            params.embedding = true;
-            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
+            params.reranking = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(common_arg(
@@ -3217,32 +3213,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
-    add_opt(common_arg(
-        {"-ctkd", "--cache-type-k-draft"}, "TYPE",
-        string_format(
-            "KV cache data type for K for the draft model\n"
-            "allowed values: %s\n"
-            "(default: %s)",
-            get_all_kv_cache_types().c_str(),
-            ggml_type_name(params.speculative.cache_type_k)
-        ),
-        [](common_params & params, const std::string & value) {
-            params.speculative.cache_type_k = kv_cache_type_from_str(value);
-        }
-    ).set_env("LLAMA_ARG_CACHE_TYPE_K_DRAFT"));
-    add_opt(common_arg(
-        {"-ctvd", "--cache-type-v-draft"}, "TYPE",
-        string_format(
-            "KV cache data type for V for the draft model\n"
-            "allowed values: %s\n"
-            "(default: %s)",
-            get_all_kv_cache_types().c_str(),
-            ggml_type_name(params.speculative.cache_type_v)
-        ),
-        [](common_params & params, const std::string & value) {
-            params.speculative.cache_type_v = kv_cache_type_from_str(value);
-        }
-    ).set_env("LLAMA_ARG_CACHE_TYPE_V_DRAFT"));
 
     add_opt(common_arg(
         {"-mv", "--model-vocoder"}, "FNAME",

common/build-info.cpp.in

@@ -1,4 +1,4 @@
-int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
+int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
 char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
 char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/chat-parser.cpp

@@ -49,7 +49,6 @@ bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::
 
     // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
     result_.tool_calls.emplace_back(tool_call);
-
     return true;
 }
 bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
@@ -379,7 +378,3 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         /* .is_partial = */ found_healing_marker,
     };
 }
-
-void common_chat_msg_parser::clear_tools() {
-    result_.tool_calls.clear();
-}

common/chat-parser.h

@@ -115,6 +115,4 @@ class common_chat_msg_parser {
         const std::vector<std::vector<std::string>> & args_paths = {},
         const std::vector<std::vector<std::string>> & content_paths = {}
     );
-
-    void clear_tools();
 };

common/chat.cpp

@@ -1838,7 +1838,7 @@ static common_chat_params common_chat_templates_apply_legacy(
     if (res < 0) {
         // if the custom "tmpl" is not supported, we throw an error
         // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
-        throw std::runtime_error("this custom template is not supported, try using --jinja");
+        throw std::runtime_error("this custom template is not supported");
     }
 
     // if it turns out that our buffer is too small, we resize it
@@ -1921,9 +1921,7 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     } catch (const common_chat_msg_partial_exception & ex) {
         LOG_DBG("Partial parse: %s\n", ex.what());
         if (!is_partial) {
-            builder.clear_tools();
-            builder.move_to(0);
-            common_chat_parse_content_only(builder);
+            throw std::runtime_error(ex.what());
         }
     }
     auto msg = builder.result();

common/cmake/build-info-gen-cpp.cmake

@@ -0,0 +1,24 @@
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
+
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
+
+# Only write the build info if it changed
+if(EXISTS ${OUTPUT_FILE})
+    file(READ ${OUTPUT_FILE} CONTENTS)
+    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_COMMIT ${CMAKE_MATCH_1})
+    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_COMPILER ${CMAKE_MATCH_1})
+    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_TARGET ${CMAKE_MATCH_1})
+    if (
+        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
+        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
+        NOT OLD_TARGET   STREQUAL BUILD_TARGET
+    )
+        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+    endif()
+else()
+    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+endif()

common/common.cpp

@@ -706,17 +706,11 @@ bool fs_validate_filename(const std::string & filename) {
         // disable C++17 deprecation warning for std::codecvt_utf8
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
-
         std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
 #endif
 
         filename_utf32 = converter.from_bytes(filename);
@@ -773,9 +767,6 @@ bool fs_validate_filename(const std::string & filename) {
     return true;
 }
 
-#include <iostream>
-
-
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -793,16 +784,9 @@ bool fs_create_directory_with_parents(const std::string & path) {
     // process path from front to back, procedurally creating directories
     while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
         const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
 
-        pos_slash += 1;
-
-        // skip the drive letter, in some systems it can return an access denied error
-        if (subpath.length() == 2 && subpath[1] == ':') {
-            continue;
-        }
-
-        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
-
+        const bool success = CreateDirectoryW(test, NULL);
         if (!success) {
             const DWORD error = GetLastError();
 
@@ -816,6 +800,8 @@ bool fs_create_directory_with_parents(const std::string & path) {
                 return false;
             }
         }
+
+        pos_slash += 1;
     }
 
     return true;
@@ -911,6 +897,34 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
+    if (params.reranking) {
+        bool ok = true;
+
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_model_free(model);
+
+            return iparams;
+        }
+    }
+
     auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -952,35 +966,6 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
     }
 
-    if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
-        bool ok = true;
-
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
-        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
-            ok = false;
-        } else if (!has_eos) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (!ok) {
-            llama_free(lctx);
-            llama_model_free(model);
-
-            return iparams;
-        }
-    }
-
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
         llama_adapter_lora_ptr lora;
@@ -1158,6 +1143,11 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
 
+    if (params.reranking) {
+        cparams.embeddings    = true;
+        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
+    }
+
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;
 
@@ -1290,9 +1280,6 @@ std::vector<llama_token> common_tokenize(
     int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
     n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
-    if (n_tokens == std::numeric_limits<int32_t>::min()) {
-        throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit");
-    }
     if (n_tokens < 0) {
         result.resize(-n_tokens);
         int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);

common/common.h

@@ -199,9 +199,6 @@ struct common_params_speculative {
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
 
-    ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
-    ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
-
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
@@ -358,7 +355,7 @@ struct common_params {
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embeddings
-    std::string cls_sep    = "\t";  // separator of classification sequences
+    bool reranking         = false; // enable reranking support on server
 
     // server params
     int32_t port           = 8080;         // server listens on this network port

convert_hf_to_gguf.py

@@ -519,7 +519,7 @@ class TextModel(ModelBase):
     def set_gguf_parameters(self):
         self.gguf_writer.add_block_count(self.block_count)
 
-        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions", "max_length"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
@@ -556,8 +556,11 @@ class TextModel(ModelBase):
             logger.info(f"gguf: experts used count = {n_experts_used}")
 
         if (head_dim := self.hparams.get("head_dim")) is not None:
-            self.gguf_writer.add_key_length(head_dim)
-            self.gguf_writer.add_value_length(head_dim)
+            # Workaround for incorrect AutoConfig value for DeepSeekV3 (is set correctly in DeepSeekV2Model class)
+            # https://github.com/huggingface/transformers/blob/19224c3642705c5b6988c9f5f4251f83323d05ae/src/transformers/models/deepseek_v3/configuration_deepseek_v3.py#L210
+            if self.hparams.get("model_type") != "deepseek_v3":
+                self.gguf_writer.add_key_length(head_dim)
+                self.gguf_writer.add_value_length(head_dim)
 
         self.gguf_writer.add_file_type(self.ftype)
         logger.info(f"gguf: file type = {self.ftype}")
@@ -1898,7 +1901,9 @@ class LlamaModel(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if (rope_dim := hparams.get("head_dim")) is None:
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -1980,8 +1985,7 @@ class LlamaModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                if (dim := self.hparams.get("head_dim")) is None:
-                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -2016,20 +2020,6 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("ArceeForCausalLM")
-class ArceeModel(LlamaModel):
-    model_arch = gguf.MODEL_ARCH.ARCEE
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self._try_set_pooling_type()
-        rope_scaling = self.hparams.get("rope_scaling") or {}
-        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
-            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
-            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
-
-
 @ModelBase.register(
     "LlavaForConditionalGeneration", # pixtral
     "Mistral3ForConditionalGeneration", # mistral small 3.1
@@ -2145,6 +2135,7 @@ class Llama4Model(LlamaModel):
 
     def set_vocab(self):
         self._set_vocab_gpt2()
+        self.gguf_writer.add_add_bos_token(True)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -2316,7 +2307,9 @@ class DeciModel(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if (rope_dim := hparams.get("head_dim")) is None:
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -2356,8 +2349,7 @@ class DeciModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                if (dim := self.hparams.get("head_dim")) is None:
-                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -3675,7 +3667,9 @@ class InternLM3Model(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if (rope_dim := hparams.get("head_dim")) is None:
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -3917,6 +3911,9 @@ class BertModel(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
 
 @ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
 class DistilBertModel(BertModel):
@@ -3958,6 +3955,8 @@ class RobertaModel(BertModel):
         bpe_tok_path = self.dir_model / "tokenizer.json"
         if bpe_tok_path.exists():
             self._set_vocab_gpt2()
+            self.gguf_writer.add_add_bos_token(True)
+            self.gguf_writer.add_add_eos_token(True)
 
             # we need this to validate the size of the token_type embeddings
             # though currently we are passing all zeros to the token_type embeddings
@@ -4063,34 +4062,6 @@ class NomicBertModel(BertModel):
         raise ValueError(f"unknown tokenizer: {toktyp}")
 
 
-@ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")
-class NeoBert(BertModel):
-    model_arch = gguf.MODEL_ARCH.NEO_BERT
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-
-        # NeoBERT uses 2/3 of the intermediate size as feed forward length
-        self.gguf_writer.add_feed_forward_length(int(2 * self.hparams["intermediate_size"] / 3))
-        self.gguf_writer.add_rope_freq_base(10000.0)  # default value for NeoBERT
-        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
-
-        f_rms_eps = self.hparams.get("norm_eps", 1e-6)  # default value for NeoBERT
-        self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
-        logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
-
-        self.gguf_writer.add_pooling_type(gguf.PoolingType.CLS) # https://huggingface.co/chandar-lab/NeoBERT#how-to-use
-
-    def modify_tensors(self, data_torch, name, bid):
-        if name.startswith("decoder."):
-            return []
-
-        if name.startswith("model."):
-            name = name[6:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
-
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
@@ -4842,6 +4813,8 @@ class JinaBertV2Model(BertModel):
             self.gguf_writer.add_token_type_count(2)
         else:
             raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel')
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
 
 
 @ModelBase.register("OpenELMForCausalLM")
@@ -5083,7 +5056,9 @@ class DeepseekModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        if (rope_dim := hparams.get("head_dim")) is None:
+        if "head_dim" in hparams:
+            rope_dim = hparams["head_dim"]
+        else:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
@@ -5287,34 +5262,6 @@ class DeepseekV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("Dots1ForCausalLM")
-class Dots1Model(Qwen2MoeModel):
-    model_arch = gguf.MODEL_ARCH.DOTS1
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.hparams["num_experts"] = self.hparams["n_routed_experts"]
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
-        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
-        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
-        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
-
-        if self.hparams["scoring_func"] == "noaux_tc":
-            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
-        else:
-            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
-        if name.endswith("e_score_correction_bias"):
-            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
-        if "shared_experts" in name:
-            return [(self.map_tensor_name(name), data_torch)]
-        return super().modify_tensors(data_torch, name, bid)
-
-
 @ModelBase.register("PLMForCausalLM")
 class PLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PLM
@@ -5443,6 +5390,9 @@ class T5Model(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
     def set_gguf_parameters(self):
         if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
             logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@@ -5580,6 +5530,9 @@ class T5EncoderModel(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
     def set_gguf_parameters(self):
         if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
             logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@@ -5967,8 +5920,7 @@ class ExaoneModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                if (dim := self.hparams.get("head_dim")) is None:
-                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -6080,8 +6032,7 @@ class BailingMoeModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        if (rope_dim := hparams.get("head_dim")) is None:
-            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
+        rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
         rope_scaling = self.hparams.get("rope_scaling") or {}
@@ -6113,8 +6064,7 @@ class BailingMoeModel(TextModel):
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
         n_embd = self.hparams["hidden_size"]
-        if (head_dim := self.hparams.get("head_dim")) is None:
-            head_dim = n_embd // n_head
+        head_dim = self.hparams.get("head_dim") or n_embd // n_head
 
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
 
@@ -6375,8 +6325,8 @@ def parse_args() -> argparse.Namespace:
         help="model is executed on big endian machine",
     )
     parser.add_argument(
-        "model", type=str,
-        help="directory containing model file or huggingface repository ID (if --remote)",
+        "model", type=Path,
+        help="directory containing model file",
         nargs="?",
     )
     parser.add_argument(
@@ -6479,20 +6429,18 @@ def main() -> None:
     else:
         logging.basicConfig(level=logging.INFO)
 
+    dir_model = args.model
+
     if args.remote:
-        hf_repo_id = args.model
         from huggingface_hub import snapshot_download
         local_dir = snapshot_download(
-            repo_id=hf_repo_id,
+            repo_id=str(dir_model),
             allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"])
         dir_model = Path(local_dir)
         logger.info(f"Downloaded config and tokenizer to {local_dir}")
-    else:
-        hf_repo_id = None
-        dir_model = Path(args.model)
 
     if not dir_model.is_dir():
-        logger.error(f'Error: {dir_model} is not a directory')
+        logger.error(f'Error: {args.model} is not a directory')
         sys.exit(1)
 
     ftype_map: dict[str, gguf.LlamaFileType] = {
@@ -6512,9 +6460,9 @@ def main() -> None:
 
     if args.outfile is not None:
         fname_out = args.outfile
-    elif hf_repo_id:
+    elif args.remote:
         # if remote, use the model ID as the output file name
-        fname_out = Path("./" + hf_repo_id.replace("/", "-") + "-{ftype}.gguf")
+        fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf")
     else:
         fname_out = dir_model
 
@@ -6543,7 +6491,7 @@ def main() -> None:
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
-                                     remote_hf_model_id=hf_repo_id)
+                                     remote_hf_model_id=str(args.model) if args.remote else None)
 
         if args.vocab_only:
             logger.info("Exporting model vocab...")

docs/build-s390x.md

@@ -1,157 +0,0 @@
-> [!IMPORTANT]
-> This build documentation is specific only to IBM Z & LinuxONE mainframes (s390x). You can find the build documentation for other architectures: [build.md](build.md).
-
-# Build llama.cpp locally (for s390x)
-
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
-
-The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
-
-**To get the code:**
-
-```bash
-git clone https://github.com/ggml-org/llama.cpp
-cd llama.cpp
-```
-
-## CPU Build with BLAS
-
-Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements.
-
-```bash
-cmake -S . -B build             \
-    -DCMAKE_BUILD_TYPE=Release  \
-    -DGGML_BLAS=ON              \
-    -DGGML_BLAS_VENDOR=OpenBLAS
-
-cmake --build build --config Release -j $(nproc)
-```
-
-**Notes**:
-- For faster repeated compilation, install [ccache](https://ccache.dev/)
-- By default, VXE/VXE2 is enabled. To disable it (not recommended):
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DGGML_VXE=OFF
-
-    cmake --build build --config Release -j $(nproc)
-    ```
-
-- For debug builds:
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Debug    \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS
-
-    cmake --build build --config Debug -j $(nproc)
-    ```
-
-- For static builds, add `-DBUILD_SHARED_LIBS=OFF`:
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DBUILD_SHARED_LIBS=OFF
-
-    cmake --build build --config Release -j $(nproc)
-    ```
-
-## Getting GGUF Models
-
-All models need to be converted to Big-Endian. You can achieve this in three cases:
-
-1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)**
-
-    You can find popular models pre-converted and verified at [s390x Ready Models](hf.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08).
-
-    These models and their respective tokenizers are verified to run correctly on IBM Z & LinuxONE.
-
-2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
-
-    ```bash
-    python3 convert_hf_to_gguf.py \
-        --outfile model-name-be.f16.gguf \
-        --outtype f16 \
-        --bigendian \
-        model-directory/
-    ```
-
-    For example,
-
-    ```bash
-    python3 convert_hf_to_gguf.py \
-        --outfile granite-3.3-2b-instruct-be.f16.gguf \
-        --outtype f16 \
-        --bigendian \
-        granite-3.3-2b-instruct/
-    ```
-
-3. **Convert existing GGUF Little-Endian model to Big-Endian**
-
-    ```bash
-    python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
-    ```
-
-    For example,
-    ```bash
-    python3 gguf-py/gguf/scripts/gguf_convert_endian.py granite-3.3-2b-instruct-le.f16.gguf BIG
-    mv granite-3.3-2b-instruct-le.f16.gguf granite-3.3-2b-instruct-be.f16.gguf
-    ```
-
-    **Notes:**
-    - The GGUF endian conversion script may not support all data types at the moment and may fail for some models/quantizations. When that happens, please try manually converting the safetensors model to GGUF Big-Endian via Step 2.
-
-## IBM Accelerators
-
-### 1. SIMD Acceleration
-
-Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14 or EC13. In such systems, the APIs can still run but will use a scalar implementation.
-
-### 2. zDNN Accelerator
-
-*Only available in IBM z16 or later system. No direction at the moment.*
-
-### 3. Spyre Accelerator
-
-*No direction at the moment.*
-
-## Performance Tuning
-
-### 1. Virtualization Setup
-
-It is strongly recommended to use only LPAR (Type-1) virtualization to get the most performance.
-
-Note: Type-2 virtualization is not supported at the moment, while you can get it running, the performance will not be the best.
-
-### 2. IFL (Core) Count
-
-It is recommended to allocate a minimum of 8 shared IFLs assigned to the LPAR. Increasing the IFL count past 8 shared IFLs will only improve Prompt Processing performance but not Token Generation.
-
-Note: IFL count does not equate to vCPU count.
-
-### 3. SMT vs NOSMT (Simultaneous Multithreading)
-
-It is strongly recommended to disable SMT via the kernel boot parameters as it negatively affects performance. Please refer to your Linux distribution's guide on disabling SMT via kernel boot parameters.
-
-### 4. BLAS vs NOBLAS
-
-IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongly recommended to use BLAS.
-
-## Getting Help on IBM Z & LinuxONE
-
-1. **Bugs, Feature Requests**
-
-    Please file an issue in llama.cpp and ensure that the title contains "s390x".
-
-2. **Other Questions**
-
-    Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com).
-

docs/build.md

@@ -1,6 +1,6 @@
 # Build llama.cpp locally
 
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
 
 The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
 

docs/function-calling.md

@@ -11,7 +11,7 @@ Function calling is supported for all models (see https://github.com/ggml-org/ll
   - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2
   - Functionary v3.1 / v3.2
   - Hermes 2/3, Qwen 2.5
-  - Qwen 2.5 Coder
+  - Qwen 2.5 Coder (WIP: https://github.com/ggml-org/llama.cpp/pull/12034)
   - Mistral Nemo
   - Firefunction v2
   - Command R7B

docs/multimodal.md

@@ -107,7 +107,3 @@ NOTE: some models may require large context window, for example: `-c 8192`
 (tool_name) -hf ggml-org/Qwen2.5-Omni-3B-GGUF
 (tool_name) -hf ggml-org/Qwen2.5-Omni-7B-GGUF
 ```
-
-## Finding more models:
-
-GGUF models on Huggingface with vision capabilities can be found here: https://huggingface.co/models?pipeline_tag=image-text-to-text&sort=trending&search=gguf

examples/embedding/embedding.cpp

@@ -133,36 +133,10 @@ int main(int argc, char ** argv) {
     // max batch size
     const uint64_t n_batch = params.n_batch;
 
-    // get added sep and eos token, if any
-    const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : "";
-    const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : "";
-
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        std::vector<llama_token> inp;
-
-        // split classification pairs and insert expected separator tokens
-        if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) {
-            std::vector<std::string> pairs = split_lines(prompt, params.cls_sep);
-            std::string final_prompt;
-
-            for (size_t i = 0; i < pairs.size(); i++) {
-                final_prompt += pairs[i];
-                if (i != pairs.size() - 1) {
-                    if (!added_eos_token.empty()) {
-                        final_prompt += added_eos_token;
-                    }
-                    if (!added_sep_token.empty()) {
-                        final_prompt += added_sep_token;
-                    }
-                }
-            }
-
-            inp = common_tokenize(ctx, final_prompt, true, true);
-        } else {
-            inp = common_tokenize(ctx, prompt, true, true);
-        }
+        auto inp = common_tokenize(ctx, prompt, true, true);
         if (inp.size() > n_batch) {
             LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -171,11 +145,11 @@ int main(int argc, char ** argv) {
         inputs.push_back(inp);
     }
 
-    // check if the last token is SEP/EOS
+    // check if the last token is SEP
     // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
-        if (inp.empty() || (inp.back() != llama_vocab_sep(vocab) && inp.back() != llama_vocab_eos(vocab))) {
-            LOG_WRN("%s: last token in the prompt is not SEP or EOS\n", __func__);
+        if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
+            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
             LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }

examples/gritlm/gritlm.cpp

@@ -41,11 +41,12 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // add input to batch (this increments n_tokens)
         for (int32_t j = 0; j < n_toks; j++) {
-            common_batch_add(batch, inputs[j], j, { 0 }, true);
+            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
         llama_memory_clear(llama_get_memory(ctx), true);
+        llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -102,6 +103,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
     llama_token eos_token = llama_vocab_eos(vocab);
 
     llama_memory_clear(llama_get_memory(ctx), true);
+    llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
@@ -164,8 +166,6 @@ int main(int argc, char * argv[]) {
     llama_model_params mparams = common_model_params_to_llama(params);
     llama_context_params cparams = common_context_params_to_llama(params);
 
-    cparams.embeddings = true;
-
     llama_backend_init();
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -213,8 +213,6 @@ int main(int argc, char * argv[]) {
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
     }
 
-    llama_set_embeddings(ctx, false);
-
     // ### Generation ###
     // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
     {

ggml/CMakeLists.txt

@@ -172,7 +172,6 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
-option(GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
@@ -368,8 +367,6 @@ if (MSVC)
         /wd4005  # Macro redefinition
         /wd4244  # Conversion from one type to another type, possible loss of data
         /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
-        /wd4305  # Conversion from 'type1' to 'type2', possible loss of data
-        /wd4566  # Conversion from 'char' to 'wchar_t', possible loss of data
         /wd4996  # Disable POSIX deprecation warnings
         /wd4702  # Unreachable code warnings
     )
@@ -389,46 +386,4 @@ if (MSVC)
     disable_msvc_warnings(ggml-cpu-skylakex)
     disable_msvc_warnings(ggml-cpu-icelake)
     disable_msvc_warnings(ggml-cpu-alderlake)
-
-    if (GGML_BUILD_EXAMPLES)
-        disable_msvc_warnings(common-ggml)
-        disable_msvc_warnings(common)
-
-        disable_msvc_warnings(mnist-common)
-        disable_msvc_warnings(mnist-eval)
-        disable_msvc_warnings(mnist-train)
-
-        disable_msvc_warnings(gpt-2-ctx)
-        disable_msvc_warnings(gpt-2-alloc)
-        disable_msvc_warnings(gpt-2-backend)
-        disable_msvc_warnings(gpt-2-sched)
-        disable_msvc_warnings(gpt-2-quantize)
-        disable_msvc_warnings(gpt-2-batched)
-
-        disable_msvc_warnings(gpt-j)
-        disable_msvc_warnings(gpt-j-quantize)
-
-        disable_msvc_warnings(magika)
-        disable_msvc_warnings(yolov3-tiny)
-        disable_msvc_warnings(sam)
-
-        disable_msvc_warnings(simple-ctx)
-        disable_msvc_warnings(simple-backend)
-    endif()
-
-    if (GGML_BUILD_TESTS)
-        disable_msvc_warnings(test-mul-mat)
-        disable_msvc_warnings(test-arange)
-        disable_msvc_warnings(test-backend-ops)
-        disable_msvc_warnings(test-cont)
-        disable_msvc_warnings(test-conv-transpose)
-        disable_msvc_warnings(test-conv-transpose-1d)
-        disable_msvc_warnings(test-conv1d)
-        disable_msvc_warnings(test-conv2d)
-        disable_msvc_warnings(test-conv2d-dw)
-        disable_msvc_warnings(test-customop)
-        disable_msvc_warnings(test-dup)
-        disable_msvc_warnings(test-opt)
-        disable_msvc_warnings(test-pool)
-    endif ()
 endif()

ggml/cmake/common.cmake

@@ -36,7 +36,8 @@ function(ggml_get_system_arch)
             (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
             CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
         set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
+    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR
+            "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
         set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)

ggml/include/ggml.h

@@ -489,7 +489,6 @@ extern "C" {
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
         GGML_OP_PAD_REFLECT_1D,
-        GGML_OP_ROLL,
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
@@ -1802,17 +1801,6 @@ extern "C" {
             int                   p0,
             int                   p1);
 
-    // Move tensor elements by an offset given for each dimension. Elements that
-    // are shifted beyond the last position are wrapped around to the beginning.
-    GGML_API struct ggml_tensor * ggml_roll(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                   shift0,
-            int                   shift1,
-            int                   shift2,
-            int                   shift3);
-
-
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
     // return: [N, dim]

ggml/src/CMakeLists.txt

@@ -286,10 +286,6 @@ function(ggml_add_cpu_backend_variant tag_name)
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
-    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
-        foreach (feat ${ARGN})
-            set(GGML_INTERNAL_${feat} ON)
-        endforeach()
     endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -315,45 +311,18 @@ if (GGML_CPU_ALL_VARIANTS)
             # MSVC doesn't support AMX
             ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
         endif()
-    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
-        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-            # Many of these features are optional so we build versions with popular
-            # combinations and name the backends based on the version they were
-            # first released with
-            ggml_add_cpu_backend_variant(armv8.0_1)
-            ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
-            ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
-            ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
-            ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
-            ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
-            ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
-            ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
-        elseif (CMAKE_SYSTEM_NAME MATCHES "Android")
-            # Android-specific backends with SoC-compatible feature sets
-            ggml_add_cpu_backend_variant(android_armv8.0_1)
-            ggml_add_cpu_backend_variant(android_armv8.2_1    DOTPROD)
-            ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
-            ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
-        elseif (APPLE)
-            ggml_add_cpu_backend_variant(apple_m1             DOTPROD)
-            ggml_add_cpu_backend_variant(apple_m2_m3          DOTPROD MATMUL_INT8)
-            ggml_add_cpu_backend_variant(apple_m4             DOTPROD MATMUL_INT8 NOSVE SME)
-        else()
-            message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
-        endif()
-    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
-        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-            ggml_add_cpu_backend_variant(power0)
-            ggml_add_cpu_backend_variant(power7_1       POWER7)
-            ggml_add_cpu_backend_variant(power7_2       POWER7  VSX)
-            ggml_add_cpu_backend_variant(power8_1       POWER8)
-            ggml_add_cpu_backend_variant(power8_2       POWER8  VSX)
-            ggml_add_cpu_backend_variant(power9         POWER9  VSX)
-            ggml_add_cpu_backend_variant(power10        POWER10 VSX)
-            ggml_add_cpu_backend_variant(power11        POWER11 VSX)
-        else()
-            message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
-        endif()
+    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM" AND CMAKE_SYSTEM_NAME MATCHES "Linux")
+        # Many of these features are optional so we build versions with popular
+        # combinations and name the backends based on the version they were
+        # first released with
+        ggml_add_cpu_backend_variant(armv8.0_1)
+        ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
+        ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+        ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
+        ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
+        ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
+        ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
+        ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
     else()
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()

ggml/src/ggml-backend-reg.cpp

@@ -69,9 +69,6 @@
 #if defined(__clang__)
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
 
 namespace fs = std::filesystem;
@@ -94,8 +91,6 @@ static std::string path_str(const fs::path & path) {
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
 #endif
 
 #ifdef _WIN32

ggml/src/ggml-cpu/CMakeLists.txt

@@ -158,48 +158,48 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
                 elseif(GGML_CPU_ALL_VARIANTS)
-                    # Begin with the lowest baseline
-                    set(ARM_MCPU "armv8-a")
-                    set(ARCH_TAGS "")
-                    set(ARCH_DEFINITIONS "")
+                    if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+                        # Begin with the lowest baseline
+                        set(ARM_MCPU "armv8-a")
+                        set(ARCH_TAGS "")
+                        set(ARCH_DEFINITIONS "")
 
-                    # When a feature is selected, bump the MCPU to the first
-                    # version that supported it
-                    if (GGML_INTERNAL_DOTPROD)
-                        set(ARM_MCPU "armv8.2-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
+                        # When a feature is selected, bump the MCPU to the first
+                        # version that supported it
+                        if (GGML_INTERNAL_DOTPROD)
+                            set(ARM_MCPU "armv8.2-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
+                        endif()
+                        if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
+                            set(ARM_MCPU "armv8.2-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+fp16")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
+                        endif()
+                        if (GGML_INTERNAL_SVE)
+                            set(ARM_MCPU "armv8.2-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+sve")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
+                        endif()
+                        if (GGML_INTERNAL_MATMUL_INT8)
+                            set(ARM_MCPU "armv8.6-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
+                        endif()
+                        if (GGML_INTERNAL_SVE2)
+                            set(ARM_MCPU "armv8.6-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+sve2")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
+                        endif()
+                        if (GGML_INTERNAL_SME)
+                            set(ARM_MCPU "armv9.2-a")
+                            set(ARCH_TAGS "${ARCH_TAGS}+sme")
+                            list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
+                        endif()
+
+                        list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
+                        ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
                     endif()
-                    if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
-                        set(ARM_MCPU "armv8.2-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+fp16")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
-                    endif()
-                    if (GGML_INTERNAL_SVE)
-                        set(ARM_MCPU "armv8.2-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+sve")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
-                    endif()
-                    if (GGML_INTERNAL_MATMUL_INT8)
-                        set(ARM_MCPU "armv8.6-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
-                    endif()
-                    if (GGML_INTERNAL_SVE2)
-                        set(ARM_MCPU "armv8.6-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+sve2")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
-                    endif()
-                    if (GGML_INTERNAL_NOSVE)
-                        set(ARCH_TAGS "${ARCH_TAGS}+nosve")
-                    endif()
-                    if (GGML_INTERNAL_SME)
-                        set(ARM_MCPU "armv9.2-a")
-                        set(ARCH_TAGS "${ARCH_TAGS}+sme")
-                        list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
-                    endif()
-                    list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
-                    ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
                 endif()
             endif()
 
@@ -388,27 +388,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
             endif()
-        elseif(GGML_CPU_ALL_VARIANTS)
-            # Begin with the lowest baseline
-            set(ARCH_DEFINITIONS "")
-
-            # When a feature is selected, bump the MCPU to the first
-            # version that supported it
-            foreach(PVER RANGE 7 11)
-                if(DEFINED GGML_INTERNAL_POWER${PVER})
-                    set(POWERPC_MCPU "power${PVER}")
-                    list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER})
-                endif()
-            endforeach()
-            if (GGML_INTERNAL_VSX)
-                list(APPEND ARCH_DEFINITIONS GGML_USE_VSX)
-                list(APPEND ARCH_FLAGS -mvsx)
-            endif()
-
-            if (DEFINED POWERPC_MCPU)
-                list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU})
-            endif()
-            ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS})
         else()
             if (GGML_CPU_POWERPC_CPUTYPE)
                 list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
@@ -486,9 +465,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.6.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
+        set(KLEIDIAI_ARCHIVE_MD5  "75b4ad68f25ab673dcc01065e5a0b05f")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)

ggml/src/ggml-cpu/arch-fallback.h

@@ -1,184 +0,0 @@
-#pragma once
-
-// Rename `_generic` functions if no native implementation is available.
-// This effectively selects the generic implementation.
-
-#if defined(GGML_CPU_GENERIC)
-// quants.c
-#define quantize_row_q8_0_generic quantize_row_q8_0
-#define quantize_row_q8_1_generic quantize_row_q8_1
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
-#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
-#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
-#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
-#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
-#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
-#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
-#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
-// repack.cpp
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__POWERPC__) || defined(__powerpc__)
-// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
-// quants.c
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__loongarch64)
-// quants.c
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__riscv)
-// quants.c
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
-#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__s390x__)
-// quants.c
-#define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#elif defined(__wasm__)
-// quants.c
-#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
-#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
-#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
-#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
-#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
-// repack.cpp
-#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
-#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
-#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
-#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
-#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
-#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
-#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
-#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
-#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
-#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
-#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
-#endif

ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp

@@ -1,82 +0,0 @@
-# include "ggml-backend-impl.h"
-
-#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
-
-#if defined(__linux__)
-#include <sys/auxv.h>
-#endif
-
-#include <string>
-
-struct powerpc_features {
-    std::string platform = "";
-    int power_version    = -1;
-
-    bool has_vsx         = false;
-
-    powerpc_features() {
-#if defined(__linux__)
-        unsigned long auxval = getauxval(AT_PLATFORM);
-        if (auxval) {
-            platform = std::string(reinterpret_cast<const char*>(auxval));
-            // TBD: Do systems exist that return this in uppercase?
-            if (platform.substr(0, 5) == "power") {
-                // Extractt a numeric suffix, if one exists
-                int vpos = -1;
-                for (int i = platform.length() - 1; i >= 0; i--) {
-                    if (std::isdigit(platform[i])) {
-                        vpos = i;
-                    } else {
-                        break;
-                    }
-                }
-                if (vpos > -1) {
-                    power_version = std::stoi(platform.substr(vpos));
-                }
-            }
-        }
-#endif
-        if (power_version >= 9) {
-            has_vsx = true;
-        }
-    }
-};
-
-static int ggml_backend_cpu_powerpc_score() {
-    int score = 1;
-    powerpc_features pf;
-
-// Platform scores
-#if defined(GGML_USE_POWER7)
-    if (pf.power_version < 7) { return 0; }
-    score += 1<<1;
-#endif
-#if defined(GGML_USE_POWER8)
-    if (pf.power_version < 8) { return 0; }
-    score += 1<<2;
-#endif
-#if defined(GGML_USE_POWER9)
-    if (pf.power_version < 9) { return 0; }
-    score += 1<<3;
-#endif
-#if defined(GGML_USE_POWER10)
-    if (pf.power_version < 10) { return 0; }
-    score += 1<<4;
-#endif
-#if defined(GGML_USE_POWER11)
-    if (pf.power_version < 11) { return 0; }
-    score += 1<<5;
-#endif
-
-// Feature scores
-#if defined(GGML_USE_VSX)
-    if (!pf.has_vsx) { return 0; }
-    score += 1<<6;
-#endif
-
-    return score;
-}
-
-GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score)
-
-#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -371,7 +371,7 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
 #endif
 
-typedef signed   char char8x16_t  __attribute__((vector_size(16)));
+typedef signed char char8x16_t __attribute__((vector_size(16)));
 typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
 
 typedef int8_t  int8x16_t __attribute__((vector_size(16)));
@@ -382,10 +382,10 @@ typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
 typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
 typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
 
-typedef float  float32x4_t  __attribute__((vector_size(16)));
-typedef double double64x2_t __attribute__((vector_size(16)));
+typedef float float32x4_t __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute((vector_size(16)));
 
-typedef signed   long long long64x2_t  __attribute__((vector_size(16)));
+typedef signed long long long64x2_t __attribute((vector_size(16)));
 typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
 
 typedef struct ggml_uint8x16x2_t {
@@ -503,9 +503,31 @@ static __m256 __lasx_xvreplfr2vr_s(const float val) {
 // TODO: move to ggml-threading
 void ggml_barrier(struct ggml_threadpool * tp);
 
-void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
-int  ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
-
 #ifdef __cplusplus
 }
 #endif
+
+#define GGML_DO_PRAGMA_(x) _Pragma (#x)
+#define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
+#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__)
+// Note for Apple targets:
+// - clang: aliases are not supported on darwin
+// - all native kernels need to be implemented in both x86 and arm files
+// - on iOS, tvOS, and visionOS, if cmake cannot determine the target architecture, all `_generic` names are replaced by defines
+# define GGML_WEAK_ALIAS(name, alias)
+#elif defined(__GNUC__)
+// GCC/Clang on *nix
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(weak name = alias) // NOLINT
+#elif defined(_MSC_VER) && defined(_WIN64)
+// MSVC
+// Note: C name mangling varies across different calling conventions
+// see https://learn.microsoft.com/en-us/cpp/build/reference/decorated-names?view=msvc-170
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:" #name "=" #alias))
+#elif defined(_MSC_VER) && defined(WIN32)
+// ref: https://github.com/ggml-org/whisper.cpp/pull/3239#issuecomment-2958224591
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:_" #name "=_" #alias))
+#else
+# error "Unsupported compiler for GGML_WEAK_ALIAS"
+#endif
+
+#define GGML_CPU_NATIVE_IMPL(name) GGML_WEAK_ALIAS(name, name ## _generic)

ggml/src/ggml-cpu/ggml-cpu.c

@@ -74,8 +74,13 @@
 
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
+    int has_neon;
+    int has_dotprod;
+    int has_i8mm;
+    int has_sve;
     int sve_cnt;
-} ggml_arm_arch_features = { 0 };
+    int has_sme;
+} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
 #endif
 
 
@@ -554,14 +559,6 @@ void ggml_barrier(struct ggml_threadpool * tp) {
 #endif
 }
 
-void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value) {
-    atomic_store_explicit(&tp->current_chunk, value, memory_order_relaxed);
-}
-
-int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value) {
-    return atomic_fetch_add_explicit(&tp->current_chunk, value, memory_order_relaxed);
-}
-
 #if defined(__gnu_linux__)
 static cpu_set_t ggml_get_numa_affinity(void) {
     cpu_set_t cpuset;
@@ -673,15 +670,87 @@ bool ggml_is_numa(void) {
 
 #if defined(__linux__) && defined(__aarch64__)
 #include <sys/auxv.h>
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
+#if !defined(HWCAP2_I8MM)
+#define HWCAP2_I8MM (1 << 13)
+#endif
+
+#if !defined(HWCAP2_SME)
+#define HWCAP2_SME (1 << 23)
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#if defined(__linux__) && defined(__aarch64__)
+    uint32_t hwcap = getauxval(AT_HWCAP);
+    uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+    ggml_arm_arch_features.has_neon    = !!(hwcap & HWCAP_ASIMD);
+    ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
+    ggml_arm_arch_features.has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
+    ggml_arm_arch_features.has_sve     = !!(hwcap & HWCAP_SVE);
+    ggml_arm_arch_features.has_sme     = !!(hwcap2 & HWCAP2_SME);
+
+#if defined(__ARM_FEATURE_SVE)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
 #endif
-}
+#elif defined(__APPLE__)
+    int oldp = 0;
+    size_t size = sizeof(oldp);
+    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_neon = oldp;
 
-#endif // __ARM_ARCH
+    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_dotprod = oldp;
+
+    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_i8mm = oldp;
+
+    if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_sme = oldp;
+
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#else
+// Run-time CPU feature detection not implemented for this platform, fallback to compile time
+#if defined(__ARM_NEON)
+    ggml_arm_arch_features.has_neon = 1;
+#else
+    ggml_arm_arch_features.has_neon = 0;
+#endif
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_arm_arch_features.has_i8mm = 1;
+#else
+    ggml_arm_arch_features.has_i8mm = 0;
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.has_sve = 1;
+    ggml_arm_arch_features.sve_cnt = 16;
+#else
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#endif
+
+#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
+    ggml_arm_arch_features.has_sme = 1;
+#else
+    ggml_arm_arch_features.has_sme = 0;
+#endif
+#endif
+}
+#endif
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
     GGML_ASSERT(!ggml_get_no_alloc(ctx));
@@ -1890,10 +1959,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_pad_reflect_1d(params, tensor);
             } break;
-        case GGML_OP_ROLL:
-            {
-                ggml_compute_forward_roll(params, tensor);
-            } break;
         case GGML_OP_ARANGE:
             {
                 ggml_compute_forward_arange(params, tensor);
@@ -2218,7 +2283,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_PAD_REFLECT_1D:
-        case GGML_OP_ROLL:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
@@ -3371,7 +3435,7 @@ int ggml_cpu_has_vxe(void) {
 
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
-    return 1;
+    return ggml_arm_arch_features.has_neon;
 #else
     return 0;
 #endif
@@ -3379,7 +3443,7 @@ int ggml_cpu_has_neon(void) {
 
 int ggml_cpu_has_dotprod(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
-    return 1;
+    return ggml_arm_arch_features.has_dotprod;
 #else
     return 0;
 #endif
@@ -3387,7 +3451,7 @@ int ggml_cpu_has_dotprod(void) {
 
 int ggml_cpu_has_sve(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
-    return 1;
+    return ggml_arm_arch_features.has_sve;
 #else
     return 0;
 #endif
@@ -3395,7 +3459,7 @@ int ggml_cpu_has_sve(void) {
 
 int ggml_cpu_has_matmul_int8(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
-    return 1;
+    return ggml_arm_arch_features.has_i8mm;
 #else
     return 0;
 #endif
@@ -3411,7 +3475,7 @@ int ggml_cpu_get_sve_cnt(void) {
 
 int ggml_cpu_has_sme(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
-    return 1;
+    return ggml_arm_arch_features.has_sme;
 #else
     return 0;
 #endif

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -53,6 +53,7 @@
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
+#include <atomic>
 #include <array>
 #include <type_traits>
 
@@ -62,7 +63,7 @@
 #define NOINLINE __attribute__((__noinline__))
 #endif
 
-#if defined(__ARM_NEON) || defined(__AVX512F__) || defined(__VXE__) || defined(__VXE2__)
+#if defined(__ARM_NEON) || defined(__AVX512F__)
 #define VECTOR_REGISTERS 32
 #else
 #define VECTOR_REGISTERS 16
@@ -109,12 +110,6 @@ inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
 inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
-#if defined(__VXE__) || defined(__VXE2__)
-inline float32x4_t add(float32x4_t x, float32x4_t y) { return vec_add(x, y); }
-inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vec_sub(x, y); }
-inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
-#endif
-
 #if defined(__MMA__)
 typedef vector unsigned char vec_t;
 typedef __vector_quad acc_t;
@@ -168,13 +163,6 @@ inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
 #endif
 #endif
 
-#if defined(__VXE__) || defined(__VXE2__)
-template <>
-inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
-    return vec_madd(a, b, c);
-}
-#endif
-
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED HORIZONTAL SUM
 
@@ -191,13 +179,6 @@ inline float hsum(float16x8_t x) {
 }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
-#if defined(__VXE__) || defined(__VXE2__)
-inline float hsum(float32x4_t x) {
-    float32x4_t tmp = x + vec_reve(x);
-    return tmp[0] + tmp[1];
-}
-#endif
-
 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 inline float hsum(__m128 x) {
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
@@ -247,21 +228,6 @@ template <> inline float32x4_t load(const ggml_fp16_t *p) {
 #endif // _MSC_VER
 #endif // __ARM_NEON
 
-#if defined(__VXE__) || defined(__VXE2__)
-template <> inline float32x4_t load(const ggml_fp16_t * p) {
-    float tmp[4];
-
-    for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(p[i]);
-    }
-
-    return vec_xl(0, (const float *)(tmp));
-}
-template <> inline float32x4_t load(const float * p) {
-    return vec_xl(0, p);
-}
-#endif
-
 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 template <> inline __m128 load(const float *p) {
     return _mm_loadu_ps(p);
@@ -428,6 +394,8 @@ class tinyBLAS {
 
     template <int RM, int RN, int BM>
     NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
+        static std::atomic<int64_t> current_chunk;
+
         GGML_ASSERT(m % (RM * BM) == 0);
         const int64_t ytiles = m / (RM * BM);
         const int64_t xtiles = (n + RN -1) / RN;
@@ -442,7 +410,7 @@ class tinyBLAS {
         if (params->ith == 0) {
             GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
             // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-            ggml_threadpool_chunk_set(params->threadpool, params->nth);
+            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
         }
 
         ggml_barrier(params->threadpool);
@@ -471,7 +439,8 @@ class tinyBLAS {
                 GGML_ASSERT(jj == jj2);
             }
 
-            job = ggml_threadpool_chunk_add(params->threadpool, 1);
+            // next step.
+            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
         }
 
         ggml_barrier(params->threadpool);
@@ -3354,14 +3323,6 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             (const float *)B, ldb,
             (float *)C, ldc};
         return tb.matmul(m, n);
-#elif defined(__VXE__) || defined(__VXE2__)
-        if (n < 4)
-            return false;
-        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
-            k, (const float *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc};
-        return tb.matmul(m, n);
 #elif defined(__MMA__)
         if (k % 8)
             return false;
@@ -3453,16 +3414,6 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                 (float *)C, ldc};
             return tb.matmul(m, n);
         }
-#elif defined(__VXE__) || defined(__VXE2__)
-        if (n < 4)
-            return false;
-        if (Btype == GGML_TYPE_F16) {
-            tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
-                k, (const ggml_fp16_t *)A, lda,
-                (const ggml_fp16_t *)B, ldb,
-                (float *)C, ldc};
-            return tb.matmul(m, n);
-        }
 #endif
         return false;
     }

ggml/src/ggml-cpu/llamafile/sgemm.h

@@ -1,11 +1,6 @@
 #pragma once
 #include <stdint.h>
 #include <stdbool.h>
-
-#if defined(__VXE__) || defined(__VXE2__)
-#include <vecintrin.h>
-#endif
-
 #ifdef __cplusplus
 extern "C" {
 #endif

ggml/src/ggml-cpu/ops.cpp

@@ -6793,73 +6793,6 @@ void ggml_compute_forward_pad_reflect_1d(
     }
 }
 
-// ggml_compute_forward_roll
-
-static int64_t ggml_wrap_index(int64_t i, int64_t ne) {
-    if (i < 0) {
-        return i + ne;
-    } else if (i >= ne) {
-        return i - ne;
-    }
-    return i;
-}
-
-static void ggml_compute_forward_roll_f32(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src_data = (const float *) src0->data;
-    float * dst_data = (float *) dst->data;
-
-    GGML_TENSOR_UNARY_OP_LOCALS
-
-    const int s0 = ggml_get_op_params_i32(dst, 0);
-    const int s1 = ggml_get_op_params_i32(dst, 1);
-    const int s2 = ggml_get_op_params_i32(dst, 2);
-    const int s3 = ggml_get_op_params_i32(dst, 3);
-
-    const int64_t total = ne1 * ne2 * ne3;
-    const int64_t per_thread = (total + params->nth) / params->nth;
-    const int64_t start = params->ith * per_thread;
-    const int64_t end   = std::min(start + per_thread, total);
-
-    for (int64_t i = start; i < end; ++i) {
-        const int64_t i1 = i % ne1;
-        const int64_t i2 = (i / ne1) % ne2;
-        const int64_t i3 = i / (ne2 * ne1);
-        float * dst_row = dst_data + (i3*nb3 + i2*nb2 + i1*nb1) / sizeof(float);
-
-        const int64_t i01 = ggml_wrap_index(i1 - s1, ne01);
-        const int64_t i02 = ggml_wrap_index(i2 - s2, ne02);
-        const int64_t i03 = ggml_wrap_index(i3 - s3, ne03);
-        const float * src_row = src_data + (i03*nb03 + i02*nb02 + i01*nb01) / sizeof(float);
-
-        const int64_t s = ggml_wrap_index(-s0, ne00);
-        const int64_t n = ne00 - s;
-        ggml_vec_cpy_f32(n, dst_row,     src_row + s);
-        ggml_vec_cpy_f32(s, dst_row + n, src_row);
-    }
-}
-
-void ggml_compute_forward_roll(
-        const ggml_compute_params * params,
-        ggml_tensor * dst) {
-
-    const ggml_tensor * src0 = dst->src[0];
-
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_roll_f32(params, dst);
-            } break;
-        default:
-            {
-                GGML_ABORT("fatal error");
-            }
-    }
-}
-
 // ggml_compute_forward_arange
 
 static void ggml_compute_forward_arange_f32(

ggml/src/ggml-cpu/ops.h

@@ -72,7 +72,6 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
 void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/quants.c

@@ -5,8 +5,6 @@
 #include "ggml-quants.h"
 #include "quants.h"
 
-#include "arch-fallback.h"
-
 #include <string.h>
 #include <assert.h>
 #include <float.h>
@@ -40,10 +38,12 @@ void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
     quantize_row_q8_0_ref(x, y, k);
 }
+GGML_CPU_NATIVE_IMPL(quantize_row_q8_0)
 
 void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
     quantize_row_q8_1_ref(x, y, k);
 }
+GGML_CPU_NATIVE_IMPL(quantize_row_q8_1)
 
 //
 // 2-6 bit quantization in super-blocks
@@ -104,6 +104,7 @@ void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy,
 void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
     quantize_row_q8_K_ref(x, y, k);
 }
+GGML_CPU_NATIVE_IMPL(quantize_row_q8_K)
 
 //===================================== Dot products =================================
 
@@ -142,6 +143,7 @@ void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_0_q8_0)
 
 // TODO: add WASM SIMD
 void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -179,6 +181,7 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_1_q8_1)
 
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
@@ -222,6 +225,7 @@ void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_0_q8_0)
 
 void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_1;
@@ -265,6 +269,7 @@ void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_1_q8_1)
 
 void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
@@ -295,6 +300,7 @@ void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q8_0_q8_0)
 
 void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
@@ -347,6 +353,7 @@ void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_tq1_0_q8_K)
 
 void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
@@ -379,6 +386,7 @@ void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_tq2_0_q8_K)
 
 void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
@@ -431,6 +439,7 @@ void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     }
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q2_K_q8_K)
 
 void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -510,6 +519,7 @@ void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     for (int l = 0; l < 8; ++l) sumf += sums[l];
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q3_K_q8_K)
 
 void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -585,6 +595,7 @@ void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     for (int l = 0; l < 8; ++l) sumf += sums[l];
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q4_K_q8_K)
 
 void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy,  size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -665,6 +676,7 @@ void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     for (int l = 0; l < 8; ++l) sumf += sums[l];
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q5_K_q8_K)
 
 void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -720,6 +732,7 @@ void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     for (int l = 0; l < 8; ++l) sumf += sums[l];
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_q6_K_q8_K)
 
 void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -762,6 +775,7 @@ void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
     *s = 0.125f * sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_xxs_q8_K)
 
 void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -812,6 +826,7 @@ void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
     *s = 0.125f * sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_xs_q8_K)
 
 void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -864,6 +879,7 @@ void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 
     *s = 0.125f * sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq2_s_q8_K)
 
 void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -908,6 +924,7 @@ void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
     *s = 0.25f * sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq3_xxs_q8_K)
 
 void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -964,6 +981,7 @@ void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq3_s_q8_K)
 
 void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -1007,6 +1025,7 @@ void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq1_s_q8_K)
 
 void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
@@ -1068,6 +1087,7 @@ void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq1_m_q8_K)
 
 void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
@@ -1097,6 +1117,7 @@ void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq4_nl_q8_0)
 
 void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(nrc == 1);
@@ -1143,6 +1164,7 @@ void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
     *s = sumf;
 }
+GGML_CPU_NATIVE_IMPL(ggml_vec_dot_iq4_xs_q8_K)
 
 // ============================ 4-bit non-linear quants
 

ggml/src/ggml-cpu/quants.h

@@ -84,6 +84,33 @@ void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
+#if defined(GGML_CPU_GENERIC)
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+#endif
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-cpu/repack.cpp

@@ -8,8 +8,6 @@
 #include "ggml-cpu-impl.h"
 #include "traits.h"
 
-#include "arch-fallback.h"
-
 #include <cmath>
 #include <cstring>
 #include <cassert>
@@ -85,6 +83,7 @@ void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GG
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x4)
 
 void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
@@ -123,6 +122,7 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_0_4x8)
 
 void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK_K == 256);
@@ -174,6 +174,7 @@ void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GG
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_quantize_mat_q8_K_4x8)
 
 } // extern "C"
 
@@ -243,6 +244,7 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
         for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x4_q8_0)
 
 void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -287,6 +289,7 @@ void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
         for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_4x8_q8_0)
 
 void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -333,6 +336,7 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_0_8x8_q8_0)
 
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
@@ -411,6 +415,7 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemv_q4_K_8x8_q8_K)
 
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -457,6 +462,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemv_iq4_nl_4x4_q8_0)
 
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -513,6 +519,7 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x4_q8_0)
 
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -567,6 +574,7 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_4x8_q8_0)
 
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -621,6 +629,7 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_0_8x8_q8_0)
 
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
@@ -710,6 +719,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemm_q4_K_8x8_q8_K)
 
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -766,6 +776,7 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
         }
     }
 }
+GGML_CPU_NATIVE_IMPL(ggml_gemm_iq4_nl_4x4_q8_0)
 
 } // extern "C"
 
@@ -1163,24 +1174,13 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         // not realy a GGML_TYPE_Q8_0 but same size.
         switch (op->op) {
             case GGML_OP_MUL_MAT:
-                {
-                    size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
-                    return true;
-                }
+                size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
+                return true;
             case GGML_OP_MUL_MAT_ID:
-                {
-                    size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
-                    size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc.
-
-                    const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert
-                    const int64_t ne12 = op->src[1]->ne[2]; // n_tokens
-
-                    const size_t sizeof_mmid_row_mapping = sizeof(int64_t);
-
-                    size += sizeof_mmid_row_mapping*ne02*(ne12 + 1);
-
-                    return true;
-                }
+                size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
+                size = GGML_PAD(size, sizeof(int64_t));  // + padding for next bloc.
+                size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2];
+                return true;
             default:
                 // GGML_ABORT("fatal error");
                 break;
@@ -1316,17 +1316,14 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             int32_t i2;
         };
 
-        GGML_ASSERT(params->wsize >=
-                (GGML_PAD(nbw3, sizeof(int64_t)) +
-                 n_as*(ne12 + 1)*sizeof(mmid_row_mapping))
-                );
+        GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) +
+                                      n_as * ne12 * sizeof(mmid_row_mapping)));
 
-        auto * wdata          = (char *)params->wdata;
-        auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t));
+        auto * wdata             = (char *)     params->wdata;
+        auto * wdata_src1_end    = (char *)     wdata + GGML_PAD(nbw3, sizeof(int64_t));
+        auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
 
-        // total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = int64_t)
-        auto * matrix_row_counts = (int64_t *) (wdata_src1_end);                                        // [n_as]
-        struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12]
+        struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as);  // [n_as][ne12]
 
         // src1: float32 => param type
         for (int64_t i12 = 0; i12 < ne12; ++i12) {
@@ -1411,45 +1408,44 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
     }
 };
 
+// instance for Q4
+static const tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
+static const tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
+static const tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+static const tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
+
+// instance for IQ4
+static const tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
+
 }  // namespace ggml::cpu::repack
 
 static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) {
-
-    // instance for Q4
-    static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
-    static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
-    static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
-    static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
-
-    // instance for IQ4
-    static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
-
     if (cur->type == GGML_TYPE_Q4_0) {
         if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
             if (cur->ne[1] % 8 == 0) {
-                return &q4_0_8x8_q8_0;
+                return &ggml::cpu::repack::q4_0_8x8_q8_0;
             }
         }
         if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
             if (cur->ne[1] % 4 == 0) {
-                return &q4_0_4x8_q8_0;
+                return &ggml::cpu::repack::q4_0_4x8_q8_0;
             }
         }
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {
-                return &q4_0_4x4_q8_0;
+                return &ggml::cpu::repack::q4_0_4x4_q8_0;
             }
         }
     } else if (cur->type == GGML_TYPE_Q4_K) {
         if (ggml_cpu_has_avx2()) {
             if (cur->ne[1] % 8 == 0) {
-                return &q4_K_8x8_q8_K;
+                return &ggml::cpu::repack::q4_K_8x8_q8_K;
             }
         }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {
-                return &iq4_nl_4x4_q8_0;
+                return &ggml::cpu::repack::iq4_nl_4x4_q8_0;
             }
         }
     }

ggml/src/ggml-cpu/repack.h

@@ -64,6 +64,10 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wro
 extern "C" {
 #endif
 
+// Workaround for clang:
+// clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous''
+// repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call)
+#if defined(GGML_CPU_CLANG_WORKAROUND) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -77,6 +81,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif // !defined(__clang__)
 
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -93,6 +98,22 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
+#if defined(GGML_CPU_GENERIC)
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#endif
+
 #if defined(__cplusplus)
 } // extern "C"
 #endif

ggml/src/ggml-cpu/simd-mappings.h

@@ -944,8 +944,10 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
     for (int i = 0; i < offset; ++i) {              \
         x[i] = vec_add(x[i], x[offset + i]);        \
     }                                               \
-    float32x4_t tmp = x[0] + vec_reve(x[0]);        \
-    res = tmp[0] + tmp[1];                          \
+    res = vec_extract(x[0], 0) +                    \
+          vec_extract(x[0], 1) +                    \
+          vec_extract(x[0], 2) +                    \
+          vec_extract(x[0], 3);                     \
 }
 
 #define GGML_F32_VEC        GGML_F32x4

ggml/src/ggml-cuda/common.cuh

@@ -19,10 +19,10 @@
 #endif
 #include "ggml-common.h"
 
+#include <cstdio>
 #include <array>
 #include <cassert>
 #include <cfloat>
-#include <cstdio>
 #include <string>
 #include <vector>
 
@@ -207,9 +207,9 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
@@ -262,11 +262,11 @@ static bool cp_async_available(const int cc) {
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
-    return 64;
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+    return __AMDGCN_WAVEFRONT_SIZE;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
 [[noreturn]]
@@ -767,7 +767,21 @@ struct ggml_backend_cuda_context {
         name(GGML_CUDA_NAME + std::to_string(device)) {
     }
 
-    ~ggml_backend_cuda_context();
+    ~ggml_backend_cuda_context() {
+        if (copy_event != nullptr) {
+            CUDA_CHECK(cudaEventDestroy(copy_event));
+        }
+        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
+            for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
+                if (streams[i][j] != nullptr) {
+                    CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
+                }
+            }
+            if (cublas_handles[i] != nullptr) {
+                CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+            }
+        }
+    }
 
     cudaStream_t stream(int device, int stream) {
         if (streams[device][stream] == nullptr) {

ggml/src/ggml-cuda/conv2d-dw.cu

@@ -1,161 +0,0 @@
-#include "conv2d-dw.cuh"
-
-struct conv_params {
-    int in_w, in_h;
-    int out_w, out_h;
-    int kernel_w, kernel_h;
-    int stride_x, stride_y;
-    int padding_x, padding_y;
-    int dilation_x, dilation_y;
-    int channels, batches;
-};
-
-struct kernel_bounds {
-    int y_min, y_max;
-    int x_min, x_max;
-};
-
-__device__ __forceinline__ kernel_bounds calculate_kernel_bounds(int out_x, int out_y, const conv_params & params) {
-    kernel_bounds bounds;
-    bounds.y_min = max(0, (params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y);
-    bounds.y_max =
-        min(params.kernel_h,
-            (params.in_h + params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y);
-    bounds.x_min = max(0, (params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x);
-    bounds.x_max =
-        min(params.kernel_w,
-            (params.in_w + params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x);
-    return bounds;
-}
-
-__device__ __forceinline__ int calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) {
-    return out_coord * stride + kern_coord * dilation - padding;
-}
-
-struct whcn_layout {
-    __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) {
-        return n * (params.channels * params.in_w * params.in_h) + c * params.in_w * params.in_h + y * params.in_w + x;
-    }
-
-    __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) {
-        return c * params.kernel_h * params.kernel_w + ky * params.kernel_w + kx;
-    }
-
-    __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) {
-        return n * (params.channels * params.out_w * params.out_h) + c * params.out_w * params.out_h +
-               y * params.out_w + x;
-    }
-
-    __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y,
-                                          int & out_x) {
-        out_x = global_idx % params.out_w;
-        out_y = (global_idx / params.out_w) % params.out_h;
-        c     = (global_idx / (params.out_w * params.out_h)) % params.channels;
-        n     = global_idx / (params.out_w * params.out_h * params.channels);
-    }
-};
-
-struct cwhn_layout {
-    __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) {
-        return n * (params.channels * params.in_w * params.in_h) + (y * params.in_w + x) * params.channels + c;
-    }
-
-    __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) {
-        return (ky * params.kernel_w + kx) * params.channels + c;
-    }
-
-    __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) {
-        return n * (params.channels * params.out_w * params.out_h) + y * (params.out_w * params.channels) +
-               x * params.channels + c;
-    }
-
-    __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y,
-                                          int & out_x) {
-        c     = global_idx % params.channels;
-        out_x = (global_idx / params.channels) % params.out_w;
-        out_y = (global_idx / (params.channels * params.out_w)) % params.out_h;
-        n     = global_idx / (params.channels * params.out_w * params.out_h);
-    }
-};
-
-template <typename T, typename Layout>
-__global__ void conv2d_dw_kernel(const T * __restrict__ input, const T * __restrict__ kernel, T * __restrict__ output,
-                                 const int in_w, const int in_h, const int out_w, const int out_h,
-                                 const int kernel_w, const int kernel_h, const int stride_x, const int stride_y,
-                                 const int padding_x, const int padding_y, const int dilation_x, const int dilation_y,
-                                 const int channels, const int batches) {
-    const int global_idx     = blockIdx.x * blockDim.x + threadIdx.x;
-    const int total_elements = batches * channels * out_h * out_w;
-
-    if (global_idx >= total_elements) {
-        return;
-    }
-
-    conv_params params = { in_w,     in_h,      out_w,     out_h,      kernel_w,   kernel_h, stride_x,
-                           stride_y, padding_x, padding_y, dilation_x, dilation_y, channels, batches };
-
-    int batch_idx, channel_idx, out_y_idx, out_x_idx;
-    Layout::unpack_indices(global_idx, params, batch_idx, channel_idx, out_y_idx, out_x_idx);
-
-    T accumulator = 0;
-    kernel_bounds bounds = calculate_kernel_bounds(out_x_idx, out_y_idx, params);
-
-    for (int kern_y = bounds.y_min; kern_y < bounds.y_max; ++kern_y) {
-        int in_y_idx = calculate_input_coord(out_y_idx, kern_y, params.stride_y, params.dilation_y, params.padding_y);
-
-        for (int kern_x = bounds.x_min; kern_x < bounds.x_max; ++kern_x) {
-            int in_x_idx = calculate_input_coord(out_x_idx, kern_x, params.stride_x, params.dilation_x, params.padding_x);
-
-            const T input_val  = input[Layout::input_index(batch_idx, channel_idx, in_y_idx, in_x_idx, params)];
-            const T kernel_val = kernel[Layout::kernel_index(channel_idx, kern_y, kern_x, params)];
-
-            accumulator += input_val * kernel_val;
-        }
-    }
-
-    output[Layout::output_index(batch_idx, channel_idx, out_y_idx, out_x_idx, params)] = accumulator;
-}
-
-void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * kernel = dst->src[0];
-    const ggml_tensor * input  = dst->src[1];
-
-    GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    const float * w_d = (const float *) kernel->data;
-    const float * x_d = (const float *) input->data;
-    float *       y_d = (float *) dst->data;
-
-    const int32_t * p          = (const int32_t *) dst->op_params;
-    const int       stride_x   = p[0];
-    const int       stride_y   = p[1];
-    const int       padding_x  = p[2];
-    const int       padding_y  = p[3];
-    const int       dilation_x = p[4];
-    const int       dilation_y = p[5];
-
-    const int in_w     = input->ne[0];
-    const int in_h     = input->ne[1];
-    const int kernel_w = kernel->ne[0];
-    const int kernel_h = kernel->ne[1];
-    const int out_w    = dst->ne[0];
-    const int out_h    = dst->ne[1];
-    const int channels = dst->ne[2];
-    const int batches  = dst->ne[3];
-
-    cudaStream_t st = ctx.stream();
-
-    const int total  = batches * channels * out_h * out_w;
-    const int blocks = (total + CUDA_CONV2D_DW_BLOCK_SIZE - 1) / CUDA_CONV2D_DW_BLOCK_SIZE;
-
-    if (ggml_is_contiguous(input)) {
-        conv2d_dw_kernel<float, whcn_layout><<<blocks, CUDA_CONV2D_DW_BLOCK_SIZE, 0, st>>>(
-            x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y,
-            dilation_x, dilation_y, channels, batches);
-    } else if (ggml_is_contiguous_channels(input)) {
-        conv2d_dw_kernel<float, cwhn_layout><<<blocks, CUDA_CONV2D_DW_BLOCK_SIZE, 0, st>>>(
-            x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y,
-            dilation_x, dilation_y, channels, batches);
-    } else {
-        GGML_ABORT("Unsupported memory layout for conv_2d_dw");
-    }
-}

ggml/src/ggml-cuda/conv2d-dw.cuh

@@ -1,5 +0,0 @@
-#pragma once
-#include "common.cuh"
-
-#define CUDA_CONV2D_DW_BLOCK_SIZE 256
-void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/conv2d-transpose.cu

@@ -1,91 +0,0 @@
-#include <algorithm>
-
-#include "conv2d-transpose.cuh"
-#include "ggml.h"
-
-__global__ void conv2d_transpose_kernel(const float * __restrict__ input, const half * __restrict__ kernel,
-                                        float * __restrict__ output, const int in_w, const int in_h, const int out_w,
-                                        const int out_h, const int kernel_w, const int kernel_h, const int stride,
-                                        const int c_in, const int c_out, const int batches) {
-    const int global_idx = blockIdx.x * blockDim.x + threadIdx.x;
-
-    const int total_elements = out_w * out_h * c_out * batches;
-
-    if (global_idx >= total_elements) {
-        return;
-    }
-
-    const int out_x_idx = global_idx % out_w;
-    const int out_y_idx = (global_idx / out_w) % out_h;
-    const int c_idx     = (global_idx / (out_w * out_h)) % c_out;
-    const int n_idx     = global_idx / (out_w * out_h * c_out);
-
-    float accumulator = 0;
-    // For each output idx, find the inputs that contribute to it by checking stride alignment and bounds
-
-    for (int c_in_idx = 0; c_in_idx < c_in; c_in_idx++) {
-        for (int kh = 0; kh < kernel_h; ++kh) {
-            int in_y = out_y_idx - kh;
-            if (in_y < 0 || in_y % stride) continue;
-            in_y /= stride;
-            if (in_y >= in_h) continue;
-
-            for (int kw = 0; kw < kernel_w; ++kw) {
-                int in_x = out_x_idx - kw;
-                if (in_x < 0 || in_x % stride) continue;
-                in_x /= stride;
-                if (in_x >= in_w) continue;
-
-                const int input_idx = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + (in_w) *in_y + in_x;
-                const int kernel_idx =
-                    (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx + (kernel_w) *kh + kw;
-
-                float input_val = input[input_idx];
-                half  kern_val  = kernel[kernel_idx];
-
-                accumulator += input_val * (float) kern_val;
-            }
-        }
-    }
-
-    output[(out_w * out_h * c_out) * n_idx + (out_w * out_h) * c_idx + (out_w) *out_y_idx + out_x_idx] = accumulator;
-}
-
-//input is (W, H, C_in, N), Kernel is (W, H, C_out, C_in)
-void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * kernel = dst->src[0];
-    const ggml_tensor * input  = dst->src[1];
-
-    GGML_ASSERT(kernel->type == GGML_TYPE_F16 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-
-    const float * input_data  = (const float *) input->data;
-    float *       output_data = (float *) dst->data;
-    const half * kernel_data = (const half *) kernel->data;
-
-    const int input_w      = input->ne[0];
-    const int input_h      = input->ne[1];
-    const int output_w     = dst->ne[0];
-    const int output_h     = dst->ne[1];
-    const int channels_in  = input->ne[2];
-    const int channels_out = kernel->ne[2];
-    const int kernel_w     = kernel->ne[0];
-    const int kernel_h     = kernel->ne[1];
-    const int stride       = dst->op_params[0];
-    const int batches      = input->ne[3];
-
-    GGML_ASSERT(channels_in == kernel->ne[3]);
-    GGML_ASSERT(stride > 0);
-
-    cudaStream_t st = ctx.stream();
-
-    GGML_ASSERT(ggml_is_contiguous(input));
-    GGML_ASSERT(ggml_is_contiguous(kernel));
-    GGML_ASSERT(ggml_is_contiguous(dst));
-
-    const int total  = (output_w * output_h * channels_out * batches);
-    const int blocks = (total + CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE;
-
-    conv2d_transpose_kernel<<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
-        input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride,
-        channels_in, channels_out, batches);
-}

ggml/src/ggml-cuda/conv2d-transpose.cuh

@@ -1,4 +0,0 @@
-#include "common.cuh"
-
-#define CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE 256
-void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -11,8 +11,6 @@
 #include "ggml-cuda/clamp.cuh"
 #include "ggml-cuda/concat.cuh"
 #include "ggml-cuda/conv-transpose-1d.cuh"
-#include "ggml-cuda/conv2d-dw.cuh"
-#include "ggml-cuda/conv2d-transpose.cuh"
 #include "ggml-cuda/convert.cuh"
 #include "ggml-cuda/count-equal.cuh"
 #include "ggml-cuda/cpy.cuh"
@@ -49,7 +47,6 @@
 #include <atomic>
 #include <charconv>
 #include <cinttypes>
-#include <condition_variable>
 #include <cstddef>
 #include <cstdint>
 #include <float.h>
@@ -57,8 +54,9 @@
 #include <map>
 #include <memory>
 #include <mutex>
-#include <stdarg.h>
+#include <stdint.h>
 #include <stdio.h>
+#include <stdarg.h>
 #include <stdlib.h>
 #include <string>
 #include <vector>
@@ -516,33 +514,6 @@ std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(i
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
-// destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
-// this lock is used to ensure that no cuBLAS handle is destroyed while a graph is being captured
-
-static std::mutex ggml_cuda_lock;
-static std::condition_variable ggml_cuda_lock_cv;
-static std::atomic<int> ggml_cuda_lock_counter;
-
-ggml_backend_cuda_context::~ggml_backend_cuda_context() {
-    std::unique_lock<std::mutex> lock(ggml_cuda_lock);
-    ggml_cuda_lock_cv.wait(lock, []{ return ggml_cuda_lock_counter.load(std::memory_order_relaxed) == 0; });
-
-    if (copy_event != nullptr) {
-        CUDA_CHECK(cudaEventDestroy(copy_event));
-    }
-    for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
-        for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
-            if (streams[i][j] != nullptr) {
-                CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
-            }
-        }
-        if (cublas_handles[i] != nullptr) {
-            CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
-        }
-    }
-}
-
-
 // cuda buffer
 
 struct ggml_backend_cuda_buffer_context {
@@ -2339,12 +2310,6 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_IM2COL:
             ggml_cuda_op_im2col(ctx, dst);
             break;
-        case GGML_OP_CONV_2D_DW:
-            ggml_cuda_op_conv2d_dw(ctx, dst);
-            break;
-        case GGML_OP_CONV_TRANSPOSE_2D:
-            ggml_cuda_conv_2d_transpose_p0(ctx, dst);
-            break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             ggml_cuda_op_conv_transpose_1d(ctx,dst);
             break;
@@ -2699,9 +2664,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                                ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft) || (integrated && ggml_backend_buft_is_cuda_host(node->src[j]->buffer->buft)));
                     }
                 }
-#else
-                GGML_UNUSED(integrated);
-#endif // NDEBUG
+#endif
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {
@@ -2720,11 +2683,6 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
 
             CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
             graph_evaluated_or_captured = true; // CUDA graph has been captured
-
-            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
-            if (ggml_cuda_lock_counter.fetch_sub(1, std::memory_order_relaxed) == 1) {
-                ggml_cuda_lock_cv.notify_all();
-            }
         } else {
             graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
         }
@@ -2800,13 +2758,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         }
     }
 
-    if (use_cuda_graph && cuda_graph_update_required) {
-        // Start CUDA graph capture
-        {
-            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
-            ggml_cuda_lock_counter.fetch_add(1, std::memory_order_relaxed);
-        }
-
+    if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
@@ -3255,8 +3207,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
         }
         case GGML_OP_IM2COL:
-        case GGML_OP_CONV_2D_DW:
-        case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:

ggml/src/ggml-cuda/ssm-scan.cu

@@ -10,8 +10,6 @@ __global__ void __launch_bounds__(splitD, 2)
                  float * __restrict__ dst, const int64_t L) {
     GGML_UNUSED(src1_nb0);
     GGML_UNUSED(src2_nb0);
-
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     const int bidx = blockIdx.x;  // split along B
     const int bidy = blockIdx.y;  // split along D
     const int tid  = threadIdx.x;
@@ -46,16 +44,16 @@ __global__ void __launch_bounds__(splitD, 2)
     if (N == 16) {
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = A_block[(wid * warp_size + i) * stride_A + wtid];
+            float value = A_block[(wid * warpSize + i) * stride_A + wtid];
             // todo: bank conflict
             // I am always confused with how to use the swizzling method to solve
             // bank conflit. Hoping somebody can tell me.
-            smem_A[(wid * warp_size + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
 #pragma unroll
         for (size_t i = 0; i < splitD / 4; i += 2) {
-            float value = s0_block[(wid * warp_size + i) * stride_s0 + wtid];
-            smem_s0[(wid * warp_size + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid];
+            smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
     }
 

ggml/src/ggml-hip/CMakeLists.txt

@@ -113,10 +113,6 @@ if (GGML_HIP_ROCWMMA_FATTN)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
 endif()
 
-if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
-    add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
-endif()
-
 if (NOT GGML_CUDA_FA)
     add_compile_definitions(GGML_CUDA_NO_FA)
 endif()

ggml/src/ggml-metal/CMakeLists.txt

@@ -44,22 +44,21 @@ if (GGML_METAL_EMBED_LIBRARY)
     set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
 
     add_custom_command(
-        OUTPUT "${METALLIB_EMBED_ASM}"
+        OUTPUT ${METALLIB_EMBED_ASM}
         COMMAND echo "Embedding Metal library"
-        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
-        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
+        COMMAND sed -e '/__embed_ggml-common.h__/r         ${METALLIB_COMMON}' -e '/__embed_ggml-common.h__/d'         < ${METALLIB_SOURCE}           > ${METALLIB_SOURCE_EMBED_TMP}
+        COMMAND sed -e '/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}'   -e '/\#include \"ggml-metal-impl.h\"/d' < ${METALLIB_SOURCE_EMBED_TMP} > ${METALLIB_SOURCE_EMBED}
+        COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
+        COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
+        COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
+        COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
+        COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
+        COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
         DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
         COMMENT "Generate assembly for embedded Metal library"
-        VERBATIM
     )
 
-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM})
 else()
     if (GGML_METAL_SHADER_DEBUG)
         # custom command to do the following:

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

@@ -498,7 +498,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COS,
     GGML_METAL_KERNEL_TYPE_NEG,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
-    GGML_METAL_KERNEL_TYPE_MEAN,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
     GGML_METAL_KERNEL_TYPE_ARGMAX,
@@ -1455,7 +1454,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                             cos,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG,                             neg,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN,                            mean,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,                 pool_2d_avg_f32,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,                 pool_2d_max_f32,                 true);
@@ -1655,7 +1653,6 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_LOG:
             return false; // TODO: implement
         case GGML_OP_SUM_ROWS:
-        case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
@@ -2403,30 +2400,11 @@ static bool ggml_metal_encode_node(
                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
             } break;
         case GGML_OP_SUM_ROWS:
-        case GGML_OP_MEAN:
             {
                 GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
 
-                id<MTLComputePipelineState> pipeline = nil;
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
 
-                switch (dst->op) {
-                    case GGML_OP_SUM_ROWS:
-                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
-                        break;
-                    case GGML_OP_MEAN:
-                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MEAN].pipeline;
-                        break;
-                    default:
-                        GGML_ABORT("fatal error");
-                }
-
-                int nth = 32; // SIMD width
-
-                while (nth < ne00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
-                    nth *= 2;
-                }
-
-                nth = MIN(nth, ne00);
 
                 ggml_metal_kargs_sum_rows args = {
                    /*.ne00 =*/ ne00,
@@ -2456,12 +2434,11 @@ static bool ggml_metal_encode_node(
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBytes:&args length:sizeof(args) atIndex:0];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&args length:sizeof(args) atIndex:2];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
             } break;
         case GGML_OP_SOFT_MAX:
             {

ggml/src/ggml-metal/ggml-metal.metal

@@ -993,61 +993,31 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
-template <bool norm>
 kernel void kernel_sum_rows(
-        constant ggml_metal_kargs_sum_rows & args,
         device const float * src0,
         device       float * dst,
-        threadgroup  float * shmem_f32 [[threadgroup(0)]],
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort  sgitg[[simdgroup_index_in_threadgroup]],
-        ushort  tiisg[[thread_index_in_simdgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
-    int64_t i3 = tgpig.z;
-    int64_t i2 = tgpig.y;
-    int64_t i1 = tgpig.x;
+        constant ggml_metal_kargs_sum_rows & args,
+        uint3 tpig[[thread_position_in_grid]]) {
+    int64_t i3 = tpig.z;
+    int64_t i2 = tpig.y;
+    int64_t i1 = tpig.x;
 
     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
         return;
     }
 
-    if (sgitg == 0) {
-        shmem_f32[tiisg] = 0.0f;
-    }
-
     device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
     device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
 
-    float sumf = 0;
+    float row_sum = 0;
 
-    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
-        sumf += src_row[i0];
+    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
+        row_sum += src_row[i0];
     }
 
-    sumf = simd_sum(sumf);
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    if (tiisg == 0) {
-        shmem_f32[sgitg] = sumf;
-    }
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    sumf = shmem_f32[tiisg];
-    sumf = simd_sum(sumf);
-
-    if (tpitg.x == 0) {
-        dst_row[0] = norm ? sumf / args.ne00 : sumf;
-    }
+    dst_row[0] = row_sum;
 }
 
-typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
-
-template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
-template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
-
 template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,

ggml/src/ggml-sycl/CMakeLists.txt

@@ -142,7 +142,7 @@ else()
         FetchContent_Declare(
             ONEMATH
             GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git
-            GIT_TAG 8efe85f5aaebb37f1d8c503b7af66315feabf142
+            GIT_TAG c255b1b4c41e2ee3059455c1f96a965d6a62568a
         )
         FetchContent_MakeAvailable(ONEMATH)
         # Create alias to match with find_package targets name

ggml/src/ggml-sycl/binbcast.cpp

@@ -225,9 +225,9 @@ struct bin_bcast_sycl {
                     dpct::has_capability_or_fail(stream->get_device(),
                                                  {sycl::aspect::fp16});
 
-                    sycl_parallel_for(
-                        stream,
-                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * sycl::range<3>(1, 1, block_size),
+                    stream->parallel_for(
+                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
+                                              sycl::range<3>(1, 1, block_size),
                                           sycl::range<3>(1, 1, block_size)),
                         [=](sycl::nd_item<3> item_ct1) {
                             k_bin_bcast_unravel<bin_op>(
@@ -246,8 +246,9 @@ struct bin_bcast_sycl {
                 dpct::has_capability_or_fail(stream->get_device(),
                                              {sycl::aspect::fp16});
 
-                sycl_parallel_for(
-                    stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                stream->parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
                                             ne2, ne3, ne10, ne11, ne12, ne13,
                                             s1, s2, s3, s01, s02, s03, s11, s12, s13,

ggml/src/ggml-sycl/common.hpp

@@ -513,9 +513,9 @@ constexpr size_t ceil_div(const size_t m, const size_t n) {
 
 bool gpu_has_xmx(sycl::device &dev);
 
-template <int N, class T> std::string debug_get_array_str(const std::string & prefix, const T array[N]) {
+template <int N, class T> void debug_print_array(const std::string & prefix, const T array[N]) {
     if (LIKELY(!g_ggml_sycl_debug)) {
-        return "";
+        return;
     }
     std::stringstream ss;
     ss << prefix << "=[";
@@ -526,26 +526,29 @@ template <int N, class T> std::string debug_get_array_str(const std::string & pr
         ss << array[N - 1];
     }
     ss << "]";
-    return ss.str();
+    GGML_SYCL_DEBUG("%s", ss.str().c_str());
 }
 
-inline std::string debug_get_tensor_str(const std::string &prefix,
-        const ggml_tensor *tensor, const std::string &suffix = "") {
-    std::stringstream ss;
-    if (LIKELY(!g_ggml_sycl_debug)) { return ss.str(); }
-    ss << prefix.c_str() << "=";
-    if (tensor) {
-        ss << "'" << tensor->name << "':type=" << ggml_type_name(tensor->type);
-        ss << debug_get_array_str<GGML_MAX_DIMS>(";ne", tensor->ne);
-        ss << debug_get_array_str<GGML_MAX_DIMS>(";nb", tensor->nb);
-
-        if (!ggml_is_contiguous(tensor)) { ss << ";strided"; }
-        if (ggml_is_permuted(tensor)) { ss << ";permuted"; }
-    } else {
-        ss << "nullptr";
+inline void debug_print_tensor(const std::string & prefix, const ggml_tensor * tensor,
+                               const std::string & suffix = "") {
+    if (LIKELY(!g_ggml_sycl_debug)) {
+        return;
     }
-    ss << suffix;
-    return ss.str();
+    GGML_SYCL_DEBUG("%s=", prefix.c_str());
+    if (tensor) {
+        GGML_SYCL_DEBUG("'%s':type=%s", tensor->name, ggml_type_name(tensor->type));
+        debug_print_array<GGML_MAX_DIMS>(";ne", tensor->ne);
+        debug_print_array<GGML_MAX_DIMS>(";nb", tensor->nb);
+        if (!ggml_is_contiguous(tensor)) {
+            GGML_SYCL_DEBUG(";strided");
+        }
+        if (ggml_is_permuted(tensor)) {
+            GGML_SYCL_DEBUG(";permuted");
+        }
+    } else {
+        GGML_SYCL_DEBUG("nullptr");
+    }
+    GGML_SYCL_DEBUG("%s", suffix.c_str());
 }
 
 // Use scope_op_debug_print to log operations coming from running a model
@@ -561,10 +564,10 @@ struct scope_op_debug_print {
             return;
         }
         GGML_SYCL_DEBUG("[SYCL][OP] call %s%s:", func.data(), func_suffix.data());
-        GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" dst", dst).c_str());
+        debug_print_tensor(" dst", dst);
         if (dst) {
             for (std::size_t i = 0; i < num_src; ++i) {
-                GGML_SYCL_DEBUG("%s", debug_get_tensor_str("\tsrc" + std::to_string(i), dst->src[i]).c_str());
+                debug_print_tensor("\tsrc" + std::to_string(i), dst->src[i]);
             }
         }
         GGML_SYCL_DEBUG("%s\n", suffix.data());

ggml/src/ggml-sycl/concat.cpp

@@ -89,24 +89,33 @@ static void concat_f32_sycl(const float *x, const float *y, float *dst,
   sycl::range<3> gridDim(ne2, ne1, num_blocks);
   switch (dim) {
   case 0:
-      sycl_parallel_for(stream,
-                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1); });
-      break;
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1);
+        });
+    break;
   case 1:
-      sycl_parallel_for(stream,
-                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1); });
-      break;
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1);
+        });
+    break;
   // dim >=2 will be dispatched to the default path
   default:
-      sycl_parallel_for(stream,
-                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1); });
-      break;
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1);
+        });
+    break;
   }
 }
 
@@ -120,29 +129,33 @@ static void concat_f32_sycl_non_cont(
     int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2,
     uint64_t nb3, int32_t dim) {
   sycl::range<3> gridDim(ne3, ne2, ne1);
-  sycl_parallel_for(stream, sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
-      int64_t i3 = item_ct1.get_group(0);
-      int64_t i2 = item_ct1.get_group(1);
-      int64_t i1 = item_ct1.get_group(2);
+  stream->parallel_for(
+      sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)),
+      [=](sycl::nd_item<3> item_ct1) {
+        int64_t i3 = item_ct1.get_group(0);
+        int64_t i2 = item_ct1.get_group(1);
+        int64_t i1 = item_ct1.get_group(2);
 
-      int64_t o[4] = { 0, 0, 0, 0 };
-      o[dim]       = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+        int64_t o[4] = {0, 0, 0, 0};
+        o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
 
-      const float * x;
+        const float *x;
 
-      for (int i0 = item_ct1.get_local_id(2); i0 < ne0; i0 += item_ct1.get_local_range(2)) {
+        for (int i0 = item_ct1.get_local_id(2); i0 < ne0;
+             i0 += item_ct1.get_local_range(2)) {
           if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-              x = (const float *) (src0 + (i3) *nb03 + (i2) *nb02 + (i1) *nb01 + (i0) *nb00);
+            x = (const float *)(src0 + (i3)*nb03 + (i2)*nb02 + (i1)*nb01 +
+                                (i0)*nb00);
           } else {
-              x = (const float *) (src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + (i1 - o[1]) * nb11 +
-                                   (i0 - o[0]) * nb10);
+            x = (const float *)(src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 +
+                                (i1 - o[1]) * nb11 + (i0 - o[0]) * nb10);
           }
 
           float *y = (float *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0);
 
           *y = *x;
-      }
-  });
+        }
+      });
 }
 
 void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {

ggml/src/ggml-sycl/conv.cpp

@@ -59,10 +59,16 @@ static void conv_transpose_1d_f32_f32_sycl(
     const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE;
     const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE);
     const sycl::range<3> block_nums(1, 1, num_blocks);
-    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-        conv_transpose_1d_kernel(s0, output_size, src0_ne0, src0_ne1, src0_ne2, src1_ne0, dst_ne0, src0, src1, dst,
-                                 item_ct1);
-    });
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv_transpose_1d_kernel(
+                s0, output_size,
+                src0_ne0, src0_ne1, src0_ne2,
+                src1_ne0, dst_ne0,
+                src0, src1, dst, item_ct1);
+        });
 }
 
 void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {

ggml/src/ggml-sycl/convert.cpp

@@ -33,11 +33,14 @@ static void dequantize_block_sycl(const void *__restrict__ vx,
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
-        sycl_parallel_for(
-            stream,
-            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block<qk, qr, dequantize_kernel>(vx, y, k, item_ct1); });
+        stream->parallel_for(
+            sycl::nd_range<3>(
+                sycl::range<3>(1, 1, num_blocks) *
+                    sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
+                sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                dequantize_block<qk, qr, dequantize_kernel>(vx, y, k, item_ct1);
+            });
     }
 }
 
@@ -50,18 +53,24 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q2_K(vx, y, item_ct1);
+                             });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q2_K(vx, y, item_ct1);
+                             });
     }
 
 #endif
@@ -76,18 +85,24 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q3_K(vx, y, item_ct1);
+                             });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q3_K(vx, y, item_ct1);
+                             });
     }
 #endif
 }
@@ -101,9 +116,12 @@ static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_0(vx, y, nb32, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_0(vx, y, nb32, item_ct1);
+                             });
     }
 }
 
@@ -117,12 +135,13 @@ static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int
     int constexpr WARP_K = WARP_SIZE * QK4_0;
     const int n_warp = (k + WARP_K - 1) / WARP_K;
     GGML_ASSERT(k % 2 == 0);
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) * sycl::range<3>(1, 1, WARP_SIZE),
-                                        sycl::range<3>(1, 1, WARP_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                          dequantize_block_q4_0_reorder(vx, y, k, item_ct1);
-                      });
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) *
+        sycl::range<3>(1, 1, WARP_SIZE),
+        sycl::range<3>(1, 1, WARP_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
+            dequantize_block_q4_0_reorder(vx, y, k, item_ct1);
+        });
+
 }
 
 template <typename dst_t>
@@ -134,9 +153,12 @@ static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_1(vx, y, nb32, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_1(vx, y, nb32, item_ct1);
+                             });
     }
 }
 
@@ -149,13 +171,14 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler &cgh) {
             sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(12), cgh);
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                [=](sycl::nd_item<3> item_ct1) {
-                    dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1);
-                });
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1);
+                             });
         });
     }
 }
@@ -168,13 +191,13 @@ static void dequantize_row_q4_K_sycl_reorder(const void * vx, dst_t * y, const i
 
     dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-    sycl_launch(stream, [&](sycl::handler & cgh) {
+    stream->submit([&](sycl::handler & cgh) {
         sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(12), cgh);
 
-        sycl_parallel_for<1>(cgh, sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)),
-                             [=](sycl::nd_item<1> item_ct1) {
-                                 dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb);
-                             });
+        cgh.parallel_for(sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)),
+                         [=](sycl::nd_item<1> item_ct1) {
+                             dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb);
+                         });
     });
 }
 
@@ -187,18 +210,24 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q5_K(vx, y, item_ct1);
+                             });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q5_K(vx, y, item_ct1);
+                             });
     }
 
 #endif
@@ -213,18 +242,24 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q6_K(vx, y, item_ct1);
+                             });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); });
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q6_K(vx, y, item_ct1);
+                             });
     }
 
 #endif
@@ -236,9 +271,9 @@ static void dequantize_row_q6_K_sycl_reorder(const void * vx, dst_t * y, const i
 
     dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
-                      [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+        [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); });
 }
 
 template <typename dst_t>
@@ -249,10 +284,15 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_s(vx, y, item_ct1, iq1s_grid_gpu); });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq1_s(
+                                     vx, y, item_ct1, iq1s_grid_gpu
+                                     );
+                             });
         });
     }
 }
@@ -265,10 +305,15 @@ static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_m(vx, y, item_ct1, iq1s_grid_gpu); });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq1_m(
+                                     vx, y, item_ct1, iq1s_grid_gpu
+                                     );
+                             });
         });
     }
 }
@@ -281,12 +326,15 @@ static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int64_t
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                [=](sycl::nd_item<3> item_ct1) {
-                    dequantize_block_iq2_xxs(vx, y, item_ct1, iq2xxs_grid, ksigns_iq2xs, kmask_iq2xs);
-                });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_xxs(
+                                     vx, y, item_ct1, iq2xxs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
         });
     }
 }
@@ -299,12 +347,15 @@ static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int64_t k
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                [=](sycl::nd_item<3> item_ct1) {
-                    dequantize_block_iq2_xs(vx, y, item_ct1, iq2xs_grid, ksigns_iq2xs, kmask_iq2xs);
-                });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_xs(
+                                     vx, y, item_ct1, iq2xs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
         });
     }
 }
@@ -317,10 +368,13 @@ static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq2_s(vx, y, item_ct1); });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_s(vx, y, item_ct1);
+                             });
         });
     }
 }
@@ -334,12 +388,15 @@ static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int64_t
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                [=](sycl::nd_item<3> item_ct1) {
-                    dequantize_block_iq3_xxs(vx, y, item_ct1, iq3xxs_grid, ksigns_iq2xs, kmask_iq2xs);
-                });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq3_xxs(
+                                     vx, y, item_ct1, iq3xxs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
         });
     }
 }
@@ -352,10 +409,14 @@ static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq3_s(vx, y, item_ct1, kmask_iq2xs, iq3s_grid); });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq3_s(
+                                     vx, y, item_ct1, kmask_iq2xs, iq3s_grid);
+                             });
         });
     }
 }
@@ -371,11 +432,14 @@ static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int64_t k
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
-                sycl_parallel_for(
-                    cgh,
-                    sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                    [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_xs(vx, y, item_ct1); });
+            stream->submit([&](sycl::handler &cgh) {
+                  cgh.parallel_for(
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                            sycl::range<3>(1, 1, 32),
+                                        sycl::range<3>(1, 1, 32)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                            dequantize_block_iq4_xs(vx, y, item_ct1);
+                      });
             });
       }
 #endif
@@ -389,11 +453,14 @@ static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int64_t k
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
-                sycl_parallel_for(
-                    cgh,
-                    sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
-                    [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_nl(vx, y, item_ct1); });
+            stream->submit([&](sycl::handler &cgh) {
+                  cgh.parallel_for(
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                            sycl::range<3>(1, 1, 32),
+                                        sycl::range<3>(1, 1, 32)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                            dequantize_block_iq4_nl(vx, y, item_ct1);
+                      });
             });
       }
 }

ggml/src/ggml-sycl/cpy.cpp

@@ -413,8 +413,7 @@ static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        sycl_parallel_for(
-            stream,
+        stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -432,8 +431,7 @@ static void ggml_cpy_f32_f32_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        sycl_parallel_for(
-            stream,
+        stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -451,8 +449,7 @@ static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        sycl_parallel_for(
-            stream,
+        stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -468,11 +465,11 @@ static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK8_0 == 0);
     const int num_blocks = ne / QK8_0;
-    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
 }
 
 static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -480,11 +477,11 @@ static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
 }
 
 static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -493,11 +490,11 @@ static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
-    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
 }
 
 static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -505,9 +502,8 @@ static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    sycl_parallel_for(
-        stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-        [=](sycl::nd_item<3> item_ct1) {
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
             cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -520,11 +516,11 @@ static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
-    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
 }
 
 static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -532,9 +528,8 @@ static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    sycl_parallel_for(
-        stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-        [=](sycl::nd_item<3> item_ct1) {
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
             cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -547,11 +542,11 @@ static void ggml_cpy_f32_q5_0_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
-    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
 }
 
 static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -559,9 +554,8 @@ static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    sycl_parallel_for(
-        stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-        [=](sycl::nd_item<3> item_ct1) {
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
             cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -574,11 +568,11 @@ static void ggml_cpy_f32_q5_1_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK5_1 == 0);
     const int num_blocks = ne / QK5_1;
-    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
 }
 
 static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -586,9 +580,8 @@ static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    sycl_parallel_for(
-        stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-        [=](sycl::nd_item<3> item_ct1) {
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
             cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -601,11 +594,11 @@ static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne,
                                      const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK4_NL == 0);
     const int num_blocks = ne / QK4_NL;
-    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
+                                                   ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
 }
 
 static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -616,8 +609,7 @@ static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        sycl_parallel_for(
-            stream,
+        stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -636,8 +628,7 @@ static void ggml_cpy_i16_i16_sycl(const char * cx, char * cdst, const int ne, co
         // dpct::has_capability_or_fail(stream->get_device(),
         //                              {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream,
+        stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -656,8 +647,7 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co
         // dpct::has_capability_or_fail(stream->get_device(),
         //                              {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream,
+        stream->parallel_for(
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -672,13 +662,11 @@ static void ggml_cpy_q8_0_q8_0(const char * cx, char * cdst, const int ne, const
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_q_q<block_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
-                                                     ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
 }
 
 
@@ -687,13 +675,11 @@ static void ggml_cpy_q5_0_q5_0(const char * cx, char * cdst, const int ne, const
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_q_q<block_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
-                                                     ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
 }
 
 
@@ -703,13 +689,11 @@ static void ggml_cpy_q5_1_q5_1(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
 
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_q_q<block_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
-                                                     ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
 }
 
 
@@ -718,13 +702,10 @@ static void ggml_cpy_q4_0_q4_0(const char * cx, char * cdst, const int ne, const
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          cpy_q_q<block_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
-                                                     ne12, nb10, nb11, nb12, nb13, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
 }
 
 
@@ -734,18 +715,16 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
 
    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
-   sycl_parallel_for(stream,
-                     sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                                       sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
-                     [=](sycl::nd_item<3> item_ct1) {
-                         cpy_q_q<block_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
-                                                    ne12, nb10, nb11, nb12, nb13, item_ct1);
-                     });
+   stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
 }
 
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try {
     // Unlike other operators ggml_sycl_cpy takes 2 distinct tensors instead of a dst ggml_tensor and rely on its src field
-    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0, debug_get_tensor_str("\tsrc0", src0));
+    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0,
+                                         std::string(" src0 type=") + ggml_type_name(src0->type));
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 

ggml/src/ggml-sycl/dmmv.cpp

@@ -208,10 +208,12 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, nrows, item_ct1);
-                          });
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols,
+                                                          nrows, item_ct1);
+            });
     }
 }
 
@@ -875,11 +877,12 @@ static void dequantize_mul_mat_vec_q4_0_sycl_reorder(const void *vx, const dfloa
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              dequantize_mul_mat_vec_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(vx, y, dst, ncols,
-                                                                                                    nrows, item_ct1);
-                          });
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
     }
 }
 
@@ -897,10 +900,12 @@ static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>(vx, y, dst, ncols, nrows, item_ct1);
-                          });
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
     }
 }
 
@@ -916,10 +921,12 @@ static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>(vx, y, dst, ncols, nrows, item_ct1);
-                          });
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
     }
 }
 
@@ -935,10 +942,12 @@ static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>(vx, y, dst, ncols, nrows, item_ct1);
-                          });
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
     }
 }
 
@@ -954,10 +963,12 @@ static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>(vx, y, dst, ncols, nrows, item_ct1);
-                          });
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
     }
 }
 
@@ -973,10 +984,12 @@ static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>(vx, y, dst, ncols, nrows, item_ct1);
-                          });
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
     }
 }
 
@@ -989,10 +1002,11 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
-    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-                          dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1);
+        });
 }
 
 static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y,
@@ -1004,10 +1018,11 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
-    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-                          dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1);
+        });
 }
 
 static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y,
@@ -1019,10 +1034,11 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
-    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-                          dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1);
+        });
 }
 
 static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y,
@@ -1031,10 +1047,11 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y,
                                              dpct::queue_ptr stream) {
     GGML_ASSERT(ncols % QK_K == 0);
     const sycl::range<3> block_dims(1, 1, QK_WARP_SIZE);
-    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
-                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-                          dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1);
+        });
 }
 
 static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
@@ -1046,10 +1063,11 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
-    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-                          dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1);
+        });
 }
 
 void ggml_sycl_op_dequantize_mul_mat_vec(

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -13,10 +13,10 @@
 #ifndef GGML_SYCL_DPCT_HELPER_HPP
 #define GGML_SYCL_DPCT_HELPER_HPP
 
-#include <map>
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
 #include <syclcompat/math.hpp>
+#include <map>
 
 #ifdef GGML_SYCL_USE_INTEL_ONEMKL
 #include <oneapi/mkl.hpp>
@@ -118,36 +118,6 @@ inline auto get_onemath_backend(sycl::queue& queue)
 #endif
 }
 
-#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
-    namespace syclex = sycl::ext::oneapi::experimental;
-#endif
-
-template <int NR, typename Func>
-__dpct_inline__ void sycl_parallel_for(sycl::handler & cgh, sycl::nd_range<NR> nd_range, Func && func) {
-#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
-    syclex::nd_launch(cgh, nd_range, func);
-#else
-    cgh.parallel_for(nd_range, func);
-#endif
-}
-
-template <int NR, typename Func>
-__dpct_inline__ void sycl_parallel_for(sycl::queue * q, sycl::nd_range<NR> nd_range, Func && func) {
-#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
-    syclex::nd_launch(*q, nd_range, func);
-#else
-    q->parallel_for(nd_range, func);
-#endif
-}
-
-template <typename Func> __dpct_inline__ void sycl_launch(sycl::queue * stream, Func && func) {
-#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
-    syclex::submit(*stream, func);
-#else
-    stream->submit(func);
-#endif
-}
-
 namespace dpct
 {
     typedef sycl::queue *queue_ptr;

ggml/src/ggml-sycl/element_wise.cpp

@@ -329,51 +329,60 @@ static void acc_f32_sycl(const float *x, const float *y, float *dst,
                          const int ne12, const int nb1, const int nb2,
                          const int offset, queue_ptr stream) {
     int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                          acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset, item_ct1);
-                      });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset,
+                    item_ct1);
+        });
 }
 
 template<typename T>
 static void gelu_sycl(const T *x, T *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { gelu(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            gelu(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void silu_sycl(const T *x, T *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { silu(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            silu(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void sgn_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
     // hard code for now
     const int num_blocks = ceil_div(k, 256);
-    sycl_parallel_for(
-        stream, sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)),
-        [=](sycl::nd_item<3> item_ct1) { sgn(x, dst, k, item_ct1); });
+    stream->parallel_for(
+            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
+            sgn(x, dst, k, item_ct1);
+            });
 }
 
 template<typename T>
 static void abs_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
     // hard code for now
     const int num_blocks = ceil_div(k, 256);
-    sycl_parallel_for(
-        stream,
-        sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)),
-        [=](sycl::nd_item<3> item_ct1) { abs_op(x, dst, k, item_ct1); });
+    stream->parallel_for(
+            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
+            abs_op(x, dst, k, item_ct1);
+            });
 }
 
 
@@ -381,20 +390,23 @@ template<typename T>
 static void elu_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
     // hard code for now
     const int num_blocks = ceil_div(k, 256);
-    sycl_parallel_for(
-        stream,
-        sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)),
-        [=](sycl::nd_item<3> item_ct1) { elu_op(x, dst, k, item_ct1); });
+    stream->parallel_for(
+            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
+            elu_op(x, dst, k, item_ct1);
+            });
 }
 
 template<typename T>
 static void gelu_quick_sycl(const T *x, T *dst, const int k,
                                 queue_ptr stream) {
     const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { gelu_quick(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            gelu_quick(x, dst, k, item_ct1);
+        });
 }
 
 
@@ -402,133 +414,169 @@ template<typename T>
 static void gelu_erf_sycl(const T *x, T *dst, const int k,
                                 queue_ptr stream) {
     const int num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { gelu_erf(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            gelu_erf(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void tanh_sycl(const T *x, T *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { tanh(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            tanh(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void relu_sycl(const T *x, T *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { relu(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            relu(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void hardsigmoid_sycl(const T *x, T *dst, const int k,
                                  queue_ptr stream) {
     const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE;
-    sycl_parallel_for(
-        stream,
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE),
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) { hardsigmoid(x, dst, k, item_ct1); });
+        [=](sycl::nd_item<3> item_ct1) {
+            hardsigmoid(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void hardswish_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE;
-    sycl_parallel_for(
-        stream,
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE),
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) { hardswish(x, dst, k, item_ct1); });
+        [=](sycl::nd_item<3> item_ct1) {
+            hardswish(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void exp_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { exp(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            exp(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void log_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { log(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            log(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void neg_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { neg(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            neg(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void step_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { step(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            step(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void sigmoid_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIGMOID_BLOCK_SIZE - 1) / SYCL_SIGMOID_BLOCK_SIZE;
-    sycl_parallel_for(
-        stream,
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE),
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) { sigmoid(x, dst, k, item_ct1); });
+        [=](sycl::nd_item<3> item_ct1) {
+            sigmoid(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void sqrt_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SQRT_BLOCK_SIZE - 1) / SYCL_SQRT_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { sqrt(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            sqrt(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void sin_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { sin(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            sin(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
 static void cos_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { cos(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            cos(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
@@ -536,20 +584,26 @@ static void leaky_relu_sycl(const T *x, T *dst, const int k,
                                 const float negative_slope,
                                 queue_ptr stream) {
     const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { leaky_relu(x, dst, k, negative_slope, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            leaky_relu(x, dst, k, negative_slope, item_ct1);
+        });
 }
 
 template<typename T>
 static void sqr_sycl(const T *x, T *dst, const int k,
                          queue_ptr stream) {
     const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { sqr(x, dst, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            sqr(x, dst, k, item_ct1);
+        });
 }
 
 template<typename T>
@@ -560,8 +614,9 @@ static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01,
     int dst_size = ne10 * ne11 * ne12 * ne13;
     int num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
     sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE);
-    sycl_parallel_for<1>(
-        stream, sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+    stream->parallel_for(
+        sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)),
+        [=](sycl::nd_item<1> item_ct1) {
             upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
         });
 }
@@ -572,10 +627,12 @@ static void pad_sycl(const T *x, T *dst, const int ne00,
                          const int ne1, const int ne2, queue_ptr stream) {
     int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE;
     sycl::range<3> gridDim(ne2, ne1, num_blocks);
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            pad(x, dst, ne0, ne00, ne01, ne02, item_ct1);
+        });
 }
 
 template<typename T>
@@ -583,10 +640,13 @@ static void clamp_sycl(const T *x, T *dst, const float min,
                            const float max, const int k,
                            queue_ptr stream) {
     const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE;
-    sycl_parallel_for(stream,
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { clamp(x, dst, min, max, k, item_ct1); });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            clamp(x, dst, min, max, k, item_ct1);
+        });
 }
 
 inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-sycl/gemm.hpp

@@ -65,9 +65,6 @@ public:
 
         dnnl::primitive_attr primitive_attr;
         primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
-#ifdef GGML_SYCL_F16
-        primitive_attr.set_fpmath_mode(dnnl::fpmath_mode::f16);
-#endif
 
         auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
         auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));

ggml/src/ggml-sycl/getrows.cpp

@@ -60,6 +60,54 @@ static void k_get_rows(
     dst_row[iybs + iqs + y_offset] = v.y();
 }
 
+template<int qk, int qr, dequantize_kernel_t_reorder dequantize_kernel_recorder, typename dst_t>
+static void k_get_rows_reorder(
+            const void * src0, const void *src0_dq, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12,
+            const sycl::nd_item<3> &item_ct1/*, size_t s13*/) {
+
+    const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) +
+                     item_ct1.get_local_id(2)) *
+                    2;
+    const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                    item_ct1.get_local_id(1);
+    const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
+                     item_ct1.get_local_id(0)) /
+                    ne12;
+    const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
+                     item_ct1.get_local_id(0)) %
+                    ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+    auto ncols = ne00;
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+
+    const int src0_off = i01 * ncols + i00;
+    const int ib = src0_off / QK4_0; // block index
+    const int iqs = (i00%qk)/qr; // x quant index
+    const int iybs = i00 - i00%qk; // dst block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel_recorder((const void *)src0_dq, ib, (const void *)src0, src0_off/2, v);
+
+    dst_row[iybs + iqs + 0] = v.x();
+    dst_row[iybs + iqs + y_offset] = v.y();
+
+    GGML_UNUSED(nb01);
+    GGML_UNUSED(nb02);
+    GGML_UNUSED(nb03);
+}
+
 template<typename src0_t, typename dst_t>
 static void k_get_rows_float(
             const src0_t * src0, const int32_t * src1, dst_t * dst,
@@ -118,15 +166,58 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
 
     GGML_ASSERT(ne00 % 2 == 0);
 
-    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-        k_get_rows<qk, qr, dq>(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, s3, nb01, nb02, nb03, s10, s11, s12,
-                               item_ct1);
-    });
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             k_get_rows<qk, qr, dq>(
+                                 src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
+                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
+                         });
 
     GGML_UNUSED(dst);
     GGML_UNUSED(ctx);
 }
 
+template <int qk, int qr, dequantize_kernel_t_reorder dq_reorder>
+static void get_rows_sycl_reorder(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                          ggml_tensor *dst, const void *src0_dd,
+                          const int32_t *src1_dd, float *dst_dd,
+                          queue_ptr stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE);
+    const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE);
+    const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    GGML_ASSERT(ne00 % 2 == 0);
+
+    const uint8_t* src0_q = (const uint8_t*)src0_dd;
+    const size_t ncols = ne00;
+    const size_t nrows = ne01;
+    const sycl::half* src0_dq = (const sycl::half*)(src0_q + nrows * ncols / 2);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
+                             k_get_rows_reorder<qk, qr, dq_reorder>(
+                                 src0_dd, src0_dq, src1_dd, dst_dd, ne00, ne12, s1, s2,
+                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
+                         });
+
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
+}
+
+
 template <typename src0_t>
 static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                 const ggml_tensor *src1, ggml_tensor *dst,
@@ -154,8 +245,9 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        sycl_parallel_for(
-            stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
                 k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
                                  s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
             });
@@ -185,8 +277,13 @@ void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
             src1_i32, (float *)dst->data, ctx.stream());
             break;
         case GGML_TYPE_Q4_0:
-            get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
-            src1_i32, (float *)dst->data, ctx.stream());
+            if (ctx.opt_feature.reorder && dst->op == GGML_OP_MUL_MAT) {
+                get_rows_sycl_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+                src1_i32, (float *)dst->data, ctx.stream());
+            } else {
+                get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+                src1_i32, (float *)dst->data, ctx.stream());
+            }
             break;
         case GGML_TYPE_Q4_1:
             get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,

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

@@ -347,7 +347,7 @@ static enum ggml_status
 ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                      ggml_tensor *tensor) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
+    debug_print_tensor(": tensor=", tensor, "\n");
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
 
     if (tensor->view_src != NULL) {
@@ -385,7 +385,7 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                 const void *data, size_t offset,
                                                 size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
+    debug_print_tensor(": tensor=", tensor);
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
     ggml_sycl_set_device(ctx->device);
@@ -413,7 +413,7 @@ static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                                 void *data, size_t offset,
                                                 size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
+    debug_print_tensor(": tensor=", tensor);
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
 
@@ -444,8 +444,8 @@ ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                     ggml_tensor *dst) try {
     bool is_cpy_supported = ggml_backend_buffer_is_sycl(src->buffer);
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
+    debug_print_tensor(": dst=", dst);
+    debug_print_tensor(" src=", src);
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context;
@@ -525,7 +525,7 @@ catch (sycl::exception const &exc) {
 static void ggml_backend_sycl_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value,
                                                    size_t offset, size_t size) {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
+    debug_print_tensor(": tensor=", tensor);
     GGML_SYCL_DEBUG(" size=%zu offset=%zu value=%u\n", size, offset, value);
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
     SYCL_CHECK(ggml_sycl_set_device(ctx->device));
@@ -805,7 +805,7 @@ static enum ggml_status
 ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                            ggml_tensor *tensor) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
+    debug_print_tensor(": tensor=", tensor, "\n");
     GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
 
     ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
@@ -891,7 +891,7 @@ ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                           ggml_tensor *tensor, const void *data,
                                           size_t offset, size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
+    debug_print_tensor(": tensor=", tensor);
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
@@ -947,7 +947,7 @@ ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                           const ggml_tensor *tensor, void *data,
                                           size_t offset, size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
+    debug_print_tensor(": tensor=", tensor);
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
@@ -1887,12 +1887,13 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
     const size_t shared_mem = ncols_pad * sizeof(int);
 
     if (order == GGML_SORT_ORDER_ASC) {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler &cgh) {
             sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
                 sycl::range<1>(shared_mem), cgh);
 
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_ASC>(
                         x, dst, ncols, ncols_pad, item_ct1,
                         dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
@@ -1900,12 +1901,13 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                 });
         });
     } else if (order == GGML_SORT_ORDER_DESC) {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler &cgh) {
             sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
                 sycl::range<1>(shared_mem), cgh);
 
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_DESC>(
                         x, dst, ncols, ncols_pad, item_ct1,
                         dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
@@ -1923,47 +1925,50 @@ static void argmax_f32_i32_sycl(const float *x, int *dst, const int ncols,
     const sycl::range<3> block_nums(1, nrows, 1);
     const size_t shared_mem = 256 * sizeof(float);
 
-    sycl_launch(stream, [&](sycl::handler & cgh) {
+    stream->submit([&](sycl::handler &cgh) {
         sycl::local_accessor<float, 1> shared_data(
             sycl::range<1>(shared_mem/sizeof(float)), cgh);
         sycl::local_accessor<int, 1> shared_indices(
             sycl::range<1>(shared_mem/sizeof(float)), cgh);
 
-        sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            const int tid = item_ct1.get_local_id(2);
-            const int row = item_ct1.get_global_id(1);
+        cgh.parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                const int tid = item_ct1.get_local_id(2);
+                const int row = item_ct1.get_global_id(1);
 
-            float max_val = -INFINITY;
-            int   max_idx = -1;
+                float max_val = -INFINITY;
+                int max_idx = -1;
 
-            for (int col = tid; col < ncols; col += 256) {
-                float val = x[row * ncols + col];
-                if (val > max_val) {
-                    max_val = val;
-                    max_idx = col;
-                }
-            }
-
-            shared_data[tid]    = max_val;
-            shared_indices[tid] = max_idx;
-            item_ct1.barrier(sycl::access::fence_space::local_space);
-
-            for (int stride = 256 / 2; stride > 0; stride >>= 1) {
-                if (tid < stride) {
-                    float val1 = shared_data[tid];
-                    float val2 = shared_data[tid + stride];
-                    if (val2 > val1) {
-                        shared_data[tid]    = val2;
-                        shared_indices[tid] = shared_indices[tid + stride];
+                for (int col = tid; col < ncols; col += 256) {
+                    float val = x[row * ncols + col];
+                    if (val > max_val) {
+                        max_val = val;
+                        max_idx = col;
                     }
                 }
-                item_ct1.barrier(sycl::access::fence_space::local_space);
-            }
 
-            if (tid == 0) {
-                dst[row] = shared_indices[0];
-            }
-        });
+                shared_data[tid] = max_val;
+                shared_indices[tid] = max_idx;
+                item_ct1.barrier(sycl::access::fence_space::local_space);
+
+                for (int stride = 256/2; stride > 0; stride >>= 1) {
+                    if (tid < stride) {
+                        float val1 = shared_data[tid];
+                        float val2 = shared_data[tid + stride];
+                        if (val2 > val1) {
+                            shared_data[tid] = val2;
+                            shared_indices[tid] = shared_indices[tid + stride];
+                        }
+                    }
+                    item_ct1.barrier(sycl::access::fence_space::local_space);
+                }
+
+
+                if (tid == 0) {
+                    dst[row] = shared_indices[0];
+                }
+            });
     });
 }
 static void diag_mask_inf_f32_sycl(const float *x, float *dst,
@@ -2122,18 +2127,21 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const sycl::half *src1_ptr = src1->type == GGML_TYPE_F16
                 ? (const sycl::half *)src1->data + src1_padded_row_size
                                          : src1_as_f16.get();
+        ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
 
 #if GGML_SYCL_DNNL
         if (!g_ggml_sycl_disable_dnn) {
             DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ptr,
                                       DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
-                                      dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
+                                      dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
+            scope_op_debug_print scope_dbg_print(__func__, "/to_fp32_sycl", dst, /*num_src=*/2,
+                                                 " : converting dst to fp32");
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
+            to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
         }
         else
 #endif
         {
-            ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
-
             const sycl::half alpha_f16 = 1.0f;
             const sycl::half beta_f16  = 0.0f;
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
@@ -2947,7 +2955,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                 void **       ptrs_dst_get = ptrs_dst.get();
                 size_t        nb12_scaled  = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half);
                 size_t        nb13_scaled  = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half);
-                sycl_parallel_for(cgh, sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     k_compute_batched_ptrs(src0_f16, src1_f16, dst_ddf, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02,
                                            nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1);
                 });
@@ -3451,7 +3459,7 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             {
                 sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u));
                 sycl::range<3> grid_dims(1, n_ids, ids->ne[1]);
-                sycl_launch(stream, [&](sycl::handler & cgh) {
+                stream->submit([&](sycl::handler &cgh) {
                     sycl::local_accessor<int, 0> src1_row_acc(cgh);
 
                     char *__restrict src1_contiguous_get =
@@ -3463,8 +3471,9 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
                     size_t ids_nb_ct6 = ids->nb[1];
                     size_t ids_nb_ct7 = ids->nb[0];
 
-                    sycl_parallel_for(
-                        cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                    cgh.parallel_for(
+                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                        [=](sycl::nd_item<3> item_ct1) {
                             k_copy_src1_to_contiguous(
                                 src1_original, src1_contiguous_get,
                                 dev_cur_src1_row_get,
@@ -3495,14 +3504,15 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             {
                 sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u));
                 sycl::range<3> grid_dims(1, 1, num_src1_rows);
-                sycl_launch(stream, [&](sycl::handler & cgh) {
+                stream->submit([&](sycl::handler &cgh) {
                     const char *__restrict dst_contiguous_get =
                         dst_contiguous.get();
                     const mmid_row_mapping *__restrict dev_row_mapping_get =
                         dev_row_mapping.get();
 
-                    sycl_parallel_for(
-                        cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                    cgh.parallel_for(
+                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                        [=](sycl::nd_item<3> item_ct1) {
                             k_copy_dst_from_contiguous(dst_original,
                                                        dst_contiguous_get,
                                                        dev_row_mapping_get,
@@ -3856,7 +3866,7 @@ static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                const void *data, size_t offset,
                                                size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
+    debug_print_tensor(": tensor=", tensor);
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -3877,7 +3887,7 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
                                                void *data, size_t offset,
                                                size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
+    debug_print_tensor(": tensor=", tensor);
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -3900,8 +3910,8 @@ static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend,
     bool is_cpy_supported                = dst->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) &&
                             ggml_backend_buffer_is_sycl(src->buffer);
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
-    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
+    debug_print_tensor(": dst=", dst);
+    debug_print_tensor(" src=", src);
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         /*

ggml/src/ggml-sycl/gla.cpp

@@ -11,13 +11,13 @@ static void gated_linear_attn_f32_kernel(const dpct::queue_ptr stream, u_int B,
     const u_int n_seq_tokens = T / B;
     sycl::range<1> block_dims((C / H));
     sycl::range<1> grid_dims((B * H));
-    sycl_launch(stream, [&](sycl::handler & cgh) {
+    stream->submit([&](sycl::handler & cgh) {
         /* local memory accessors*/
         auto _k  = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
         auto _r  = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
         auto _td = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
 
-        sycl_parallel_for<1>(cgh, sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) {
+        cgh.parallel_for(sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) {
             u_int tid = item.get_local_id(0);
             u_int bid = item.get_group(0);
 

ggml/src/ggml-sycl/im2col.cpp

@@ -70,7 +70,7 @@ static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t I
 
     const int64_t CHW = IC * KH * KW;
 
-    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
+    stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
         im2col_kernel<T>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1,
                          p0, p1, d0, d1, item_ct1);
     });

ggml/src/ggml-sycl/mmq.cpp

@@ -1818,7 +1818,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
@@ -1829,8 +1829,9 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1852,7 +1853,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
@@ -1863,8 +1864,9 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1931,7 +1933,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
@@ -1942,8 +1944,9 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1965,7 +1968,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
@@ -1976,8 +1979,9 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2044,7 +2048,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
@@ -2055,8 +2059,9 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2078,7 +2083,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
@@ -2089,8 +2094,9 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2157,7 +2163,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
@@ -2168,8 +2174,9 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2191,7 +2198,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
@@ -2202,8 +2209,9 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2270,7 +2278,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
@@ -2281,8 +2289,9 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q8_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2304,7 +2313,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
@@ -2315,8 +2324,9 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q8_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2383,7 +2393,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
@@ -2396,8 +2406,9 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q2_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2420,7 +2431,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
@@ -2433,8 +2444,9 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q2_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2504,7 +2516,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
@@ -2519,8 +2531,9 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q3_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2544,7 +2557,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
@@ -2559,8 +2572,9 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q3_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2630,7 +2644,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
@@ -2643,8 +2657,9 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2667,7 +2682,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
@@ -2680,8 +2695,9 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2749,7 +2765,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
@@ -2762,8 +2778,9 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2786,7 +2803,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
@@ -2799,8 +2816,9 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2868,7 +2886,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
@@ -2881,8 +2899,9 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q6_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2905,7 +2924,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            sycl_launch(stream, [&](sycl::handler & cgh) {
+            stream->submit([&](sycl::handler &cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
@@ -2918,8 +2937,9 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                sycl_parallel_for(
-                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q6_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,

ggml/src/ggml-sycl/mmvq.cpp

@@ -544,12 +544,12 @@ static void reorder_mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy,
     const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, (block_num_y * WARP_SIZE));
     const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
-    sycl_launch(stream, [&](sycl::handler & cgh) {
-        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
-                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0>>(vx, vy, dst, ncols, nrows,
-                                                                                            nd_item);
-                          });
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
+                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0>>(vx, vy, dst, ncols, nrows,
+                                                                                           nd_item);
+                         });
     });
 }
 
@@ -561,12 +561,12 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float *
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
 
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+        stream->submit([&](sycl::handler & cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                 mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
+                                     vx, vy, dst, ncols, nrows, item_ct1);
+                             });
         });
     }
 }
@@ -580,12 +580,17 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK4_0, QI4_1, block_q4_1,
+                                      VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -599,12 +604,17 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK5_0, QI5_0, block_q5_0,
+                                      VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -618,12 +628,17 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK5_1, QI5_1, block_q5_1,
+                                      VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -637,12 +652,17 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK8_0, QI8_0, block_q8_0,
+                                      VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -656,12 +676,17 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI2_K, block_q2_K,
+                                      VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -675,12 +700,17 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI3_K, block_q3_K,
+                                      VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -694,12 +724,17 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI4_K, block_q4_K,
+                                      VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -715,12 +750,12 @@ static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy,
     const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
     const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
-    sycl_launch(stream, [&](sycl::handler & cgh) {
-        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
-                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols, nrows,
-                                                                                            nd_item);
-                          });
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
+                            [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols,
+                                                                                            nrows, nd_item);
+                            });
     });
 }
 
@@ -734,12 +769,17 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI5_K, block_q5_K,
+                                      VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -754,12 +794,12 @@ static void reorder_mul_mat_vec_q6_k_q8_1_sycl(const void * vx, const void * vy,
     const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
     const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
-    sycl_launch(stream, [&](sycl::handler & cgh) {
-        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
-                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K>>(vx, vy, dst, ncols, nrows,
-                                                                                            nd_item);
-                          });
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
+                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K>>(vx, vy, dst, ncols, nrows,
+                                                                                           nd_item);
+                         });
     });
 }
 static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
@@ -771,12 +811,17 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
-                                      vx, vy, dst, ncols, nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI6_K, block_q6_K,
+                                      VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -791,12 +836,14 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS / 2, block_iq2_xxs, 1>(vx, vy, dst, ncols,
-                                                                                                  nrows, item_ct1);
-                              });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS/2, block_iq2_xxs, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -810,12 +857,14 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS / 2, block_iq2_xs, 1>(vx, vy, dst, ncols,
-                                                                                               nrows, item_ct1);
-                              });
+        stream->submit([&](sycl::handler & cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS/2, block_iq2_xs, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -829,12 +878,15 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S / 2, block_iq2_s, 1>(vx, vy, dst, ncols, nrows,
-                                                                                            item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S/2, block_iq2_s, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -848,12 +900,15 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS / 2, block_iq3_xxs, 1>(vx, vy, dst, ncols,
-                                                                                                  nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS/2, block_iq3_xxs, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -867,12 +922,15 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S / 2, block_iq3_s, 1>(vx, vy, dst, ncols, nrows,
-                                                                                            item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S/2, block_iq3_s, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -886,12 +944,15 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(vx, vy, dst, ncols, nrows,
-                                                                                        item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -905,12 +966,14 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(vx, vy, dst, ncols, nrows,
-                                                                                        item_ct1);
-                              });
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -924,12 +987,15 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(vx, vy, dst, ncols, nrows,
-                                                                                             item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }
@@ -943,12 +1009,15 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS / 4, block_iq4_xs, 1>(vx, vy, dst, ncols,
-                                                                                               nrows, item_ct1);
-                              });
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS/4, block_iq4_xs, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
         });
     }
 }

ggml/src/ggml-sycl/norm.cpp

@@ -254,13 +254,14 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
-                                           nullptr, WARP_SIZE);
-                              });
-        });
+        stream->submit([&](sycl::handler& cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
+                });
+            });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -271,15 +272,16 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<sycl::float2, 1> s_sum_acc_ct1(
                             sycl::range<1>(work_group_size / WARP_SIZE), cgh);
-            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
-                                           get_pointer(s_sum_acc_ct1), work_group_size);
-                              });
-        });
+            cgh.parallel_for(
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
+                });
+            });
     }
 }
 
@@ -288,14 +290,18 @@ static void group_norm_f32_sycl(const float* x, float* dst,
     const int ne_elements, queue_ptr stream, int device) {
     if (group_size < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             const float eps_ct4 = eps;
-            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, nullptr,
-                                                 WARP_SIZE);
-                              });
-        });
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    group_norm_f32(
+                        x, dst, group_size, ne_elements, eps_ct4, item_ct1,
+                        nullptr, WARP_SIZE);
+                });
+            });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -307,18 +313,22 @@ static void group_norm_f32_sycl(const float* x, float* dst,
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
 
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
 
             const float eps_ct4 = eps;
 
-            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1,
-                                                 get_pointer(s_sum_acc_ct1), work_group_size);
-                              });
-        });
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    group_norm_f32(x, dst, group_size, ne_elements,
+                        eps_ct4, item_ct1,
+                        get_pointer(s_sum_acc_ct1), work_group_size);
+                });
+            });
     }
 }
 
@@ -330,13 +340,14 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
-                                               nullptr, WARP_SIZE);
-                              });
-        });
+        stream->submit([&](sycl::handler& cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
+                });
+            });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -347,15 +358,16 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
-            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
-                                               get_pointer(s_sum_acc_ct1), work_group_size);
-                              });
-        });
+            cgh.parallel_for(
+                sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
+                });
+            });
     }
 }
 
@@ -366,12 +378,16 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        sycl_launch(stream, [&](sycl::handler & cgh) {
-            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  l2_norm_f32(x, dst, ncols, eps, item_ct1, nullptr, WARP_SIZE);
-                              });
-        });
+        stream->submit([&](sycl::handler& cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    l2_norm_f32(x, dst, ncols, eps, item_ct1,
+                        nullptr, WARP_SIZE);
+                });
+            });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -382,15 +398,18 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
-            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
-                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                  l2_norm_f32(x, dst, ncols, eps, item_ct1, get_pointer(s_sum_acc_ct1),
-                                              work_group_size);
-                              });
-        });
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    l2_norm_f32(x, dst, ncols, eps, item_ct1,
+                        get_pointer(s_sum_acc_ct1), work_group_size);
+                });
+            });
     }
 }
 

ggml/src/ggml-sycl/rope.cpp

@@ -235,22 +235,20 @@ static void rope_norm_sycl(const T * x, T * dst, const int ne0, const int ne1, c
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) {
-                              rope_norm<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-                                                  attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
-                          });
+        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            rope_norm<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                                theta_scale, freq_factors, item_ct1);
+        });
     } else {
         /*
         DPCT1049:41: The work-group size passed to the SYCL kernel may exceed
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) {
-                              rope_norm<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-                                                 attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
-                          });
+        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            rope_norm<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                               theta_scale, freq_factors, item_ct1);
+        });
     }
 }
 
@@ -269,17 +267,15 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c
     dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
     if (freq_factors == nullptr) {
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) {
-                              rope_neox<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-                                                  attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
-                          });
+        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            rope_neox<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                                theta_scale, freq_factors, item_ct1);
+        });
     } else {
-        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                          [=](sycl::nd_item<3> item_ct1) {
-                              rope_neox<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
-                                                 attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
-                          });
+        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            rope_neox<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                               theta_scale, freq_factors, item_ct1);
+        });
     }
 }
 
@@ -302,12 +298,12 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1,
     }
     // launch kernel
     if (freq_factors == nullptr) {
-        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                   corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
     } else {
-        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
@@ -337,12 +333,12 @@ static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1,
     }
     // launch kernel
     if (freq_factors == nullptr) {
-        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_vision<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                   corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
     } else {
-        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_vision<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });

ggml/src/ggml-sycl/softmax.cpp

@@ -127,11 +127,11 @@ static void soft_max_f32_submitter(const float * x, const T * mask, float * dst,
                                    const int nrows_y, const float scale, const float max_bias, const float m0,
                                    const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
                                    const size_t n_local_scratch, queue_ptr stream) {
-    sycl_launch(stream, [&](sycl::handler & cgh) {
+    stream->submit([&](sycl::handler &cgh) {
         sycl::local_accessor<float, 1> local_buf_acc(n_local_scratch, cgh);
 
-        sycl_parallel_for(
-            cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        cgh.parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, dst, ncols_par,
                                                                              nrows_y, scale, max_bias, m0,

ggml/src/ggml-sycl/tsembd.cpp

@@ -45,9 +45,14 @@ static void timestep_embedding_f32_sycl(
     int num_blocks = (half_ceil + SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE;
     sycl::range<3> block_dims(1, 1, SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE);
     sycl::range<3> gridDim(1, ne00, num_blocks);
-    sycl_parallel_for(stream, sycl::nd_range<3>(gridDim * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-        timestep_embedding_f32(x, dst, nb1, dim, max_period, item_ct1);
-    });
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            gridDim * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            timestep_embedding_f32(
+                x, dst, nb1, dim, max_period, item_ct1
+            );
+        });
 }
 
 void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-sycl/wkv.cpp

@@ -207,11 +207,12 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     // Submit kernel
     if (C / H == WKV_BLOCK_SIZE) {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
 
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv6_f32_kernel<WKV_BLOCK_SIZE>(
                         B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
                         item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -219,11 +220,12 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
                 });
         });
     } else {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
 
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv6_f32_kernel<WKV_BLOCK_SIZE * 2>(
                         B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
                         item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -262,11 +264,12 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     // Submit kernel
     if (C / H == WKV_BLOCK_SIZE) {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
 
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv7_f32_kernel<WKV_BLOCK_SIZE>(
                         B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d,
                         item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -274,11 +277,12 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
                 });
         });
     } else {
-        sycl_launch(stream, [&](sycl::handler & cgh) {
+        stream->submit([&](sycl::handler& cgh) {
             sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
 
-            sycl_parallel_for(
-                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv7_f32_kernel<WKV_BLOCK_SIZE * 2>(
                         B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d,
                         item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -49,7 +49,15 @@ if (Vulkan_FOUND)
                              ../../include/ggml-vulkan.h
                             )
 
-    set(VULKAN_SHADER_GEN_CMAKE_ARGS "")
+    set(VULKAN_SHADER_GEN_CMAKE_ARGS
+        -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}
+        -DCMAKE_RUNTIME_OUTPUT_DIRECTORY=${CMAKE_RUNTIME_OUTPUT_DIRECTORY}
+    )
+
+    set(VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS "")
+    if (CMAKE_BUILD_TYPE AND CMAKE_BUILD_TYPE MATCHES "Debug|Release|MinSizeRel|RelWithDebInfo")
+        list(APPEND VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS --config=${CMAKE_BUILD_TYPE})
+    endif()
 
     # Test all shader extensions
     test_shader_extension_support(
@@ -128,45 +136,42 @@ if (Vulkan_FOUND)
         set(HOST_CMAKE_TOOLCHAIN_FILE "")
     endif()
 
+    # Always use ExternalProject_Add approach
     include(ExternalProject)
 
+    # Add toolchain file if cross-compiling
     if (CMAKE_CROSSCOMPILING)
         list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE})
         message(STATUS "vulkan-shaders-gen toolchain file: ${HOST_CMAKE_TOOLCHAIN_FILE}")
     endif()
 
+    # Native build through ExternalProject_Add
     ExternalProject_Add(
         vulkan-shaders-gen
         SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders
-        CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}/$<CONFIG>
-                   -DCMAKE_INSTALL_BINDIR=.
-                   -DCMAKE_BUILD_TYPE=$<CONFIG>
-                   ${VULKAN_SHADER_GEN_CMAKE_ARGS}
-
-        BUILD_COMMAND   ${CMAKE_COMMAND} --build   . --config $<CONFIG>
-
-        # NOTE: When DESTDIR is set using Makefile generators and
-        # "make install" triggers the build step, vulkan-shaders-gen
-        # would be installed into the DESTDIR prefix, so it is unset
-        # to ensure that does not happen.
-
-        INSTALL_COMMAND ${CMAKE_COMMAND} -E env --unset=DESTDIR
-                        ${CMAKE_COMMAND} --install . --config $<CONFIG>
+        CMAKE_ARGS ${VULKAN_SHADER_GEN_CMAKE_ARGS}
+        BUILD_COMMAND ${CMAKE_COMMAND} --build . ${VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS}
+        INSTALL_COMMAND ${CMAKE_COMMAND} --install .
+        INSTALL_DIR ${CMAKE_BINARY_DIR}
     )
+    ExternalProject_Add_StepTargets(vulkan-shaders-gen build install)
 
     set (_ggml_vk_host_suffix $<IF:$<STREQUAL:${CMAKE_HOST_SYSTEM_NAME},Windows>,.exe,>)
-    set (_ggml_vk_genshaders_dir "${CMAKE_BINARY_DIR}/$<CONFIG>")
-    set (_ggml_vk_genshaders_cmd "${_ggml_vk_genshaders_dir}/vulkan-shaders-gen${_ggml_vk_host_suffix}")
-    set (_ggml_vk_header     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp")
-    set (_ggml_vk_source     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp")
-    set (_ggml_vk_input_dir  "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders")
-    set (_ggml_vk_output_dir "${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv")
+    set (_ggml_vk_genshaders_cmd ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/vulkan-shaders-gen${_ggml_vk_host_suffix})
+    set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
+    set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
+    set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
+    set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
 
-    file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp")
+    file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
+    set (_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen)
+
+    # Add build and install dependencies for all builds
+    set(_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen-build vulkan-shaders-gen-install)
 
     add_custom_command(
         OUTPUT ${_ggml_vk_header}
-               ${_ggml_vk_source}
+                ${_ggml_vk_source}
 
         COMMAND ${_ggml_vk_genshaders_cmd}
             --glslc      ${Vulkan_GLSLC_EXECUTABLE}
@@ -176,9 +181,7 @@ if (Vulkan_FOUND)
             --target-cpp ${_ggml_vk_source}
             --no-clean
 
-        DEPENDS ${_ggml_vk_shader_files}
-                vulkan-shaders-gen
-
+        DEPENDS ${_ggml_vk_shader_deps}
         COMMENT "Generate vulkan shaders"
     )
 

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

@@ -168,11 +168,6 @@ struct vk_command_pool {
     vk_queue *q;
 };
 
-// Prevent simultaneous submissions to the same queue.
-// This could be per vk_queue if we stopped having two vk_queue structures
-// sharing the same vk::Queue.
-static std::mutex queue_mutex;
-
 struct vk_queue {
     uint32_t queue_family_index;
     vk::Queue queue;
@@ -1271,7 +1266,6 @@ static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_command
 static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
     if (ctx->seqs.empty()) {
         if (fence) {
-            std::lock_guard<std::mutex> guard(queue_mutex);
             ctx->p->q->queue.submit({}, fence);
         }
         return;
@@ -1341,7 +1335,6 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
         }
     }
 
-    std::lock_guard<std::mutex> guard(queue_mutex);
     ctx->p->q->queue.submit(submit_infos, fence);
 
     ctx->seqs.clear();
@@ -9495,12 +9488,6 @@ static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer
     UNUSED(buft);
 }
 
-static size_t ggml_backend_vk_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    return vk_instance.devices[0]->suballocation_block_size;
-
-    UNUSED(buft);
-}
-
 // Should be changed to return device-specific host buffer type
 // but that probably requires changes in llama.cpp
 ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
@@ -9509,7 +9496,7 @@ ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
             /* .get_name         = */ ggml_backend_vk_host_buffer_type_name,
             /* .alloc_buffer     = */ ggml_backend_vk_host_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_vk_host_buffer_type_get_alignment,
-            /* .get_max_size     = */ ggml_backend_vk_host_buffer_type_get_max_size,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },

ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt

@@ -25,3 +25,15 @@ add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
+
+# Configure output directories for MSVC builds
+if(MSVC)
+    # Get the main project's runtime output directory if possible
+    if(DEFINED CMAKE_RUNTIME_OUTPUT_DIRECTORY)
+        foreach(CONFIG ${CMAKE_CONFIGURATION_TYPES})
+            string(TOUPPER ${CONFIG} CONFIG)
+            set_target_properties(${TARGET} PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY_${CONFIG} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
+        endforeach()
+    endif()
+endif()

ggml/src/ggml.c

@@ -888,6 +888,12 @@ struct ggml_context {
     struct ggml_object * objects_end;
 };
 
+struct ggml_context_container {
+    bool used;
+
+    struct ggml_context context;
+};
+
 //
 // data types
 //
@@ -955,7 +961,6 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "UPSCALE",
     "PAD",
     "PAD_REFLECT_1D",
-    "ROLL",
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
@@ -986,7 +991,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "OPT_STEP_ADAMW",
 };
 
-static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1051,7 +1056,6 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "upscale(x)",
     "pad(x)",
     "pad_reflect_1d(x)",
-    "roll(x)",
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
@@ -1082,7 +1086,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "adamw(x)",
 };
 
-static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -4343,34 +4347,6 @@ struct ggml_tensor * ggml_pad_reflect_1d(
     return result;
 }
 
-// ggml_roll
-
-struct ggml_tensor * ggml_roll(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        int                   shift0,
-        int                   shift1,
-        int                   shift2,
-        int                   shift3) {
-    GGML_ASSERT(a->nb[0] == ggml_type_size(a->type));
-    GGML_ASSERT(abs(shift0) < a->ne[0]);
-    GGML_ASSERT(abs(shift1) < a->ne[1]);
-    GGML_ASSERT(abs(shift2) < a->ne[2]);
-    GGML_ASSERT(abs(shift3) < a->ne[3]);
-
-    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
-
-    ggml_set_op_params_i32(result, 0, shift0);
-    ggml_set_op_params_i32(result, 1, shift1);
-    ggml_set_op_params_i32(result, 2, shift2);
-    ggml_set_op_params_i32(result, 3, shift3);
-
-    result->op     = GGML_OP_ROLL;
-    result->src[0] = a;
-
-    return result;
-}
-
 // ggml_arange
 
 struct ggml_tensor * ggml_arange(

gguf-py/gguf/constants.py

@@ -198,7 +198,6 @@ class Keys:
         MASK_ID              = "tokenizer.ggml.mask_token_id"
         ADD_BOS              = "tokenizer.ggml.add_bos_token"
         ADD_EOS              = "tokenizer.ggml.add_eos_token"
-        ADD_SEP              = "tokenizer.ggml.add_sep_token"
         ADD_PREFIX           = "tokenizer.ggml.add_space_prefix"
         REMOVE_EXTRA_WS      = "tokenizer.ggml.remove_extra_whitespaces"
         PRECOMPILED_CHARSMAP = "tokenizer.ggml.precompiled_charsmap"
@@ -292,7 +291,6 @@ class MODEL_ARCH(IntEnum):
     BERT             = auto()
     NOMIC_BERT       = auto()
     NOMIC_BERT_MOE   = auto()
-    NEO_BERT         = auto()
     JINA_BERT_V2     = auto()
     BLOOM            = auto()
     STABLELM         = auto()
@@ -345,8 +343,6 @@ class MODEL_ARCH(IntEnum):
     WAVTOKENIZER_DEC = auto()
     PLM              = auto()
     BAILINGMOE       = auto()
-    DOTS1            = auto()
-    ARCEE            = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -575,7 +571,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.BERT:             "bert",
     MODEL_ARCH.NOMIC_BERT:       "nomic-bert",
     MODEL_ARCH.NOMIC_BERT_MOE:   "nomic-bert-moe",
-    MODEL_ARCH.NEO_BERT:         "neo-bert",
     MODEL_ARCH.JINA_BERT_V2:     "jina-bert-v2",
     MODEL_ARCH.BLOOM:            "bloom",
     MODEL_ARCH.STABLELM:         "stablelm",
@@ -628,8 +623,6 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
     MODEL_ARCH.PLM:              "plm",
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
-    MODEL_ARCH.DOTS1:            "dots1",
-    MODEL_ARCH.ARCEE:            "arcee",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1084,18 +1077,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
-    MODEL_ARCH.NEO_BERT: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_QKV,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_NORM,
-        MODEL_TENSOR.FFN_DOWN,
-        MODEL_TENSOR.FFN_UP,
-        MODEL_TENSOR.ENC_OUTPUT_NORM,
-        MODEL_TENSOR.CLS,
-        MODEL_TENSOR.CLS_OUT,
-    ],
     MODEL_ARCH.JINA_BERT_V2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
@@ -2063,45 +2044,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
-    MODEL_ARCH.DOTS1: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.OUTPUT,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_Q,
-        MODEL_TENSOR.ATTN_Q_NORM,
-        MODEL_TENSOR.ATTN_K,
-        MODEL_TENSOR.ATTN_K_NORM,
-        MODEL_TENSOR.ATTN_V,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_EXP_PROBS_B,
-        MODEL_TENSOR.FFN_NORM,
-        MODEL_TENSOR.FFN_GATE,
-        MODEL_TENSOR.FFN_GATE_EXP,
-        MODEL_TENSOR.FFN_GATE_INP,
-        MODEL_TENSOR.FFN_GATE_SHEXP,
-        MODEL_TENSOR.FFN_DOWN,
-        MODEL_TENSOR.FFN_DOWN_EXP,
-        MODEL_TENSOR.FFN_DOWN_SHEXP,
-        MODEL_TENSOR.FFN_UP,
-        MODEL_TENSOR.FFN_UP_EXP,
-        MODEL_TENSOR.FFN_UP_SHEXP,
-    ],
-    MODEL_ARCH.ARCEE: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.OUTPUT,
-        MODEL_TENSOR.ROPE_FREQS,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_Q,
-        MODEL_TENSOR.ATTN_K,
-        MODEL_TENSOR.ATTN_V,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.ATTN_ROT_EMBD,
-        MODEL_TENSOR.FFN_NORM,
-        MODEL_TENSOR.FFN_DOWN,
-        MODEL_TENSOR.FFN_UP,
-    ],
     # TODO
 }
 

gguf-py/gguf/gguf_writer.py

@@ -271,7 +271,7 @@ class GGUFWriter:
 
     def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None:
         if any(key in kv_data for kv_data in self.kv_data):
-            logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}')
+            raise ValueError(f'Duplicated key name {key!r}')
 
         self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)
 
@@ -891,9 +891,6 @@ class GGUFWriter:
     def add_add_eos_token(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_EOS, value)
 
-    def add_add_sep_token(self, value: bool) -> None:
-        self.add_bool(Keys.Tokenizer.ADD_SEP, value)
-
     def add_add_space_prefix(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_PREFIX, value)
 

gguf-py/gguf/tensor_mapping.py

@@ -31,7 +31,6 @@ class TensorNameMap:
             "model.embeddings",                          # rwkv7
             "model.word_embeddings",                     # bailingmoe
             "language_model.model.embed_tokens",         # llama4
-            "encoder",                                   # neobert
         ),
 
         # Token type embeddings
@@ -135,7 +134,6 @@ class TensorNameMap:
             "rwkv.blocks.{bid}.ln1",                                # rwkv6
             "model.layers.{bid}.ln1",                               # rwkv7
             "model.layers.{bid}.input_layernorm",                   # llama4
-            "transformer_encoder.{bid}.attention_norm",             # neobert
         ),
 
         # Attention norm 2
@@ -163,7 +161,6 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.qkv_proj",                               # phi3
             "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
             "transformer.layers.{bid}.attn.qkv_proj",                              # openelm
-            "transformer_encoder.{bid}.qkv",                                       # neobert
         ),
 
         # Attention query
@@ -239,7 +236,6 @@ class TensorNameMap:
             "transformer.layers.{bid}.attn.out_proj",                       # openelm
             "transformer.h.{bid}.attn.attention.out_proj",                  # exaone
             "model.layers.{bid}.self_attn.o_proj",                          # llama4
-            "transformer_encoder.{bid}.wo",                                 # neobert
         ),
 
         # Attention output norm
@@ -280,7 +276,6 @@ class TensorNameMap:
             "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
             "transformer.layers.{bid}.ffn_norm",                             # openelm
             "model.layers.{bid}.post_attention_layernorm",                   # llama4
-            "transformer_encoder.{bid}.ffn_norm",                            # neobert
         ),
 
         # Post feed-forward norm
@@ -310,7 +305,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
-            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
+            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
         ),
 
         # Feed-forward up
@@ -345,7 +340,6 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
             "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
             "model.layers.{bid}.feed_forward.up_proj",                # llama4
-            "transformer_encoder.{bid}.ffn.w12",                      # neobert
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -428,7 +422,6 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
             "model.layers.h.{bid}.mlp.c_proj",                        # exaone
             "model.layers.{bid}.feed_forward.down_proj",              # llama4
-            "transformer_encoder.{bid}.ffn.w3",                       # neobert
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -839,14 +832,12 @@ class TensorNameMap:
         # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
             "encoder.final_layer_norm", # t5
-            "layer_norm",               # neobert
         ),
 
         MODEL_TENSOR.CLS: (
             "classifier",       # jina
             "classifier.dense", # roberta
             "pre_classifier",   # distillbert
-            "dense",            # neobert
         ),
 
         MODEL_TENSOR.CLS_OUT: (

gguf-py/gguf/vocab.py

@@ -7,10 +7,7 @@ import os
 from pathlib import Path
 from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable
 
-try:
-    from sentencepiece import SentencePieceProcessor
-except ImportError:
-    SentencePieceProcessor = None
+from sentencepiece import SentencePieceProcessor
 
 import gguf
 
@@ -119,7 +116,6 @@ class SpecialVocab:
         logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping')
 
     def _try_load_from_tokenizer_json(self, path: Path) -> bool:
-        tokenizer = None
         tokenizer_file = path / 'tokenizer.json'
         if tokenizer_file.is_file():
             with open(tokenizer_file, encoding = 'utf-8') as f:
@@ -153,87 +149,11 @@ class SpecialVocab:
             added_tokens = tokenizer.get('added_tokens', {})
         else:
             added_tokens = {}
-        tokenizer_config = None
         tokenizer_config_file = path / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, encoding = 'utf-8') as f:
-                tokenizer_config = json.load(f)
-        if tokenizer:
-            special_bos = (tokenizer_config or {}).get('bos_token')
-            special_cls = (tokenizer_config or {}).get('cls_token')
-            special_eos = (tokenizer_config or {}).get('eos_token')
-            special_sep = (tokenizer_config or {}).get('sep_token')
-            if not special_bos and special_cls and tokenizer_config:
-                tokenizer_config['bos_token'] = special_bos = special_cls
-            if not special_eos and special_sep and tokenizer_config:
-                tokenizer_config['eos_token'] = special_eos = special_sep
-            post_processor = tokenizer.get('post_processor', {})
-            for processor in post_processor.get('processors', [post_processor]):
-                if processor.get('type') == 'RobertaProcessing':
-                    self.add_special_token['bos'] = True
-                    self.add_special_token['eos'] = True
-                    self.add_special_token['sep'] = True
-                    if not special_cls and tokenizer_config:
-                        special_cls = processor.get('cls', [special_bos])[0]
-                        tokenizer_config['cls_token'] = special_cls
-                    if not special_sep and tokenizer_config:
-                        special_sep = processor.get('sep', [special_eos])[0]
-                        tokenizer_config['sep_token'] = special_sep
-                    continue
-                # Crude parsing of TemplateProcessing to determine if BOS/SEP/EOS should be added
-                # Only works with simple templates, **will** get it wrong on unusual sequences
-                if processor.get('type') == 'TemplateProcessing':
-                    tmpl_single = processor.get('single', [])
-                    tmpl_pair = processor.get('pair', [])
-                    special_first = None
-                    special_last = None
-                    if len(tmpl_single) > 1:
-                        if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'):
-                            if not tokenizer_config:
-                                special_bos = special_first
-                            self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False
-                            if special_first not in (special_bos, special_cls):
-                                logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing<single>')
-                        if special_last := tmpl_single[-1].get('SpecialToken', {}).get('id'):
-                            if not tokenizer_config:
-                                special_eos = special_last
-                            self.add_special_token['eos'] = True if special_last == special_eos else False
-                            if special_last != special_eos:
-                                logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing<single>')
-                    if tmpl_pair:
-                        seq_start = 1 if tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0
-                        seq_stop = -1 if tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None
-                        if seq_start == 0 or seq_stop is None:
-                            logger.warning('TemplateProcessing<single> leading/trailing special tokens do not match TemplateProcessing<pair>')
-                        if tmpl_pair := tmpl_pair[slice(seq_start, seq_stop)]:
-                            tmpl_a = tmpl_pair[0].get('Sequence', {}).get('id')
-                            tmpl_b = tmpl_pair[-1].get('Sequence', {}).get('id')
-                            if tmpl_a != 'A' or tmpl_b != 'B':
-                                logger.warning(f'Unknown sequence {tmpl_a}...{tmpl_b} in TemplateProcessing<pair>')
-                            # A [sep] [eos] B
-                            if tmpl_a == 'A' and tmpl_b == 'B' and (tmpl_pair := tmpl_pair[1:-1]):
-                                add_sep = False
-                                if special_entry := tmpl_pair[0].get('SpecialToken', {}).get('id'):
-                                    if special_entry in (special_sep, special_eos) and not special_last:
-                                        add_sep = True
-                                    if special_entry not in (special_sep, special_eos):
-                                        logger.warning(f'Unknown separator token {special_entry!r} in TemplateProcessing<pair>')
-                                else:
-                                    logger.warning(f'Unknown middle sequence {tmpl_pair[0]!r} in TemplateProcessing<pair>')
-                                if len(tmpl_pair) == 2:
-                                    if special_entry := tmpl_pair[1].get('SpecialToken', {}).get('id'):
-                                        if special_entry in (special_sep, special_eos):
-                                            add_sep = True
-                                        if special_entry not in (special_sep, special_eos):
-                                            logger.warning(f'Unknown second separator token {special_entry!r} in TemplateProcessing<pair>')
-                                    else:
-                                        logger.warning(f'Unknown second middle sequence {tmpl_pair[1]!r} in TemplateProcessing<pair>')
-                                self.add_special_token['sep'] = add_sep
-                                if add_sep and not special_sep and tokenizer_config:
-                                    tokenizer_config['sep_token'] = special_eos
-                    continue
-        if not tokenizer_config:
+        if not tokenizer_config_file.is_file():
             return True
+        with open(tokenizer_config_file, encoding = 'utf-8') as f:
+            tokenizer_config = json.load(f)
         chat_template_alt = None
         chat_template_file = path / 'chat_template.json'
         if chat_template_file.is_file():
@@ -382,9 +302,6 @@ class SentencePieceVocab(Vocab):
     name = "spm"
 
     def __init__(self, base_path: Path):
-        if SentencePieceProcessor is None:
-            raise RuntimeError("sentencepiece is not installed")
-
         added_tokens: dict[str, int] = {}
         if (fname_tokenizer := base_path / 'tokenizer.model').exists():
             # normal location

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.17.1"
+version = "0.17.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [
@@ -22,7 +22,7 @@ python = ">=3.8"
 numpy = ">=1.17"
 tqdm = ">=4.27"
 pyyaml = ">=5.1"
-sentencepiece = { version = ">=0.1.98,<=0.2.0", optional = true }
+sentencepiece = ">=0.1.98,<=0.2.0"
 PySide6 = { version = "^6.9", python = ">=3.9,<3.14", optional = true }
 
 [tool.poetry.dev-dependencies]

include/llama.h

@@ -243,21 +243,18 @@ extern "C" {
 
     typedef bool (*llama_progress_callback)(float progress, void * user_data);
 
-    // Input data for llama_encode/llama_decode
+    // Input data for llama_decode
     // A llama_batch object can contain input about one or many sequences
     // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
     //
     // - token  : the token ids of the input (used when embd is NULL)
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
-    //            (if set to NULL, the token position will be tracked automatically by llama_encode/llama_decode)
+    //            (if set to NULL, the token position will be tracked automatically by llama_decode)
     // - seq_id : the sequence to which the respective token belongs
     //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
-    //            (if set to NULL:
-    //               - if embeddings: all tokens are output
-    //               - if not:        only the last token is output
-    //            )
+    //            (if set to NULL, only the logits for last token will be returned)
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -265,8 +262,8 @@ extern "C" {
         llama_token  *  token;
         float        *  embd;
         llama_pos    *  pos;
-        int32_t      *  n_seq_id;
-        llama_seq_id ** seq_id;
+        int32_t      *  n_seq_id; // TODO: remove, should belong to only 1 sequence
+        llama_seq_id ** seq_id;   // TODO: become llama_seq_id * seq_id;
         int8_t       *  logits;   // TODO: rename this to "output"
     } llama_batch;
 
@@ -964,8 +961,8 @@ extern "C" {
     // Get the number of threads used for prompt and batch processing (multiple token).
     LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx);
 
-    // Set whether the context outputs embeddings or not
-    // TODO: rename to avoid confusion with llama_get_embeddings()
+    // Set whether the model is in embeddings mode or not
+    // If true, embeddings will be returned but logits will not
     LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
 
     // Set whether to use causal attention or not
@@ -1044,7 +1041,6 @@ extern "C" {
 
     LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab);
     LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab);
-    LLAMA_API bool llama_vocab_get_add_sep(const struct llama_vocab * vocab);
 
     LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab);
     LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab);
@@ -1088,7 +1084,6 @@ extern "C" {
     /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
     /// @return Returns the number of tokens on success, no more than n_tokens_max
     /// @return Returns a negative number on failure - the number of tokens that would have been returned
-    /// @return Returns INT32_MIN on overflow (e.g., tokenization result size exceeds int32_t limit)
     /// @param add_special Allow to add BOS and EOS tokens if model is configured to do so.
     /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated
     ///                      as plaintext. Does not insert a leading space.

requirements/requirements-compare-llama-bench.txt

@@ -1,3 +1,2 @@
 tabulate~=0.9.0
 GitPython~=3.1.43
-matplotlib~=3.10.0

scripts/compare-llama-bench.py

@@ -19,7 +19,6 @@ except ImportError as e:
     print("the following Python libraries are required: GitPython, tabulate.") # noqa: NP100
     raise e
 
-
 logger = logging.getLogger("compare-llama-bench")
 
 # All llama-bench SQL fields
@@ -123,15 +122,11 @@ help_s = (
 parser.add_argument("--check", action="store_true", help="check if all required Python libraries are installed")
 parser.add_argument("-s", "--show", help=help_s)
 parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
-parser.add_argument("--plot", help="generate a performance comparison plot and save to specified file (e.g., plot.png)")
-parser.add_argument("--plot_x", help="parameter to use as x axis for plotting (default: n_depth)", default="n_depth")
-parser.add_argument("--plot_log_scale", action="store_true", help="use log scale for x axis in plots (off by default)")
 
 known_args, unknown_args = parser.parse_known_args()
 
 logging.basicConfig(level=logging.DEBUG if known_args.verbose else logging.INFO)
 
-
 if known_args.check:
     # Check if all required Python libraries are installed. Would have failed earlier if not.
     sys.exit(0)
@@ -504,6 +499,7 @@ else:
 
 name_compare = bench_data.get_commit_name(hexsha8_compare)
 
+
 # If the user provided columns to group the results by, use them:
 if known_args.show is not None:
     show = known_args.show.split(",")
@@ -548,14 +544,6 @@ else:
             show.remove(prop)
         except ValueError:
             pass
-
-    # Add plot_x parameter to parameters to show if it's not already present:
-    if known_args.plot:
-        for k, v in PRETTY_NAMES.items():
-            if v == known_args.plot_x and k not in show:
-                show.append(k)
-                break
-
     rows_show = bench_data.get_rows(show, hexsha8_baseline, hexsha8_compare)
 
 if not rows_show:
@@ -612,161 +600,6 @@ if "gpu_info" in show:
 headers  = [PRETTY_NAMES[p] for p in show]
 headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
 
-if known_args.plot:
-    def create_performance_plot(table_data: list[list[str]], headers: list[str], baseline_name: str, compare_name: str, output_file: str, plot_x_param: str, log_scale: bool = False):
-        try:
-            import matplotlib.pyplot as plt
-            import matplotlib
-            matplotlib.use('Agg')
-        except ImportError as e:
-            logger.error("matplotlib is required for --plot.")
-            raise e
-
-        data_headers = headers[:-4] # Exclude the last 4 columns (Test, baseline t/s, compare t/s, Speedup)
-        plot_x_index = None
-        plot_x_label = plot_x_param
-
-        if plot_x_param not in ["n_prompt", "n_gen", "n_depth"]:
-            pretty_name = PRETTY_NAMES.get(plot_x_param, plot_x_param)
-            if pretty_name in data_headers:
-                plot_x_index = data_headers.index(pretty_name)
-                plot_x_label = pretty_name
-            elif plot_x_param in data_headers:
-                plot_x_index = data_headers.index(plot_x_param)
-                plot_x_label = plot_x_param
-            else:
-                logger.error(f"Parameter '{plot_x_param}' not found in current table columns. Available columns: {', '.join(data_headers)}")
-                return
-
-        grouped_data = {}
-
-        for i, row in enumerate(table_data):
-            group_key_parts = []
-            test_name = row[-4]
-
-            base_test = ""
-            x_value = None
-
-            if plot_x_param in ["n_prompt", "n_gen", "n_depth"]:
-                for j, val in enumerate(row[:-4]):
-                    header_name = data_headers[j]
-                    if val is not None and str(val).strip():
-                        group_key_parts.append(f"{header_name}={val}")
-
-                if plot_x_param == "n_prompt" and "pp" in test_name:
-                    base_test = test_name.split("@")[0]
-                    x_value = base_test
-                elif plot_x_param == "n_gen" and "tg" in test_name:
-                    x_value = test_name.split("@")[0]
-                elif plot_x_param == "n_depth" and "@d" in test_name:
-                    base_test = test_name.split("@d")[0]
-                    x_value = int(test_name.split("@d")[1])
-                else:
-                    base_test = test_name
-
-                if base_test.strip():
-                    group_key_parts.append(f"Test={base_test}")
-            else:
-                for j, val in enumerate(row[:-4]):
-                    if j != plot_x_index:
-                        header_name = data_headers[j]
-                        if val is not None and str(val).strip():
-                            group_key_parts.append(f"{header_name}={val}")
-                    else:
-                        x_value = val
-
-                group_key_parts.append(f"Test={test_name}")
-
-            group_key = tuple(group_key_parts)
-
-            if group_key not in grouped_data:
-                grouped_data[group_key] = []
-
-            grouped_data[group_key].append({
-                'x_value': x_value,
-                'baseline': float(row[-3]),
-                'compare': float(row[-2]),
-                'speedup': float(row[-1])
-            })
-
-        if not grouped_data:
-            logger.error("No data available for plotting")
-            return
-
-        def make_axes(num_groups, max_cols=2, base_size=(8, 4)):
-            from math import ceil
-            cols = 1 if num_groups == 1 else min(max_cols, num_groups)
-            rows = ceil(num_groups / cols)
-
-            # Scale figure size by grid dimensions
-            w, h = base_size
-            fig, ax_arr = plt.subplots(rows, cols,
-                                       figsize=(w * cols, h * rows),
-                                       squeeze=False)
-
-            axes = ax_arr.flatten()[:num_groups]
-            return fig, axes
-
-        num_groups = len(grouped_data)
-        fig, axes = make_axes(num_groups)
-
-        plot_idx = 0
-
-        for group_key, points in grouped_data.items():
-            if plot_idx >= len(axes):
-                break
-            ax = axes[plot_idx]
-
-            try:
-                points_sorted = sorted(points, key=lambda p: float(p['x_value']) if p['x_value'] is not None else 0)
-                x_values = [float(p['x_value']) if p['x_value'] is not None else 0 for p in points_sorted]
-            except ValueError:
-                points_sorted = sorted(points, key=lambda p: group_key)
-                x_values = [p['x_value'] for p in points_sorted]
-
-            baseline_vals = [p['baseline'] for p in points_sorted]
-            compare_vals = [p['compare'] for p in points_sorted]
-
-            ax.plot(x_values, baseline_vals, 'o-', color='skyblue',
-                    label=f'{baseline_name}', linewidth=2, markersize=6)
-            ax.plot(x_values, compare_vals, 's--', color='lightcoral', alpha=0.8,
-                    label=f'{compare_name}', linewidth=2, markersize=6)
-
-            if log_scale:
-                ax.set_xscale('log', base=2)
-                unique_x = sorted(set(x_values))
-                ax.set_xticks(unique_x)
-                ax.set_xticklabels([str(int(x)) for x in unique_x])
-
-            title_parts = []
-            for part in group_key:
-                if '=' in part:
-                    key, value = part.split('=', 1)
-                    title_parts.append(f"{key}: {value}")
-
-            title = ', '.join(title_parts) if title_parts else "Performance comparison"
-
-            ax.set_xlabel(plot_x_label, fontsize=12, fontweight='bold')
-            ax.set_ylabel('Tokens per second (t/s)', fontsize=12, fontweight='bold')
-            ax.set_title(title, fontsize=12, fontweight='bold')
-            ax.legend(loc='best', fontsize=10)
-            ax.grid(True, alpha=0.3)
-
-            plot_idx += 1
-
-        for i in range(plot_idx, len(axes)):
-            axes[i].set_visible(False)
-
-        fig.suptitle(f'Performance comparison: {compare_name} vs. {baseline_name}',
-                     fontsize=14, fontweight='bold')
-        fig.subplots_adjust(top=1)
-
-        plt.tight_layout()
-        plt.savefig(output_file, dpi=300, bbox_inches='tight')
-        plt.close()
-
-    create_performance_plot(table, headers, name_baseline, name_compare, known_args.plot, known_args.plot_x, known_args.plot_log_scale)
-
 print(tabulate( # noqa: NP100
     table,
     headers=headers,

scripts/sync-ggml.last

@@ -1 +1 @@
-9e4bee1c5afc2d677a5b32ecb90cbdb483e81fff
+6a7d170c04789f6ebcf320ed03c1b16973f93bd7

src/CMakeLists.txt

@@ -22,9 +22,8 @@ add_library(llama
             llama-io.cpp
             llama-kv-cache-unified.cpp
             llama-kv-cache-unified-iswa.cpp
+            llama-kv-cache-recurrent.cpp
             llama-memory.cpp
-            llama-memory-hybrid.cpp
-            llama-memory-recurrent.cpp
             llama-mmap.cpp
             llama-model-loader.cpp
             llama-model-saver.cpp

src/llama-arch.cpp

@@ -20,7 +20,6 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
     { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
-    { LLM_ARCH_NEO_BERT,         "neo-bert"         },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
@@ -73,8 +72,6 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-    { LLM_ARCH_DOTS1,            "dots1"            },
-    { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -147,7 +144,6 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
-    { LLM_KV_ATTENTION_LAYER_INDICES,                "%s.attention.layer_indices"                },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -198,7 +194,6 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_MASK_ID,              "tokenizer.ggml.mask_token_id"            },
     { LLM_KV_TOKENIZER_ADD_BOS,              "tokenizer.ggml.add_bos_token"            },
     { LLM_KV_TOKENIZER_ADD_EOS,              "tokenizer.ggml.add_eos_token"            },
-    { LLM_KV_TOKENIZER_ADD_SEP,              "tokenizer.ggml.add_sep_token"            },
     { LLM_KV_TOKENIZER_ADD_PREFIX,           "tokenizer.ggml.add_space_prefix"         },
     { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,      "tokenizer.ggml.remove_extra_whitespaces" },
     { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap"     },
@@ -248,24 +243,6 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
-    {
-        LLM_ARCH_ARCEE,
-        {
-            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
-            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
-            { LLM_TENSOR_OUTPUT,          "output" },
-            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
-            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
-            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
-            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
-            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
-            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
-            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-        },
-    },
     {
         LLM_ARCH_LLAMA4,
         {
@@ -517,21 +494,6 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
-    {
-        LLM_ARCH_NEO_BERT,
-        {
-            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
-            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
-            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
-            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-            { LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" },
-            { LLM_TENSOR_CLS,             "cls" },
-            { LLM_TENSOR_CLS_OUT,         "cls.output" },
-        },
-    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {
@@ -1593,34 +1555,6 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
-    {
-        LLM_ARCH_DOTS1,
-        {
-            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
-            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
-            { LLM_TENSOR_OUTPUT,             "output" },
-            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
-            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
-            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
-            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
-            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
-            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
-            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
-            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
-            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
-            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
-            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
-            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
-            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
-            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
-            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
-            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
-            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
-            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
-            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
-            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
-        }
-    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1818,25 +1752,3 @@ llm_arch llm_arch_from_string(const std::string & name) {
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
     return LLM_TENSOR_INFOS.at(tensor);
 }
-
-bool llm_arch_is_recurrent(const llm_arch & arch) {
-    switch (arch) {
-        case LLM_ARCH_MAMBA:
-        case LLM_ARCH_RWKV6:
-        case LLM_ARCH_RWKV6QWEN2:
-        case LLM_ARCH_RWKV7:
-        case LLM_ARCH_ARWKV7:
-            return true;
-        default:
-            return false;
-    }
-}
-
-bool llm_arch_is_hybrid(const llm_arch & arch) {
-    // TODO: There are currently no hybrid models! Once there are, this will be
-    //  the place to identify them
-    switch (arch) {
-        default:
-            return false;
-    }
-}

src/llama-arch.h

@@ -24,7 +24,6 @@ enum llm_arch {
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
     LLM_ARCH_NOMIC_BERT_MOE,
-    LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -77,8 +76,6 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
-    LLM_ARCH_DOTS1,
-    LLM_ARCH_ARCEE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -151,7 +148,6 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
-    LLM_KV_ATTENTION_LAYER_INDICES,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -194,7 +190,6 @@ enum llm_kv {
     LLM_KV_TOKENIZER_MASK_ID,
     LLM_KV_TOKENIZER_ADD_BOS,
     LLM_KV_TOKENIZER_ADD_EOS,
-    LLM_KV_TOKENIZER_ADD_SEP,
     LLM_KV_TOKENIZER_ADD_PREFIX,
     LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
     LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
@@ -441,6 +436,3 @@ const char * llm_arch_name(llm_arch arch);
 llm_arch llm_arch_from_string(const std::string & name);
 
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
-
-bool llm_arch_is_recurrent(const llm_arch & arch);
-bool llm_arch_is_hybrid   (const llm_arch & arch);

src/llama-batch.h

@@ -2,146 +2,91 @@
 
 #include "llama.h"
 
-#include "llama-cparams.h"
-
 #include <array>
 #include <vector>
-#include <set>
-#include <bitset>
-#include <unordered_map>
 
-// keep this struct lightweight
-// it points to data in `llama_batch_allocr`
+// very similar to llama_batch,
+// but has more metadata about sequences
 struct llama_ubatch {
     bool equal_seqs;
     // TODO: whole_seqs for embeddings?
 
     uint32_t n_tokens;     // total tokens (n_seq_tokens * n_seqs)
-    uint32_t n_seq_tokens; // tokens per sequence set
-    uint32_t n_seqs;       // sequence sets in the ubatch
-    uint32_t n_seqs_unq;   // unique sequence ids in the ubatch
+    uint32_t n_seq_tokens; // tokens per sequence
+    uint32_t n_seqs;
 
-    // seq_id_unq: unique sequence ids in the ubatch
-    // seq_idx:    indices of the unique sequence ids in the ubatch in [0, n_seqs_unq)
-    //             used for extracting sequence pooled embeddings
-
-    //                          // size               | idx | val
-    llama_token  *  token;      // [n_tokens]         | i   | id, token
-    float        *  embd;       // [n_embd, n_tokens] | i   | embd
-    llama_pos    *  pos;        // [n_tokens]         | i   | pos
-    int32_t      *  n_seq_id;   // [n_tokens]         | i   | -
-    llama_seq_id ** seq_id;     // [n_tokens]         | s   | s0, s1, seq_id
-    llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id
-    int32_t      *  seq_idx;    // [LLAMA_MAX_SEQ]    | -   | seq_idx
-    int8_t       *  output;     // [n_tokens]         | i   | -
+    llama_token  *  token;    // [n_tokens]
+    float        *  embd;     // [n_embd, n_tokens]
+    llama_pos    *  pos;      // [n_tokens]
+    int32_t      *  n_seq_id; // [n_seqs] // TODO: remove, should belong to only 1 sequence
+    llama_seq_id ** seq_id;   // [n_seqs] // TODO: become llama_seq_id * seq_id;
+    int8_t       *  output;   // [n_tokens]
 };
 
-// a helper for sanitizing, fulfilling and splitting a batch
-class llama_batch_allocr {
-public:
-    llama_batch_allocr(uint32_t n_pos_per_embd);
+struct llama_sbatch_seq {
+    int32_t n_seq_id;
 
-    // sanitize and auto-gen missing data in the input batch
-    // memory is optional. if provided will be used to check for sequence continuity and to determine the positions
-    bool init(
-            const llama_batch & batch_inp,
-            const llama_vocab & vocab,
-            const llama_memory_i * memory,
-            uint32_t n_embd,
-            bool output_all);
+    llama_seq_id * seq_id;
 
-    const llama_batch & get_batch() const;
+    size_t offset;
+    size_t length;
+};
 
-    uint32_t get_n_tokens()  const;
-    uint32_t get_n_outputs() const;
+// sequence-length-aware batch splitting
+struct llama_sbatch {
+    // tokens left in this batch
+    size_t n_tokens;
 
-    // the array of output indices in the order they were encountered during the ubatch splitting
-    std::vector<int32_t> & get_out_ids();
-
-    // min/max positions of each sequence in the current ubatch
-    llama_pos seq_pos_min(llama_seq_id seq_id) const;
-    llama_pos seq_pos_max(llama_seq_id seq_id) const;
-
-    // call once before splitting the batch to reset the internal state
-    void split_reset();
-
-    // simple split, unknown number of sequence sets of unequal lengths
-    llama_ubatch split_simple(uint32_t n_ubatch);
-
-    // make ubatches of equal-length sequences sets
-    llama_ubatch split_equal(uint32_t n_ubatch);
-
-    // sequence-set-wise split - each ubatch contains a single sequence-set
-    llama_ubatch split_seq(uint32_t n_ubatch);
-
-    // a helper method for creating a well-defined ubatch of tokens
-    // TODO: support embeddings if needed in the future
-    llama_ubatch ubatch_reserve(uint32_t n_seq_tokens, uint32_t n_seqs);
-
-private:
-    void clear();
-
-    // create the next ubatch based on the provided batch indices (idxs) and the number of sequence sets (n_seqs)
-    // return llama_ubatch.n_tokens == 0 if the entire batch was consumed
-    llama_ubatch ubatch_add(const std::vector<int32_t> & idxs, uint32_t n_seqs, bool equal_seqs);
-
-    // for debugging, start with LLAMA_BATCH_DEBUG=2
-    void ubatch_print(const llama_ubatch & ubatch, int debug);
-
-    llama_batch batch;
-
-    // only for debugging purposes
-    const llama_vocab * vocab;
-
-    // TODO: this is more of a temporary solution until we have a better way to handle multiple positions per token/embd
-    //       ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762
-    const uint32_t n_pos_per_embd;
-
-    uint32_t n_embd;
-    uint32_t n_outputs;
-
-    std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
-
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id *> seq_id;
-    std::vector<llama_seq_id>   seq_id_unq;
-    std::vector<int32_t>        seq_idx;
-    std::vector<int8_t>         output;
-
-    using pos_set_t = std::set<llama_pos>;
-    using seq_cpl_t = std::vector<bool>;
-
-    std::vector<pos_set_t> seq_pos; // seq_pos[s]: the set of positions in sequence s
-    std::vector<seq_cpl_t> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
-
-    using idx_vec_t = std::vector<int32_t>;
-    using seq_set_t = std::bitset<LLAMA_MAX_SEQ>;
-
-    std::vector<seq_set_t> seq_set; // seq_set[i]: the sequence set of token i
-
-    std::unordered_map<seq_set_t, idx_vec_t> seq_set_map; // the indices at which the sequence set appears
+    size_t n_embd;
 
+    // sorted indices into the batch
+    std::vector<int64_t> ids;
     // batch indices of the output
-    std::vector<int32_t> out_ids;
+    std::vector<int64_t> out_ids;
+    std::vector<llama_sbatch_seq> seq;
 
-    // used[i] indicates if token i has already been used in a previous ubatch
-    std::vector<bool> used;
+    const llama_batch * batch = nullptr;
 
-    // llama_ubatch points to this data:
-    struct ubatch {
+    // buffers for the ubatches
+    // TODO: very hacky, this needs a complete rework
+    struct ubatch_data {
         std::vector<llama_token>    token;
         std::vector<float>          embd;
         std::vector<llama_pos>      pos;
         std::vector<int32_t>        n_seq_id;
         std::vector<llama_seq_id *> seq_id;
-        std::vector<llama_seq_id>   seq_id_unq;
-        std::vector<int32_t>        seq_idx;
         std::vector<int8_t>         output;
     };
 
-    // current splitting state:
-    std::vector<ubatch> ubatches;
+    std::vector<ubatch_data> udatas;
 
-    int debug;
+    llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false);
+
+    void add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length);
+
+    // simple split, unknown number of sequences of unequal lengths
+    llama_ubatch split_simple(size_t n_ubatch);
+
+    // make batches of equal-length sequences
+    llama_ubatch split_equal(size_t n_ubatch);
+
+    // sequence-wise split
+    llama_ubatch split_seq(size_t n_ubatch);
+
+    llama_sbatch() = default;
+    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false);
+};
+
+// temporary allocate memory for the input batch if needed
+struct llama_batch_allocr {
+    struct llama_batch batch;
+
+    std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id *> seq_id;
+    std::vector<int8_t>         output;
+
+    // optionally fulfill the batch returned by llama_batch_get_one
+    llama_batch_allocr(struct llama_batch in_batch, llama_pos p0);
 };

src/llama-chat.cpp

@@ -183,8 +183,6 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_BAILING;
     } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
         return LLM_CHAT_TEMPLATE_LLAMA4;
-    } else if (tmpl_contains("<|endofuserprompt|>")) {
-        return LLM_CHAT_TEMPLATE_DOTS1;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -333,7 +331,7 @@ int32_t llm_chat_apply_template(
             std::string role(message->role);
             if (role == "system") {
                 // there is no system message for gemma, but we will merge it with user prompt, so nothing is broken
-                system_prompt += trim(message->content);
+                system_prompt = trim(message->content);
                 continue;
             }
             // in gemma, "assistant" is "model"
@@ -355,7 +353,7 @@ int32_t llm_chat_apply_template(
             std::string role(message->role);
             if (role == "system") {
                 // there is no system message support, we will merge it with user prompt
-                system_prompt += message->content;
+                system_prompt = message->content;
                 continue;
             } else if (role == "user") {
                 ss << "Human: ";
@@ -645,21 +643,6 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "Assistant:";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_DOTS1) {
-        // dots.llm1.inst (DOTS1)
-        for (auto message : chat) {
-            std::string role(message->role);
-            if (role == "system") {
-                ss << "<|system|>" << message->content << "<|endofsystem|>";
-            } else if (role == "user") {
-                ss << "<|userprompt|>" << message->content << "<|endofuserprompt|>";
-            } else {
-                ss << "<|response|>" << message->content << "<|endofresponse|>";
-            }
-        }
-        if (add_ass) {
-            ss << "<|response|>";
-        }
     } else {
         // template not supported
         return -1;

src/llama-chat.h

@@ -43,7 +43,6 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_SMOLVLM,
-    LLM_CHAT_TEMPLATE_DOTS1,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

src/llama-context.cpp

@@ -1,7 +1,6 @@
 #include "llama-context.h"
 
 #include "llama-impl.h"
-#include "llama-batch.h"
 #include "llama-io.h"
 #include "llama-memory.h"
 #include "llama-mmap.h"
@@ -19,8 +18,7 @@
 llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
-    model(model),
-    balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
+    model(model) {
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -29,8 +27,8 @@ llama_context::llama_context(
     const auto & hparams = model.hparams;
 
     cparams.n_seq_max = std::max(1u, params.n_seq_max);
-    if (cparams.n_seq_max > LLAMA_MAX_SEQ) {
-        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_SEQ));
+    if (cparams.n_seq_max > LLAMA_MAX_PARALLEL_SEQUENCES) {
+        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_PARALLEL_SEQUENCES));
     }
 
     cparams.n_threads        = params.n_threads;
@@ -496,7 +494,7 @@ float * llama_context::get_logits() {
 }
 
 float * llama_context::get_logits_ith(int32_t i) {
-    int64_t j = -1;
+    int32_t j = -1;
 
     try {
         if (logits == nullptr) {
@@ -519,7 +517,7 @@ float * llama_context::get_logits_ith(int32_t i) {
         }
         if (j >= n_outputs) {
             // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
+            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
         }
 
         return logits + j*model.vocab.n_tokens();
@@ -538,7 +536,7 @@ float * llama_context::get_embeddings() {
 }
 
 float * llama_context::get_embeddings_ith(int32_t i) {
-    int64_t j = -1;
+    int32_t j = -1;
 
     try {
         if (embd == nullptr) {
@@ -561,7 +559,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
         }
         if (j >= n_outputs) {
             // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
+            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
         }
 
         return embd + j*model.hparams.n_embd;
@@ -721,30 +719,40 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch,
     return res;
 }
 
-int llama_context::encode(const llama_batch & batch_inp) {
-    GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT
-
-    if (batch_inp.n_tokens == 0) {
+int llama_context::encode(llama_batch & inp_batch) {
+    if (inp_batch.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
+    // temporary allocate memory for the input batch if needed
+    // note: during encode, we always pass the full sequence starting from pos = 0
+    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : 0);
+
+    const llama_batch & batch = batch_allocr.batch;
+    const int32_t n_tokens = batch.n_tokens;
+
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd = hparams.n_embd;
+    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    // note: during encode, we always pass the full sequence starting from pos = 0
-    if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, true)) {
-        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
-        return -1;
+    // TODO: move the validation to the llama_batch_allocr
+    if (batch.token) {
+        for (int32_t i = 0; i < n_tokens; ++i) {
+            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
+                return -1;
+            }
+
+            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
+                LLAMA_LOG_ERROR("%s: invalid seq_id[%d] = %d > %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
+                throw -1;
+            }
+        }
     }
 
-    const uint32_t n_tokens = balloc->get_n_tokens();
-
-    const llama_ubatch ubatch = balloc->split_simple(n_tokens);
-
     // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
-    GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");
+    GGML_ASSERT(cparams.n_ubatch >= (uint32_t) n_tokens && "encoder requires n_ubatch >= n_tokens");
 
     if (t_compute_start_us == 0) {
         t_compute_start_us = ggml_time_us();
@@ -755,13 +763,19 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     n_queued_tokens += n_tokens;
 
+    const int64_t n_embd = hparams.n_embd;
+
+    llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true);
+
+    const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
+
     // reserve output buffer
     if (output_reserve(n_tokens) < n_tokens) {
         LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens);
         return -2;
     };
 
-    for (uint32_t i = 0; i < n_tokens; ++i) {
+    for (int32_t i = 0; i < n_tokens; ++i) {
         output_ids[i] = i;
     }
 
@@ -813,28 +827,32 @@ int llama_context::encode(const llama_batch & batch_inp) {
                 {
                     // extract sequence embeddings
                     auto & embd_seq_out = embd_seq;
+                    embd_seq_out.clear();
 
-                    for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
-                        const llama_seq_id seq_id  = ubatch.seq_id_unq[s];
-                        const int32_t      seq_idx = ubatch.seq_idx[seq_id];
+                    GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits
 
+                    for (int32_t i = 0; i < n_tokens; i++) {
+                        const llama_seq_id seq_id = ubatch.seq_id[i][0];
+                        if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                            continue;
+                        }
                         embd_seq_out[seq_id].resize(n_embd);
-                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_idx)*sizeof(float), n_embd*sizeof(float));
+                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
                     }
                 } break;
             case LLAMA_POOLING_TYPE_RANK:
                 {
                     // extract the rerank score - n_cls_out floats per sequence
                     auto & embd_seq_out = embd_seq;
-
                     const uint32_t n_cls_out = hparams.n_cls_out;
 
-                    for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
-                        const llama_seq_id seq_id  = ubatch.seq_id_unq[s];
-                        const int32_t      seq_idx = ubatch.seq_idx[seq_id];
-
+                    for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
+                        const llama_seq_id seq_id = ubatch.seq_id[s][0];
+                        if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                            continue;
+                        }
                         embd_seq_out[seq_id].resize(n_cls_out);
-                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_idx)*sizeof(float), n_cls_out*sizeof(float));
+                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_id)*sizeof(float), n_cls_out*sizeof(float));
                     }
                 } break;
             case LLAMA_POOLING_TYPE_UNSPECIFIED:
@@ -859,16 +877,14 @@ int llama_context::encode(const llama_batch & batch_inp) {
         cross.v_embd.resize(cross.n_embd*cross.n_enc);
         memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd));
 
-        const auto & batch = balloc->get_batch();
-
         // remember the sequence ids used during the encoding - needed for cross attention later
+        // TODO: the seuqence indexing here is likely not correct in the general case
+        //       probably works only for split_simple
         cross.seq_ids_enc.resize(n_tokens);
-        for (uint32_t i = 0; i < n_tokens; i++) {
+        for (int32_t i = 0; i < n_tokens; i++) {
             cross.seq_ids_enc[i].clear();
-
-            for (int s = 0; s < batch.n_seq_id[i]; s++) {
-                const llama_seq_id seq_id = batch.seq_id[i][s];
-
+            for (int s = 0; s < ubatch.n_seq_id[i]; s++) {
+                llama_seq_id seq_id = ubatch.seq_id[i][s];
                 cross.seq_ids_enc[i].insert(seq_id);
             }
         }
@@ -877,40 +893,68 @@ int llama_context::encode(const llama_batch & batch_inp) {
     return 0;
 }
 
-int llama_context::decode(const llama_batch & batch_inp) {
-    GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT
-
+int llama_context::decode(llama_batch & inp_batch) {
     if (!memory) {
         LLAMA_LOG_DEBUG("%s: cannot decode batches with this context (calling encode() instead)\n", __func__);
-        return encode(batch_inp);
+        return encode(inp_batch);
     }
 
-    if (batch_inp.n_tokens == 0) {
+    if (inp_batch.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
+    if (!inp_batch.pos) {
+        if (inp_batch.seq_id) {
+            LLAMA_LOG_ERROR("%s: pos == NULL, but seq_id != NULL\n", __func__);
+            return -1;
+        }
+    }
+
+    // temporary allocate memory for the input batch if needed
+    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : memory->seq_pos_max(0) + 1);
+
+    const llama_batch & batch = batch_allocr.batch;
+
     const auto & vocab   = model.vocab;
     const auto & hparams = model.hparams;
 
     const int32_t n_vocab = vocab.n_tokens();
-    const int64_t n_embd  = hparams.n_embd;
 
-    // when computing embeddings, all tokens are output
-    const bool output_all = cparams.embeddings;
+    const int64_t n_tokens_all = batch.n_tokens;
+    const int64_t n_embd       = hparams.n_embd;
 
-    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, output_all)) {
-        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
-        return -1;
+    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
+
+    // TODO: move the validation to the llama_batch_allocr
+    if (batch.token) {
+        for (int64_t i = 0; i < n_tokens_all; ++i) {
+            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
+                LLAMA_LOG_ERROR("%s: invalid token[%" PRId64 "] = %d\n", __func__, i, batch.token[i]);
+                return -1;
+            }
+
+            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
+                LLAMA_LOG_ERROR("%s: invalid seq_id[%" PRId64 "] = %d >= %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
+                return -1;
+            }
+        }
     }
 
-    const uint32_t n_tokens_all  = balloc->get_n_tokens();
-    const uint32_t n_outputs_all = balloc->get_n_outputs();
+    // this indicates we are doing pooled embedding
+    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
 
-    if (output_all) {
+    int64_t n_outputs_all = 0;
+
+    // count outputs
+    for (uint32_t i = 0; i < n_tokens_all; ++i) {
+        n_outputs_all += batch.logits[i] != 0;
+    }
+
+    if (embd_pooled) {
         // require that all tokens are output
         if (n_outputs_all != n_tokens_all) {
-            LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %d, n_tokens_all = %d)\n",
+            LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %" PRId64 ", n_tokens_all = %" PRId64 ")\n",
                     __func__, n_outputs_all, n_tokens_all);
             return -1;
         }
@@ -936,7 +980,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     llama_memory_state_ptr mstate;
 
     while (true) {
-        mstate = memory->init_batch(*balloc, cparams.n_ubatch, output_all);
+        mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled);
         if (!mstate) {
             return -2;
         }
@@ -957,19 +1001,19 @@ int llama_context::decode(const llama_batch & batch_inp) {
                         did_optimize = true;
 
                         if (kv_self_update(true)) {
-                            LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, balloc->get_n_tokens());
+                            LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, batch.n_tokens);
 
                             continue;
                         }
                     }
 
-                    LLAMA_LOG_WARN("%s: failed to find a memory slot for batch of size %d\n", __func__, balloc->get_n_tokens());
+                    LLAMA_LOG_WARN("%s: failed to find a memory slot for batch of size %d\n", __func__, batch.n_tokens);
 
                     return 1;
                 }
             case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
                 {
-                    LLAMA_LOG_ERROR("%s: compute failed while preparing batch of size %d\n", __func__, balloc->get_n_tokens());
+                    LLAMA_LOG_ERROR("%s: compute failed while preparing batch of size %d\n", __func__, batch.n_tokens);
 
                     return -2;
                 }
@@ -980,7 +1024,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
     // reserve output buffer
     if (output_reserve(n_outputs_all) < n_outputs_all) {
-        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
+        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
         return -2;
     };
 
@@ -996,6 +1040,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
             if (n_outputs_all == n_tokens_all) {
                 n_outputs_new = ubatch.n_tokens;
             } else {
+                GGML_ASSERT(ubatch.output);
                 for (uint32_t i = 0; i < ubatch.n_tokens; i++) {
                     n_outputs_new += (int32_t) (ubatch.output[i] != 0);
                 }
@@ -1013,19 +1058,18 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
         if (!res) {
             // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
-            llama_pos pos_min[LLAMA_MAX_SEQ];
-            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            llama_pos pos_min[LLAMA_MAX_PARALLEL_SEQUENCES];
+            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
                 pos_min[s] = std::numeric_limits<llama_pos>::max();
             }
 
-            // TODO: fix sequence indexing
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 const auto & seq_id = ubatch.seq_id[i][0];
 
                 pos_min[seq_id] = std::min(pos_min[seq_id], ubatch.pos[i]);
             }
 
-            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
                 if (pos_min[s] == std::numeric_limits<llama_pos>::max()) {
                     continue;
                 }
@@ -1048,7 +1092,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
         //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
         //}
 
-        auto * t_logits = res->get_logits();
+        auto * t_logits = cparams.embeddings ? nullptr         : res->get_logits();
         auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
 
         if (t_embd && res->get_embd_pooled()) {
@@ -1095,27 +1139,27 @@ int llama_context::decode(const llama_batch & batch_inp) {
                         // extract sequence embeddings (cleared before processing each batch)
                         auto & embd_seq_out = embd_seq;
 
-                        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
-                            const llama_seq_id seq_id  = ubatch.seq_id_unq[s];
-                            const int32_t      seq_idx = ubatch.seq_idx[seq_id];
-
+                        for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
+                            const llama_seq_id seq_id = ubatch.seq_id[s][0];
+                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                                continue;
+                            }
                             embd_seq_out[seq_id].resize(n_embd);
-                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_idx)*sizeof(float), n_embd*sizeof(float));
+                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
                         }
                     } break;
                 case LLAMA_POOLING_TYPE_RANK:
                     {
-                        // extract the rerank score - n_cls_out floats per sequence
+                        // extract the rerank score - a single float per sequence
                         auto & embd_seq_out = embd_seq;
 
-                        const uint32_t n_cls_out = hparams.n_cls_out;
-
-                        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
-                            const llama_seq_id seq_id  = ubatch.seq_id_unq[s];
-                            const int32_t      seq_idx = ubatch.seq_idx[seq_id];
-
-                            embd_seq_out[seq_id].resize(n_cls_out);
-                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_idx)*sizeof(float), n_cls_out*sizeof(float));
+                        for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
+                            const llama_seq_id seq_id = ubatch.seq_id[s][0];
+                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                                continue;
+                            }
+                            embd_seq_out[seq_id].resize(1);
+                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float));
                         }
                     } break;
                 case LLAMA_POOLING_TYPE_UNSPECIFIED:
@@ -1132,14 +1176,14 @@ int llama_context::decode(const llama_batch & batch_inp) {
     n_outputs = n_outputs_all;
 
     // set output mappings
-    if (n_outputs > 0) {
+    {
         bool sorted_output = true;
 
-        auto & out_ids = balloc->get_out_ids();
+        auto & out_ids = mstate->out_ids();
 
-        GGML_ASSERT(out_ids.size() == (size_t) n_outputs);
+        GGML_ASSERT(out_ids.size() == (size_t) n_outputs_all);
 
-        for (int64_t i = 0; i < n_outputs; ++i) {
+        for (int64_t i = 0; i < n_outputs_all; ++i) {
             int64_t out_id = out_ids[i];
             output_ids[out_id] = i;
             if (out_id != i) {
@@ -1151,22 +1195,20 @@ int llama_context::decode(const llama_batch & batch_inp) {
         // note: this is mostly relevant for recurrent models atm
         if (!sorted_output) {
             const uint32_t n_vocab = model.vocab.n_tokens();
-            const uint64_t n_embd  = model.hparams.n_embd;
+            const uint32_t n_embd  = model.hparams.n_embd;
 
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
             // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort)
-            for (uint32_t i = 0; i < n_outputs - 1; ++i) {
-                uint32_t j_min = i;
-                for (uint32_t j = i + 1; j < n_outputs; ++j) {
+            for (int32_t i = 0; i < n_outputs - 1; ++i) {
+                int32_t j_min = i;
+                for (int32_t j = i + 1; j < n_outputs; ++j) {
                     if (out_ids[j] < out_ids[j_min]) {
                         j_min = j;
                     }
                 }
-                if (j_min == i) {
-                    continue;
-                }
+                if (j_min == i) { continue; }
                 std::swap(out_ids[i], out_ids[j_min]);
                 if (logits_size > 0) {
                     for (uint32_t k = 0; k < n_vocab; k++) {
@@ -1179,10 +1221,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     }
                 }
             }
-
             std::fill(output_ids.begin(), output_ids.end(), -1);
-
-            for (uint32_t i = 0; i < n_outputs; ++i) {
+            for (int32_t i = 0; i < n_outputs; ++i) {
                 output_ids[out_ids[i]] = i;
             }
         }
@@ -1202,7 +1242,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
 // output
 //
 
-uint32_t llama_context::output_reserve(int32_t n_outputs) {
+int32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
@@ -1212,8 +1252,9 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto n_vocab = vocab.n_tokens();
     const auto n_embd  = hparams.n_embd;
 
-    bool has_logits = true;
-    bool has_embd   = cparams.embeddings;
+    // TODO: use a per-batch flag for logits presence instead
+    bool has_logits = !cparams.embeddings;
+    bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
 
     // TODO: hacky enc-dec support
     if (model.arch == LLM_ARCH_T5) {
@@ -1267,7 +1308,8 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
     // set all ids as invalid (negative)
     std::fill(output_ids.begin(), output_ids.end(), -1);
 
-    this->n_outputs = 0;
+    this->n_outputs     = 0;
+    this->n_outputs_max = n_outputs_max;
 
     return n_outputs_max;
 }
@@ -1308,8 +1350,8 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
 
     this->n_outputs = n_outputs;
 
-    llama_batch_allocr balloc(model.hparams.n_pos_per_embd());
-    llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs);
+    llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
+    llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
 
     auto * gf = graph_init();
     auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate);
@@ -1758,12 +1800,14 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
 
         std::vector<int32_t> w_output_pos;
 
+        GGML_ASSERT(n_outputs <= n_outputs_max);
+
         w_output_pos.resize(n_outputs);
 
         // build a more compact representation of the output ids
         for (size_t i = 0; i < n_batch(); ++i) {
             // map an output id to a position in the batch
-            int64_t pos = output_ids[i];
+            int32_t pos = output_ids[i];
             if (pos >= 0) {
                 GGML_ASSERT(pos < n_outputs);
                 w_output_pos[pos] = i;
@@ -2029,20 +2073,18 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, true)) {
-            LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
-            return;
-        }
-
-        const uint32_t n_tokens_all = balloc->get_n_tokens();
+        const auto n_tokens_all = batch.n_tokens;
 
         n_queued_tokens += n_tokens_all;
 
+        // this indicates we are doing pooled embedding
+        const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
+
         embd_seq.clear();
 
-        uint32_t n_outputs_all = n_tokens_all;
+        int64_t n_outputs_all = n_tokens_all;
 
-        auto mstate = memory->init_batch(*balloc, cparams.n_ubatch, true);
+        auto mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled);
         if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
             LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__);
             break;
@@ -2050,7 +2092,7 @@ void llama_context::opt_epoch_iter(
 
         // reserve output buffer
         if (output_reserve(n_outputs_all) < n_outputs_all) {
-            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
+            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
             GGML_ABORT("TODO: handle this error");
         };
 

src/llama-context.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "llama.h"
+#include "llama-batch.h"
 #include "llama-cparams.h"
 #include "llama-graph.h"
 #include "llama-adapter.h"
@@ -12,7 +13,6 @@
 #include <vector>
 
 struct llama_model;
-class llama_batch_allocr;
 
 class llama_io_read_i;
 class llama_io_write_i;
@@ -102,8 +102,8 @@ struct llama_context {
             llama_memory_state_i * mstate,
                      ggml_status & ret);
 
-    int encode(const llama_batch & batch_inp);
-    int decode(const llama_batch & batch_inp);
+    int encode(llama_batch & inp_batch);
+    int decode(llama_batch & inp_batch);
 
     //
     // state save/load
@@ -181,7 +181,7 @@ private:
 
     // Make sure enough space is available for outputs.
     // Returns max number of outputs for which space was reserved.
-    uint32_t output_reserve(int32_t n_outputs);
+    int32_t output_reserve(int32_t n_outputs);
 
     //
     // graph
@@ -246,10 +246,8 @@ private:
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
 
-    // reuse the batch_allocr to avoid unnecessary memory allocations
-    std::unique_ptr<llama_batch_allocr> balloc;
-
-    uint32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch
+    int32_t n_outputs     = 0; // number of actually-used outputs in the current ubatch or last logical batch
+    int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers
 
     std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 

src/llama-cparams.cpp

@@ -1,5 +1,5 @@
 #include "llama-cparams.h"
 
 size_t llama_max_parallel_sequences(void) {
-    return LLAMA_MAX_SEQ;
+    return LLAMA_MAX_PARALLEL_SEQUENCES;
 }

src/llama-cparams.h

@@ -4,7 +4,7 @@
 
 #include <cstdint>
 
-#define LLAMA_MAX_SEQ 64
+#define LLAMA_MAX_PARALLEL_SEQUENCES 64
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference

src/llama-graph.cpp

@@ -6,8 +6,7 @@
 
 #include "llama-kv-cache-unified.h"
 #include "llama-kv-cache-unified-iswa.h"
-#include "llama-memory-hybrid.h"
-#include "llama-memory-recurrent.h"
+#include "llama-kv-cache-recurrent.h"
 
 #include <cassert>
 #include <cmath>
@@ -92,28 +91,36 @@ void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) {
-    GGML_ASSERT(out_ids);
+    if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        //GGML_ASSERT(out_ids && "every model that can must skip unused outputs");
 
-    const int64_t n_tokens = ubatch->n_tokens;
+        if (!out_ids) {
+            LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__);
+        } else {
+            const int64_t n_tokens = ubatch->n_tokens;
 
-    GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer));
-    int32_t * data = (int32_t *) out_ids->data;
+            GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer));
+            int32_t * data = (int32_t *) out_ids->data;
 
-    if (n_outputs == n_tokens) {
-        for (int i = 0; i < n_tokens; ++i) {
-            data[i] = i;
-        }
-
-        return;
-    }
-
-    GGML_ASSERT(ubatch->output);
-
-    int n_outputs = 0;
-
-    for (int i = 0; i < n_tokens; ++i) {
-        if (ubatch->output[i]) {
-            data[n_outputs++] = i;
+            if (n_outputs == n_tokens) {
+                for (int i = 0; i < n_tokens; ++i) {
+                    data[i] = i;
+                }
+            } else if (ubatch->output) {
+                int32_t n_outputs = 0;
+                for (int i = 0; i < n_tokens; ++i) {
+                    if (ubatch->output[i]) {
+                        data[n_outputs++] = i;
+                    }
+                }
+                // the graph needs to have been passed the correct number of outputs
+                GGML_ASSERT(n_outputs == n_outputs);
+            } else if (n_outputs == 1) {
+                // only keep last output
+                data[0] = n_tokens - 1;
+            } else {
+                GGML_ASSERT(n_outputs == 0);
+            }
         }
     }
 }
@@ -122,114 +129,123 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
         const int64_t n_tokens     = ubatch->n_tokens;
         const int64_t n_seq_tokens = ubatch->n_seq_tokens;
-        const int64_t n_seqs_unq   = ubatch->n_seqs_unq;
+        const int64_t n_seqs       = ubatch->n_seqs;
 
         GGML_ASSERT(mean);
         GGML_ASSERT(ggml_backend_buffer_is_host(mean->buffer));
 
         float * data = (float *) mean->data;
-        memset(mean->data, 0, n_tokens*n_seqs_unq*ggml_element_size(mean));
+        memset(mean->data, 0, n_tokens * n_tokens * ggml_element_size(mean));
 
-        std::vector<uint64_t> sums(n_seqs_unq, 0);
-        for (int i = 0; i < n_tokens; i += n_seq_tokens) {
-            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
-                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
-                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
+        std::vector<uint64_t> sum(n_tokens, 0);
 
-                sums[seq_idx] += ubatch->n_seq_tokens;
+        for (int s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = ubatch->seq_id[s][0];
+
+            // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
+
+            sum[seq_id] += ubatch->n_seq_tokens;
+        }
+
+        std::vector<float> div(n_tokens, 0.0f);
+        for (int i = 0; i < n_tokens; ++i) {
+            const uint64_t s = sum[i];
+            if (s > 0) {
+                div[i] = 1.0f/float(s);
             }
         }
 
-        std::vector<float> div(n_seqs_unq, 0.0f);
-        for (int s = 0; s < n_seqs_unq; ++s) {
-            const uint64_t sum = sums[s];
-            if (sum > 0) {
-                div[s] = 1.0f/float(sum);
-            }
-        }
+        for (int s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
-        for (int i = 0; i < n_tokens; i += n_seq_tokens) {
-            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
-                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
-                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
-
-                for (int j = 0; j < n_seq_tokens; ++j) {
-                    data[seq_idx*n_tokens + i + j] = div[seq_idx];
-                }
+            for (int i = 0; i < n_seq_tokens; ++i) {
+                data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id];
             }
         }
     }
 }
 
 void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
-    const int64_t n_tokens     = ubatch->n_tokens;
-    const int64_t n_seq_tokens = ubatch->n_seq_tokens;
-    const int64_t n_seqs_unq   = ubatch->n_seqs_unq;
-
     if (cparams.embeddings && (
-            cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
-            cparams.pooling_type == LLAMA_POOLING_TYPE_RANK
-        )) {
+                cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
+                cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) {
+        const int64_t n_tokens     = ubatch->n_tokens;
+        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+        const int64_t n_seqs       = ubatch->n_seqs;
+
         GGML_ASSERT(cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer));
 
         uint32_t * data = (uint32_t *) cls->data;
-        memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls));
+        memset(cls->data, 0, n_tokens * ggml_element_size(cls));
 
-        for (int i = 0; i < n_tokens; i += n_seq_tokens) {
-            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
-                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
-                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
+        for (int s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
-                data[seq_idx] = i;
+            // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK");
+
+            for (int i = 0; i < n_seq_tokens; ++i) {
+                const llama_pos pos = ubatch->pos[s*n_seq_tokens + i];
+
+                if (pos == 0) {
+                    data[seq_id] = s*n_seq_tokens + i;
+                }
             }
         }
     }
 
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) {
+        const int64_t n_tokens     = ubatch->n_tokens;
+        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+        const int64_t n_seqs       = ubatch->n_seqs;
+
         GGML_ASSERT(cls);
         GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer));
 
         uint32_t * data = (uint32_t *) cls->data;
-        memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls));
+        memset(cls->data, 0, n_tokens * ggml_element_size(cls));
 
-        std::vector<int> last_pos(n_seqs_unq, -1);
-        std::vector<int> last_row(n_seqs_unq, -1);
+        std::vector<int> last_pos(n_tokens, -1);
+        std::vector<int> last_row(n_tokens, -1);
 
-        for (int i = 0; i < n_tokens; ++i) {
-            const llama_pos pos = ubatch->pos[i];
+        for (int s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
-            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
-                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
-                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
+            // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST");
 
-                if (pos >= last_pos[seq_idx]) {
-                    last_pos[seq_idx] = pos;
-                    last_row[seq_idx] = i;
+            for (int i = 0; i < n_seq_tokens; ++i) {
+                const llama_pos pos = ubatch->pos[s*n_seq_tokens + i];
+
+                if (pos >= last_pos[seq_id]) {
+                    last_pos[seq_id] = pos;
+                    last_row[seq_id] = s*n_seq_tokens + i;
                 }
             }
         }
 
-        for (int s = 0; s < n_seqs_unq; ++s) {
-            if (last_row[s] >= 0) {
-                data[s] = last_row[s];
+        for (int i = 0; i < n_tokens; ++i) {
+            if (last_row[i] >= 0) {
+                data[i] = last_row[i];
             }
         }
     }
 }
 
-void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
-    const int64_t n_rs = mem_state->get_n_rs();
+    const int64_t n_kv = kv_state->get_n_kv();
 
     if (s_copy) {
         GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
         int32_t * data = (int32_t *) s_copy->data;
 
         // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
-        for (uint32_t i = 0; i < n_rs; ++i) {
-            data[i] = mem_state->s_copy(i);
+        for (uint32_t i = 0; i < n_kv; ++i) {
+            data[i] = kv_state->s_copy(i);
         }
     }
 }
@@ -245,36 +261,87 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
-    const int64_t n_kv     = ubatch->n_tokens;
-    const int64_t n_tokens = ubatch->n_tokens;
+    if (kq_mask) {
+        if (cparams.causal_attn) {
+            const int64_t n_kv         = ubatch->n_tokens;
+            const int64_t n_tokens     = ubatch->n_tokens;
+            const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+            const int64_t n_seqs       = ubatch->n_seqs;
 
-    GGML_ASSERT(kq_mask);
-    GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));
+            GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));
+            float * data = (float *) kq_mask->data;
 
-    float * data = (float *) kq_mask->data;
+            for (int h = 0; h < 1; ++h) {
+                for (int s1 = 0; s1 < n_seqs; ++s1) {
+                    const llama_seq_id seq_id = ubatch->seq_id[s1][0];
 
-    for (int h = 0; h < 1; ++h) {
-        for (int i1 = 0; i1 < n_tokens; ++i1) {
-            const llama_seq_id s1 = ubatch->seq_id[i1][0];
+                    for (int j = 0; j < n_seq_tokens; ++j) {
+                        const int32_t tj = s1*n_seq_tokens + j;
 
-            for (int i0 = 0; i0 < n_tokens; ++i0) {
-                float f = -INFINITY;
+                        for (int s0 = 0; s0 < n_seqs; ++s0) {
+                            for (int i = 0; i < n_seq_tokens; ++i) {
+                                const int32_t ti = s0*n_seq_tokens + i;
+                                float f = -INFINITY;
 
-                for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
-                    const llama_seq_id s0 = ubatch->seq_id[i0][0];
+                                for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
+                                    if (ubatch->seq_id[s0][s] == seq_id && ubatch->pos[ti] <= ubatch->pos[tj]) {
+                                        if (hparams.use_alibi) {
+                                            f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]);
+                                        } else {
+                                            f = 0.0f;
+                                        }
+                                        break;
+                                    }
+                                }
 
-                    // TODO: reimplement this like in llama_kv_cache_unified
-                    if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) {
-                        if (hparams.use_alibi) {
-                            f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
-                        } else {
-                            f = 0.0f;
+                                data[h*(n_kv*n_tokens) + tj*n_kv + ti] = f;
+                            }
                         }
-                        break;
                     }
                 }
+            }
+        } else {
+            const int64_t n_tokens     = ubatch->n_tokens;
+            const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+            const int64_t n_seqs       = ubatch->n_seqs;
+            const int64_t n_stride     = ubatch->n_tokens;
 
-                data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
+            GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));
+
+            float * data = (float *) kq_mask->data;
+
+            for (int h = 0; h < 1; ++h) {
+                for (int s1 = 0; s1 < n_seqs; ++s1) {
+                    const llama_seq_id seq_id = ubatch->seq_id[s1][0];
+
+                    for (int j = 0; j < n_seq_tokens; ++j) {
+                        const int32_t tj = s1*n_seq_tokens + j;
+
+                        for (int s0 = 0; s0 < n_seqs; ++s0) {
+                            for (int i = 0; i < n_seq_tokens; ++i) {
+                                const int32_t ti = s0*n_seq_tokens + i;
+                                float f = -INFINITY;
+
+                                for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
+                                    if (ubatch->seq_id[s0][s] == seq_id) {
+                                        if (hparams.use_alibi) {
+                                            f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]);
+                                        } else {
+                                            f = 0.0f;
+                                        }
+                                        break;
+                                    }
+                                }
+
+                                data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f;
+                            }
+                        }
+
+                        for (int i = n_tokens; i < n_stride; ++i) {
+                            data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY;
+                        }
+                    }
+                }
             }
         }
     }
@@ -297,59 +364,38 @@ void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch
 }
 
 void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
-    GGML_ASSERT(cross_kq_mask);
+    if (cross_kq_mask) {
+        const int64_t n_enc    = cross_kq_mask->ne[0];
+        const int64_t n_tokens = ubatch->n_tokens;
 
-    const int64_t n_enc    = cross_kq_mask->ne[0];
-    const int64_t n_tokens = ubatch->n_tokens;
+        GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
+        GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
 
-    GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
-    GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
+        float * data = (float *) cross_kq_mask->data;
 
-    float * data = (float *) cross_kq_mask->data;
-
-    for (int h = 0; h < 1; ++h) {
-        for (int i = 0; i < n_tokens; ++i) {
-            for (int j = 0; j < n_enc; ++j) {
-                float f = -INFINITY;
-
-                for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
-                    const llama_seq_id seq_id = ubatch->seq_id[i][s];
-
-                    if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) {
-                        f = 0.0f;
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                for (int i = 0; i < n_enc; ++i) {
+                    float f = -INFINITY;
+                    for (int s = 0; s < ubatch->n_seq_id[j]; ++s) {
+                        const llama_seq_id seq_id = ubatch->seq_id[j][s];
+                        if (cross->seq_ids_enc[i].find(seq_id) != cross->seq_ids_enc[i].end()) {
+                            f = 0.0f;
+                        }
                     }
+                    data[h*(n_enc*n_tokens) + j*n_enc + i] = f;
                 }
-
-                data[h*(n_enc*n_tokens) + i*n_enc + j] = f;
             }
-        }
 
-        for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
-            for (int j = 0; j < n_enc; ++j) {
-                data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY;
+            for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+                for (int j = 0; j < n_enc; ++j) {
+                    data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY;
+                }
             }
         }
     }
 }
 
-void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        mem_state->get_state_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
-
-    const int64_t n_rs = mem_state->get_state_recr()->get_n_rs();
-
-    if (s_copy) {
-        GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
-        int32_t * data = (int32_t *) s_copy->data;
-
-        // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
-        for (uint32_t i = 0; i < n_rs; ++i) {
-            data[i] = mem_state->get_state_recr()->s_copy(i);
-        }
-    }
-}
-
 //
 // llm_graph_context
 //
@@ -395,6 +441,10 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     res              (std::make_unique<llm_graph_result>()) {
     }
 
+int64_t llm_graph_context::n_pos_per_embd() const {
+    return hparams.rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+}
+
 void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const {
     if (cb_func) {
         cb_func(ubatch, cur, name, il);
@@ -839,11 +889,11 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos() const {
-    auto inp = std::make_unique<llm_graph_input_pos>(hparams.n_pos_per_embd());
+    auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_embd());
 
     auto & cur = inp->pos;
 
-    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, (int64_t)n_tokens*hparams.n_pos_per_embd());
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_embd());
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -866,14 +916,6 @@ ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_out_ids() const {
-    // note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls,
-    //       but this would make the graph topology depend on the number of output tokens, which can interere with
-    //       features that require constant topology such as pipline parallelism
-    //       ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471
-    //if (n_outputs < n_tokens) {
-    //    return nullptr;
-    //}
-
     auto inp = std::make_unique<llm_graph_input_out_ids>(hparams, cparams, n_outputs);
 
     auto & cur = inp->out_ids;
@@ -891,7 +933,7 @@ ggml_tensor * llm_graph_context::build_inp_mean() const {
 
     auto & cur = inp->mean;
 
-    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, ubatch.n_seqs_unq);
+    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -904,7 +946,24 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
 
     auto & cur = inp->cls;
 
-    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_seqs_unq);
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+ggml_tensor * llm_graph_context::build_inp_s_copy() const {
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+
+    auto inp = std::make_unique<llm_graph_input_s_copy>(kv_state);
+
+    const auto n_kv = kv_state->get_n_kv();
+
+    auto & cur = inp->s_copy;
+
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv);
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -981,33 +1040,6 @@ ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_t
     return pos_bias;
 }
 
-llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
-    const auto * mem_state = static_cast<const llama_memory_hybrid_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(hparams, cparams, mem_state);
-
-    {
-        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers");
-
-        const auto n_kv = inp->mem_state->get_state_attn()->get_n_kv();
-
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
-        ggml_set_input(inp->self_kq_mask);
-
-        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
-    }
-
-    {
-        const auto n_rs = mem_state->get_state_recr()->get_n_rs();
-
-        inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
-        ggml_set_input(inp->s_copy);
-    }
-
-    return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
-}
-
 ggml_tensor * llm_graph_context::build_attn_mha(
          ggml_cgraph * gf,
          ggml_tensor * q,
@@ -1252,6 +1284,36 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
+llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
+    const auto * kv_state = static_cast<const llama_kv_cache_unified_iswa_state *>(mstate);
+
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, kv_state);
+
+    {
+        const auto n_kv = kv_state->get_base()->get_n_kv();
+
+        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        ggml_set_input(inp->self_kq_mask);
+
+        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+    }
+
+    {
+        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
+
+        const auto n_kv = kv_state->get_swa()->get_n_kv();
+
+        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
+        ggml_set_input(inp->self_kq_mask_swa);
+
+        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
+    }
+
+    return (llm_graph_input_attn_kv_unified_iswa *) res->add_input(std::move(inp));
+}
+
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_kv_unified_iswa * inp,
         ggml_cgraph * gf,
@@ -1361,99 +1423,20 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_mem_hybrid * inp,
-        ggml_cgraph * gf,
-        ggml_tensor * wo,
-        ggml_tensor * wo_b,
-        ggml_tensor * q_cur,
-        ggml_tensor * k_cur,
-        ggml_tensor * v_cur,
-        ggml_tensor * kq_b,
-        ggml_tensor * v_mla,
-            float     kq_scale,
-            int       il) const {
-    // these nodes are added to the graph together so that they are not reordered
-    // by doing so, the number of splits in the graph is reduced
-    ggml_build_forward_expand(gf, q_cur);
-    ggml_build_forward_expand(gf, k_cur);
-    ggml_build_forward_expand(gf, v_cur);
+ggml_tensor * llm_graph_context::build_recurrent_state(
+         ggml_cgraph * gf,
+         ggml_tensor * s,
+         ggml_tensor * state_copy,
+             int32_t   state_size,
+             int32_t   n_seqs,
+                bool   avoid_copies) const {
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
-    const auto * kv_state = static_cast<const llama_memory_hybrid_state *>(mstate)->get_state_attn();
+    const auto n_kv    = kv_state->get_n_kv();
+    const auto kv_head = kv_state->get_head();
+    const auto rs_zero = kv_state->get_rs_z();
 
-    // store to KV cache
-    {
-        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
-    }
-
-    const auto & kq_mask = inp->get_kq_mask();
-
-    ggml_tensor * q = q_cur;
-    ggml_tensor * k = kv_state->get_k(ctx0, il);
-    ggml_tensor * v = kv_state->get_v(ctx0, il);
-
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
-    cb(cur, "kqv_out", il);
-
-    if (wo) {
-        cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
-            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
-        }
-    }
-
-    if (wo_b) {
-        cur = ggml_add(ctx0, cur, wo_b);
-    }
-
-    return cur;
-}
-
-llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_iswa_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, kv_state);
-
-    {
-        const auto n_kv = kv_state->get_base()->get_n_kv();
-
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
-        ggml_set_input(inp->self_kq_mask);
-
-        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
-    }
-
-    {
-        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
-
-        const auto n_kv = kv_state->get_swa()->get_n_kv();
-
-        inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
-        ggml_set_input(inp->self_kq_mask_swa);
-
-        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
-    }
-
-    return (llm_graph_input_attn_kv_unified_iswa *) res->add_input(std::move(inp));
-}
-
-ggml_tensor * llm_graph_context::build_rs(
-        ggml_cgraph * gf,
-        ggml_tensor * s,
-        ggml_tensor * state_copy,
-            int32_t   state_size,
-            int32_t   n_seqs,
-           uint32_t   n_kv,
-           uint32_t   kv_head,
-           uint32_t   kv_size,
-            int32_t   rs_zero,
-               bool   avoid_copies) const {
-
-    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_size);
+    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_state->get_size());
 
     // Clear a single state which will then be copied to the other cleared states.
     // Note that this is a no-op when the view is zero-sized.
@@ -1484,59 +1467,22 @@ ggml_tensor * llm_graph_context::build_rs(
     return output_states;
 }
 
-llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
-    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_rs>(kv_state);
-
-    const auto n_rs = kv_state->get_n_rs();
-
-    inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
-    ggml_set_input(inp->s_copy);
-
-    return (llm_graph_input_rs *) res->add_input(std::move(inp));
-}
-
-ggml_tensor * llm_graph_context::build_rs(
-        llm_graph_input_rs * inp,
-        ggml_cgraph * gf,
-        ggml_tensor * s,
-            int32_t   state_size,
-            int32_t   n_seqs,
-               bool   avoid_copies) const {
-    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
-
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), avoid_copies);
-}
-
-ggml_tensor * llm_graph_context::build_rs(
-        llm_graph_input_mem_hybrid * inp,
-        ggml_cgraph * gf,
-        ggml_tensor * s,
-            int32_t   state_size,
-            int32_t   n_seqs,
-               bool   avoid_copies) const {
-    const auto * kv_state = static_cast<const llama_memory_hybrid_state *>(mstate)->get_state_recr();
-
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), avoid_copies);
-}
-
 ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
-    llm_graph_input_rs * inp,
-           ggml_cgraph * gf,
-    const llama_ubatch & ubatch,
+         ggml_cgraph * gf,
+         ggml_tensor * state_copy,
+  const llama_ubatch & ubatch,
                  int   il) const {
-    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
     const auto token_shift_count = hparams.token_shift_count;
 
     const int64_t n_seqs  = ubatch.n_seqs;
 
-    ggml_tensor * token_shift_all = kv_state->get_r_l(il);
+    ggml_tensor * token_shift_all = kv_state->get_k_l(il);
 
-    ggml_tensor * token_shift = build_rs(
-            inp, gf, token_shift_all,
-            hparams.n_embd_r(), n_seqs);
+    ggml_tensor * token_shift = build_recurrent_state(
+            gf, token_shift_all, state_copy,
+            hparams.n_embd_k_s(), n_seqs);
 
     token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);
 
@@ -1547,7 +1493,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
          ggml_tensor * token_shift,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
+    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
 
     const auto token_shift_count = hparams.token_shift_count;
     const auto n_embd = hparams.n_embd;
@@ -1559,7 +1505,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
     return ggml_cpy(
         ctx0,
         ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0),
-        ggml_view_1d(ctx0, kv_state->get_r_l(il), hparams.n_embd_r()*n_seqs, hparams.n_embd_r()*kv_head*ggml_element_size(kv_state->get_r_l(il)))
+        ggml_view_1d(ctx0, kv_state->get_k_l(il), hparams.n_embd_k_s()*n_seqs, hparams.n_embd_k_s()*kv_head*ggml_element_size(kv_state->get_k_l(il)))
     );
 }
 
@@ -1610,30 +1556,23 @@ void llm_graph_context::build_pooling(
                 ggml_tensor * inp_cls = build_inp_cls();
                 inp = ggml_get_rows(ctx0, inp, inp_cls);
 
-                if (cls) {
+                if (cls != nullptr && cls_b != nullptr) {
                     // classification head
                     // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                    cur = ggml_mul_mat(ctx0, cls, inp);
-                    if (cls_b) {
-                        cur = ggml_add(ctx0, cur, cls_b);
-                    }
+                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
                     cur = ggml_tanh(ctx0, cur);
 
                     // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
                     // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
                     if (cls_out) {
-                        cur = ggml_mul_mat(ctx0, cls_out, cur);
-                        if (cls_out_b) {
-                            cur = ggml_add(ctx0, cur, cls_out_b);
-                        }
+                        GGML_ASSERT(cls_out_b != nullptr);
+                        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
                     }
                 } else if (cls_out) {
                     // Single layer classification head (direct projection)
                     // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
-                    cur = ggml_mul_mat(ctx0, cls_out, inp);
-                    if (cls_out_b) {
-                        cur = ggml_add(ctx0, cur, cls_out_b);
-                    }
+                    GGML_ASSERT(cls_out_b != nullptr);
+                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, inp), cls_out_b);
                 } else {
                     GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
                 }

src/llama-graph.h

@@ -21,8 +21,7 @@ struct llama_memory_state_i;
 
 class llama_kv_cache_unified_state;
 class llama_kv_cache_unified_iswa_state;
-class llama_memory_recurrent_state;
-class llama_memory_hybrid_state;
+class llama_kv_cache_recurrent_state;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -95,14 +94,14 @@ public:
 
 class llm_graph_input_pos : public llm_graph_input_i {
 public:
-    llm_graph_input_pos(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
+    llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
     virtual ~llm_graph_input_pos() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * pos = nullptr; // I32 [n_batch]
 
-    const uint32_t n_pos_per_embd = 1;
+    const int64_t n_pos_per_embd = 1;
 };
 
 // temperature tuning, used by llama4
@@ -189,16 +188,16 @@ public:
     const llama_cparams & cparams;
 };
 
-class llm_graph_input_rs : public llm_graph_input_i {
+class llm_graph_input_s_copy : public llm_graph_input_i {
 public:
-    llm_graph_input_rs(const llama_memory_recurrent_state * mem_state) : mem_state(mem_state) {}
-    virtual ~llm_graph_input_rs() = default;
+    llm_graph_input_s_copy(const llama_kv_cache_recurrent_state * kv_state) : kv_state(kv_state) {}
+    virtual ~llm_graph_input_s_copy() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
-    const llama_memory_recurrent_state * mem_state;
+    const llama_kv_cache_recurrent_state * kv_state;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -301,33 +300,6 @@ public:
     const llama_cross * cross = nullptr;
 };
 
-class llm_graph_input_mem_hybrid : public llm_graph_input_i {
-public:
-    llm_graph_input_mem_hybrid(
-            const llama_hparams & hparams,
-            const llama_cparams & cparams,
-            const llama_memory_hybrid_state * mem_state) :
-        hparams(hparams),
-        cparams(cparams),
-        mem_state(mem_state) {
-    }
-    virtual ~llm_graph_input_mem_hybrid() = default;
-
-    void set_input(const llama_ubatch * ubatch) override;
-
-    ggml_tensor * s_copy; // I32 [kv_size]
-
-    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
-
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
-
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
-
-    const llama_memory_hybrid_state * mem_state;
-};
-
 //
 // llm_graph_result
 //
@@ -406,7 +378,7 @@ struct llm_graph_params {
     const llama_memory_state_i * mstate;
     const llama_cross          * cross;
 
-    uint32_t n_outputs;
+    int32_t n_outputs;
 
     const llm_graph_cb & cb;
 };
@@ -440,8 +412,8 @@ struct llm_graph_context {
     const float norm_eps;
     const float norm_rms_eps;
 
-    const int64_t n_tokens;
-    const int64_t n_outputs;
+    const int32_t n_tokens;
+    const int32_t n_outputs;
     const int32_t n_ctx_orig; // yarn
 
     const enum llama_pooling_type pooling_type;
@@ -464,6 +436,8 @@ struct llm_graph_context {
 
     llm_graph_context(const llm_graph_params & params);
 
+    int64_t n_pos_per_embd() const;
+
     void cb(ggml_tensor * cur, const char * name, int il) const;
 
     //
@@ -534,14 +508,13 @@ struct llm_graph_context {
     ggml_tensor * build_inp_out_ids() const;
     ggml_tensor * build_inp_mean() const;
     ggml_tensor * build_inp_cls() const;
+    ggml_tensor * build_inp_s_copy() const;
 
     ggml_tensor * build_inp_cross_embd() const;
     ggml_tensor * build_inp_pos_bucket_enc() const;
     ggml_tensor * build_inp_pos_bucket_dec() const;
     ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const;
 
-    llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
-
     //
     // attention
     //
@@ -616,62 +589,22 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
-    ggml_tensor * build_attn(
-            llm_graph_input_mem_hybrid * inp,
-            ggml_cgraph * gf,
-            ggml_tensor * wo,
-            ggml_tensor * wo_b,
-            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
-            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
-            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
-            ggml_tensor * kq_b,
-            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                  float   kq_scale,
-                    int   il) const;
     //
     // recurrent
     //
 
-    // TODO: avoid notion of "kv"
-    // TODO: move this implementation to llama_memory_recurrent.
-    //       this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
-    //       when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
-    //         implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
-    //         `llama_memory_recurrent`
-    ggml_tensor * build_rs(
-            ggml_cgraph * gf,
-            ggml_tensor * s,
-            ggml_tensor * state_copy,
-                int32_t   state_size,
-                int32_t   n_seqs,
-               uint32_t   n_kv,
-               uint32_t   kv_head,
-               uint32_t   kv_size,
-                int32_t   rs_zero,
-                   bool   avoid_copies = false) const;
-
-    llm_graph_input_rs * build_rs_inp() const;
-
-    ggml_tensor * build_rs(
-            llm_graph_input_rs * inp,
-            ggml_cgraph * gf,
-            ggml_tensor * s,
-                int32_t   state_size,
-                int32_t   n_seqs,
-                   bool   avoid_copies = false) const;
-
-    ggml_tensor * build_rs(
-            llm_graph_input_mem_hybrid * inp,
-            ggml_cgraph * gf,
-            ggml_tensor * s,
-                int32_t   state_size,
-                int32_t   n_seqs,
-                   bool   avoid_copies = false) const;
+    ggml_tensor * build_recurrent_state(
+             ggml_cgraph * gf,
+             ggml_tensor * s,
+             ggml_tensor * state_copy,
+                 int32_t   state_size,
+                 int32_t   n_seqs,
+                    bool   avoid_copies = false) const;
 
     ggml_tensor * build_rwkv_token_shift_load(
-        llm_graph_input_rs * inp,
-               ggml_cgraph * gf,
-        const llama_ubatch & ubatch,
+             ggml_cgraph * gf,
+             ggml_tensor * state_copy,
+      const llama_ubatch & ubatch,
                      int   il) const;
 
     ggml_tensor * build_rwkv_token_shift_store(