ggml-cpu : remove unnecesary arm feature detection (#14281)

Support for Arm runtime feature detection has now been added to GGML_CPU_ALL_VARIANTS. This removes the old and not very functional code.

.devops/intel.Dockerfile

@@ -49,19 +49,23 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
-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
+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
 
+ENV PATH="/opt/venv/bin:$PATH"
 
 ENTRYPOINT ["/app/tools.sh"]
 

.github/workflows/build.yml

@@ -306,6 +306,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
+          export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 3600
 
@@ -692,7 +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: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+            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'
           - 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'
@@ -777,6 +778,7 @@ 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

@@ -89,6 +89,14 @@ 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})
@@ -155,6 +163,8 @@ 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()
@@ -204,10 +214,6 @@ 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) / [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)
+[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)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
@@ -18,7 +18,6 @@ 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=OFF"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"

common/CMakeLists.txt

@@ -7,8 +7,8 @@ llama_add_compile_flags()
 # Build info header
 #
 
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
+    set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
 
     # Is git submodule
     if(NOT IS_DIRECTORY "${GIT_DIR}")
@@ -18,36 +18,26 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         if (SLASH_POS EQUAL 0)
             set(GIT_DIR "${REAL_GIT_DIR}")
         else()
-            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+            set(GIT_DIR "${PROJECT_SOURCE_DIR}/${REAL_GIT_DIR}")
         endif()
     endif()
 
     if(EXISTS "${GIT_DIR}/index")
-        set(GIT_INDEX "${GIT_DIR}/index")
+        # For build-info.cpp below
+        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${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()
 
-# 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 "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
-    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
-    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
-    VERBATIM
-)
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
+configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+
 set(TARGET build_info)
-add_library(${TARGET} OBJECT build-info.cpp)
+add_library(${TARGET} OBJECT ${OUTPUT_FILE})
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

common/arg.cpp

@@ -988,10 +988,6 @@ 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",
@@ -2747,9 +2743,10 @@ 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)", params.reranking ? "enabled" : "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
         [](common_params & params) {
-            params.reranking = true;
+            params.embedding = true;
+            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(common_arg(
@@ -3213,6 +3210,32 @@ 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 = @BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
 char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
 char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/chat-parser.cpp

@@ -49,6 +49,7 @@ 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) {
@@ -378,3 +379,7 @@ 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

@@ -115,4 +115,6 @@ 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");
+        throw std::runtime_error("this custom template is not supported, try using --jinja");
     }
 
     // if it turns out that our buffer is too small, we resize it
@@ -1921,7 +1921,9 @@ 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) {
-            throw std::runtime_error(ex.what());
+            builder.clear_tools();
+            builder.move_to(0);
+            common_chat_parse_content_only(builder);
         }
     }
     auto msg = builder.result();

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

@@ -1,24 +0,0 @@
-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

@@ -466,7 +466,7 @@ size_t string_find_partial_stop(const std::string_view & str, const std::string_
 
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
-    return std::regex_replace(s, special_chars, "\\$0");
+    return std::regex_replace(s, special_chars, "\\$&");
 }
 
 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
@@ -706,11 +706,17 @@ 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);
@@ -767,6 +773,9 @@ 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
@@ -784,9 +793,16 @@ 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();
 
-        const bool success = CreateDirectoryW(test, NULL);
+        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);
+
         if (!success) {
             const DWORD error = GetLastError();
 
@@ -800,8 +816,6 @@ bool fs_create_directory_with_parents(const std::string & path) {
                 return false;
             }
         }
-
-        pos_slash += 1;
     }
 
     return true;
@@ -897,34 +911,6 @@ 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);
@@ -966,6 +952,35 @@ 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;
@@ -1143,11 +1158,6 @@ 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;
 

common/common.h

@@ -199,6 +199,9 @@ 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;
 
@@ -355,7 +358,6 @@ 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
-    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"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions", "max_length"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
@@ -1898,9 +1898,7 @@ class LlamaModel(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
+        if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -1982,7 +1980,8 @@ 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)
-                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+                if (dim := self.hparams.get("head_dim")) is None:
+                    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)
@@ -2017,6 +2016,20 @@ 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
@@ -2304,9 +2317,7 @@ class DeciModel(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
+        if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -2346,7 +2357,8 @@ 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)
-                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+                if (dim := self.hparams.get("head_dim")) is None:
+                    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)
@@ -3664,9 +3676,7 @@ class InternLM3Model(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
+        if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -4059,6 +4069,34 @@ 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
@@ -4798,25 +4836,6 @@ class OlmoeModel(TextModel):
 class JinaBertV2Model(BertModel):
     model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.intermediate_size = self.hparams["intermediate_size"]
-
-    def get_tensors(self):
-        for name, data in super().get_tensors():
-            if 'gated_layer' in name:
-                d1 = data[:self.intermediate_size, :]
-                name1 = name.replace('gated_layers', 'gated_layers_w')
-                name1 = name1.replace('up_gated_layer', 'gated_layers_v')
-                d2 = data[self.intermediate_size:, :]
-                name2 = name.replace('gated_layers', 'gated_layers_v')
-                name2 = name2.replace('up_gated_layer', 'gated_layers_w')
-                yield name1, d1
-                yield name2, d2
-                continue
-
-            yield name, data
-
     def set_vocab(self):
         tokenizer_class = 'BertTokenizer'
         with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
@@ -4832,14 +4851,6 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "bert.", remove the prefix
-        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
-        if name.startswith("bert."):
-            name = name[5:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
 
 @ModelBase.register("OpenELMForCausalLM")
 class OpenELMModel(TextModel):
@@ -5080,9 +5091,7 @@ class DeepseekModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
+        if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
@@ -5286,6 +5295,34 @@ 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
@@ -5944,7 +5981,8 @@ 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)
-                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+                if (dim := self.hparams.get("head_dim")) is None:
+                    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)
@@ -6056,7 +6094,8 @@ class BailingMoeModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"]
+        if (rope_dim := hparams.get("head_dim")) is None:
+            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
         rope_scaling = self.hparams.get("rope_scaling") or {}
@@ -6088,7 +6127,8 @@ 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"]
-        head_dim = self.hparams.get("head_dim") or n_embd // n_head
+        if (head_dim := self.hparams.get("head_dim")) is None:
+            head_dim = n_embd // n_head
 
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
 
@@ -6349,8 +6389,8 @@ def parse_args() -> argparse.Namespace:
         help="model is executed on big endian machine",
     )
     parser.add_argument(
-        "model", type=Path,
-        help="directory containing model file",
+        "model", type=str,
+        help="directory containing model file or huggingface repository ID (if --remote)",
         nargs="?",
     )
     parser.add_argument(
@@ -6453,18 +6493,20 @@ 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=str(dir_model),
+            repo_id=hf_repo_id,
             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: {args.model} is not a directory')
+        logger.error(f'Error: {dir_model} is not a directory')
         sys.exit(1)
 
     ftype_map: dict[str, gguf.LlamaFileType] = {
@@ -6484,9 +6526,9 @@ def main() -> None:
 
     if args.outfile is not None:
         fname_out = args.outfile
-    elif args.remote:
+    elif hf_repo_id:
         # if remote, use the model ID as the output file name
-        fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf")
+        fname_out = Path("./" + hf_repo_id.replace("/", "-") + "-{ftype}.gguf")
     else:
         fname_out = dir_model
 
@@ -6515,7 +6557,7 @@ 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=str(args.model) if args.remote else None)
+                                     remote_hf_model_id=hf_repo_id)
 
         if args.vocab_only:
             logger.info("Exporting model vocab...")

docs/build-s390x.md

@@ -0,0 +1,157 @@
+> [!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/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 (WIP: https://github.com/ggml-org/llama.cpp/pull/12034)
+  - Qwen 2.5 Coder
   - Mistral Nemo
   - Firefunction v2
   - Command R7B

docs/multimodal.md

@@ -107,3 +107,7 @@ 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/gritlm/gritlm.cpp

@@ -41,12 +41,11 @@ 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 }, j >= n_inst);
+            common_batch_add(batch, inputs[j], j, { 0 }, true);
         }
 
         // 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
@@ -103,7 +102,6 @@ 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);
@@ -166,6 +164,8 @@ 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);
@@ -213,6 +213,8 @@ 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

@@ -172,6 +172,7 @@ 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)
@@ -367,6 +368,8 @@ 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
     )
@@ -386,4 +389,46 @@ 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

@@ -36,8 +36,7 @@ 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} " STREQUAL "ppc64le " OR
-            "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
         set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)

ggml/src/CMakeLists.txt

@@ -270,17 +270,23 @@ endfunction()
 function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
-    foreach (feat NATIVE
-                  SSE42
-                  AVX AVX2 BMI2 AVX_VNNI FMA F16C
-                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
-                  AMX_TILE AMX_INT8 AMX_BF16)
-        set(GGML_${feat} OFF)
-    endforeach()
+    if (GGML_SYSTEM_ARCH STREQUAL "x86")
+        foreach (feat NATIVE
+                      SSE42
+                      AVX AVX2 BMI2 AVX_VNNI FMA F16C
+                      AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
+                      AMX_TILE AMX_INT8 AMX_BF16)
+            set(GGML_${feat} OFF)
+        endforeach()
 
-    foreach (feat ${ARGN})
-        set(GGML_${feat} ON)
-    endforeach()
+        foreach (feat ${ARGN})
+            set(GGML_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "ARM")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
 endfunction()
@@ -290,6 +296,8 @@ ggml_add_backend(CPU)
 if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
+    elseif (GGML_CPU_ARM_ARCH)
+        message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
     endif()
     if (GGML_SYSTEM_ARCH STREQUAL "x86")
         ggml_add_cpu_backend_variant(x64)
@@ -303,8 +311,34 @@ 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()
     else()
-        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported on ${GGML_SYSTEM_ARCH}")
+        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")

ggml/src/ggml-backend-reg.cpp

@@ -69,6 +69,9 @@
 #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;
@@ -91,6 +94,8 @@ 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

@@ -1,3 +1,17 @@
+function(ggml_add_cpu_backend_features cpu_name arch)
+    # The feature detection code is compiled as a separate target so that
+    # it can be built without the architecture flags
+    # Since multiple variants of the CPU backend may be included in the same
+    # build, using set_source_files_properties() to set the arch flags is not possible
+    set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
+    add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
+    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
+    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
+    set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME})
+endfunction()
+
 function(ggml_add_cpu_backend_variant_impl tag_name)
     if (tag_name)
         set(GGML_CPU_NAME ggml-cpu-${tag_name})
@@ -143,6 +157,49 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
+                elseif(GGML_CPU_ALL_VARIANTS)
+                    # Begin with the lowest baseline
+                    set(ARM_MCPU "armv8-a")
+                    set(ARCH_TAGS "")
+                    set(ARCH_DEFINITIONS "")
+
+                    # When a feature is selected, bump the MCPU to the first
+                    # version that supported it
+                    if (GGML_INTERNAL_DOTPROD)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${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_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()
 
@@ -306,18 +363,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
                 message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
             endif()
-
-            # The feature detection code is compiled as a separate target so that
-            # it can be built without the architecture flags
-            # Since multiple variants of the CPU backend may be included in the same
-            # build, using set_source_files_properties() to set the arch flags is not possible
-            set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
-            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/x86/cpu-feats.cpp)
-            target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
-            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
-            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
-            set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-            target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} x86 ${ARCH_DEFINITIONS})
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
         message(STATUS "PowerPC detected")

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

@@ -0,0 +1,184 @@
+#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/arm/cpu-feats.cpp

@@ -0,0 +1,94 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__aarch64__)
+
+#if defined(__linux__)
+#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
+
+struct aarch64_features {
+    // has_neon not needed, aarch64 has NEON guaranteed
+    bool has_dotprod     = false;
+    bool has_fp16_va     = false;
+    bool has_sve         = false;
+    bool has_sve2        = false;
+    bool has_i8mm        = false;
+    bool has_sme         = false;
+
+    aarch64_features() {
+#if defined(__linux__)
+        uint32_t hwcap = getauxval(AT_HWCAP);
+        uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+        has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
+        has_fp16_va = !!(hwcap & HWCAP_FPHP);
+        has_sve     = !!(hwcap & HWCAP_SVE);
+        has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
+        has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
+        has_sme     = !!(hwcap2 & HWCAP2_SME);
+#elif defined(__APPLE__)
+        int oldp = 0;
+        size_t size = sizeof(oldp);
+
+        if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) == 0) {
+            has_dotprod = static_cast<bool>(oldp);
+        }
+
+        if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) == 0) {
+            has_i8mm = static_cast<bool>(oldp);
+        }
+
+        if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) == 0) {
+            has_sme = static_cast<bool>(oldp);
+        }
+
+        // Apple apparently does not implement SVE yet
+#endif
+    }
+};
+
+static int ggml_backend_cpu_aarch64_score() {
+    int score = 1;
+    aarch64_features af;
+
+#ifdef GGML_USE_DOTPROD
+    if (!af.has_dotprod) { return 0; }
+    score += 1<<1;
+#endif
+#ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
+    if (!af.has_fp16_va) { return 0; }
+    score += 1<<2;
+#endif
+#ifdef GGML_USE_SVE
+    if (!af.has_sve) { return 0; }
+    score += 1<<3;
+#endif
+#ifdef GGML_USE_MATMUL_INT8
+    if (!af.has_i8mm) { return 0; }
+    score += 1<<4;
+#endif
+#ifdef GGML_USE_SVE2
+    if (!af.has_sve2) { return 0; }
+    score += 1<<5;
+#endif
+#ifdef GGML_USE_SME
+    if (!af.has_sme) { return 0; }
+    score += 1<<6;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_aarch64_score)
+
+# endif // defined(__aarch64__)

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 {
@@ -503,31 +503,9 @@ 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

@@ -74,13 +74,8 @@
 
 #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;
-    int has_sme;
-} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
+} ggml_arm_arch_features = { 0 };
 #endif
 
 
@@ -559,6 +554,14 @@ 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;
@@ -670,87 +673,15 @@ 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__)
-    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)
+#if defined(__linux__) && defined(__aarch64__) && 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;
-
-    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
+
+#endif // __ARM_ARCH
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
     GGML_ASSERT(!ggml_get_no_alloc(ctx));
@@ -3435,7 +3366,7 @@ int ggml_cpu_has_vxe(void) {
 
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
-    return ggml_arm_arch_features.has_neon;
+    return 1;
 #else
     return 0;
 #endif
@@ -3443,7 +3374,7 @@ int ggml_cpu_has_neon(void) {
 
 int ggml_cpu_has_dotprod(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
-    return ggml_arm_arch_features.has_dotprod;
+    return 1;
 #else
     return 0;
 #endif
@@ -3451,7 +3382,7 @@ int ggml_cpu_has_dotprod(void) {
 
 int ggml_cpu_has_sve(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
-    return ggml_arm_arch_features.has_sve;
+    return 1;
 #else
     return 0;
 #endif
@@ -3459,7 +3390,7 @@ int ggml_cpu_has_sve(void) {
 
 int ggml_cpu_has_matmul_int8(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
-    return ggml_arm_arch_features.has_i8mm;
+    return 1;
 #else
     return 0;
 #endif
@@ -3475,7 +3406,7 @@ int ggml_cpu_get_sve_cnt(void) {
 
 int ggml_cpu_has_sme(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
-    return ggml_arm_arch_features.has_sme;
+    return 1;
 #else
     return 0;
 #endif

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

@@ -53,7 +53,6 @@
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
-#include <atomic>
 #include <array>
 #include <type_traits>
 
@@ -63,7 +62,7 @@
 #define NOINLINE __attribute__((__noinline__))
 #endif
 
-#if defined(__ARM_NEON) || defined(__AVX512F__)
+#if defined(__ARM_NEON) || defined(__AVX512F__) || defined(__VXE__) || defined(__VXE2__)
 #define VECTOR_REGISTERS 32
 #else
 #define VECTOR_REGISTERS 16
@@ -110,6 +109,12 @@ 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;
@@ -163,6 +168,13 @@ 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
 
@@ -179,6 +191,13 @@ 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__)
@@ -228,6 +247,21 @@ 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);
@@ -394,8 +428,6 @@ 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;
@@ -410,7 +442,7 @@ 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.
-            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
+            ggml_threadpool_chunk_set(params->threadpool, params->nth);
         }
 
         ggml_barrier(params->threadpool);
@@ -439,8 +471,7 @@ class tinyBLAS {
                 GGML_ASSERT(jj == jj2);
             }
 
-            // next step.
-            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
+            job = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
 
         ggml_barrier(params->threadpool);
@@ -3323,6 +3354,14 @@ 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;
@@ -3414,6 +3453,16 @@ 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

@@ -1,6 +1,11 @@
 #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/quants.c

@@ -5,6 +5,8 @@
 #include "ggml-quants.h"
 #include "quants.h"
 
+#include "arch-fallback.h"
+
 #include <string.h>
 #include <assert.h>
 #include <float.h>
@@ -38,12 +40,10 @@ 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
@@ -104,7 +104,6 @@ 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 =================================
 
@@ -143,7 +142,6 @@ 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) {
@@ -181,7 +179,6 @@ 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;
@@ -225,7 +222,6 @@ 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;
@@ -269,7 +265,6 @@ 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;
@@ -300,7 +295,6 @@ 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);
@@ -353,7 +347,6 @@ 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);
@@ -386,7 +379,6 @@ 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);
@@ -439,7 +431,6 @@ 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);
@@ -519,7 +510,6 @@ 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);
@@ -595,7 +585,6 @@ 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);
@@ -676,7 +665,6 @@ 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);
@@ -732,7 +720,6 @@ 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);
@@ -775,7 +762,6 @@ 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);
@@ -826,7 +812,6 @@ 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);
@@ -879,7 +864,6 @@ 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);
@@ -924,7 +908,6 @@ 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);
@@ -981,7 +964,6 @@ 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);
@@ -1025,7 +1007,6 @@ 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);
@@ -1087,7 +1068,6 @@ 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);
@@ -1117,7 +1097,6 @@ 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);
@@ -1164,7 +1143,6 @@ 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

@@ -84,33 +84,6 @@ 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

@@ -8,6 +8,8 @@
 #include "ggml-cpu-impl.h"
 #include "traits.h"
 
+#include "arch-fallback.h"
+
 #include <cmath>
 #include <cstring>
 #include <cassert>
@@ -83,7 +85,6 @@ 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);
@@ -122,7 +123,6 @@ 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);
@@ -174,7 +174,6 @@ 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"
 
@@ -244,7 +243,6 @@ 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;
@@ -289,7 +287,6 @@ 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;
@@ -336,7 +333,6 @@ 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;
@@ -415,7 +411,6 @@ 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;
@@ -462,7 +457,6 @@ 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;
@@ -519,7 +513,6 @@ 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;
@@ -574,7 +567,6 @@ 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;
@@ -629,7 +621,6 @@ 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;
@@ -719,7 +710,6 @@ 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;
@@ -776,7 +766,6 @@ 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"
 

ggml/src/ggml-cpu/repack.h

@@ -64,10 +64,6 @@ 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);
@@ -81,7 +77,6 @@ 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);
@@ -98,22 +93,6 @@ 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

@@ -944,10 +944,8 @@ 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]);        \
     }                                               \
-    res = vec_extract(x[0], 0) +                    \
-          vec_extract(x[0], 1) +                    \
-          vec_extract(x[0], 2) +                    \
-          vec_extract(x[0], 3);                     \
+    float32x4_t tmp = x[0] + vec_reve(x[0]);        \
+    res = tmp[0] + tmp[1];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4

ggml/src/ggml-cuda/common.cuh

@@ -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(RDNA4))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && 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__)
-    return __AMDGCN_WAVEFRONT_SIZE;
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+    return 64;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
 }
 
 [[noreturn]]

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

@@ -2664,7 +2664,9 @@ 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)));
                     }
                 }
-#endif
+#else
+                GGML_UNUSED(integrated);
+#endif // NDEBUG
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {

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

@@ -10,6 +10,8 @@ __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;
@@ -44,16 +46,16 @@ __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 * warpSize + i) * stride_A + wtid];
+            float value = A_block[(wid * warp_size + 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 * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            smem_A[(wid * warp_size + 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 * warpSize + i) * stride_s0 + wtid];
-            smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
+            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;
         }
     }
 

ggml/src/ggml-hip/CMakeLists.txt

@@ -113,6 +113,10 @@ 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

@@ -44,21 +44,22 @@ 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

@@ -498,6 +498,7 @@ 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,
@@ -1454,6 +1455,7 @@ 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);
@@ -1653,6 +1655,7 @@ 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]);
@@ -2400,11 +2403,30 @@ 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 = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                id<MTLComputePipelineState> pipeline = nil;
 
+                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,
@@ -2434,11 +2456,12 @@ static bool ggml_metal_encode_node(
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [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 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 dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_SOFT_MAX:
             {

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

@@ -993,31 +993,61 @@ 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,
-        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;
+        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;
 
     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 row_sum = 0;
+    float sumf = 0;
 
-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
-        row_sum += src_row[i0];
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
     }
 
-    dst_row[0] = row_sum;
+    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;
+    }
 }
 
+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-opencl/CMakeLists.txt

@@ -80,6 +80,7 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
     mul_mv_q6_k
+    mul_mv_id_q4_0_f32_8x_flat
     mul
     norm
     relu

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

@@ -321,6 +321,7 @@ struct ggml_backend_opencl_context {
     cl_program program_upscale;
     cl_program program_concat;
     cl_program program_tsembd;
+    cl_program program_mul_mv_id_q4_0_f32_8x_flat;
 
     cl_kernel kernel_add, kernel_add_row;
     cl_kernel kernel_mul, kernel_mul_row;
@@ -366,6 +367,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_concat_f32_contiguous;
     cl_kernel kernel_concat_f32_non_contiguous;
     cl_kernel kernel_timestep_embedding;
+    cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // Transpose kernels
@@ -1112,7 +1114,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
-        // repeat
+    // repeat
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
         const std::string kernel_src {
@@ -1256,6 +1258,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         }
     }
 
+    // mul_mv_id_q4_0_f32_8x_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q4_0_f32_8x_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q4_0_f32_8x_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat = clCreateKernel(backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat, "kernel_mul_mv_id_q4_0_f32_8x_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -2178,6 +2196,13 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             }
             return false;
+        case GGML_OP_MUL_MAT_ID:
+            if (op->src[0]->type == GGML_TYPE_Q4_0) {
+                if (op->src[1]->type == GGML_TYPE_F32) {
+                    return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+                }
+            }
+            return false;
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
@@ -5536,6 +5561,136 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     }
 }
 
+static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const ggml_tensor * src2 = dst->src[2];
+    GGML_ASSERT(src2);
+    GGML_ASSERT(src2->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+#ifdef GGML_OPENCL_SOA_Q
+    ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+#endif
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+
+    const int ne20 = src2->ne[0];
+    const int ne21 = src2->ne[1];
+
+    const cl_ulong nb21 = src2->nb[1];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+
+    const int r2 = ne12/ne02;
+    const int r3 = ne13/ne03;
+    const int dst_rows = ne20*ne21; // ne20 = n_used_experts, ne21 = n_rows
+
+    GGML_ASSERT(ne00 == ne10);
+
+    int sgs   = 32; // subgroup size
+    int nsg   = 1;  // number of subgroups
+    int nrows = 1;  // number of row in src1
+    int ndst  = 4;  // number of values produced by each subgroup
+
+    cl_kernel kernel;
+
+    // subgroup mat vec
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0: {
+            kernel = backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 1;
+                ndst = 8;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 1;
+                ndst = 8;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &r3));
+
+            break;
+        }
+        default:
+            GGML_ASSERT(false && "not implemented");;
+    }
+
+    int _ne1 = 1;
+    int ne123 = dst_rows;
+
+    size_t global_work_size[] = {(size_t)(ne01+ndst*nsg-1)/(ndst*nsg)*sgs, (size_t)(_ne1+nrows-1)/nrows*nsg, (size_t)ne123};
+    size_t local_work_size[] = {(size_t)sgs, (size_t)nsg, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+#endif
+}
+
 static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -6444,6 +6599,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_mul_mat;
             break;
+        case GGML_OP_MUL_MAT_ID:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mul_mat_id;
+            break;
         case GGML_OP_SCALE:
             if (!any_on_device) {
                 return false;

ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl

@@ -0,0 +1,283 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global char  * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q4_0_f32_8x_flat(
+        global char  *  src0_q,
+        global half  *  src0_d,
+        global float *  src1,
+        ulong           offset1,
+        global char  *  src2,
+        ulong           offset2,
+        global float *  dst,
+        ulong           offsetd,
+        int             ne00,
+        int             ne01,
+        int             ne02,
+        ulong           nb00,
+        ulong           nb02,
+        int             ne10,
+        int             ne11,
+        int             ne12,
+        ulong           nb11,
+        ulong           nb12,
+        int             ne20,
+        int             ne21,
+        ulong           nb21,
+        int             ne0,
+        int             ne1,
+        int             r2,
+        int             r3
+) {
+    src1 = (global float *)((global char *)src1 + offset1);
+    src2 = (global char  *)((global char *)src2 + offset2);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    const int i02 = ((global int *)(src2 + iid1*nb21))[idx];
+
+    const int i11 = idx%ne11;
+    const int i12 = iid1;
+
+    const int i1 = idx;
+    const int i2 = i12;
+
+    global char  * src0_q_cur = src0_q + (i02*nb02/nb00)*(QK4_0/2);
+    global half  * src0_d_cur = src0_d + (i02*nb02/nb00);
+    global float * src1_cur   = (global float *)((global char *) src1  + i11*nb11 + i12*nb12);
+    global float * dst_cur    = dst + i1*ne0 + i2*ne1*ne0;
+
+    mul_vec_q_n_f32_8x_flat(src0_q_cur, src0_d_cur, src1_cur, dst_cur, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-sycl/CMakeLists.txt

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

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> void debug_print_array(const std::string & prefix, const T array[N]) {
+template <int N, class T> std::string debug_get_array_str(const std::string & prefix, const T array[N]) {
     if (LIKELY(!g_ggml_sycl_debug)) {
-        return;
+        return "";
     }
     std::stringstream ss;
     ss << prefix << "=[";
@@ -526,29 +526,26 @@ template <int N, class T> void debug_print_array(const std::string & prefix, con
         ss << array[N - 1];
     }
     ss << "]";
-    GGML_SYCL_DEBUG("%s", ss.str().c_str());
+    return ss.str();
 }
 
-inline void debug_print_tensor(const std::string & prefix, const ggml_tensor * tensor,
-                               const std::string & suffix = "") {
-    if (LIKELY(!g_ggml_sycl_debug)) {
-        return;
-    }
-    GGML_SYCL_DEBUG("%s=", prefix.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) {
-        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");
-        }
+        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 {
-        GGML_SYCL_DEBUG("nullptr");
+        ss << "nullptr";
     }
-    GGML_SYCL_DEBUG("%s", suffix.c_str());
+    ss << suffix;
+    return ss.str();
 }
 
 // Use scope_op_debug_print to log operations coming from running a model
@@ -564,10 +561,10 @@ struct scope_op_debug_print {
             return;
         }
         GGML_SYCL_DEBUG("[SYCL][OP] call %s%s:", func.data(), func_suffix.data());
-        debug_print_tensor(" dst", dst);
+        GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" dst", dst).c_str());
         if (dst) {
             for (std::size_t i = 0; i < num_src; ++i) {
-                debug_print_tensor("\tsrc" + std::to_string(i), dst->src[i]);
+                GGML_SYCL_DEBUG("%s", debug_get_tensor_str("\tsrc" + std::to_string(i), dst->src[i]).c_str());
             }
         }
         GGML_SYCL_DEBUG("%s\n", suffix.data());

ggml/src/ggml-sycl/cpy.cpp

@@ -723,8 +723,7 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const
 
 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,
-                                         std::string(" src0 type=") + ggml_type_name(src0->type));
+    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0, debug_get_tensor_str("\tsrc0", src0));
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 

ggml/src/ggml-sycl/gemm.hpp

@@ -65,6 +65,9 @@ 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

@@ -60,54 +60,6 @@ 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,
@@ -177,47 +129,6 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
     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,
@@ -277,13 +188,8 @@ 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:
-            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());
-            }
+            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__);
-    debug_print_tensor(": tensor=", tensor, "\n");
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
     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__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     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__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     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__);
-    debug_print_tensor(": dst=", dst);
-    debug_print_tensor(" src=", src);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
     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__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     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__);
-    debug_print_tensor(": tensor=", tensor, "\n");
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
     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__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     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__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     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);
@@ -2127,21 +2127,18 @@ 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_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);
+                                      dst_dd_i, DnnlGemmWrapper::to_dt<float>(), 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(
@@ -3866,7 +3863,7 @@ 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__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     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;
@@ -3887,7 +3884,7 @@ 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__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     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;
@@ -3910,8 +3907,8 @@ 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__);
-    debug_print_tensor(": dst=", dst);
-    debug_print_tensor(" src=", src);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         /*

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -49,15 +49,7 @@ if (Vulkan_FOUND)
                              ../../include/ggml-vulkan.h
                             )
 
-    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()
+    set(VULKAN_SHADER_GEN_CMAKE_ARGS "")
 
     # Test all shader extensions
     test_shader_extension_support(
@@ -136,42 +128,45 @@ 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 ${VULKAN_SHADER_GEN_CMAKE_ARGS}
-        BUILD_COMMAND ${CMAKE_COMMAND} --build . ${VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS}
-        INSTALL_COMMAND ${CMAKE_COMMAND} --install .
-        INSTALL_DIR ${CMAKE_BINARY_DIR}
+        CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${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>
     )
-    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_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 ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
+    set (_ggml_vk_genshaders_dir "${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")
 
-    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 ${_ggml_vk_shader_deps} vulkan-shaders-gen-build vulkan-shaders-gen-install)
+    file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp")
 
     add_custom_command(
         OUTPUT ${_ggml_vk_header}
-                ${_ggml_vk_source}
+               ${_ggml_vk_source}
 
         COMMAND ${_ggml_vk_genshaders_cmd}
             --glslc      ${Vulkan_GLSLC_EXECUTABLE}
@@ -181,7 +176,9 @@ if (Vulkan_FOUND)
             --target-cpp ${_ggml_vk_source}
             --no-clean
 
-        DEPENDS ${_ggml_vk_shader_deps}
+        DEPENDS ${_ggml_vk_shader_files}
+                vulkan-shaders-gen
+
         COMMENT "Generate vulkan shaders"
     )
 

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

@@ -78,7 +78,7 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 #define VK_VENDOR_ID_INTEL 0x8086
 #define VK_VENDOR_ID_NVIDIA 0x10de
 
-#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 32
+#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256
 
 #define GGML_VK_MAX_NODES 8192
 
@@ -102,25 +102,11 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 
 struct ggml_backend_vk_context;
 
-struct vk_queue {
-    uint32_t queue_family_index;
-    vk::Queue queue;
-    vk::CommandPool pool;
-    uint32_t cmd_buffer_idx;
-    std::vector<vk::CommandBuffer> cmd_buffers;
-
-    vk::PipelineStageFlags stage_flags;
-
-    bool transfer_only;
-};
+#define MAX_PARAMETER_COUNT 8
 
 struct vk_pipeline_struct {
     std::string name;
     vk::ShaderModule shader_module;
-    vk::DescriptorSetLayout dsl;
-    std::vector<vk::DescriptorPool> descriptor_pools;
-    std::vector<vk::DescriptorSet> descriptor_sets;
-    uint32_t descriptor_set_idx;
     vk::PipelineLayout layout;
     vk::Pipeline pipeline;
     uint32_t push_constant_size;
@@ -167,6 +153,45 @@ struct ggml_backend_vk_buffer_type_context {
     vk_device device;
 };
 
+struct vk_queue;
+
+// Stores command pool/buffers. There's an instance of this
+// for each (context,queue) pair and for each (device,queue) pair.
+struct vk_command_pool {
+    void init(vk_device& device, vk_queue *q_);
+    void destroy(vk::Device& device);
+
+    vk::CommandPool pool;
+    uint32_t cmd_buffer_idx;
+    std::vector<vk::CommandBuffer> cmd_buffers;
+
+    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;
+
+    vk_command_pool cmd_pool;
+
+    vk::PipelineStageFlags stage_flags;
+
+    bool transfer_only;
+
+    // copy everything except the cmd_pool
+    void copyFrom(vk_queue &other) {
+        queue_family_index = other.queue_family_index;
+        queue = other.queue;
+        stage_flags = other.stage_flags;
+        transfer_only = other.transfer_only;
+    }
+};
+
 static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft);
 static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
 static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft);
@@ -341,6 +366,8 @@ struct vk_device_struct {
     // set to true to indicate that some shaders need to be compiled after the dryrun
     bool need_compiles {};
 
+    vk::DescriptorSetLayout dsl;
+
     vk_matmul_pipeline pipeline_matmul_f32 {};
     vk_matmul_pipeline pipeline_matmul_f32_f16 {};
     vk_matmul_pipeline pipeline_matmul_bf16 {};
@@ -458,7 +485,6 @@ struct vk_device_struct {
     vk_pipeline pipeline_flash_attn_split_k_reduce;
 
     std::unordered_map<std::string, vk_pipeline_ref> pipelines;
-    std::unordered_map<std::string, uint64_t> pipeline_descriptor_set_requirements;
 
     std::vector<std::tuple<void*, size_t, vk_buffer>> pinned_memory;
 
@@ -483,10 +509,8 @@ struct vk_device_struct {
 
         ggml_vk_destroy_buffer(sync_staging);
 
-        device.destroyCommandPool(compute_queue.pool);
-        if (!single_queue) {
-            device.destroyCommandPool(transfer_queue.pool);
-        }
+        compute_queue.cmd_pool.destroy(device);
+        transfer_queue.cmd_pool.destroy(device);
 
         for (auto& pipeline : pipelines) {
             if (pipeline.second.expired()) {
@@ -498,10 +522,26 @@ struct vk_device_struct {
         }
         pipelines.clear();
 
+        device.destroyDescriptorSetLayout(dsl);
+
         device.destroy();
     }
 };
 
+void vk_command_pool::init(vk_device& device, vk_queue *q_) {
+    cmd_buffer_idx = 0;
+    q = q_;
+
+    vk::CommandPoolCreateInfo command_pool_create_info(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), q->queue_family_index);
+    pool = device->device.createCommandPool(command_pool_create_info);
+}
+
+void vk_command_pool::destroy(vk::Device& device) {
+    device.destroyCommandPool(pool);
+    pool = nullptr;
+    cmd_buffers.clear();
+}
+
 struct vk_buffer_struct {
     vk::Buffer buffer = VK_NULL_HANDLE;
     vk::DeviceMemory device_memory = VK_NULL_HANDLE;
@@ -819,7 +859,7 @@ struct vk_context_struct {
     std::vector<vk_staging_memcpy> in_memcpys;
     std::vector<vk_staging_memcpy> out_memcpys;
 
-    vk_queue * q;
+    vk_command_pool * p {};
 };
 typedef std::shared_ptr<vk_context_struct> vk_context;
 typedef std::weak_ptr<vk_context_struct> vk_context_ref;
@@ -930,6 +970,14 @@ struct ggml_backend_vk_context {
     vk_context_ref transfer_ctx;
 
     std::vector<vk_context_ref> tensor_ctxs;
+
+    std::vector<vk::DescriptorPool> descriptor_pools;
+    std::vector<vk::DescriptorSet> descriptor_sets;
+    uint32_t descriptor_set_idx {};
+    uint32_t pipeline_descriptor_set_requirements {};
+
+    vk_command_pool compute_cmd_pool;
+    vk_command_pool transfer_cmd_pool;
 };
 
 static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000;  // NOLINT
@@ -1060,39 +1108,19 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
                  ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " <<
                  disable_robustness << ", " << require_full_subgroups << ", " << required_subgroup_size << ")");
     GGML_ASSERT(parameter_count > 0);
+    GGML_ASSERT(parameter_count <= MAX_PARAMETER_COUNT);
     GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
 
     vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast<const uint32_t *>(spv_data));
     pipeline->shader_module = device->device.createShaderModule(shader_module_create_info);
 
-    std::vector<vk::DescriptorSetLayoutBinding> dsl_binding;
-    std::vector<vk::DescriptorBindingFlags> dsl_binding_flags;
-    for (uint32_t i = 0; i < parameter_count; i++) {
-        dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute});
-        dsl_binding_flags.push_back({});
-    }
-
-    vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags };
-
     vk::PushConstantRange pcr(
         vk::ShaderStageFlagBits::eCompute,
         0,
         pipeline->push_constant_size
     );
 
-    vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
-        {},
-        dsl_binding);
-    descriptor_set_layout_create_info.setPNext(&dslbfci);
-    pipeline->dsl = device->device.createDescriptorSetLayout(descriptor_set_layout_create_info);
-
-    vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
-    vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
-    pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
-
-    pipeline->descriptor_set_idx = 0;
-
-    vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), pipeline->dsl, pcr);
+    vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), device->dsl, pcr);
     pipeline->layout = device->device.createPipelineLayout(pipeline_layout_create_info);
 
     std::vector<vk::SpecializationMapEntry> specialization_entries(specialization_constants.size());
@@ -1167,15 +1195,6 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
 
 static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) {
     VK_LOG_DEBUG("ggml_pipeline_destroy_pipeline(" << pipeline->name << ")");
-    for (auto& pool : pipeline->descriptor_pools) {
-        device.destroyDescriptorPool(pool);
-    }
-    pipeline->descriptor_pools.clear();
-    pipeline->descriptor_sets.clear();
-    pipeline->descriptor_set_idx = 0;
-
-    device.destroyDescriptorSetLayout(pipeline->dsl);
-
     device.destroyPipelineLayout(pipeline->layout);
 
     device.destroyShaderModule(pipeline->shader_module);
@@ -1183,97 +1202,77 @@ static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline)
     device.destroyPipeline(pipeline->pipeline);
 }
 
-static void ggml_pipeline_request_descriptor_sets(vk_device& device, vk_pipeline& pipeline, uint32_t n) {
+static void ggml_pipeline_request_descriptor_sets(ggml_backend_vk_context *ctx, vk_pipeline& pipeline, uint32_t n) {
     VK_LOG_DEBUG("ggml_pipeline_request_descriptor_sets(" << pipeline->name << ", " << n << ")");
-    device->pipeline_descriptor_set_requirements[pipeline->name] += n;
+    ctx->pipeline_descriptor_set_requirements += n;
     if (!pipeline->compiled) {
         pipeline->needed = true;
-        device->need_compiles = true;
+        ctx->device->need_compiles = true;
     }
 }
 
-static void ggml_pipeline_allocate_descriptor_sets(vk_device& device) {
-    std::lock_guard<std::mutex> guard(device->mutex);
+static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx) {
 
-    for (auto& pair : device->pipeline_descriptor_set_requirements) {
-        vk_pipeline pipeline = device->pipelines.at(pair.first).lock();
-        const uint64_t n = pair.second;
+    if (ctx->descriptor_sets.size() >= ctx->pipeline_descriptor_set_requirements) {
+        // Enough descriptors are available
+        return;
+    }
 
-        VK_LOG_DEBUG("ggml_pipeline_allocate_descriptor_sets(" << pipeline->name << ", " << n << ")");
+    vk_device& device = ctx->device;
 
-        if (pipeline->descriptor_sets.size() >= pipeline->descriptor_set_idx + n) {
-            // Enough descriptors are available
-            continue;
+    uint32_t to_alloc = ctx->pipeline_descriptor_set_requirements - ctx->descriptor_sets.size();
+    uint32_t pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE - ctx->descriptor_sets.size() % VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+    uint32_t pool_idx = ctx->descriptor_sets.size() / VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+
+    while (to_alloc > 0) {
+        const uint32_t alloc_count = std::min(pool_remaining, to_alloc);
+        to_alloc -= alloc_count;
+        pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+
+        if (pool_idx >= ctx->descriptor_pools.size()) {
+            vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, MAX_PARAMETER_COUNT * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
+            vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
+            ctx->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
         }
 
-        uint32_t to_alloc = pipeline->descriptor_set_idx + n - pipeline->descriptor_sets.size();
-        uint32_t pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE - pipeline->descriptor_sets.size() % VK_DEVICE_DESCRIPTOR_POOL_SIZE;
-        uint32_t pool_idx = pipeline->descriptor_sets.size() / VK_DEVICE_DESCRIPTOR_POOL_SIZE;
-
-        while (to_alloc > 0) {
-            const uint32_t alloc_count = std::min(pool_remaining, to_alloc);
-            to_alloc -= alloc_count;
-            pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE;
-
-            if (pool_idx >= pipeline->descriptor_pools.size()) {
-                vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
-                vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
-                pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
-            }
-
-            std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
-            for (uint32_t i = 0; i < alloc_count; i++) {
-                layouts[i] = pipeline->dsl;
-            }
-            vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[pool_idx], alloc_count, layouts.data());
-            std::vector<vk::DescriptorSet> sets = device->device.allocateDescriptorSets(descriptor_set_alloc_info);
-            pipeline->descriptor_sets.insert(pipeline->descriptor_sets.end(), sets.begin(), sets.end());
-
-            pool_idx++;
+        std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
+        for (uint32_t i = 0; i < alloc_count; i++) {
+            layouts[i] = device->dsl;
         }
+        vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(ctx->descriptor_pools[pool_idx], alloc_count, layouts.data());
+        std::vector<vk::DescriptorSet> sets = device->device.allocateDescriptorSets(descriptor_set_alloc_info);
+        ctx->descriptor_sets.insert(ctx->descriptor_sets.end(), sets.begin(), sets.end());
+
+        pool_idx++;
     }
 }
 
-static void ggml_pipeline_cleanup(vk_pipeline& pipeline) {
-    VK_LOG_DEBUG("ggml_pipeline_cleanup(" << pipeline->name << ")");
-    pipeline->descriptor_set_idx = 0;
-}
-
-static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_queue& q) {
+static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_command_pool& p) {
     VK_LOG_DEBUG("ggml_vk_create_cmd_buffer()");
-    std::lock_guard<std::mutex> guard(device->mutex);
 
-    if (q.cmd_buffers.size() > q.cmd_buffer_idx) {
+    if (p.cmd_buffers.size() > p.cmd_buffer_idx) {
         // Reuse command buffer
-        return q.cmd_buffers[q.cmd_buffer_idx++];
+        return p.cmd_buffers[p.cmd_buffer_idx++];
     }
 
     vk::CommandBufferAllocateInfo command_buffer_alloc_info(
-        q.pool,
+        p.pool,
         vk::CommandBufferLevel::ePrimary,
         1);
     const std::vector<vk::CommandBuffer> cmd_buffers = device->device.allocateCommandBuffers(command_buffer_alloc_info);
     auto buf = cmd_buffers.front();
 
-    q.cmd_buffers.push_back(buf);
-    q.cmd_buffer_idx++;
+    p.cmd_buffers.push_back(buf);
+    p.cmd_buffer_idx++;
 
     return buf;
 }
 
-static vk_submission ggml_vk_create_submission(vk_device& device, vk_queue& q, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) {
-    VK_LOG_DEBUG("ggml_vk_create_submission()");
-    vk_submission s;
-    s.buffer = ggml_vk_create_cmd_buffer(device, q);
-    s.wait_semaphores = std::move(wait_semaphores);
-    s.signal_semaphores = std::move(signal_semaphores);
-    return s;
-}
-
 static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
     if (ctx->seqs.empty()) {
         if (fence) {
-            ctx->q->queue.submit({}, fence);
+            std::lock_guard<std::mutex> guard(queue_mutex);
+            ctx->p->q->queue.submit({}, fence);
         }
         return;
     }
@@ -1312,7 +1311,7 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
             tl_signal_vals.push_back({});
             tl_signal_semaphores.push_back({});
             for (size_t i = 0; i < submission.wait_semaphores.size(); i++) {
-                stage_flags[idx].push_back(ctx->q->stage_flags);
+                stage_flags[idx].push_back(ctx->p->q->stage_flags);
                 tl_wait_vals[idx].push_back(submission.wait_semaphores[i].value);
                 tl_wait_semaphores[idx].push_back(submission.wait_semaphores[i].s);
             }
@@ -1342,7 +1341,8 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
         }
     }
 
-    ctx->q->queue.submit(submit_infos, fence);
+    std::lock_guard<std::mutex> guard(queue_mutex);
+    ctx->p->q->queue.submit(submit_infos, fence);
 
     ctx->seqs.clear();
 }
@@ -1400,28 +1400,25 @@ static void ggml_vk_create_queue(vk_device& device, vk_queue& q, uint32_t queue_
     q.queue_family_index = queue_family_index;
     q.transfer_only = transfer_only;
 
-    vk::CommandPoolCreateInfo command_pool_create_info_compute(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), queue_family_index);
-    q.pool = device->device.createCommandPool(command_pool_create_info_compute);
-
-    q.cmd_buffer_idx = 0;
+    q.cmd_pool.init(device, &q);
 
     q.queue = device->device.getQueue(queue_family_index, queue_index);
 
     q.stage_flags = stage_flags;
 }
 
-static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) {
+static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_command_pool& p) {
     vk_context result = std::make_shared<vk_context_struct>();
     VK_LOG_DEBUG("ggml_vk_create_context(" << result << ")");
     ctx->gc.contexts.emplace_back(result);
-    result->q = &q;
+    result->p = &p;
     return result;
 }
 
-static vk_context ggml_vk_create_temporary_context(vk_queue& q) {
+static vk_context ggml_vk_create_temporary_context(vk_command_pool& p) {
     vk_context result = std::make_shared<vk_context_struct>();
     VK_LOG_DEBUG("ggml_vk_create_temporary_context(" << result << ")");
-    result->q = &q;
+    result->p = &p;
     return result;
 }
 
@@ -1454,15 +1451,29 @@ static vk::Event ggml_vk_create_event(ggml_backend_vk_context * ctx) {
     return ctx->gc.events[ctx->event_idx++];
 }
 
-static void ggml_vk_queue_cleanup(vk_device& device, vk_queue& q) {
-    VK_LOG_DEBUG("ggml_vk_queue_cleanup()");
-    std::lock_guard<std::mutex> guard(device->mutex);
+static void ggml_vk_command_pool_cleanup(vk_device& device, vk_command_pool& p) {
+    VK_LOG_DEBUG("ggml_vk_command_pool_cleanup()");
 
     // Requires command buffers to be done
-    device->device.resetCommandPool(q.pool);
-    q.cmd_buffer_idx = 0;
+    device->device.resetCommandPool(p.pool);
+    p.cmd_buffer_idx = 0;
 }
 
+static void ggml_vk_queue_command_pools_cleanup(vk_device& device) {
+    VK_LOG_DEBUG("ggml_vk_queue_command_pools_cleanup()");
+
+    // Arbitrary frequency to cleanup/reuse command buffers
+    static constexpr uint32_t cleanup_frequency = 10;
+
+    if (device->compute_queue.cmd_pool.cmd_buffer_idx >= cleanup_frequency) {
+        ggml_vk_command_pool_cleanup(device, device->compute_queue.cmd_pool);
+    }
+    if (device->transfer_queue.cmd_pool.cmd_buffer_idx >= cleanup_frequency) {
+        ggml_vk_command_pool_cleanup(device, device->transfer_queue.cmd_pool);
+    }
+}
+
+
 static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
     for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
         vk::MemoryType memory_type = mem_props->memoryTypes[i];
@@ -1481,8 +1492,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
         throw vk::OutOfDeviceMemoryError("Requested buffer size exceeds device memory allocation limit");
     }
 
-    std::lock_guard<std::mutex> guard(device->mutex);
-
     vk_buffer buf = std::make_shared<vk_buffer_struct>();
 
     if (size == 0) {
@@ -1611,11 +1620,11 @@ static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) {
 static void ggml_vk_sync_buffers(vk_context& ctx) {
     VK_LOG_DEBUG("ggml_vk_sync_buffers()");
 
-    const bool transfer_queue = ctx->q->transfer_only;
+    const bool transfer_queue = ctx->p->q->transfer_only;
 
     ctx->s->buffer.pipelineBarrier(
-        ctx->q->stage_flags,
-        ctx->q->stage_flags,
+        ctx->p->q->stage_flags,
+        ctx->p->q->stage_flags,
         {},
         { {
           { !transfer_queue ? (vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) : (vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) },
@@ -1634,8 +1643,8 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
 
     ctx->s->buffer.waitEvents(
         events,
-        ctx->q->stage_flags,
-        ctx->q->stage_flags,
+        ctx->p->q->stage_flags,
+        ctx->p->q->stage_flags,
         {},
         {},
         {}
@@ -3369,6 +3378,22 @@ static vk_device ggml_vk_get_device(size_t idx) {
             }
         }
 
+
+        std::vector<vk::DescriptorSetLayoutBinding> dsl_binding;
+        std::vector<vk::DescriptorBindingFlags> dsl_binding_flags;
+        for (uint32_t i = 0; i < MAX_PARAMETER_COUNT; i++) {
+            dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute});
+            dsl_binding_flags.push_back({});
+        }
+
+        vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags };
+
+        vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
+            {},
+            dsl_binding);
+        descriptor_set_layout_create_info.setPNext(&dslbfci);
+        device->dsl = device->device.createDescriptorSetLayout(descriptor_set_layout_create_info);
+
         ggml_vk_load_shaders(device);
 
         if (!device->single_queue) {
@@ -3376,7 +3401,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
             ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true);
         } else {
             // TODO: Use pointer or reference to avoid copy
-            device->transfer_queue = device->compute_queue;
+            device->transfer_queue.copyFrom(device->compute_queue);
+            device->transfer_queue.cmd_pool.init(device, &device->transfer_queue);
         }
 
         device->buffer_type = {
@@ -3742,6 +3768,9 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->fence = ctx->device->device.createFence({});
     ctx->almost_ready_fence = ctx->device->device.createFence({});
 
+    ctx->compute_cmd_pool.init(ctx->device, &ctx->device->compute_queue);
+    ctx->transfer_cmd_pool.init(ctx->device, &ctx->device->transfer_queue);
+
 #ifdef GGML_VULKAN_CHECK_RESULTS
     const char* skip_checks = getenv("GGML_VULKAN_SKIP_CHECKS");
     vk_skip_checks = (skip_checks == NULL ? 0 : atoi(skip_checks));
@@ -4107,9 +4136,9 @@ static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf
     }
 }
 
-static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bool one_time = true) {
+static vk_submission ggml_vk_begin_submission(vk_device& device, vk_command_pool& p, bool one_time = true) {
     vk_submission s;
-    s.buffer = ggml_vk_create_cmd_buffer(device, q);
+    s.buffer = ggml_vk_create_cmd_buffer(device, p);
     if (one_time) {
         s.buffer.begin({ vk::CommandBufferUsageFlagBits::eOneTimeSubmit });
     } else {
@@ -4154,10 +4183,10 @@ static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context&
         std::cerr << "(" << buffer.buffer << ", " << buffer.offset << ", " << buffer.range << "), ";
     }
     std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))");
-    GGML_ASSERT(pipeline->descriptor_set_idx < pipeline->descriptor_sets.size());
-    GGML_ASSERT(descriptor_buffer_infos.size() == pipeline->parameter_count);
+    GGML_ASSERT(ctx->descriptor_set_idx < ctx->descriptor_sets.size());
+    GGML_ASSERT(descriptor_buffer_infos.size() <= MAX_PARAMETER_COUNT);
 
-    vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->descriptor_set_idx++];
+    vk::DescriptorSet& descriptor_set = ctx->descriptor_sets[ctx->descriptor_set_idx++];
     vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
     ctx->device->device.updateDescriptorSets({ write_descriptor_set }, {});
 
@@ -4194,7 +4223,7 @@ static void ggml_vk_ctx_begin(vk_device& device, vk_context& subctx) {
         ggml_vk_ctx_end(subctx);
     }
 
-    subctx->seqs.push_back({ ggml_vk_begin_submission(device, *subctx->q) });
+    subctx->seqs.push_back({ ggml_vk_begin_submission(device, *subctx->p) });
     subctx->s = subctx->seqs[subctx->seqs.size() - 1].data();
 }
 
@@ -4395,7 +4424,9 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
             memcpy((uint8_t *)dst->ptr + offset + i * width, (const uint8_t *) src + i * spitch, width);
         }
     } else {
-        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
+        std::lock_guard<std::mutex> guard(dst->device->mutex);
+
+        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool);
         ggml_vk_ctx_begin(dst->device, subctx);
         ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, spitch, width, height, true);
         ggml_vk_ctx_end(subctx);
@@ -4407,6 +4438,7 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
         ggml_vk_submit(subctx, dst->device->fence);
         VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
         dst->device->device.resetFences({ dst->device->fence });
+        ggml_vk_queue_command_pools_cleanup(dst->device);
     }
 }
 
@@ -4483,7 +4515,9 @@ static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_
 
         memcpy(dst, (uint8_t *) src->ptr + offset, size);
     } else {
-        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
+        std::lock_guard<std::mutex> guard(src->device->mutex);
+
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool);
         ggml_vk_ctx_begin(src->device, subctx);
         ggml_vk_buffer_read_async(subctx, src, offset, dst, size, true);
         ggml_vk_ctx_end(subctx);
@@ -4491,6 +4525,7 @@ static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_
         ggml_vk_submit(subctx, src->device->fence);
         VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read waitForFences");
         src->device->device.resetFences({ src->device->fence });
+        ggml_vk_queue_command_pools_cleanup(src->device);
 
         for (auto& cpy : subctx->out_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
@@ -4510,15 +4545,17 @@ static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t ds
 
 static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
     if (src->device == dst->device) {
+        std::lock_guard<std::mutex> guard(src->device->mutex);
         VK_LOG_DEBUG("ggml_vk_buffer_copy(SINGLE_DEVICE, " << size << ")");
         // Copy within the device
-        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool);
         ggml_vk_ctx_begin(src->device, subctx);
         ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size);
         ggml_vk_ctx_end(subctx);
         ggml_vk_submit(subctx, src->device->fence);
         VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences");
         src->device->device.resetFences({ src->device->fence });
+        ggml_vk_queue_command_pools_cleanup(src->device);
     } else {
         VK_LOG_DEBUG("ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")");
         // Copy device to device
@@ -4543,7 +4580,8 @@ static void ggml_vk_buffer_memset_async(vk_context& ctx, vk_buffer& dst, size_t
 static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
     VK_LOG_DEBUG("ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")");
 
-    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
+    std::lock_guard<std::mutex> guard(dst->device->mutex);
+    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool);
     ggml_vk_ctx_begin(dst->device, subctx);
     subctx->s->buffer.fillBuffer(dst->buffer, offset, size, c);
     ggml_vk_ctx_end(subctx);
@@ -4551,6 +4589,7 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
     ggml_vk_submit(subctx, dst->device->fence);
     VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences");
     dst->device->device.resetFences({ dst->device->fence });
+    ggml_vk_queue_command_pools_cleanup(dst->device);
 }
 
 static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int n, int k, const vk_pipeline& pipeline) {
@@ -4964,18 +5003,18 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         }
 
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (qx_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
         }
         if (qy_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
         if (quantize_y) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_q8_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
         }
         if (split_k > 1) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, 1);
         }
         return;
     }
@@ -5157,12 +5196,12 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
         // Request descriptor sets
         if (qx_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
         }
         if (qy_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
-        ggml_pipeline_request_descriptor_sets(ctx->device, dmmv, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
         return;
     }
 
@@ -5295,7 +5334,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
 
     if (dryrun) {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1);
         return;
     }
 
@@ -5384,7 +5423,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     if (dryrun) {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1);
         return;
     }
 
@@ -5571,12 +5610,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         }
 
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (qx_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
         }
         if (qy_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
         return;
     }
@@ -5765,12 +5804,12 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
         // Request descriptor sets
         if (qx_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
         }
         if (qy_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
-        ggml_pipeline_request_descriptor_sets(ctx->device, dmmv, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
         return;
     }
 
@@ -6090,9 +6129,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     if (dryrun) {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (split_k > 1) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
         }
         return;
     }
@@ -6655,7 +6694,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     }
 
     if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         return;
     }
 
@@ -7036,7 +7075,7 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_ASSERT(pipeline != nullptr);
 
     if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         return;
     }
 
@@ -7175,7 +7214,7 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
     GGML_ASSERT(pipeline != nullptr);
 
     if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         return;
     }
 
@@ -7853,9 +7892,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         }
     }
 
-    ggml_pipeline_request_descriptor_sets(ctx->device, p, num_it);
+    ggml_pipeline_request_descriptor_sets(ctx, p, num_it);
     if (split_k > 1) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, num_it);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it);
 
         if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) {
             // Resize buffer
@@ -7870,7 +7909,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         ggml_vk_load_shaders(ctx->device);
     }
 
-    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 
     vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
     vk_buffer d_Y = ggml_vk_create_buffer_check(ctx->device, sizeof(Y_TYPE) * y_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
@@ -7912,7 +7951,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     ggml_vk_buffer_write(d_X, 0, x, sizeof(X_TYPE) * k * m * batch);
     ggml_vk_buffer_write(d_Y, 0, y, sizeof(Y_TYPE) * k * n * batch);
 
-    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
     ggml_vk_ctx_begin(ctx->device, subctx);
     for (size_t i = 0; i < num_it; i++) {
         ggml_vk_matmul(
@@ -7928,6 +7967,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     ggml_vk_submit(subctx, ctx->fence);
     VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences");
     ctx->device->device.resetFences({ ctx->fence });
+    ggml_vk_queue_command_pools_cleanup(ctx->device);
 
     auto end = std::chrono::high_resolution_clock::now();
     double time = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0;
@@ -8029,16 +8069,13 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
 
     free(d_chk);
 
-    ggml_vk_queue_cleanup(ctx->device, ctx->device->transfer_queue);
-    ggml_vk_queue_cleanup(ctx->device, ctx->device->compute_queue);
+    ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
+    ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
 
     ggml_vk_destroy_buffer(d_X);
     ggml_vk_destroy_buffer(d_Y);
     ggml_vk_destroy_buffer(d_D);
 
-    ggml_pipeline_cleanup(p);
-    ggml_pipeline_cleanup(ctx->device->pipeline_matmul_split_k_reduce);
-
     free(x);
     free(y);
     free(d);
@@ -8116,17 +8153,17 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
     ggml_vk_quantize_data(x, qx, ne, quant);
     ggml_vk_dequantize_data(qx, x_ref, ne, quant);
 
-    ggml_pipeline_request_descriptor_sets(ctx->device, p, 1);
+    ggml_pipeline_request_descriptor_sets(ctx, p, 1);
 
     if (ctx->device->need_compiles) {
         ggml_vk_load_shaders(ctx->device);
     }
 
-    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
 
-    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
     ggml_vk_ctx_begin(ctx->device, subctx);
     const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne };
     ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc, { (uint32_t)ne, 1, 1});
@@ -8137,6 +8174,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
     ggml_vk_submit(subctx, ctx->fence);
     VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
     ctx->device->device.resetFences({ ctx->fence });
+    ggml_vk_queue_command_pools_cleanup(ctx->device);
 
     auto end = std::chrono::high_resolution_clock::now();
 
@@ -8216,17 +8254,17 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 //
 //     vk_pipeline p = ggml_vk_get_quantize_pipeline(ctx, quant);
 //
-//     ggml_pipeline_request_descriptor_sets(ctx->device, p, 1);
+//     ggml_pipeline_request_descriptor_sets(ctx, p, 1);
 //
 //     if (ctx->device->need_compiles) {
 //         ggml_vk_load_shaders(ctx->device);
 //     }
 //
-//     ggml_pipeline_allocate_descriptor_sets(ctx->device);
+//     ggml_pipeline_allocate_descriptor_sets(ctx);
 //
 //     ggml_vk_buffer_write(x_buf, 0, x, x_sz);
 //
-//     vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+//     vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
 //     ggml_vk_ctx_begin(ctx->device, subctx);
 //     ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(x_buf), ggml_vk_subbuffer(qx_buf), ne);
 //     ggml_vk_ctx_end(subctx);
@@ -8236,6 +8274,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 //     ggml_vk_submit(subctx, ctx->fence);
 //     VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_quantize waitForFences");
 //     ctx->device->device.resetFences({ ctx->fence });
+//     ggml_vk_queue_command_pools_cleanup(ctx->device);
 //
 //     auto end = std::chrono::high_resolution_clock::now();
 //
@@ -8375,9 +8414,9 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         // y[i] = i % k;
     }
 
-    ggml_pipeline_request_descriptor_sets(ctx->device, p, num_it);
+    ggml_pipeline_request_descriptor_sets(ctx, p, num_it);
     if (split_k > 1) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, num_it);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it);
 
         if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) {
             // Resize buffer
@@ -8388,19 +8427,19 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         }
     }
     if (mmq) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_quantize_q8_1, num_it);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_quantize_q8_1, num_it);
     }
 
     if (ctx->device->need_compiles) {
         ggml_vk_load_shaders(ctx->device);
     }
 
-    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
     ggml_vk_buffer_write(y_buf, 0, y, y_sz);
 
-    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
     ggml_vk_ctx_begin(ctx->device, subctx);
     if (mmq) {
         for (size_t i = 0; i < num_it; i++) {
@@ -8429,6 +8468,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     ggml_vk_submit(subctx, ctx->fence);
     VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
     ctx->device->device.resetFences({ ctx->fence });
+    ggml_vk_queue_command_pools_cleanup(ctx->device);
 
     auto end = std::chrono::high_resolution_clock::now();
 
@@ -8743,7 +8783,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
     if (!dryrun) {
         if (ctx->compute_ctx.expired()) {
-            compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+            compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
             ctx->compute_ctx = compute_ctx;
             ggml_vk_ctx_begin(ctx->device, compute_ctx);
         } else {
@@ -8797,7 +8837,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
                 // These operations all go through ggml_vk_op_f32, so short-circuit and
                 // do the only thing needed for the dryrun.
                 vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op);
-                ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+                ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
                 return false;
             }
         default:
@@ -9189,19 +9229,8 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
     }
     ctx->gc.temp_buffers.clear();
 
-    for (auto& dsr : ctx->device->pipeline_descriptor_set_requirements) {
-        vk_pipeline_ref plr = ctx->device->pipelines[dsr.first];
-
-        if (plr.expired()) {
-            continue;
-        }
-
-        vk_pipeline pl = plr.lock();
-        ggml_pipeline_cleanup(pl);
-    }
-
-    ggml_vk_queue_cleanup(ctx->device, ctx->device->compute_queue);
-    ggml_vk_queue_cleanup(ctx->device, ctx->device->transfer_queue);
+    ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
+    ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
 
     for (size_t i = 0; i < ctx->gc.semaphores.size(); i++) {
         ctx->device->device.destroySemaphore({ ctx->gc.semaphores[i].s });
@@ -9222,7 +9251,8 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
 
     ctx->tensor_ctxs.clear();
     ctx->gc.contexts.clear();
-    ctx->device->pipeline_descriptor_set_requirements.clear();
+    ctx->pipeline_descriptor_set_requirements = 0;
+    ctx->descriptor_set_idx = 0;
 }
 
 // Clean up on backend free
@@ -9249,6 +9279,15 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
 
     ctx->device->device.destroyFence(ctx->fence);
     ctx->device->device.destroyFence(ctx->almost_ready_fence);
+
+    for (auto& pool : ctx->descriptor_pools) {
+        ctx->device->device.destroyDescriptorPool(pool);
+    }
+    ctx->descriptor_pools.clear();
+    ctx->descriptor_sets.clear();
+
+    ctx->compute_cmd_pool.destroy(ctx->device->device);
+    ctx->transfer_cmd_pool.destroy(ctx->device->device);
 }
 
 static int ggml_vk_get_device_count() {
@@ -9456,6 +9495,12 @@ 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() {
@@ -9464,7 +9509,7 @@ 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     = */ NULL, // defaults to SIZE_MAX
+            /* .get_max_size     = */ ggml_backend_vk_host_buffer_type_get_max_size,
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
@@ -9515,7 +9560,7 @@ static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor
 
     if (ctx->transfer_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
         ctx->transfer_ctx = transfer_ctx;
         ggml_vk_ctx_begin(ctx->device, transfer_ctx);
     } else {
@@ -9538,7 +9583,7 @@ static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_
 
     if (ctx->transfer_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
         ctx->transfer_ctx = transfer_ctx;
         ggml_vk_ctx_begin(ctx->device, transfer_ctx);
     } else {
@@ -9561,7 +9606,7 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_
 
         if (ctx->transfer_ctx.expired()) {
             // Initialize new transfer context
-            transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+            transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
             ctx->transfer_ctx = transfer_ctx;
             ggml_vk_ctx_begin(ctx->device, transfer_ctx);
         } else {
@@ -9622,7 +9667,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ggml_vk_load_shaders(ctx->device);
     }
     ggml_vk_preallocate_buffers(ctx);
-    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 
     int last_node = cgraph->n_nodes - 1;
 
@@ -9654,7 +9699,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ctx->device->device.resetQueryPool(ctx->device->query_pool, 0, cgraph->n_nodes+1);
 
         GGML_ASSERT(ctx->compute_ctx.expired());
-        compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
         ctx->compute_ctx = compute_ctx;
         ggml_vk_ctx_begin(ctx->device, compute_ctx);
         compute_ctx->s->buffer.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->device->query_pool, 0);
@@ -9689,7 +9734,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
         if (vk_perf_logger_enabled) {
             if (ctx->compute_ctx.expired()) {
-                compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+                compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
                 ctx->compute_ctx = compute_ctx;
                 ggml_vk_ctx_begin(ctx->device, compute_ctx);
             } else {

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

@@ -25,15 +25,3 @@ 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

@@ -888,12 +888,6 @@ struct ggml_context {
     struct ggml_object * objects_end;
 };
 
-struct ggml_context_container {
-    bool used;
-
-    struct ggml_context context;
-};
-
 //
 // data types
 //

gguf-py/gguf/constants.py

@@ -291,6 +291,7 @@ class MODEL_ARCH(IntEnum):
     BERT             = auto()
     NOMIC_BERT       = auto()
     NOMIC_BERT_MOE   = auto()
+    NEO_BERT         = auto()
     JINA_BERT_V2     = auto()
     BLOOM            = auto()
     STABLELM         = auto()
@@ -343,6 +344,8 @@ class MODEL_ARCH(IntEnum):
     WAVTOKENIZER_DEC = auto()
     PLM              = auto()
     BAILINGMOE       = auto()
+    DOTS1            = auto()
+    ARCEE            = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -571,6 +574,7 @@ 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",
@@ -623,6 +627,8 @@ 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] = {
@@ -1077,6 +1083,18 @@ 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,
@@ -2044,6 +2062,45 @@ 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):
-            raise ValueError(f'Duplicated key name {key!r}')
+            logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}')
 
         self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)
 

gguf-py/gguf/tensor_mapping.py

@@ -31,6 +31,7 @@ class TensorNameMap:
             "model.embeddings",                          # rwkv7
             "model.word_embeddings",                     # bailingmoe
             "language_model.model.embed_tokens",         # llama4
+            "encoder",                                   # neobert
         ),
 
         # Token type embeddings
@@ -134,6 +135,7 @@ 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
@@ -161,6 +163,7 @@ 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
@@ -236,6 +239,7 @@ 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
@@ -276,6 +280,7 @@ 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
@@ -305,7 +310,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
-            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
+            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
         ),
 
         # Feed-forward up
@@ -333,11 +338,14 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "encoder.layers.{bid}.mlp.fc1",                           # nomic-bert-moe
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
-            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
+            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2 (split up/gate, no longer used)
+            "encoder.layer.{bid}.mlp.gated_layers",                   # jina-bert-v2 (GEGLU)
+            "encoder.layer.{bid}.mlp.up_gated_layer",                 # jina-v2-code (GEGLU)
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
             "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: (
@@ -370,7 +378,7 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
             "model.layers.{bid}.feed_forward.w1",         # internlm2
             "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
-            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2
+            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2 (split up/gate, no longer used)
             "transformer.h.{bid}.mlp.linear_1",           # refact
             "model.layers.{bid}.residual_mlp.w1",         # arctic
             "transformer.h.{bid}.mlp.c_fc_0",             # exaone
@@ -420,6 +428,7 @@ 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: (
@@ -830,12 +839,14 @@ 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

@@ -7,7 +7,10 @@ import os
 from pathlib import Path
 from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable
 
-from sentencepiece import SentencePieceProcessor
+try:
+    from sentencepiece import SentencePieceProcessor
+except ImportError:
+    SentencePieceProcessor = None
 
 import gguf
 
@@ -302,6 +305,9 @@ 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 = "0.17.0"
+version = "0.17.1"
 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 = ">=0.1.98,<=0.2.0"
+sentencepiece = { version = ">=0.1.98,<=0.2.0", optional = true }
 PySide6 = { version = "^6.9", python = ">=3.9,<3.14", optional = true }
 
 [tool.poetry.dev-dependencies]

include/llama.h

@@ -243,18 +243,21 @@ extern "C" {
 
     typedef bool (*llama_progress_callback)(float progress, void * user_data);
 
-    // Input data for llama_decode
+    // Input data for llama_encode/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_decode)
+    //            (if set to NULL, the token position will be tracked automatically by llama_encode/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, only the logits for last token will be returned)
+    //            (if set to NULL:
+    //               - if embeddings: all tokens are output
+    //               - if not:        only the last token is output
+    //            )
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -262,8 +265,8 @@ extern "C" {
         llama_token  *  token;
         float        *  embd;
         llama_pos    *  pos;
-        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;
+        int32_t      *  n_seq_id;
+        llama_seq_id ** seq_id;
         int8_t       *  logits;   // TODO: rename this to "output"
     } llama_batch;
 
@@ -961,8 +964,8 @@ 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 model is in embeddings mode or not
-    // If true, embeddings will be returned but logits will not
+    // Set whether the context outputs embeddings or not
+    // TODO: rename to avoid confusion with llama_get_embeddings()
     LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
 
     // Set whether to use causal attention or not

requirements/requirements-compare-llama-bench.txt

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

scripts/compare-llama-bench.py

@@ -19,6 +19,7 @@ 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
@@ -122,11 +123,15 @@ 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)
@@ -499,7 +504,6 @@ 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(",")
@@ -544,6 +548,14 @@ 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:
@@ -600,6 +612,161 @@ 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 @@
-6a7d170c04789f6ebcf320ed03c1b16973f93bd7
+8cda0a3c19f2c7dc493887353c42f6956bc268b1

src/CMakeLists.txt

@@ -22,8 +22,9 @@ 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

@@ -20,6 +20,7 @@ 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"         },
@@ -72,6 +73,8 @@ 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)"        },
 };
 
@@ -144,6 +147,7 @@ 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"              },
@@ -243,6 +247,24 @@ 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,
         {
@@ -494,6 +516,21 @@ 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,
         {
@@ -1555,6 +1592,34 @@ 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,
         {
@@ -1752,3 +1817,25 @@ 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

@@ -24,6 +24,7 @@ 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,
@@ -76,6 +77,8 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
+    LLM_ARCH_DOTS1,
+    LLM_ARCH_ARCEE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -148,6 +151,7 @@ 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,
@@ -436,3 +440,6 @@ 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.cpp

@@ -1,8 +1,14 @@
 #include "llama-batch.h"
 
+#include "llama-impl.h"
+#include "llama-cparams.h"
+#include "llama-vocab.h"
+#include "llama-memory.h"
+
 #include <cassert>
 #include <cstring>
 #include <algorithm>
+#include <sstream>
 
 llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
     // clear empty sequences
@@ -105,12 +111,7 @@ void llama_sbatch::add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & s
             ubatch.seq_id = batch->seq_id + seq.offset;
         }
     }
-    if (logits_all) {
-        for (size_t i = 0; i < length; ++i) {
-            ubatch.output[ubatch.n_tokens + i] = 1;
-            out_ids.push_back(ids[seq.offset + i]);
-        }
-    } else if (batch->logits) {
+    if (batch->logits) {
         if (ubatch.equal_seqs) {
             for (size_t i = 0; i < length; ++i) {
                 size_t id = ids[seq.offset + i];
@@ -197,11 +198,10 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
     return ubatch;
 }
 
-llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split) {
     GGML_ASSERT(batch.n_tokens >= 0);
     this->batch = &batch;
     this->n_embd = n_embd;
-    this->logits_all = logits_all;
 
     n_tokens = batch.n_tokens;
     ids.resize(n_tokens);
@@ -285,17 +285,56 @@ llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple
             );
 }
 
-llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0) {
-    batch = in_batch;
-    GGML_ASSERT(batch.n_tokens > 0);
-    if (!batch.pos) {
-        assert(p0 >= 0);
-        pos.resize(batch.n_tokens);
-        for (int32_t i = 0; i < batch.n_tokens; i++) {
-            pos[i] = p0 + i;
-        }
-        batch.pos = pos.data();
+llama_batch_allocr::llama_batch_allocr() {
+    const char * LLAMA_BATCH_DEBUG = getenv("LLAMA_BATCH_DEBUG");
+    debug = LLAMA_BATCH_DEBUG ? atoi(LLAMA_BATCH_DEBUG) : 0;
+
+    seq_pos.resize(LLAMA_MAX_SEQ);
+    seq_cpl.resize(LLAMA_MAX_SEQ);
+    for (auto & cur : seq_cpl) {
+        cur.resize(LLAMA_MAX_SEQ);
     }
+}
+
+bool llama_batch_allocr::init(
+        const llama_batch & batch_inp,
+        const llama_vocab & vocab,
+        const llama_memory_i * memory,
+        bool embd_all) {
+    clear();
+
+    batch = batch_inp;
+
+    GGML_ASSERT(batch.n_tokens > 0);
+
+    //
+    // validate input batch
+    //
+
+    if (batch.token) {
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= vocab.n_tokens()) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
+                return false;
+            }
+        }
+    }
+
+    if (batch.seq_id) {
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= LLAMA_MAX_SEQ)) {
+                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], LLAMA_MAX_SEQ);
+                    return false;
+                }
+            }
+        }
+    }
+
+    //
+    // auto-generate missing fields
+    //
+
     if (!batch.n_seq_id) {
         n_seq_id.resize(batch.n_tokens);
         for (int32_t i = 0; i < batch.n_tokens; i++) {
@@ -303,6 +342,7 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
         }
         batch.n_seq_id = n_seq_id.data();
     }
+
     if (!batch.seq_id) {
         seq_id.resize(batch.n_tokens + 1);
         seq_id[batch.n_tokens] = NULL;
@@ -311,10 +351,221 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
         }
         batch.seq_id = seq_id.data();
     }
+
+    if (!batch.pos) {
+        pos.resize(batch.n_tokens);
+
+        // initialize the starting position for each sequence based on the positions in the memory
+        llama_pos p0[LLAMA_MAX_SEQ];
+        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (!memory) {
+                p0[s] = 0;
+            } else {
+                p0[s] = memory->seq_pos_max(s) + 1;
+            }
+        }
+
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            const llama_seq_id seq_id = batch.seq_id[i][0];
+
+            pos[i] = p0[seq_id];
+
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                p0[batch.seq_id[i][s]] = pos[i] + 1;
+            }
+        }
+
+        batch.pos = pos.data();
+    }
+
     if (!batch.logits) {
-        logits.resize(batch.n_tokens);
-        logits[logits.size() - 1] = true;
-        batch.logits = logits.data();
+        if (embd_all) {
+            // return the output for all tokens
+            output.resize(batch.n_tokens, true);
+        } else {
+            // return the output only for the last token
+            output.resize(batch.n_tokens, false);
+            output[output.size() - 1] = true;
+        }
+
+        batch.logits = output.data();
+    } else if (embd_all) {
+        bool warn = false;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            if (batch.logits[i] == 0) {
+                warn = true;
+            }
+        }
+
+        if (warn) {
+            LLAMA_LOG_WARN("%s: embeddings required but some input tokens were not marked as outputs -> overriding\n", __func__);
+
+            output.resize(batch.n_tokens, true);
+            batch.logits = output.data();
+        }
+    }
+
+    //
+    // compute stats
+    //
+
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        n_outputs += batch.logits[i] != 0;
+    }
+
+    // determine coupled sequences
+    // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+            seq_pos[batch.seq_id[i][s]].insert(batch.pos[i]);
+
+            if (s > 0) {
+                const llama_seq_id s0 = batch.seq_id[i][0];
+                const llama_seq_id s1 = batch.seq_id[i][s];
+
+                // mark that sequence s1 is coupled to s0
+                seq_cpl[s1][s0] = true;
+
+                // note: the other way around is not necessary for now
+                //seq_cpl[s0][s1] = true;
+            }
+        }
+    }
+
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__);
+        LLAMA_LOG_DEBUG("%s:   n_tokens  = %d\n", __func__,          batch.n_tokens);
+        LLAMA_LOG_DEBUG("%s:   token     = %p\n", __func__, (void *) batch.token);
+        LLAMA_LOG_DEBUG("%s:   embd      = %p\n", __func__, (void *) batch.embd);
+        LLAMA_LOG_DEBUG("%s:   pos       = %p\n", __func__, (void *) batch.pos);
+        LLAMA_LOG_DEBUG("%s:   n_seq_id  = %p\n", __func__, (void *) batch.n_seq_id);
+        LLAMA_LOG_DEBUG("%s:   seq_id    = %p\n", __func__, (void *) batch.seq_id);
+        LLAMA_LOG_DEBUG("%s:   logits    = %p\n", __func__, (void *) batch.logits);
+        LLAMA_LOG_DEBUG("%s:   n_outputs = %d\n", __func__, n_outputs);
+
+        if (debug > 1) {
+            int seq_id_max = 0;
+            for (int32_t i = 0; i < batch.n_tokens; ++i) {
+                for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                    for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                        seq_id_max = std::max(seq_id_max, batch.seq_id[i][s]);
+                    }
+                }
+            }
+            ++seq_id_max;
+
+            LLAMA_LOG_DEBUG("%s:   token     = [\n", __func__);
+            for (int32_t i = 0; i < batch.n_tokens; ++i) {
+                std::vector<int8_t> seq_id(seq_id_max);
+
+                for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                    seq_id[batch.seq_id[i][s]] = 1;
+                }
+
+                std::stringstream ss;
+                for (int s = 0; s < seq_id_max; ++s) {
+                    if (seq_id[s]) {
+                        ss << s%10;
+                    } else {
+                        ss << ".";
+                    }
+                }
+
+                LLAMA_LOG_DEBUG("%s:  %4d: id = %6d (%16s), pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n",
+                        __func__, i, batch.token[i], vocab.token_to_piece(batch.token[i]).c_str(),
+                        batch.pos[i], batch.n_seq_id[i], ss.str().c_str(), batch.logits[i]);
+            }
+            LLAMA_LOG_DEBUG("%s:   ]\n", __func__);
+
+            LLAMA_LOG_DEBUG("%s:   seq       = [\n", __func__);
+            for (int s0 = 0; s0 < (int) seq_pos.size(); ++s0) {
+                if (seq_pos[s0].empty()) {
+                    continue;
+                }
+
+                std::stringstream ss;
+                for (int s1 = 0; s1 < (int) seq_cpl[s0].size(); ++s1) {
+                    if (seq_cpl[s0][s1]) {
+                        ss << s1 << " ";
+                    }
+                }
+
+                LLAMA_LOG_DEBUG("%s:  %4d: pos = [%4d, %4d], cpl = %s\n",
+                        __func__, s0, seq_pos_min(s0), seq_pos_max(s0), ss.str().empty() ? "-" : ss.str().c_str());
+            }
+            LLAMA_LOG_DEBUG("%s:   ]\n", __func__);
+        }
+    }
+
+    //
+    // consistency checks
+    //
+
+    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        if (seq_pos[s].empty()) {
+            continue;
+        }
+
+        if (memory && seq_pos_min(s) != memory->seq_pos_max(s) + 1) {
+            LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s);
+            return false;
+        }
+
+        if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
+            LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
+            return false;
+        }
+    }
+
+    if (memory) {
+        for (int32_t s0 = 0; s0 < LLAMA_MAX_SEQ; ++s0) {
+            for (int32_t s1 = 0; s1 < LLAMA_MAX_SEQ; ++s1) {
+                if (seq_cpl[s0][s1]) {
+                    if (memory->seq_pos_min(s0) != memory->seq_pos_min(s1) ||
+                        memory->seq_pos_max(s0) != memory->seq_pos_max(s1)) {
+                        LLAMA_LOG_ERROR("%s: sequence %d is coupled to %d in the input batch, but have divereged\n", __func__, s0, s1);
+                        return false;
+                    }
+                }
+            }
+        }
+    }
+
+    return true;
+}
+
+const llama_batch & llama_batch_allocr::get_batch() const {
+    return batch;
+}
+
+uint32_t llama_batch_allocr::get_n_outputs() const {
+    return n_outputs;
+}
+
+llama_pos llama_batch_allocr::seq_pos_min(llama_seq_id seq_id) const {
+    return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].begin();
+}
+
+llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
+    return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].rbegin();
+}
+
+void llama_batch_allocr::clear() {
+    n_outputs = 0;
+
+    batch = {};
+    pos.clear();
+    n_seq_id.clear();
+    seq_id.clear();
+    output.clear();
+
+    for (auto & cur : seq_pos) {
+        cur.clear();
+    }
+
+    for (auto & cur : seq_cpl) {
+        std::fill(cur.begin(), cur.end(), false);
     }
 }
 

src/llama-batch.h

@@ -4,6 +4,7 @@
 
 #include <array>
 #include <vector>
+#include <set>
 
 // very similar to llama_batch,
 // but has more metadata about sequences
@@ -18,8 +19,8 @@ struct llama_ubatch {
     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;
+    int32_t      *  n_seq_id; // [n_seqs]
+    llama_seq_id ** seq_id;   // [n_seqs]
     int8_t       *  output;   // [n_tokens]
 };
 
@@ -39,8 +40,6 @@ struct llama_sbatch {
 
     size_t n_embd;
 
-    bool logits_all; // TODO: remove once lctx.logits_all is removed too
-
     // sorted indices into the batch
     std::vector<int64_t> ids;
     // batch indices of the output
@@ -76,19 +75,45 @@ struct llama_sbatch {
     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, bool logits_all = false);
+    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;
+// a helper for sanitizing and fulfilling a batch
+class llama_batch_allocr {
+public:
+    llama_batch_allocr();
+
+    // 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,
+            bool embd_all);
+
+    const llama_batch & get_batch() const;
+
+    uint32_t get_n_outputs() const;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const;
+
+private:
+    void clear();
+
+    llama_batch batch;
+
+    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<int8_t>         logits;
+    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);
+    std::vector<std::set<llama_pos>> seq_pos; // seq_pos[s]: the set of positions in sequence s
+    std::vector<std::vector<bool>>   seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
+
+    int debug;
 };

src/llama-chat.cpp

@@ -183,6 +183,8 @@ 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;
 }
@@ -331,7 +333,7 @@ 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"
@@ -353,7 +355,7 @@ 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: ";
@@ -643,6 +645,21 @@ 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

@@ -43,6 +43,7 @@ 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

@@ -1,6 +1,7 @@
 #include "llama-context.h"
 
 #include "llama-impl.h"
+#include "llama-batch.h"
 #include "llama-io.h"
 #include "llama-memory.h"
 #include "llama-mmap.h"
@@ -18,7 +19,8 @@
 llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
-    model(model) {
+    model(model),
+    batch_allocr(std::make_unique<llama_batch_allocr>()) {
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -27,8 +29,8 @@ 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_PARALLEL_SEQUENCES) {
-        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_PARALLEL_SEQUENCES));
+    if (cparams.n_seq_max > LLAMA_MAX_SEQ) {
+        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_SEQ));
     }
 
     cparams.n_threads        = params.n_threads;
@@ -494,7 +496,7 @@ float * llama_context::get_logits() {
 }
 
 float * llama_context::get_logits_ith(int32_t i) {
-    int32_t j = -1;
+    int64_t j = -1;
 
     try {
         if (logits == nullptr) {
@@ -517,7 +519,7 @@ 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=%d, n_outputs=%d)", j, n_outputs));
+            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
         return logits + j*model.vocab.n_tokens();
@@ -536,7 +538,7 @@ float * llama_context::get_embeddings() {
 }
 
 float * llama_context::get_embeddings_ith(int32_t i) {
-    int32_t j = -1;
+    int64_t j = -1;
 
     try {
         if (embd == nullptr) {
@@ -559,7 +561,7 @@ 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=%d, n_outputs=%d)", j, n_outputs));
+            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
         return embd + j*model.hparams.n_embd;
@@ -719,52 +721,41 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch,
     return res;
 }
 
-int llama_context::encode(llama_batch & inp_batch) {
-    if (inp_batch.n_tokens == 0) {
+int llama_context::encode(const llama_batch & batch_inp) {
+    if (batch_inp.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);
+    if (!batch_allocr->init(batch_inp, model.vocab, nullptr, true)) {
+        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
+        return -1;
+    }
 
-    const llama_batch & batch = batch_allocr.batch;
-    const int32_t n_tokens = batch.n_tokens;
+    const llama_batch & batch = batch_allocr->get_batch();
 
-    const auto & hparams = model.hparams;
+    const uint32_t n_tokens = batch.n_tokens;
 
     GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
-    // 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;
-            }
-        }
-    }
-
     // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
-    GGML_ASSERT(cparams.n_ubatch >= (uint32_t) n_tokens && "encoder requires n_ubatch >= n_tokens");
+    GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");
 
     if (t_compute_start_us == 0) {
         t_compute_start_us = ggml_time_us();
     }
 
+    // TODO: this clear of the buffer can easily be forgotten - need something better
     embd_seq.clear();
 
     n_queued_tokens += n_tokens;
 
+    const auto & hparams = model.hparams;
+
     const int64_t n_embd = hparams.n_embd;
 
-    llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
+    llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true);
 
     const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
 
@@ -774,7 +765,7 @@ int llama_context::encode(llama_batch & inp_batch) {
         return -2;
     };
 
-    for (int32_t i = 0; i < n_tokens; ++i) {
+    for (uint32_t i = 0; i < n_tokens; ++i) {
         output_ids[i] = i;
     }
 
@@ -830,7 +821,8 @@ int llama_context::encode(llama_batch & inp_batch) {
 
                     GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits
 
-                    for (int32_t i = 0; i < n_tokens; i++) {
+                    // TODO: fix indexing [UBATCH_IDX]
+                    for (uint32_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;
@@ -845,6 +837,7 @@ int llama_context::encode(llama_batch & inp_batch) {
                     auto & embd_seq_out = embd_seq;
                     const uint32_t n_cls_out = hparams.n_cls_out;
 
+                    // TODO: fix indexing [UBATCH_IDX]
                     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()) {
@@ -878,10 +871,10 @@ int llama_context::encode(llama_batch & inp_batch) {
 
         // remember the sequence ids used during the encoding - needed for cross attention later
         cross.seq_ids_enc.resize(n_tokens);
-        for (int32_t i = 0; i < n_tokens; i++) {
+        for (uint32_t i = 0; i < n_tokens; i++) {
             cross.seq_ids_enc[i].clear();
-            for (int s = 0; s < ubatch.n_seq_id[i]; s++) {
-                llama_seq_id seq_id = ubatch.seq_id[i][s];
+            for (int s = 0; s < batch.n_seq_id[i]; s++) {
+                llama_seq_id seq_id = batch.seq_id[i][s];
                 cross.seq_ids_enc[i].insert(seq_id);
             }
         }
@@ -890,51 +883,45 @@ int llama_context::encode(llama_batch & inp_batch) {
     return 0;
 }
 
-int llama_context::decode(llama_batch & inp_batch) {
+int llama_context::decode(const llama_batch & batch_inp) {
     if (!memory) {
         LLAMA_LOG_DEBUG("%s: cannot decode batches with this context (calling encode() instead)\n", __func__);
-        return encode(inp_batch);
+        return encode(batch_inp);
     }
 
-    if (inp_batch.n_tokens == 0) {
+    if (batch_inp.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;
-        }
+    // when computing embeddings, all tokens are output
+    const bool embd_all = cparams.embeddings;
+
+    if (!batch_allocr->init(batch_inp, model.vocab, memory.get(), embd_all)) {
+        LLAMA_LOG_ERROR("%s: failed to initialize batch\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 llama_batch & batch = batch_allocr->get_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;
 
-    const int64_t n_tokens_all = batch.n_tokens;
-    const int64_t n_embd       = hparams.n_embd;
+    const uint32_t n_tokens_all = batch.n_tokens;
 
     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;
-            }
+    const uint32_t n_outputs_all = batch_allocr->get_n_outputs();
 
-            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;
-            }
+    if (embd_all) {
+        // 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",
+                    __func__, n_outputs_all, n_tokens_all);
+            return -1;
         }
     }
 
@@ -947,25 +934,9 @@ int llama_context::decode(llama_batch & inp_batch) {
     }
     n_queued_tokens += n_tokens_all;
 
-    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
-    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
+    // TODO: this clear of the buffer can easily be forgotten - need something better
     embd_seq.clear();
 
-    int64_t n_outputs_all = 0;
-
-    // count outputs
-    if (batch.logits && !embd_pooled) {
-        for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            n_outputs_all += batch.logits[i] != 0;
-        }
-    } else if (embd_pooled) {
-        n_outputs_all = n_tokens_all;
-    } else {
-        // keep last output only
-        n_outputs_all = 1;
-    }
-
     bool did_optimize = false;
 
     // handle any pending defrags/shifts
@@ -974,7 +945,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     llama_memory_state_ptr mstate;
 
     while (true) {
-        mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ n_outputs_all == n_tokens_all);
+        mstate = memory->init_batch(batch, cparams.n_ubatch, embd_all);
         if (!mstate) {
             return -2;
         }
@@ -1018,7 +989,7 @@ int llama_context::decode(llama_batch & inp_batch) {
 
     // reserve output buffer
     if (output_reserve(n_outputs_all) < n_outputs_all) {
-        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
+        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
         return -2;
     };
 
@@ -1027,7 +998,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     do {
         const auto & ubatch = mstate->get_ubatch();
 
-        // count the outputs in this u_batch
+        // count the outputs in this ubatch
         {
             int32_t n_outputs_new = 0;
 
@@ -1052,18 +1023,19 @@ int llama_context::decode(llama_batch & inp_batch) {
 
         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_PARALLEL_SEQUENCES];
-            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+            llama_pos pos_min[LLAMA_MAX_SEQ];
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++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_PARALLEL_SEQUENCES; ++s) {
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
                 if (pos_min[s] == std::numeric_limits<llama_pos>::max()) {
                     continue;
                 }
@@ -1086,7 +1058,7 @@ int llama_context::decode(llama_batch & inp_batch) {
         //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
         //}
 
-        auto * t_logits = cparams.embeddings ? nullptr         : res->get_logits();
+        auto * t_logits = res->get_logits();
         auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
 
         if (t_embd && res->get_embd_pooled()) {
@@ -1170,14 +1142,14 @@ int llama_context::decode(llama_batch & inp_batch) {
     n_outputs = n_outputs_all;
 
     // set output mappings
-    {
+    if (n_outputs > 0) {
         bool sorted_output = true;
 
         auto & out_ids = mstate->out_ids();
 
-        GGML_ASSERT(out_ids.size() == (size_t) n_outputs_all);
+        GGML_ASSERT(out_ids.size() == (size_t) n_outputs);
 
-        for (int64_t i = 0; i < n_outputs_all; ++i) {
+        for (int64_t i = 0; i < n_outputs; ++i) {
             int64_t out_id = out_ids[i];
             output_ids[out_id] = i;
             if (out_id != i) {
@@ -1189,20 +1161,22 @@ int llama_context::decode(llama_batch & inp_batch) {
         // note: this is mostly relevant for recurrent models atm
         if (!sorted_output) {
             const uint32_t n_vocab = model.vocab.n_tokens();
-            const uint32_t n_embd  = model.hparams.n_embd;
+            const uint64_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 (int32_t i = 0; i < n_outputs - 1; ++i) {
-                int32_t j_min = i;
-                for (int32_t j = i + 1; j < n_outputs; ++j) {
+            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) {
                     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++) {
@@ -1215,8 +1189,10 @@ int llama_context::decode(llama_batch & inp_batch) {
                     }
                 }
             }
+
             std::fill(output_ids.begin(), output_ids.end(), -1);
-            for (int32_t i = 0; i < n_outputs; ++i) {
+
+            for (uint32_t i = 0; i < n_outputs; ++i) {
                 output_ids[out_ids[i]] = i;
             }
         }
@@ -1236,7 +1212,7 @@ int llama_context::decode(llama_batch & inp_batch) {
 // output
 //
 
-int32_t llama_context::output_reserve(int32_t n_outputs) {
+uint32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
@@ -1246,9 +1222,8 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto n_vocab = vocab.n_tokens();
     const auto n_embd  = hparams.n_embd;
 
-    // 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);
+    bool has_logits = true;
+    bool has_embd   = cparams.embeddings;
 
     // TODO: hacky enc-dec support
     if (model.arch == LLM_ARCH_T5) {
@@ -1302,8 +1277,7 @@ int32_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_max = n_outputs_max;
+    this->n_outputs = 0;
 
     return n_outputs_max;
 }
@@ -1332,7 +1306,7 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
 
     if (n_tokens % n_seqs != 0) {
-        n_tokens = (n_tokens / n_seqs) * n_seqs;
+        n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
         n_outputs = std::min(n_outputs, n_tokens);
 
         LLAMA_LOG_DEBUG("%s: making n_tokens a multiple of n_seqs - n_tokens = %u, n_seqs = %u, n_outputs = %u\n", __func__, n_tokens, n_seqs, n_outputs);
@@ -1794,14 +1768,12 @@ 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
-            int32_t pos = output_ids[i];
+            int64_t pos = output_ids[i];
             if (pos >= 0) {
                 GGML_ASSERT(pos < n_outputs);
                 w_output_pos[pos] = i;
@@ -2071,14 +2043,11 @@ void llama_context::opt_epoch_iter(
 
         n_queued_tokens += n_tokens_all;
 
-        // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
-        const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
         embd_seq.clear();
 
-        int64_t n_outputs_all = n_tokens_all;
+        uint32_t n_outputs_all = n_tokens_all;
 
-        auto mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ true);
+        auto mstate = memory->init_batch(batch, cparams.n_ubatch, true);
         if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
             LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__);
             break;
@@ -2086,7 +2055,7 @@ 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 %" PRId64 " outputs\n", __func__, n_outputs_all);
+            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
             GGML_ABORT("TODO: handle this error");
         };
 

src/llama-context.h

@@ -1,7 +1,6 @@
 #pragma once
 
 #include "llama.h"
-#include "llama-batch.h"
 #include "llama-cparams.h"
 #include "llama-graph.h"
 #include "llama-adapter.h"
@@ -13,6 +12,7 @@
 #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(llama_batch & inp_batch);
-    int decode(llama_batch & inp_batch);
+    int encode(const llama_batch & batch_inp);
+    int decode(const llama_batch & batch_inp);
 
     //
     // 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.
-    int32_t output_reserve(int32_t n_outputs);
+    uint32_t output_reserve(int32_t n_outputs);
 
     //
     // graph
@@ -246,8 +246,10 @@ private:
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
 
-    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
+    // reuse the batch_allocr to avoid unnecessary memory allocations
+    std::unique_ptr<llama_batch_allocr> batch_allocr;
+
+    uint32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch
 
     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_PARALLEL_SEQUENCES;
+    return LLAMA_MAX_SEQ;
 }

src/llama-cparams.h

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

src/llama-graph.cpp

@@ -6,7 +6,8 @@
 
 #include "llama-kv-cache-unified.h"
 #include "llama-kv-cache-unified-iswa.h"
-#include "llama-kv-cache-recurrent.h"
+#include "llama-memory-hybrid.h"
+#include "llama-memory-recurrent.h"
 
 #include <cassert>
 #include <cmath>
@@ -139,6 +140,7 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
 
         std::vector<uint64_t> sum(n_tokens, 0);
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -156,6 +158,7 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
             }
         }
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -180,6 +183,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
         uint32_t * data = (uint32_t *) cls->data;
         memset(cls->data, 0, n_tokens * ggml_element_size(cls));
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -210,6 +214,7 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
         std::vector<int> last_pos(n_tokens, -1);
         std::vector<int> last_row(n_tokens, -1);
 
+        // TODO: fix indexing [UBATCH_IDX]
         for (int s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
@@ -234,34 +239,18 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
-    const int64_t n_kv = kv_state->get_n_kv();
+    const int64_t n_rs = mem_state->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_kv; ++i) {
-            data[i] = kv_state->s_copy(i);
-        }
-    }
-}
-
-void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
-    GGML_UNUSED(ubatch);
-
-    const int64_t n_kv = kv_state->get_n_kv();
-
-    if (s_mask) {
-        GGML_ASSERT(ggml_backend_buffer_is_host(s_mask->buffer));
-        float * data = (float *) s_mask->data;
-
-        // clear unused states
-        for (int i = 0; i < n_kv; ++i) {
-            data[i] = kv_state->s_mask(i);
+        for (uint32_t i = 0; i < n_rs; ++i) {
+            data[i] = mem_state->s_copy(i);
         }
     }
 }
@@ -299,6 +288,7 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
+                                // TODO: fix indexing [UBATCH_IDX]
                                 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) {
@@ -338,6 +328,7 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                                 const int32_t ti = s0*n_seq_tokens + i;
                                 float f = -INFINITY;
 
+                                // TODO: fix indexing [UBATCH_IDX]
                                 for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
                                     if (ubatch->seq_id[s0][s] == seq_id) {
                                         if (hparams.use_alibi) {
@@ -393,6 +384,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
             for (int j = 0; j < n_tokens; ++j) {
                 for (int i = 0; i < n_enc; ++i) {
                     float f = -INFINITY;
+                    // TODO: fix indexing [UBATCH_IDX]
                     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()) {
@@ -412,6 +404,24 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+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
 //
@@ -650,6 +660,7 @@ ggml_tensor * llm_graph_context::build_ffn(
             {
                 // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
                 int64_t split_point = cur->ne[0] / 2;
+                // TODO: these conts should not be needed, see https://github.com/ggml-org/llama.cpp/pull/14090#discussion_r2137437217
                 ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
                 ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
 
@@ -663,7 +674,7 @@ ggml_tensor * llm_graph_context::build_ffn(
             {
                 // Split into two equal parts
                 int64_t split_point = cur->ne[0] / 2;
-                // TODO: these conts should not be needed
+                // TODO: these conts should not be needed, see https://github.com/ggml-org/llama.cpp/pull/14090#discussion_r2137437217
                 ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
                 ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
 
@@ -969,40 +980,6 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
     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));
-
-    return cur;
-}
-
-ggml_tensor * llm_graph_context::build_inp_s_mask() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_s_mask>(kv_state);
-
-    const auto n_kv = kv_state->get_n_kv();
-
-    auto & cur = inp->s_mask;
-
-    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_kv);
-    ggml_set_input(cur);
-
-    res->add_input(std::move(inp));
-
-    return cur;
-}
-
 ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     auto inp = std::make_unique<llm_graph_input_cross_embd>(cross);
 
@@ -1072,6 +1049,33 @@ 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,
@@ -1316,36 +1320,6 @@ 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,
@@ -1455,56 +1429,182 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_copy_mask_state(
-         ggml_cgraph * gf,
-         ggml_tensor * s,
-         ggml_tensor * state_copy,
-         ggml_tensor * state_mask,
-             int32_t   n_state,
-             int32_t   n_seqs) const {
-    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+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);
 
-    const auto n_kv    = kv_state->get_n_kv();
-    const auto kv_head = kv_state->get_head();
+    const auto * kv_state = static_cast<const llama_memory_hybrid_state *>(mstate)->get_state_attn();
 
-    ggml_tensor * states = ggml_reshape_2d(ctx0, s, n_state, kv_state->get_size());
+    // 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));
+    }
 
-    // copy states
-    // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
-    // this shrinks the tensors's ne[1] to n_kv
-    states = ggml_get_rows(ctx0, states, state_copy);
+    const auto & kq_mask = inp->get_kq_mask();
 
-    // clear states of sequences which are starting at the beginning of this batch
-    // FIXME: zero-out NANs?
-    states = ggml_mul(ctx0, states, state_mask);
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = kv_state->get_k(ctx0, il);
+    ggml_tensor * v = kv_state->get_v(ctx0, il);
 
-    // copy states which won't be changed further (between n_seqs and n_kv)
+    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);
+
+    // 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.
+    ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0));
+    ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));
+
+    ggml_tensor * output_states;
+
+    if (!avoid_copies) {
+        // copy states
+        // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
+        // {state_size, kv_size} -> {state_size, n_seqs}
+        output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
+        ggml_build_forward_expand(gf, output_states);
+    } else {
+        // FIXME: make the gathering operation happen before the copy below
+        //        (maybe with an optional lambda function passed as a parameter instead of `avoid_copies`?)
+        output_states = states;
+    }
+
+    // copy extra states which won't be changed further (between n_seqs and n_kv)
+    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0]));
     ggml_build_forward_expand(gf,
         ggml_cpy(ctx0,
-            ggml_view_1d(ctx0, states, n_state*(n_kv - n_seqs), (n_seqs          )*n_state*ggml_element_size(states)),
-            ggml_view_1d(ctx0, s,      n_state*(n_kv - n_seqs), (kv_head + n_seqs)*n_state*ggml_element_size(s))));
+            states_extra,
+            ggml_view_1d(ctx0, s, state_size*(n_kv - n_seqs), (kv_head + n_seqs)*state_size*ggml_element_size(s))));
 
-    // the part of the states that will be used and modified
-    return ggml_view_2d(ctx0, states, n_state, n_seqs, states->nb[1], 0);
+    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(
-         ggml_cgraph * gf,
-         ggml_tensor * state_copy,
-         ggml_tensor * state_mask,
-  const llama_ubatch & ubatch,
+    llm_graph_input_rs * inp,
+           ggml_cgraph * gf,
+    const llama_ubatch & ubatch,
                  int   il) const {
-    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+    const auto * kv_state = static_cast<const llama_memory_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_k_l(il);
+    ggml_tensor * token_shift_all = kv_state->get_r_l(il);
 
-    ggml_tensor * token_shift = build_copy_mask_state(
-            gf, token_shift_all, state_copy, state_mask,
-            hparams.n_embd_k_s(), n_seqs);
+    ggml_tensor * token_shift = build_rs(
+            inp, gf, token_shift_all,
+            hparams.n_embd_r(), n_seqs);
 
     token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);
 
@@ -1515,7 +1615,7 @@ 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_kv_cache_recurrent_state *>(mstate);
+    const auto * kv_state = static_cast<const llama_memory_recurrent_state *>(mstate);
 
     const auto token_shift_count = hparams.token_shift_count;
     const auto n_embd = hparams.n_embd;
@@ -1527,7 +1627,7 @@ 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_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)))
+        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)))
     );
 }
 
@@ -1578,23 +1678,30 @@ void llm_graph_context::build_pooling(
                 ggml_tensor * inp_cls = build_inp_cls();
                 inp = ggml_get_rows(ctx0, inp, inp_cls);
 
-                if (cls != nullptr && cls_b != nullptr) {
+                if (cls) {
                     // classification head
                     // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
+                    cur = ggml_mul_mat(ctx0, cls, inp);
+                    if (cls_b) {
+                        cur = ggml_add(ctx0, cur, 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) {
-                        GGML_ASSERT(cls_out_b != nullptr);
-                        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
+                        cur = ggml_mul_mat(ctx0, cls_out, cur);
+                        if (cls_out_b) {
+                            cur = ggml_add(ctx0, 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
-                    GGML_ASSERT(cls_out_b != nullptr);
-                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, inp), cls_out_b);
+                    cur = ggml_mul_mat(ctx0, cls_out, inp);
+                    if (cls_out_b) {
+                        cur = ggml_add(ctx0, cur, cls_out_b);
+                    }
                 } else {
                     GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
                 }

src/llama-graph.h

@@ -21,7 +21,8 @@ struct llama_memory_state_i;
 
 class llama_kv_cache_unified_state;
 class llama_kv_cache_unified_iswa_state;
-class llama_kv_cache_recurrent_state;
+class llama_memory_recurrent_state;
+class llama_memory_hybrid_state;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -188,28 +189,16 @@ public:
     const llama_cparams & cparams;
 };
 
-class llm_graph_input_s_copy : public llm_graph_input_i {
+class llm_graph_input_rs : public llm_graph_input_i {
 public:
-    llm_graph_input_s_copy(const llama_kv_cache_recurrent_state * kv_state) : kv_state(kv_state) {}
-    virtual ~llm_graph_input_s_copy() = default;
+    llm_graph_input_rs(const llama_memory_recurrent_state * mem_state) : mem_state(mem_state) {}
+    virtual ~llm_graph_input_rs() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
-    const llama_kv_cache_recurrent_state * kv_state;
-};
-
-class llm_graph_input_s_mask : public llm_graph_input_i {
-public:
-    llm_graph_input_s_mask(const llama_kv_cache_recurrent_state * kv_state) : kv_state(kv_state) {}
-    virtual ~llm_graph_input_s_mask() = default;
-
-    void set_input(const llama_ubatch * ubatch) override;
-
-    ggml_tensor * s_mask; // F32 [1, n_kv]
-
-    const llama_kv_cache_recurrent_state * kv_state;
+    const llama_memory_recurrent_state * mem_state;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -312,6 +301,33 @@ 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
 //
@@ -390,7 +406,7 @@ struct llm_graph_params {
     const llama_memory_state_i * mstate;
     const llama_cross          * cross;
 
-    int32_t n_outputs;
+    uint32_t n_outputs;
 
     const llm_graph_cb & cb;
 };
@@ -424,8 +440,8 @@ struct llm_graph_context {
     const float norm_eps;
     const float norm_rms_eps;
 
-    const int32_t n_tokens;
-    const int32_t n_outputs;
+    const int64_t n_tokens;
+    const int64_t n_outputs;
     const int32_t n_ctx_orig; // yarn
 
     const enum llama_pooling_type pooling_type;
@@ -520,14 +536,14 @@ 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_s_mask() 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
     //
@@ -602,23 +618,62 @@ 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
     //
 
-    ggml_tensor * build_copy_mask_state(
-             ggml_cgraph * gf,
-             ggml_tensor * s,
-             ggml_tensor * state_copy,
-             ggml_tensor * state_mask,
-                 int32_t   n_state,
-                 int32_t   n_seqs) const;
+    // 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_rwkv_token_shift_load(
-             ggml_cgraph * gf,
-             ggml_tensor * state_copy,
-             ggml_tensor * state_mask,
-      const llama_ubatch & ubatch,
+        llm_graph_input_rs * inp,
+               ggml_cgraph * gf,
+        const llama_ubatch & ubatch,
                      int   il) const;
 
     ggml_tensor * build_rwkv_token_shift_store(

src/llama-hparams.cpp

@@ -65,7 +65,7 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     return n_embd_head_v * n_head_kv;
 }
 
-uint32_t llama_hparams::n_embd_k_s() const {
+uint32_t llama_hparams::n_embd_r() const {
     if (wkv_head_size != 0) {
         // for RWKV models
         return token_shift_count * n_embd;
@@ -76,7 +76,7 @@ uint32_t llama_hparams::n_embd_k_s() const {
     return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
 }
 
-uint32_t llama_hparams::n_embd_v_s() const {
+uint32_t llama_hparams::n_embd_s() const {
     if (wkv_head_size != 0) {
         // corresponds to RWKV's wkv_states size
         return n_embd * wkv_head_size;
@@ -86,6 +86,10 @@ uint32_t llama_hparams::n_embd_v_s() const {
     return ssm_d_state * ssm_d_inner;
 }
 
+bool llama_hparams::is_recurrent(uint32_t il) const {
+    return recurrent_layer_arr[il];
+}
+
 bool llama_hparams::is_swa(uint32_t il) const {
     if (il < n_layer) {
         return swa_layers[il];

src/llama-hparams.h

@@ -115,6 +115,9 @@ struct llama_hparams {
     uint32_t ssm_d_state = 0;
     uint32_t ssm_dt_rank = 0;
 
+    // for hybrid state space models
+    std::array<bool, LLAMA_MAX_LAYERS> recurrent_layer_arr;
+
     bool ssm_dt_b_c_rms = false;
 
     float f_clamp_kqv      = 0.0f;
@@ -181,10 +184,13 @@ struct llama_hparams {
 
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
-    uint32_t n_embd_k_s() const;
+    uint32_t n_embd_r() const;
 
     // dimension of the recurrent state embeddings
-    uint32_t n_embd_v_s() const;
+    uint32_t n_embd_s() const;
+
+    // whether or not the given layer is recurrent (for hybrid models)
+    bool is_recurrent(uint32_t il) const;
 
     bool is_swa(uint32_t il) const;
 };

src/llama-kv-cache-unified-iswa.cpp

@@ -95,36 +95,69 @@ llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const {
     return kv_swa->seq_pos_max(seq_id);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled, bool logits_all) {
-    GGML_UNUSED(embd_pooled);
+llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) {
+    GGML_UNUSED(embd_all);
 
-    // TODO: if we fail with split_simple, we should attempt different splitting strategies
+    // first try simple split
+    do {
+        auto sbatch = llama_sbatch(batch, hparams.n_embd, true);
+
+        std::vector<llama_ubatch> ubatches;
+
+        while (sbatch.n_tokens > 0) {
+            auto ubatch = sbatch.split_simple(n_ubatch);
+
+            ubatches.push_back(ubatch);
+        }
+
+        auto heads_base = kv_base->prepare(ubatches);
+        if (heads_base.empty()) {
+            break;
+        }
+
+        auto heads_swa = kv_swa->prepare(ubatches);
+        if (heads_swa.empty()) {
+            break;
+        }
+
+        assert(heads_base.size() == heads_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_state>(
+                this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    } while (false);
+
+    // if it fails, try equal split
+    do {
+        auto sbatch = llama_sbatch(batch, hparams.n_embd, false);
+
+        std::vector<llama_ubatch> ubatches;
+
+        while (sbatch.n_tokens > 0) {
+            auto ubatch = sbatch.split_equal(n_ubatch);
+
+            ubatches.push_back(ubatch);
+        }
+
+        auto heads_base = kv_base->prepare(ubatches);
+        if (heads_base.empty()) {
+            break;
+        }
+
+        auto heads_swa = kv_swa->prepare(ubatches);
+        if (heads_swa.empty()) {
+            break;
+        }
+
+        assert(heads_base.size() == heads_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_state>(
+                this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    } while (false);
+
+    // TODO: if we fail again, we should attempt different splitting strategies
     //       but to do that properly, we first have to refactor the batches to be more flexible
 
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, true, logits_all);
-
-    std::vector<llama_ubatch> ubatches;
-
-    while (sbatch.n_tokens > 0) {
-        auto ubatch = sbatch.split_simple(n_ubatch);
-
-        ubatches.push_back(ubatch);
-    }
-
-    auto heads_base = kv_base->prepare(ubatches);
-    if (heads_base.empty()) {
-        return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
-
-    auto heads_swa = kv_swa->prepare(ubatches);
-    if (heads_swa.empty()) {
-        return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
-
-    assert(heads_base.size() == heads_swa.size());
-
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(
-            this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
 llama_memory_state_ptr llama_kv_cache_unified_iswa::init_full() {
@@ -164,21 +197,19 @@ llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_swa() const {
 llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(llama_memory_status status) : status(status) {}
 
 llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
-        llama_kv_cache_unified_iswa * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS) {
-    state_base = kv->get_base()->init_full();
-    state_swa  = kv->get_swa ()->init_full();
-
-    status = llama_memory_status_combine(state_base->get_status(), state_swa->get_status());
+        llama_kv_cache_unified_iswa * kv) :
+    state_base(kv->get_base()->init_full()),
+    state_swa (kv->get_swa ()->init_full()),
+    status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) {
 }
 
 llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
         llama_kv_cache_unified_iswa * kv,
         llama_context * lctx,
-        bool optimize) : status(LLAMA_MEMORY_STATUS_SUCCESS) {
-    state_base = kv->get_base()->init_update(lctx, optimize);
-    state_swa  = kv->get_swa ()->init_update(lctx, optimize);
-
-    status = llama_memory_status_combine(state_base->get_status(), state_swa->get_status());
+        bool optimize) :
+    state_base(kv->get_base()->init_update(lctx, optimize)),
+    state_swa (kv->get_swa ()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) {
 }
 
 llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
@@ -186,15 +217,13 @@ llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
         llama_sbatch sbatch,
         std::vector<uint32_t> heads_base,
         std::vector<uint32_t> heads_swa,
-        std::vector<llama_ubatch> ubatches)
-        : status(LLAMA_MEMORY_STATUS_SUCCESS),
-        sbatch(std::move(sbatch)),
-        ubatches(std::move(ubatches)) {
+        std::vector<llama_ubatch> ubatches) :
+    sbatch(std::move(sbatch)),
+    ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    state_base.reset(new llama_kv_cache_unified_state(kv->get_base(), {}, std::move(heads_base), this->ubatches));
-    state_swa .reset(new llama_kv_cache_unified_state(kv->get_swa (), {}, std::move(heads_swa),  this->ubatches));
-
-    status = llama_memory_status_combine(state_base->get_status(), state_swa->get_status());
+    state_base(new llama_kv_cache_unified_state(kv->get_base(), {}, std::move(heads_base), this->ubatches)),
+    state_swa (new llama_kv_cache_unified_state(kv->get_swa (), {}, std::move(heads_swa),  this->ubatches)),
+    status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) {
 }
 
 llama_kv_cache_unified_iswa_state:: ~llama_kv_cache_unified_iswa_state() = default;

src/llama-kv-cache-unified-iswa.h

@@ -34,8 +34,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 
@@ -118,8 +117,6 @@ public:
     const llama_kv_cache_unified_state * get_swa()  const;
 
 private:
-    llama_memory_status status;
-
     //llama_kv_cache_unified_iswa * kv;
 
     llama_sbatch sbatch;
@@ -129,6 +126,8 @@ private:
 
     std::vector<llama_ubatch> ubatches;
 
-    llama_memory_state_ptr state_base;
-    llama_memory_state_ptr state_swa;
+    const llama_memory_state_ptr state_base;
+    const llama_memory_state_ptr state_swa;
+
+    const llama_memory_status status;
 };

src/llama-kv-cache-unified.cpp

@@ -68,8 +68,8 @@ llama_kv_cache_unified::llama_kv_cache_unified(
             continue;
         }
 
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
-        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
+        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
         const char * dev_name = "CPU";
 
@@ -127,6 +127,9 @@ llama_kv_cache_unified::llama_kv_cache_unified(
                 ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
                 ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
     }
+
+    const char * LLAMA_KV_CACHE_DEBUG = getenv("LLAMA_KV_CACHE_DEBUG");
+    debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
 }
 
 void llama_kv_cache_unified::clear(bool data) {
@@ -307,24 +310,27 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const {
 llama_memory_state_ptr llama_kv_cache_unified::init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) {
-    GGML_UNUSED(embd_pooled);
+            bool embd_all) {
+    GGML_UNUSED(embd_all);
 
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, true, logits_all);
+    do {
+        auto sbatch = llama_sbatch(batch, hparams.n_embd, true);
 
-    std::vector<llama_ubatch> ubatches;
-    while (sbatch.n_tokens > 0) {
-        ubatches.push_back(sbatch.split_simple(n_ubatch));
-    }
+        std::vector<llama_ubatch> ubatches;
+        while (sbatch.n_tokens > 0) {
+            ubatches.push_back(sbatch.split_simple(n_ubatch));
+        }
 
-    auto heads = prepare(ubatches);
-    if (heads.empty()) {
-        return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
+        auto heads = prepare(ubatches);
+        if (heads.empty()) {
+            break;
+        }
 
-    return std::make_unique<llama_kv_cache_unified_state>(
-            this, std::move(sbatch), std::move(heads), std::move(ubatches));
+        return std::make_unique<llama_kv_cache_unified_state>(
+                this, std::move(sbatch), std::move(heads), std::move(ubatches));
+    } while (false);
+
+    return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
 llama_memory_state_ptr llama_kv_cache_unified::init_full() {
@@ -512,43 +518,68 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
         head_cur = 0;
     }
 
-    // otherwise, one cell per token.
-
     if (n_tokens > cells.size()) {
         LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size());
         return -1;
     }
 
-//#define FIND_SLOT_DEBUG 1
-#if FIND_SLOT_DEBUG
-    LLAMA_LOG_WARN("begin: n = %5d, used = %5d, head = %5d, n_swa = %5d\n", cells.used_max_p1(), cells.get_used(), head, n_swa);
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n", __func__, cells.used_max_p1(), cells.get_used(), head, get_size(), n_swa);
 
-    // for debugging
-    {
-        std::string ss;
-        if (n_swa > 0) {
+        if ((debug == 2 && n_swa > 0) || debug > 2) {
+            std::string ss;
             for (uint32_t i = 0; i < cells.size(); ++i) {
                 if (cells.is_empty(i)) {
                     ss += '.';
                 } else {
-                    ss += std::to_string(cells.seq_get(i));
+                    assert(cells.seq_count(i) >= 1);
+
+                    if (cells.seq_count(i) == 1) {
+                        ss += std::to_string(cells.seq_get(i));
+                    } else {
+                        ss += 'M';
+                    }
                 }
                 if (i%256 == 255) {
+                    ss += " *";
                     ss += '\n';
                 }
             }
-        }
-        LLAMA_LOG_WARN("\n%s\n", ss.c_str());
-    }
-
-    for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
-        if (cells.seq_pos_min(s) < 0) {
-            continue;
+            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
         }
 
-        LLAMA_LOG_WARN("kv_cells: n_swa = %4d, min[%d] = %5d, max[%d] = %5d\n", n_swa, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
+        if ((debug == 2 && n_swa > 0) || debug > 2) {
+            std::string ss;
+            for (uint32_t i = 0; i < cells.size(); ++i) {
+                std::string cur;
+                if (cells.is_empty(i)) {
+                    cur = '.';
+                } else {
+                    cur = std::to_string(cells.pos_get(i));
+                }
+                const int n = cur.size();
+                for (int j = 0; j < 5 - n; ++j) {
+                    cur += ' ';
+                }
+                ss += cur;
+                if (i%256 == 255) {
+                    ss += " *";
+                }
+                if (i%64 == 63) {
+                    ss += '\n';
+                }
+            }
+            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
+        }
+
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (cells.seq_pos_min(s) < 0) {
+                continue;
+            }
+
+            LLAMA_LOG_DEBUG("%s: min[%d] = %5d, max[%d] = %5d\n", __func__, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
+        }
     }
-#endif
 
     uint32_t n_tested = 0;
 
@@ -559,21 +590,15 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
             continue;
         }
 
-        // keep track of what the minimum sequence positions would be if we accept the ubatch
-        llama_seq_id seq_pos_min[LLAMA_MAX_PARALLEL_SEQUENCES];
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
-            seq_pos_min[s] = cells.seq_pos_min(s);
-        }
-
         bool found = true;
         for (uint32_t i = 0; i < n_tokens; i++) {
-            const llama_pos    pos    = ubatch.pos[i];
-            const llama_seq_id seq_id = ubatch.seq_id[i][0];
+            //const llama_pos    pos    = ubatch.pos[i];
+            //const llama_seq_id seq_id = ubatch.seq_id[i][0];
 
             // can we use this cell? either:
             //  - the cell is empty
             //  - the cell is occupied only by one sequence:
-            //    - mask causally, if the sequence is the same as the one we are inserting
+            //    - (disabled) mask causally, if the sequence is the same as the one we are inserting
             //    - mask SWA, using current max pos for that sequence in the cache
             //                always insert in the cell with minimum pos
             bool can_use = cells.is_empty(head_cur + i);
@@ -581,21 +606,17 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
             if (!can_use && cells.seq_count(head_cur + i) == 1) {
                 const llama_pos pos_cell = cells.pos_get(head_cur + i);
 
-                // causal mask
-                if (cells.seq_has(head_cur + i, seq_id)) {
-                    can_use = pos_cell >= pos;
-                }
+                // (disabled) causal mask
+                // note: it's better to purge any "future" tokens beforehand
+                //if (cells.seq_has(head_cur + i, seq_id)) {
+                //    can_use = pos_cell >= pos;
+                //}
 
                 if (!can_use) {
                     const llama_seq_id seq_id_cell = cells.seq_get(head_cur + i);
 
                     // SWA mask
-                    // note: we insert only in the cell with minimum pos in order to preserve the invariant that
-                    //       all positions between [pos_min, pos_max] for each sequence will be present in the cache
-                    //       ref: https://github.com/ggml-org/llama.cpp/pull/13746#issuecomment-2916057092
-                    if (pos_cell == seq_pos_min[seq_id_cell] &&
-                        is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
-                        seq_pos_min[seq_id_cell]++;
+                    if (is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
                         can_use = true;
                     }
                 }
@@ -623,18 +644,58 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
 }
 
 void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch) {
-    for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
-        if (!cells.is_empty(head_cur + i)) {
-            cells.rm(head_cur + i);
-        }
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: ubatch info:\n", __func__);
+        LLAMA_LOG_DEBUG("%s:   n_tokens = %d, equal_seqs = %d\n", __func__, ubatch.n_tokens, ubatch.equal_seqs);
+        LLAMA_LOG_DEBUG("%s:   n_seq_tokens = %d, n_seqs = %d\n", __func__, ubatch.n_seq_tokens, ubatch.n_seqs);
+    }
 
-        cells.pos_set(head_cur + i, ubatch.pos[i]);
+    // keep track of the max sequence position that we would overwrite with this ubatch
+    // for non-SWA cache, this would be always empty
+    llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ];
+    for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        seq_pos_max_rm[s] = -1;
+    }
 
-        for (int32_t j = 0; j < ubatch.n_seq_id[i]; j++) {
-            cells.seq_add(head_cur + i, ubatch.seq_id[i][j]);
+    for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
+        for (uint32_t j = 0; j < ubatch.n_seq_tokens; ++j) {
+            const uint32_t idx = s*ubatch.n_seq_tokens + j;
+
+            if (!cells.is_empty(head_cur + idx)) {
+                assert(cells.seq_count(head_cur + idx) == 1);
+
+                const llama_seq_id seq_id = cells.seq_get(head_cur + idx);
+                const llama_pos    pos    = cells.pos_get(head_cur + idx);
+
+                seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos);
+
+                cells.rm(head_cur + idx);
+            }
+
+            cells.pos_set(head_cur + idx, ubatch.pos[idx]);
+
+            // TODO: fix indexing [UBATCH_IDX]
+            for (int32_t i = 0; i < ubatch.n_seq_id[s]; i++) {
+                cells.seq_add(head_cur + idx, ubatch.seq_id[s][i]);
+            }
         }
     }
 
+    // note: we want to preserve the invariant that all positions between [pos_min, pos_max] for each sequence
+    //       will be present in the cache. so we have to purge any position which is less than those we would overwrite
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/13746#issuecomment-2916057092
+    for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        if (seq_pos_max_rm[s] == -1) {
+            continue;
+        }
+
+        if (cells.seq_pos_min(s) <= seq_pos_max_rm[s]) {
+            LLAMA_LOG_DEBUG("%s: purging positions [%d, %d] of sequence %d from KV cache\n",
+                    __func__, cells.seq_pos_min(s), seq_pos_max_rm[s], s);
+
+            seq_rm(s, cells.seq_pos_min(s), seq_pos_max_rm[s] + 1);
+        }
+    }
     // move the head at the end of the slot
     head = head_cur + ubatch.n_tokens;
 }
@@ -731,14 +792,14 @@ ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_
 }
 
 void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
-    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 uint32_t n_tokens     = ubatch->n_tokens;
+    const uint32_t n_seq_tokens = ubatch->n_seq_tokens;
+    const uint32_t n_seqs       = ubatch->n_seqs;
 
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
     float * data = (float *) dst->data;
 
-    const auto n_kv = dst->ne[0];
+    const int64_t n_kv = dst->ne[0];
 
     // Use only the previous KV cells of the correct sequence for each token of the ubatch.
     // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
@@ -752,12 +813,14 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
     //      xxxxx-----
     //      xxxxx-----
     // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615
-    for (int h = 0; h < 1; ++h) {
-        for (int s = 0; s < n_seqs; ++s) {
+    for (uint32_t h = 0; h < 1; ++h) {
+        for (uint32_t s = 0; s < n_seqs; ++s) {
             const llama_seq_id seq_id = ubatch->seq_id[s][0];
 
-            for (int j = 0; j < n_seq_tokens; ++j) {
-                const llama_pos p1 = ubatch->pos[s*n_seq_tokens + j];
+            for (uint32_t j = 0; j < n_seq_tokens; ++j) {
+                const uint32_t idx = s*n_seq_tokens + j;
+
+                const llama_pos p1 = ubatch->pos[idx];
 
                 for (uint32_t i = 0; i < n_kv; ++i) {
                     float f = 0.0f;
@@ -787,16 +850,16 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
                         f = -INFINITY;
                     }
 
-                    data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
+                    data[h*(n_kv*n_tokens) + idx*n_kv + i] = f;
                 }
             }
         }
 
         // mask padded tokens
         if (data) {
-            for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
-                for (uint32_t j = 0; j < n_kv; ++j) {
-                    data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
+            for (uint32_t j = n_tokens; j < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++j) {
+                for (uint32_t i = 0; i < n_kv; ++i) {
+                    data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
                 }
             }
         }
@@ -1367,7 +1430,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
     for (const auto & layer : layers) {
         const uint32_t il = layer.il;
 
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
 
         // Write key type
         const int32_t k_type_i = (int32_t)layer.k->type;
@@ -1389,7 +1452,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
 
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             // Write value type
             const int32_t v_type_i = (int32_t)layer.v->type;
@@ -1413,7 +1476,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
 
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             // Write value type
             const int32_t v_type_i = (int32_t)layer.v->type;
@@ -1447,9 +1510,11 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
         seq_rm(dest_seq_id, -1, -1);
 
         llama_sbatch sbatch;
-        llama_ubatch batch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false);
+        llama_ubatch ubatch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false);
 
-        batch.n_tokens = cell_count;
+        ubatch.n_tokens = cell_count;
+        ubatch.n_seq_tokens = cell_count;
+        ubatch.n_seqs = 1;
 
         for (uint32_t i = 0; i < cell_count; ++i) {
             llama_pos pos;
@@ -1469,18 +1534,18 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
                 io.read_to(&seq_id, sizeof(seq_id));
             }
 
-            batch.pos[i]      = pos;
-            batch.n_seq_id[i] = n_seq_id;
-            batch.seq_id[i]   = &dest_seq_id;
+            ubatch.pos[i]      = pos;
+            ubatch.n_seq_id[i] = n_seq_id;
+            ubatch.seq_id[i]   = &dest_seq_id;
         }
 
-        const auto head_cur = find_slot(batch);
+        const auto head_cur = find_slot(ubatch);
         if (head_cur < 0) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
         }
 
-        apply_ubatch(head_cur, batch);
+        apply_ubatch(head_cur, ubatch);
 
         // keep the head at the old position because we will read the KV data into it in state_read_data()
         head = head_cur;
@@ -1488,8 +1553,8 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
         // DEBUG CHECK: head_cur should be our first cell, head_cur + cell_count - 1 should be our last cell (verify seq_id and pos values)
         // Assume that this is one contiguous block of cells
         GGML_ASSERT(head_cur + cell_count <= cells.size());
-        GGML_ASSERT(cells.pos_get(head_cur)                  == batch.pos[0]);
-        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == batch.pos[cell_count - 1]);
+        GGML_ASSERT(cells.pos_get(head_cur)                  == ubatch.pos[0]);
+        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == ubatch.pos[cell_count - 1]);
         GGML_ASSERT(cells.seq_has(head_cur,                  dest_seq_id));
         GGML_ASSERT(cells.seq_has(head_cur + cell_count - 1, dest_seq_id));
     } else {
@@ -1556,7 +1621,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
     for (const auto & layer : layers) {
         const uint32_t il = layer.il;
 
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
 
         // Read type of key
         int32_t k_type_i_ref;
@@ -1586,7 +1651,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
 
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             // Read type of value
             int32_t v_type_i_ref;
@@ -1616,7 +1681,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
 
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             // Read type of value
             int32_t v_type_i_ref;
@@ -1674,7 +1739,7 @@ llama_kv_cache_unified_state::llama_kv_cache_unified_state(
         llama_context * lctx,
         bool do_shift,
         defrag_info dinfo) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), lctx(lctx), do_shift(do_shift), dinfo(std::move(dinfo)) {
-    if (!do_shift && dinfo.empty()) {
+    if (!do_shift && this->dinfo.empty()) {
         status = LLAMA_MEMORY_STATUS_NO_UPDATE;
     }
 }

src/llama-kv-cache-unified.h

@@ -59,8 +59,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 
@@ -158,6 +157,8 @@ private:
     // SWA
     const uint32_t n_swa = 0;
 
+    int debug = 0;
+
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<ggml_context_ptr>        ctxs;

src/llama-kv-cells.h

@@ -23,7 +23,7 @@ public:
 
         used.clear();
 
-        for (uint32_t s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (uint32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
             seq_pos[s].clear();
         }
     }
@@ -240,7 +240,7 @@ public:
     llama_seq_id seq_get(uint32_t i) const {
         assert(seq[i].count() == 1);
 
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 return s;
             }
@@ -253,7 +253,7 @@ public:
     // return -1 if the sequence is not present
     llama_pos seq_pos_min(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
@@ -266,7 +266,7 @@ public:
     // return -1 if the sequence is not present
     llama_pos seq_pos_max(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
@@ -384,20 +384,20 @@ private:
     //
     std::vector<llama_pos> shift;
 
-    using bits_t = std::bitset<LLAMA_MAX_PARALLEL_SEQUENCES>;
+    using bits_t = std::bitset<LLAMA_MAX_SEQ>;
 
     // the bitset seq[i] tells us which sequences are currently occupying the i-th cell
     std::vector<bits_t> seq;
 
     // the set seq_pos[s] tells us which positions are currently present for sequence s
     // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache
-    std::set<llama_pos> seq_pos[LLAMA_MAX_PARALLEL_SEQUENCES];
+    std::set<llama_pos> seq_pos[LLAMA_MAX_SEQ];
 
     // helper functions for updating `seq_pos`, once cell at a time:
 
     // remove cell i
     void seq_pos_rm(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 seq_pos[s].erase(pos[i]);
             }
@@ -406,7 +406,7 @@ private:
 
     // add cell i
     void seq_pos_add(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 seq_pos[s].insert(pos[i]);
             }

src/llama-memory-hybrid.cpp

@@ -0,0 +1,247 @@
+#include "llama-memory-hybrid.h"
+
+#include "llama-impl.h"
+#include "llama-model.h"
+#include "llama-context.h"
+
+//
+// llama_memory_hybrid
+//
+
+llama_memory_hybrid::llama_memory_hybrid(
+    const llama_model & model,
+                         /* attn */
+            ggml_type    type_k,
+            ggml_type    type_v,
+                 bool    v_trans,
+             uint32_t    kv_size,
+             uint32_t    n_pad,
+             uint32_t    n_swa,
+       llama_swa_type    swa_type,
+                         /* recurrent */
+            ggml_type    type_r,
+            ggml_type    type_s,
+             uint32_t    rs_size,
+                         /* common */
+             uint32_t    n_seq_max,
+                 bool    offload,
+                         /* layer filters */
+      layer_filter_cb && filter_attn,
+      layer_filter_cb && filter_recr) :
+    hparams(model.hparams),
+    mem_attn(new llama_kv_cache_unified(
+        model,
+        filter_attn == nullptr ?
+            [&](int32_t il) { return !model.hparams.is_recurrent(il); }
+            : filter_attn,
+        type_k,
+        type_v,
+        v_trans,
+        offload,
+        kv_size,
+        n_seq_max,
+        n_pad,
+        n_swa,
+        swa_type
+    )),
+    mem_recr(new llama_memory_recurrent(
+        model,
+        filter_recr == nullptr ?
+            [&](int32_t il) { return model.hparams.is_recurrent(il); }
+            : filter_recr,
+        type_r,
+        type_s,
+        offload,
+        rs_size,
+        n_seq_max
+    )) {}
+
+llama_memory_state_ptr llama_memory_hybrid::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled) {
+
+    // since this includes a recurrent cache, we cannot use split_simple
+    auto sbatch = llama_sbatch(batch, hparams.n_embd, false);
+
+    // follow the recurrent pattern for creating the ubatch splits
+    std::vector<llama_ubatch> ubatches;
+    while (sbatch.n_tokens > 0) {
+        llama_ubatch ubatch;
+
+        if (embd_pooled) {
+            // Pooled embeddings cannot be split across ubatches (yet)
+            ubatch = sbatch.split_seq(n_ubatch);
+        } else {
+            ubatch = sbatch.split_equal(n_ubatch);
+        }
+
+        ubatches.push_back(ubatch);
+    }
+
+    // prepare the recurrent batches first
+    if (!mem_recr->prepare(ubatches)) {
+        // TODO: will the recurrent cache be in an undefined state at this point?
+        LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);
+        return std::make_unique<llama_memory_hybrid_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+    }
+
+    // prepare the attention cache
+    auto heads_attn = mem_attn->prepare(ubatches);
+    if (heads_attn.empty()) {
+        LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__);
+        return std::make_unique<llama_memory_hybrid_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+    }
+
+    return std::make_unique<llama_memory_hybrid_state>(
+        this, std::move(sbatch), std::move(heads_attn), std::move(ubatches));
+}
+
+llama_memory_state_ptr llama_memory_hybrid::init_full() {
+    return std::make_unique<llama_memory_hybrid_state>(this);
+}
+
+llama_memory_state_ptr llama_memory_hybrid::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_memory_hybrid_state>(this, lctx, optimize);
+}
+
+bool llama_memory_hybrid::get_can_shift() const {
+    // Shifting is trivially supported for recurrent
+    return mem_attn->get_can_shift();
+}
+
+void llama_memory_hybrid::clear(bool data) {
+    mem_attn->clear(data);
+    mem_recr->clear(data);
+}
+
+bool llama_memory_hybrid::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    // Try removing from the recurrent cache first since it may fail. If it does
+    // fail, the cache will not have been mutated.
+    if (!mem_recr->seq_rm(seq_id, p0, p1)) {
+        return false;
+    }
+    return mem_attn->seq_rm(seq_id, p0, p1);
+}
+
+void llama_memory_hybrid::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    mem_attn->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    mem_recr->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_memory_hybrid::seq_keep(llama_seq_id seq_id) {
+    mem_attn->seq_keep(seq_id);
+    mem_recr->seq_keep(seq_id);
+}
+
+void llama_memory_hybrid::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    mem_attn->seq_add(seq_id, p0, p1, shift);
+    mem_recr->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_memory_hybrid::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    mem_attn->seq_div(seq_id, p0, p1, d);
+    mem_recr->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_memory_hybrid::seq_pos_min(llama_seq_id seq_id) const {
+    // the min of the total cache is the max of the two caches' min values
+    return std::max(mem_attn->seq_pos_min(seq_id), mem_recr->seq_pos_min(seq_id));
+}
+
+llama_pos llama_memory_hybrid::seq_pos_max(llama_seq_id seq_id) const {
+    // the max of the total cache is the min of the two caches' max values
+    return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
+}
+
+void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+    mem_attn->state_write(io, seq_id);
+    mem_recr->state_write(io, seq_id);
+}
+
+void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+    mem_attn->state_read(io, seq_id);
+    mem_recr->state_read(io, seq_id);
+}
+
+llama_kv_cache_unified * llama_memory_hybrid::get_mem_attn() const {
+    return mem_attn.get();
+}
+
+llama_memory_recurrent * llama_memory_hybrid::get_mem_recr() const {
+    return mem_recr.get();
+}
+
+llama_memory_hybrid_state::llama_memory_hybrid_state(llama_memory_status status) : status(status) {}
+
+llama_memory_hybrid_state::llama_memory_hybrid_state(llama_memory_hybrid * mem) :
+    state_attn(mem->get_mem_attn()->init_full()),
+    state_recr(mem->get_mem_recr()->init_full()),
+    status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) {
+}
+
+llama_memory_hybrid_state::llama_memory_hybrid_state(
+        llama_memory_hybrid * mem,
+              llama_context * lctx,
+                       bool   optimize) :
+    state_attn(mem->get_mem_attn()->init_update(lctx, optimize)),
+    state_recr(mem->get_mem_recr()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) {
+}
+
+llama_memory_hybrid_state::llama_memory_hybrid_state(
+              llama_memory_hybrid * mem,
+                     llama_sbatch   sbatch,
+            std::vector<uint32_t>   heads_attn,
+        std::vector<llama_ubatch>   ubatches) :
+    sbatch(std::move(sbatch)),
+    ubatches(std::move(ubatches)),
+    // note: here we copy the ubatches. not sure if this is ideal
+    state_attn(new llama_kv_cache_unified_state(mem->get_mem_attn(), {}, std::move(heads_attn), this->ubatches)),
+    state_recr(new llama_memory_recurrent_state(mem->get_mem_recr(), {},                        this->ubatches)),
+    status(LLAMA_MEMORY_STATUS_SUCCESS) {
+}
+
+bool llama_memory_hybrid_state::next() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    state_attn->next();
+    state_recr->next();
+
+    if (++i_next >= ubatches.size()) {
+        return false;
+    }
+
+    return true;
+}
+
+bool llama_memory_hybrid_state::apply() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    bool res = true;
+
+    res = res & state_attn->apply();
+    res = res & state_recr->apply();
+
+    return res;
+}
+
+std::vector<int64_t> & llama_memory_hybrid_state::out_ids() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return sbatch.out_ids;
+}
+
+llama_memory_status llama_memory_hybrid_state::get_status() const {
+    return status;
+}
+
+const llama_ubatch & llama_memory_hybrid_state::get_ubatch() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+    return ubatches[i_next];
+}
+
+const llama_kv_cache_unified_state * llama_memory_hybrid_state::get_state_attn() const {
+    return static_cast<const llama_kv_cache_unified_state *>(state_attn.get());
+}
+
+const llama_memory_recurrent_state * llama_memory_hybrid_state::get_state_recr() const {
+    return static_cast<const llama_memory_recurrent_state *>(state_recr.get());
+}

src/llama-memory-hybrid.h

@@ -0,0 +1,143 @@
+#pragma once
+
+#include "llama-batch.h"
+#include "llama-graph.h"
+#include "llama-kv-cache-unified.h"
+#include "llama-memory.h"
+#include "llama-memory-recurrent.h"
+
+#include <memory>
+#include <vector>
+
+//
+// llama_memory_hybrid
+//
+
+// utilizes instances of llama_memory_recurrent and llama_kv_cache_unified to
+//   support models where each layer may be either attention-based or recurrent
+
+class llama_memory_hybrid : public llama_memory_i {
+public:
+
+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
+    llama_memory_hybrid(
+        const llama_model & model,
+                            /* attn */
+                ggml_type    type_k,
+                ggml_type    type_v,
+                     bool    v_trans,
+                 uint32_t    kv_size,
+                 uint32_t    n_pad,
+                 uint32_t    n_swa,
+           llama_swa_type    swa_type,
+                             /* recurrent */
+                ggml_type    type_r,
+                ggml_type    type_s,
+                 uint32_t    rs_size,
+                             /* common */
+                 uint32_t    n_seq_max,
+                     bool    offload,
+                             /* layer filters */
+          layer_filter_cb && filter_attn = nullptr,
+          layer_filter_cb && filter_recr = nullptr);
+
+    ~llama_memory_hybrid() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    llama_memory_state_ptr init_batch(
+            const llama_batch & batch,
+            uint32_t n_ubatch,
+            bool embd_pooled) override;
+
+    llama_memory_state_ptr init_full() override;
+
+    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id)                                                          override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+
+    //
+    // llama_memory_hybrid specific API
+    //
+
+    llama_kv_cache_unified * get_mem_attn() const;
+    llama_memory_recurrent * get_mem_recr() const;
+
+private:
+    const llama_hparams & hparams;
+
+    const std::unique_ptr<llama_kv_cache_unified> mem_attn;
+    const std::unique_ptr<llama_memory_recurrent> mem_recr;
+};
+
+class llama_memory_hybrid_state : public llama_memory_state_i {
+public:
+    // init failure
+    explicit llama_memory_hybrid_state(llama_memory_status status);
+
+    // init full
+    explicit llama_memory_hybrid_state(llama_memory_hybrid * mem);
+
+    // init update
+    explicit llama_memory_hybrid_state(
+        llama_memory_hybrid * mem,
+              llama_context * lctx,
+                       bool   optimize);
+
+    // init success
+    llama_memory_hybrid_state(
+              llama_memory_hybrid * mem,
+                     llama_sbatch   sbatch,
+            std::vector<uint32_t>   heads_attn,
+        std::vector<llama_ubatch>   ubatches);
+
+    ~llama_memory_hybrid_state() = default;
+
+    bool next()  override;
+    bool apply() override;
+
+    std::vector<int64_t> & out_ids() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
+
+    //
+    // llama_memory_hybrid_state
+    //
+
+    const llama_kv_cache_unified_state * get_state_attn() const;
+    const llama_memory_recurrent_state * get_state_recr() const;
+
+private:
+    llama_sbatch sbatch;
+
+    // the index of the next ubatch to process
+    size_t i_next = 0;
+
+    std::vector<llama_ubatch> ubatches;
+
+    const llama_memory_state_ptr state_attn;
+    const llama_memory_state_ptr state_recr;
+
+    const llama_memory_status status;
+};

src/llama-memory-recurrent.cpp

@@ -1,4 +1,4 @@
-#include "llama-kv-cache-recurrent.h"
+#include "llama-memory-recurrent.h"
 
 #include "llama-impl.h"
 #include "llama-io.h"
@@ -12,27 +12,28 @@
 #include <stdexcept>
 
 //
-// llama_kv_cache_recurrent
+// llama_memory_recurrent
 //
 
-llama_kv_cache_recurrent::llama_kv_cache_recurrent(
-        const llama_model & model,
-                ggml_type   type_k,
-                ggml_type   type_v,
-                     bool   offload,
-                 uint32_t   kv_size,
-                 uint32_t   n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
+llama_memory_recurrent::llama_memory_recurrent(
+        const llama_model &  model,
+          layer_filter_cb && filter,
+                ggml_type    type_r,
+                ggml_type    type_s,
+                     bool    offload,
+                 uint32_t    mem_size,
+                 uint32_t    n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
     const int32_t n_layer = hparams.n_layer;
 
-    LLAMA_LOG_INFO("%s: kv_size = %u, n_seq_max = %u, type_k = '%s', type_v = '%s', n_layer = %d\n",
-            __func__, kv_size, n_seq_max, ggml_type_name(type_k), ggml_type_name(type_v), n_layer);
+    LLAMA_LOG_INFO("%s: mem_size = %u, n_seq_max = %u, type_r = '%s', type_s = '%s', n_layer = %d\n",
+            __func__, mem_size, n_seq_max, ggml_type_name(type_r), ggml_type_name(type_s), n_layer);
 
     head = 0;
-    size = kv_size;
+    size = mem_size;
     used = 0;
 
     cells.clear();
-    cells.resize(kv_size);
+    cells.resize(mem_size);
 
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
@@ -59,12 +60,14 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
         return it->second;
     };
 
-    k_l.reserve(n_layer);
-    v_l.reserve(n_layer);
+    r_l.resize(n_layer);
+    s_l.resize(n_layer);
 
     for (int i = 0; i < n_layer; i++) {
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
-        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
+        if (filter && !filter(i)) {
+            LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, i);
+            continue;
+        }
 
         const char * dev_name = "CPU";
 
@@ -84,12 +87,12 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
             throw std::runtime_error("failed to create ggml context for kv cache");
         }
 
-        ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
-        ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
-        ggml_format_name(k, "cache_k_l%d", i);
-        ggml_format_name(v, "cache_v_l%d", i);
-        k_l.push_back(k);
-        v_l.push_back(v);
+        ggml_tensor * r = ggml_new_tensor_1d(ctx, type_r, hparams.n_embd_r()*mem_size);
+        ggml_tensor * s = ggml_new_tensor_1d(ctx, type_s, hparams.n_embd_s()*mem_size);
+        ggml_format_name(r, "cache_r_l%d", i);
+        ggml_format_name(s, "cache_s_l%d", i);
+        r_l[i] = r;
+        s_l[i] = s;
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
@@ -107,17 +110,17 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
     }
 
     {
-        const size_t memory_size_k = size_k_bytes();
-        const size_t memory_size_v = size_v_bytes();
+        const size_t memory_size_r = size_r_bytes();
+        const size_t memory_size_s = size_s_bytes();
 
-        LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
-                (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f),
-                ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
-                ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
+        LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, R (%s): %7.2f MiB, S (%s): %7.2f MiB\n", __func__,
+                (float)(memory_size_r + memory_size_s) / (1024.0f * 1024.0f),
+                ggml_type_name(type_r), (float)memory_size_r / (1024.0f * 1024.0f),
+                ggml_type_name(type_s), (float)memory_size_s / (1024.0f * 1024.0f));
     }
 }
 
-void llama_kv_cache_recurrent::clear(bool data) {
+void llama_memory_recurrent::clear(bool data) {
     for (int32_t i = 0; i < (int32_t) size; ++i) {
         cells[i].pos = -1;
         cells[i].seq_id.clear();
@@ -135,7 +138,7 @@ void llama_kv_cache_recurrent::clear(bool data) {
     }
 }
 
-bool llama_kv_cache_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
     uint32_t new_head = size;
 
     if (p0 < 0) {
@@ -154,7 +157,7 @@ bool llama_kv_cache_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_p
     if (0 <= seq_id) {
         int32_t & tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
-            const kv_cell & cell = cells[tail_id];
+            const auto & cell = cells[tail_id];
             // partial intersection is invalid
             if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) {
                 return false;
@@ -202,7 +205,7 @@ bool llama_kv_cache_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_p
     return true;
 }
 
-void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+void llama_memory_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     if (seq_id_src == seq_id_dst) {
         return;
     }
@@ -216,11 +219,11 @@ void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_
     }
 
     if ((uint32_t) seq_id_dst < size && (uint32_t) seq_id_src < size) {
-        kv_cell & tail_src = cells[seq_id_src];
-        kv_cell & tail_dst = cells[seq_id_dst];
+        auto & tail_src = cells[seq_id_src];
+        auto & tail_dst = cells[seq_id_dst];
         if (tail_dst.tail >= 0) {
             // clear destination seq_id if it wasn't empty
-            kv_cell & cell_dst = cells[tail_dst.tail];
+            auto & cell_dst = cells[tail_dst.tail];
 
             cell_dst.seq_id.erase(seq_id_dst);
             tail_dst.tail = -1;
@@ -231,7 +234,7 @@ void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_
             }
         }
         if (tail_src.tail >= 0) {
-            kv_cell & cell_src = cells[tail_src.tail];
+            auto & cell_src = cells[tail_src.tail];
 
             cell_src.seq_id.insert(seq_id_dst);
             tail_dst.tail = tail_src.tail;
@@ -239,7 +242,7 @@ void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_
     }
 }
 
-void llama_kv_cache_recurrent::seq_keep(llama_seq_id seq_id) {
+void llama_memory_recurrent::seq_keep(llama_seq_id seq_id) {
     uint32_t new_head = size;
 
     for (uint32_t i = 0; i < size; ++i) {
@@ -271,7 +274,7 @@ void llama_kv_cache_recurrent::seq_keep(llama_seq_id seq_id) {
     }
 }
 
-void llama_kv_cache_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+void llama_memory_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     if (shift == 0) {
         return;
     }
@@ -293,7 +296,7 @@ void llama_kv_cache_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_
     if (0 <= seq_id && seq_id < (int64_t) size) {
         const int32_t tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
-            kv_cell & cell = cells[tail_id];
+            auto & cell = cells[tail_id];
             if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) {
                 cell.pos += shift;
             }
@@ -301,7 +304,7 @@ void llama_kv_cache_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_
     }
 }
 
-void llama_kv_cache_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+void llama_memory_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     if (d == 1) {
         return;
     }
@@ -323,7 +326,7 @@ void llama_kv_cache_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_
     if (0 <= seq_id && seq_id < (int64_t) size) {
         const int32_t tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
-            kv_cell & cell = cells[tail_id];
+            auto & cell = cells[tail_id];
             if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) {
                 cell.pos /= d;
             }
@@ -331,7 +334,7 @@ void llama_kv_cache_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_
     }
 }
 
-llama_pos llama_kv_cache_recurrent::seq_pos_min(llama_seq_id seq_id) const {
+llama_pos llama_memory_recurrent::seq_pos_min(llama_seq_id seq_id) const {
     llama_pos result = std::numeric_limits<llama_pos>::max();
 
     for (uint32_t i = 0; i < size; ++i) {
@@ -347,7 +350,7 @@ llama_pos llama_kv_cache_recurrent::seq_pos_min(llama_seq_id seq_id) const {
     return result;
 }
 
-llama_pos llama_kv_cache_recurrent::seq_pos_max(llama_seq_id seq_id) const {
+llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
     llama_pos result = -1;
 
     for (uint32_t i = 0; i < size; ++i) {
@@ -359,18 +362,16 @@ llama_pos llama_kv_cache_recurrent::seq_pos_max(llama_seq_id seq_id) const {
     return result;
 }
 
-llama_memory_state_ptr llama_kv_cache_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled, bool logits_all) {
-    GGML_UNUSED(embd_pooled);
-
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, false, logits_all);
+llama_memory_state_ptr llama_memory_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) {
+    auto sbatch = llama_sbatch(batch, hparams.n_embd, false);
 
     std::vector<llama_ubatch> ubatches;
 
     while (sbatch.n_tokens > 0) {
         llama_ubatch ubatch;
 
-        if (embd_pooled) {
-            // Pooled embeddings cannot be split across ubatches (yet)
+        if (embd_all) {
+            // if all tokens are output, split by sequence
             ubatch = sbatch.split_seq(n_ubatch);
         } else {
             ubatch = sbatch.split_equal(n_ubatch);
@@ -380,24 +381,24 @@ llama_memory_state_ptr llama_kv_cache_recurrent::init_batch(const llama_batch &
     }
 
     if (!prepare(ubatches)) {
-        return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        return std::make_unique<llama_memory_recurrent_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
     }
 
-    return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_SUCCESS, this, std::move(sbatch), std::move(ubatches));
+    return std::make_unique<llama_memory_recurrent_state>(this, std::move(sbatch), std::move(ubatches));
 }
 
-llama_memory_state_ptr llama_kv_cache_recurrent::init_full() {
-    return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_SUCCESS, this);
+llama_memory_state_ptr llama_memory_recurrent::init_full() {
+    return std::make_unique<llama_memory_recurrent_state>(this);
 }
 
-llama_memory_state_ptr llama_kv_cache_recurrent::init_update(llama_context * lctx, bool optimize) {
+llama_memory_state_ptr llama_memory_recurrent::init_update(llama_context * lctx, bool optimize) {
     GGML_UNUSED(lctx);
     GGML_UNUSED(optimize);
 
-    return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_NO_UPDATE);
+    return std::make_unique<llama_memory_recurrent_state>(LLAMA_MEMORY_STATUS_NO_UPDATE);
 }
 
-bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatches) {
+bool llama_memory_recurrent::prepare(const std::vector<llama_ubatch> & ubatches) {
     // simply remember the full state because it is very small for this type of cache
     // TODO: optimize
     auto org_cells = cells;
@@ -406,21 +407,12 @@ bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatche
 
     bool success = true;
 
-    // TODO: here we have to verify that all ubatches can fit in the cells
-    //       however, the current implementation is broken because it relies on s_copy() and s_mask() to update the cells
-    //         during the compute of each ubatch. to reproduce, uncomment the following loop and run:
-    //
-    //           $ llama-parallel -m ./mamba-130m/ggml-model-f16.gguf -np 5 -ns 8
-    //
-    //       recovery from failures when the batch does not fit in the KV cache will not work correctly until this is fixed
-    //
-    GGML_UNUSED(ubatches);
-    //for (const auto & ubatch : ubatches) {
-    //    if (!find_slot(ubatch)) {
-    //        success = false;
-    //        break;
-    //    }
-    //}
+    for (const auto & ubatch : ubatches) {
+        if (!find_slot(ubatch)) {
+            success = false;
+            break;
+        }
+    }
 
     // restore the original state
     cells = std::move(org_cells);
@@ -430,15 +422,14 @@ bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatche
     return success;
 }
 
-bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
-    const uint32_t n_tokens = ubatch.n_tokens;
-    const uint32_t n_seqs   = ubatch.n_seqs;
+bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) {
+    const uint32_t n_seqs = ubatch.n_seqs;
 
     const uint32_t n_seq_tokens = ubatch.n_seq_tokens;
 
     // if we have enough unused cells before the current head ->
     //   better to start searching from the beginning of the cache, hoping to fill it
-    if (head > used + 2*n_tokens) {
+    if (head > used + 2*n_seqs) {
         head = 0;
     }
 
@@ -465,9 +456,9 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
                 return false;
             }
             if (j > 0) {
-                kv_cell & seq = cells[seq_id];
+                auto & seq = cells[seq_id];
                 if (seq.tail >= 0) {
-                    kv_cell & cell = cells[seq.tail];
+                    auto & cell = cells[seq.tail];
                     // clear cells from seq_ids that become shared
                     // (should not normally happen, but let's handle it anyway)
                     cell.seq_id.erase(seq_id);
@@ -487,7 +478,7 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
         std::vector<int32_t> tails_verif;
         tails_verif.assign(size, -1);
         for (uint32_t i = 0; i < size; ++i) {
-            kv_cell & cell = cells[i];
+            auto & cell = cells[i];
             for (llama_seq_id seq_id : cell.seq_id) {
                 if (tails_verif[seq_id] != -1) {
                     LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]);
@@ -508,7 +499,7 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     for (uint32_t i = 0; i < size; ++i) {
         if (next_empty_cell >= size) { next_empty_cell -= size; }
-        kv_cell & cell = cells[next_empty_cell];
+        auto & cell = cells[next_empty_cell];
         if (cell.is_empty()) { break; }
         next_empty_cell += 1;
     }
@@ -516,34 +507,34 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
     // find usable cell range
     for (uint32_t s = 0; s < n_seqs; ++s) {
         const llama_seq_id seq_id = ubatch.seq_id[s][0];
-        kv_cell & seq_meta = cells[seq_id];
+        auto & seq_meta = cells[seq_id];
         bool has_cell = false;
         if (seq_meta.tail >= 0) {
-            kv_cell & cell = cells[seq_meta.tail];
+            auto & cell = cells[seq_meta.tail];
             GGML_ASSERT(cell.has_seq_id(seq_id));
             // does this seq_id "own" the cell?
             if (cell.seq_id.size() == 1) { has_cell = true; }
         }
         if (!has_cell) {
-            kv_cell & empty_cell = cells[next_empty_cell];
+            auto & empty_cell = cells[next_empty_cell];
             GGML_ASSERT(empty_cell.is_empty());
             // copy old tail into the empty cell
             if (seq_meta.tail >= 0) {
-                kv_cell & orig_cell = cells[seq_meta.tail];
+                auto & orig_cell = cells[seq_meta.tail];
                 empty_cell.pos = orig_cell.pos;
                 empty_cell.src = orig_cell.src;
                 orig_cell.seq_id.erase(seq_id);
                 empty_cell.seq_id.insert(seq_id); // will be overwritten
+                GGML_ASSERT(!orig_cell.is_empty()); // has at least one remaining seq_id
             }
             seq_meta.tail = next_empty_cell;
             // find next empty cell
             if (s + 1 < n_seqs) {
-                next_empty_cell += 1;
                 for (uint32_t i = 0; i < size; ++i) {
-                    if (next_empty_cell >= size) { next_empty_cell -= size; }
-                    kv_cell & cell = cells[next_empty_cell];
-                    if (cell.is_empty()) { break; }
                     next_empty_cell += 1;
+                    if (next_empty_cell >= size) { next_empty_cell -= size; }
+                    auto & cell = cells[next_empty_cell];
+                    if (cell.is_empty()) { break; }
                 }
             }
         }
@@ -553,22 +544,24 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     // gather and re-order
     for (uint32_t s = 0; s < n_seqs; ++s) {
-        int32_t dst_id = s + min;
-        int32_t src_id = cells[ubatch.seq_id[s][0]].tail;
+        const int32_t dst_id = s + min;
+        const int32_t src_id = cells[ubatch.seq_id[s][0]].tail;
         if (dst_id != src_id) {
-            kv_cell & dst_cell = cells[dst_id];
-            kv_cell & src_cell = cells[src_id];
+            auto & dst_cell = cells[dst_id];
+            auto & src_cell = cells[src_id];
 
             std::swap(dst_cell.pos, src_cell.pos);
             std::swap(dst_cell.src, src_cell.src);
             std::swap(dst_cell.seq_id, src_cell.seq_id);
 
-            // swap tails (assuming they NEVER overlap)
-            for (const llama_seq_id seq_id : src_cell.seq_id) {
-                cells[seq_id].tail = src_id;
-            }
-            for (const llama_seq_id seq_id : dst_cell.seq_id) {
-                cells[seq_id].tail = dst_id;
+            // swap tails
+            for (uint32_t i = 0; i < size; ++i) {
+                int32_t & tail = cells[i].tail;
+                if (tail == src_id) {
+                    tail = dst_id;
+                } else if (tail == dst_id) {
+                    tail = src_id;
+                }
             }
         }
     }
@@ -576,8 +569,8 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
     // update the pos of the used seqs
     for (uint32_t s = 0; s < n_seqs; ++s) {
         const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1];
-        int32_t cell_id = s + min;
-        kv_cell & cell = cells[cell_id];
+        const int32_t cell_id = s + min;
+        auto & cell = cells[cell_id];
 
         if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) {
             // What should happen when the pos backtracks or skips a value?
@@ -594,61 +587,54 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
         }
     }
 
+    // Find first cell without src refs, to use as the zero-ed state
+    {
+        // TODO: bake-in src refcounts in the cell metadata
+        std::vector<int32_t> refcounts(size, 0);
+        for (size_t i = 0; i < size; ++i) {
+            const int32_t src = cells[i].src;
+            if (src >= 0) {
+                refcounts[src] += 1;
+            }
+        }
+
+        rs_z = -1;
+        for (int i = min; i <= max; ++i) {
+            if (refcounts[i] == 0) {
+                rs_z = i;
+                break;
+            }
+        }
+
+        for (int i = min; i <= max; ++i) {
+            if (cells[i].src < 0) {
+                GGML_ASSERT(rs_z >= 0);
+                cells[i].src0 = rs_z;
+            } else {
+                // Stage the source ids for all used cells to allow correct seq_* behavior
+                // and still make these values available when setting the inputs
+                cells[i].src0 = cells[i].src;
+            }
+            cells[i].src = i; // avoid moving or clearing twice
+        }
+    }
+
     // allow getting the range of used cells, from head to head + n
     head = min;
     n    = max - min + 1;
     used = std::count_if(cells.begin(), cells.end(),
-        [](const kv_cell & cell){ return !cell.is_empty(); });
+        [](const mem_cell & cell){ return !cell.is_empty(); });
 
     // sanity check
     return n >= n_seqs;
 }
 
-bool llama_kv_cache_recurrent::get_can_shift() const {
-    return false;
+bool llama_memory_recurrent::get_can_shift() const {
+    // shifting the pos is trivial for recurrent models
+    return true;
 }
 
-int32_t llama_kv_cache_recurrent::s_copy(int i) const {
-    const uint32_t cell_id = i + head;
-
-    //////////////////////////////////////////////
-    // TODO: this should not mutate the KV cache !
-    kv_cell & cell = const_cast<kv_cell &>(cells[cell_id]);
-
-    // prevent out-of-bound sources
-    if (cell.src < 0 || (uint32_t) cell.src >= size) {
-        cell.src = cell_id;
-    }
-
-    int32_t res = cell.src;
-
-    // TODO: do not mutate the KV cache
-    // ensure copy only happens once
-    if (cell.src != (int32_t) cell_id) {
-        cell.src = cell_id;
-    }
-
-    return res;
-}
-
-float llama_kv_cache_recurrent::s_mask(int i) const {
-    const uint32_t cell_id = i + head;
-
-    //////////////////////////////////////////////
-    // TODO: this should not mutate the KV cache !
-    kv_cell & cell = const_cast<kv_cell &>(cells[cell_id]);
-
-    float res = (float) (cell.src >= 0);
-
-    // only clear once
-    if (cell.src < 0) {
-        cell.src = cell_id;
-    }
-
-    return res;
-}
-
-size_t llama_kv_cache_recurrent::total_size() const {
+size_t llama_memory_recurrent::total_size() const {
     size_t size = 0;
     for (const auto & buf : bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
@@ -657,27 +643,31 @@ size_t llama_kv_cache_recurrent::total_size() const {
     return size;
 }
 
-size_t llama_kv_cache_recurrent::size_k_bytes() const {
-    size_t size_k_bytes = 0;
+size_t llama_memory_recurrent::size_r_bytes() const {
+    size_t size_r_bytes = 0;
 
-    for (const auto & k : k_l) {
-        size_k_bytes += ggml_nbytes(k);
+    for (const auto & r : r_l) {
+        if (r != nullptr) {
+            size_r_bytes += ggml_nbytes(r);
+        }
     }
 
-    return size_k_bytes;
+    return size_r_bytes;
 }
 
-size_t llama_kv_cache_recurrent::size_v_bytes() const {
-    size_t size_v_bytes = 0;
+size_t llama_memory_recurrent::size_s_bytes() const {
+    size_t size_s_bytes = 0;
 
-    for (const auto & v : v_l) {
-        size_v_bytes += ggml_nbytes(v);
+    for (const auto & s : s_l) {
+        if (s != nullptr) {
+            size_s_bytes += ggml_nbytes(s);
+        }
     }
 
-    return size_v_bytes;
+    return size_s_bytes;
 }
 
-void llama_kv_cache_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
     std::vector<std::pair<uint32_t, uint32_t>> cell_ranges; // ranges, from inclusive, to exclusive
     uint32_t cell_count = 0;
 
@@ -715,7 +705,7 @@ void llama_kv_cache_recurrent::state_write(llama_io_write_i & io, llama_seq_id s
     state_write_data(io, cell_ranges);
 }
 
-void llama_kv_cache_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
     uint32_t cell_count;
     io.read_to(&cell_count, sizeof(cell_count));
 
@@ -734,7 +724,7 @@ void llama_kv_cache_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq
     }
 }
 
-void llama_kv_cache_recurrent::state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id) const {
+void llama_memory_recurrent::state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id) const {
     for (const auto & range : cell_ranges) {
         for (uint32_t i = range.first; i < range.second; ++i) {
             const auto & cell = cells[i];
@@ -753,87 +743,85 @@ void llama_kv_cache_recurrent::state_write_meta(llama_io_write_i & io, const std
     }
 }
 
-void llama_kv_cache_recurrent::state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const {
-    const uint32_t v_trans = 0;
+void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const {
+    const uint32_t s_trans = 0;
     const uint32_t n_layer = hparams.n_layer;
 
-    io.write(&v_trans, sizeof(v_trans));
-    io.write(&n_layer, sizeof(n_layer));
+    io.write(&s_trans, sizeof(s_trans));
+    io.write(&n_layer,   sizeof(n_layer));
 
     std::vector<uint8_t> tmp_buf;
 
     // Iterate and write all the keys first, each row is a cell
     // Get whole range at a time
     for (uint32_t il = 0; il < n_layer; ++il) {
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
 
         // Write key type
-        const int32_t k_type_i = (int32_t)k_l[il]->type;
-        io.write(&k_type_i, sizeof(k_type_i));
+        const int32_t r_type_i = (int32_t)r_l[il]->type;
+        io.write(&r_type_i, sizeof(r_type_i));
 
         // Write row size of key
-        const uint64_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa);
-        io.write(&k_size_row, sizeof(k_size_row));
+        const uint64_t r_size_row = ggml_row_size(r_l[il]->type, hparams.n_embd_r());
+        io.write(&r_size_row, sizeof(r_size_row));
 
         // Read each range of cells of k_size length each into tmp_buf and write out
         for (const auto & range : cell_ranges) {
             const size_t range_size = range.second - range.first;
-            const size_t buf_size = range_size * k_size_row;
-            io.write_tensor(k_l[il], range.first * k_size_row, buf_size);
+            const size_t buf_size = range_size * r_size_row;
+            io.write_tensor(r_l[il], range.first * r_size_row, buf_size);
         }
     }
 
-    if (!v_trans) {
+    if (!s_trans) {
         for (uint32_t il = 0; il < n_layer; ++il) {
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
 
             // Write value type
-            const int32_t v_type_i = (int32_t)v_l[il]->type;
-            io.write(&v_type_i, sizeof(v_type_i));
+            const int32_t s_type_i = (int32_t)s_l[il]->type;
+            io.write(&s_type_i, sizeof(s_type_i));
 
             // Write row size of value
-            const uint64_t v_size_row = ggml_row_size(v_l[il]->type, n_embd_v_gqa);
-            io.write(&v_size_row, sizeof(v_size_row));
+            const uint64_t s_size_row = ggml_row_size(s_l[il]->type, hparams.n_embd_s());
+            io.write(&s_size_row, sizeof(s_size_row));
 
-            // Read each range of cells of v_size length each into tmp_buf and write out
+            // Read each range of cells of s_size length each into tmp_buf and write out
             for (const auto & range : cell_ranges) {
                 const size_t range_size = range.second - range.first;
-                const size_t buf_size = range_size * v_size_row;
-                io.write_tensor(v_l[il], range.first * v_size_row, buf_size);
+                const size_t buf_size = range_size * s_size_row;
+                io.write_tensor(s_l[il], range.first * s_size_row, buf_size);
             }
         }
     } else {
         // When v is transposed, we also need the element size and get the element ranges from each row
-        const uint32_t kv_size = size;
+        const uint32_t mem_size = size;
         for (uint32_t il = 0; il < n_layer; ++il) {
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_s = hparams.n_embd_s();
 
             // Write value type
-            const int32_t v_type_i = (int32_t)v_l[il]->type;
-            io.write(&v_type_i, sizeof(v_type_i));
+            const int32_t s_type_i = (int32_t)s_l[il]->type;
+            io.write(&s_type_i, sizeof(s_type_i));
 
             // Write element size
-            const uint32_t v_size_el = ggml_type_size(v_l[il]->type);
-            io.write(&v_size_el, sizeof(v_size_el));
+            const uint32_t s_size_el = ggml_type_size(s_l[il]->type);
+            io.write(&s_size_el, sizeof(s_size_el));
 
             // Write GQA embedding size
-            io.write(&n_embd_v_gqa, sizeof(n_embd_v_gqa));
+            io.write(&n_embd_s, sizeof(n_embd_s));
 
             // For each row, we get the element values of each cell
-            for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+            for (uint32_t j = 0; j < n_embd_s; ++j) {
                 // Read each range of cells of v_size_el length each into tmp_buf and write out
                 for (const auto & range : cell_ranges) {
                     const size_t range_size = range.second - range.first;
-                    const size_t src_offset = (range.first + j * kv_size) * v_size_el;
-                    const size_t buf_size = range_size * v_size_el;
-                    io.write_tensor(v_l[il], src_offset, buf_size);
+                    const size_t src_offset = (range.first + j * mem_size) * s_size_el;
+                    const size_t buf_size = range_size * s_size_el;
+                    io.write_tensor(s_l[il], src_offset, buf_size);
                 }
             }
         }
     }
 }
 
-bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) {
+bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) {
     if (dest_seq_id != -1) {
         // single sequence
 
@@ -886,7 +874,7 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
         clear(true);
 
         for (uint32_t i = 0; i < cell_count; ++i) {
-            kv_cell & cell = cells[i];
+            auto & cell = cells[i];
 
             llama_pos pos;
             uint32_t  n_seq_id;
@@ -900,7 +888,7 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
                 llama_seq_id seq_id;
                 io.read_to(&seq_id, sizeof(seq_id));
 
-                // TODO: llama_kv_cache_recurrent should have a notion of max sequences
+                // TODO: llama_memory_recurrent should have a notion of max sequences
                 //if (seq_id < 0 || (uint32_t) seq_id >= llama_n_seq_max(ctx)) {
                 if (seq_id < 0) {
                     //LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, llama_n_seq_max(ctx));
@@ -932,10 +920,10 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
     return true;
 }
 
-bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell_count) {
-    uint32_t v_trans;
+bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell_count) {
+    uint32_t s_trans;
     uint32_t n_layer;
-    io.read_to(&v_trans, sizeof(v_trans));
+    io.read_to(&s_trans, sizeof(s_trans));
     io.read_to(&n_layer, sizeof(n_layer));
 
     if (n_layer != hparams.n_layer) {
@@ -946,102 +934,100 @@ bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t ce
         LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, size);
         return false;
     }
-    if (false != (bool) v_trans) {
-        LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__);
+    if (false != (bool) s_trans) {
+        LLAMA_LOG_ERROR("%s: incompatible s transposition\n", __func__);
         return false;
     }
 
     // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block
     for (uint32_t il = 0; il < n_layer; ++il) {
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
 
         // Read type of key
-        int32_t k_type_i_ref;
-        io.read_to(&k_type_i_ref, sizeof(k_type_i_ref));
-        const int32_t k_type_i = (int32_t) k_l[il]->type;
-        if (k_type_i != k_type_i_ref) {
-            LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il);
+        int32_t r_type_i_ref;
+        io.read_to(&r_type_i_ref, sizeof(r_type_i_ref));
+        const int32_t r_type_i = (int32_t) r_l[il]->type;
+        if (r_type_i != r_type_i_ref) {
+            LLAMA_LOG_ERROR("%s: mismatched r type (%d != %d, layer %d)\n", __func__, r_type_i, r_type_i_ref, il);
             return false;
         }
 
         // Read row size of key
-        uint64_t k_size_row_ref;
-        io.read_to(&k_size_row_ref, sizeof(k_size_row_ref));
-        const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa);
-        if (k_size_row != k_size_row_ref) {
-            LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, (size_t) k_size_row_ref, il);
+        uint64_t r_size_row_ref;
+        io.read_to(&r_size_row_ref, sizeof(r_size_row_ref));
+        const size_t r_size_row = ggml_row_size(r_l[il]->type, hparams.n_embd_r());
+        if (r_size_row != r_size_row_ref) {
+            LLAMA_LOG_ERROR("%s: mismatched r row size (%zu != %zu, layer %d)\n", __func__, r_size_row, (size_t) r_size_row_ref, il);
             return false;
         }
 
         if (cell_count) {
             // Read and set the keys for the whole cell range
-            ggml_backend_tensor_set(k_l[il], io.read(cell_count * k_size_row), head * k_size_row, cell_count * k_size_row);
+            ggml_backend_tensor_set(r_l[il], io.read(cell_count * r_size_row), head * r_size_row, cell_count * r_size_row);
         }
     }
 
-    if (!v_trans) {
+    if (!s_trans) {
         for (uint32_t il = 0; il < n_layer; ++il) {
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
 
             // Read type of value
-            int32_t v_type_i_ref;
-            io.read_to(&v_type_i_ref, sizeof(v_type_i_ref));
-            const int32_t v_type_i = (int32_t)v_l[il]->type;
-            if (v_type_i != v_type_i_ref) {
-                LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
+            int32_t s_type_i_ref;
+            io.read_to(&s_type_i_ref, sizeof(s_type_i_ref));
+            const int32_t s_type_i = (int32_t)s_l[il]->type;
+            if (s_type_i != s_type_i_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s type (%d != %d, layer %d)\n", __func__, s_type_i, s_type_i_ref, il);
                 return false;
             }
 
             // Read row size of value
-            uint64_t v_size_row_ref;
-            io.read_to(&v_size_row_ref, sizeof(v_size_row_ref));
-            const size_t v_size_row = ggml_row_size(v_l[il]->type, n_embd_v_gqa);
-            if (v_size_row != v_size_row_ref) {
-                LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, (size_t) v_size_row_ref, il);
+            uint64_t s_size_row_ref;
+            io.read_to(&s_size_row_ref, sizeof(s_size_row_ref));
+            const size_t s_size_row = ggml_row_size(s_l[il]->type, hparams.n_embd_s());
+            if (s_size_row != s_size_row_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s row size (%zu != %zu, layer %d)\n", __func__, s_size_row, (size_t) s_size_row_ref, il);
                 return false;
             }
 
             if (cell_count) {
                 // Read and set the values for the whole cell range
-                ggml_backend_tensor_set(v_l[il], io.read(cell_count * v_size_row), head * v_size_row, cell_count * v_size_row);
+                ggml_backend_tensor_set(s_l[il], io.read(cell_count * s_size_row), head * s_size_row, cell_count * s_size_row);
             }
         }
     } else {
         // For each layer, read the values for each cell (transposed)
         for (uint32_t il = 0; il < n_layer; ++il) {
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_s = hparams.n_embd_s();
 
             // Read type of value
-            int32_t v_type_i_ref;
-            io.read_to(&v_type_i_ref, sizeof(v_type_i_ref));
-            const int32_t v_type_i = (int32_t)v_l[il]->type;
-            if (v_type_i != v_type_i_ref) {
-                LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
+            int32_t s_type_i_ref;
+            io.read_to(&s_type_i_ref, sizeof(s_type_i_ref));
+            const int32_t s_type_i = (int32_t)s_l[il]->type;
+            if (s_type_i != s_type_i_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s type (%d != %d, layer %d)\n", __func__, s_type_i, s_type_i_ref, il);
                 return false;
             }
 
             // Read element size of value
-            uint32_t v_size_el_ref;
-            io.read_to(&v_size_el_ref, sizeof(v_size_el_ref));
-            const size_t v_size_el = ggml_type_size(v_l[il]->type);
-            if (v_size_el != v_size_el_ref) {
-                LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, (size_t) v_size_el_ref, il);
+            uint32_t s_size_el_ref;
+            io.read_to(&s_size_el_ref, sizeof(s_size_el_ref));
+            const size_t s_size_el = ggml_type_size(s_l[il]->type);
+            if (s_size_el != s_size_el_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s element size (%zu != %zu, layer %d)\n", __func__, s_size_el, (size_t) s_size_el_ref, il);
                 return false;
             }
 
-            // Read GQA embedding size
-            uint32_t n_embd_v_gqa_ref;
-            io.read_to(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref));
-            if (n_embd_v_gqa != n_embd_v_gqa_ref) {
-                LLAMA_LOG_ERROR("%s: mismatched GQA embedding size (%u != %u, layer %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref, il);
+            // Read state embedding size
+            uint32_t n_embd_s_ref;
+            io.read_to(&n_embd_s_ref, sizeof(n_embd_s_ref));
+            if (n_embd_s != n_embd_s_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s embedding size (%u != %u, layer %d)\n", __func__, n_embd_s, n_embd_s_ref, il);
                 return false;
             }
 
             if (cell_count) {
                 // For each row in the transposed matrix, read the values for the whole cell range
-                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                    const size_t dst_offset = (head + j * size) * v_size_el;
-                    ggml_backend_tensor_set(v_l[il], io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el);
+                for (uint32_t j = 0; j < n_embd_s; ++j) {
+                    const size_t dst_offset = (head + j * size) * s_size_el;
+                    ggml_backend_tensor_set(s_l[il], io.read(cell_count * s_size_el), dst_offset, cell_count * s_size_el);
                 }
             }
         }
@@ -1051,25 +1037,23 @@ bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t ce
 }
 
 //
-// llama_kv_cache_recurrent_state
+// llama_memory_recurrent_state
 //
 
-llama_kv_cache_recurrent_state::llama_kv_cache_recurrent_state(llama_memory_status status) : status(status) {}
+llama_memory_recurrent_state::llama_memory_recurrent_state(llama_memory_status status) : status(status) {}
 
-llama_kv_cache_recurrent_state::llama_kv_cache_recurrent_state(
-        llama_memory_status status,
-        llama_kv_cache_recurrent * kv) : status(status), kv(kv), is_full(true) {
+llama_memory_recurrent_state::llama_memory_recurrent_state(
+        llama_memory_recurrent * mem) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), is_full(true) {
 }
 
-llama_kv_cache_recurrent_state::llama_kv_cache_recurrent_state(
-        llama_memory_status status,
-        llama_kv_cache_recurrent * kv,
+llama_memory_recurrent_state::llama_memory_recurrent_state(
+        llama_memory_recurrent * mem,
         llama_sbatch sbatch,
-        std::vector<llama_ubatch> ubatches) : status(status), kv(kv), sbatch(std::move(sbatch)), ubatches(std::move(ubatches)) {}
+        std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), sbatch(std::move(sbatch)), ubatches(std::move(ubatches)) {}
 
-llama_kv_cache_recurrent_state::~llama_kv_cache_recurrent_state() = default;
+llama_memory_recurrent_state::~llama_memory_recurrent_state() = default;
 
-bool llama_kv_cache_recurrent_state::next() {
+bool llama_memory_recurrent_state::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     if (++i_next >= ubatches.size()) {
@@ -1079,54 +1063,54 @@ bool llama_kv_cache_recurrent_state::next() {
     return true;
 }
 
-bool llama_kv_cache_recurrent_state::apply() {
+bool llama_memory_recurrent_state::apply() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    kv->find_slot(ubatches[i_next]);
+    mem->find_slot(ubatches[i_next]);
 
     return true;
 }
 
-std::vector<int64_t> & llama_kv_cache_recurrent_state::out_ids() {
+std::vector<int64_t> & llama_memory_recurrent_state::out_ids() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     return sbatch.out_ids;
 }
 
-llama_memory_status llama_kv_cache_recurrent_state::get_status() const {
+llama_memory_status llama_memory_recurrent_state::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_kv_cache_recurrent_state::get_ubatch() const {
+const llama_ubatch & llama_memory_recurrent_state::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     return ubatches[i_next];
 }
 
-uint32_t llama_kv_cache_recurrent_state::get_n_kv() const {
-    return is_full ? kv->size : kv->n;
+uint32_t llama_memory_recurrent_state::get_n_rs() const {
+    return is_full ? mem->size : mem->n;
 }
 
-uint32_t llama_kv_cache_recurrent_state::get_head() const {
-    return is_full ? 0 : kv->head;
+uint32_t llama_memory_recurrent_state::get_head() const {
+    return is_full ? 0 : mem->head;
 }
 
-uint32_t llama_kv_cache_recurrent_state::get_size() const {
-    return kv->size;
+int32_t llama_memory_recurrent_state::get_rs_z() const {
+    return is_full ? 0 : mem->rs_z;
 }
 
-ggml_tensor * llama_kv_cache_recurrent_state::get_k_l(int32_t il) const {
-    return kv->k_l[il];
+uint32_t llama_memory_recurrent_state::get_size() const {
+    return mem->size;
 }
 
-ggml_tensor * llama_kv_cache_recurrent_state::get_v_l(int32_t il) const {
-    return kv->v_l[il];
+ggml_tensor * llama_memory_recurrent_state::get_r_l(int32_t il) const {
+    return mem->r_l[il];
 }
 
-int32_t llama_kv_cache_recurrent_state::s_copy(int i) const {
-    return kv->s_copy(i);
+ggml_tensor * llama_memory_recurrent_state::get_s_l(int32_t il) const {
+    return mem->s_l[il];
 }
 
-float llama_kv_cache_recurrent_state::s_mask(int i) const {
-    return kv->s_mask(i);
+int32_t llama_memory_recurrent_state::s_copy(int i) const {
+    return  mem->cells[i + mem->head].src0;
 }

src/llama-memory-recurrent.h

@@ -8,22 +8,27 @@
 #include <vector>
 
 //
-// llama_kv_cache_recurrent
+// llama_memory_recurrent
 //
 
-// TODO: extract the KV cache state used for graph computation into llama_kv_cache_recurrent_state_i
+// TODO: extract the cache state used for graph computation into llama_memory_recurrent_state_i
 //       see the implementation of llama_kv_cache_unified_state_i for an example how to do it
-class llama_kv_cache_recurrent : public llama_memory_i {
+class llama_memory_recurrent : public llama_memory_i {
 public:
-    llama_kv_cache_recurrent(
-            const llama_model & model,
-                    ggml_type   type_k,
-                    ggml_type   type_v,
-                         bool   offload,
-                     uint32_t   kv_size,
-                     uint32_t   n_seq_max);
 
-    ~llama_kv_cache_recurrent() = default;
+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
+    llama_memory_recurrent(
+            const llama_model &  model,
+              layer_filter_cb && filter,
+                    ggml_type    type_r,
+                    ggml_type    type_s,
+                         bool    offload,
+                     uint32_t    mem_size,
+                     uint32_t    n_seq_max);
+
+    ~llama_memory_recurrent() = default;
 
     //
     // llama_memory_i
@@ -32,8 +37,7 @@ public:
     llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
+            bool embd_all) override;
 
     llama_memory_state_ptr init_full() override;
 
@@ -52,15 +56,11 @@ public:
 
     bool prepare(const std::vector<llama_ubatch> & ubatches);
 
-    // find a contiguous slot of kv cells and emplace the ubatch there
+    // find a contiguous slot of memory cells and emplace the ubatch there
     bool find_slot(const llama_ubatch & ubatch);
 
     bool get_can_shift() const override;
 
-    // TODO: temporary methods - they are not really const as they do const_cast<>, fix this
-    int32_t s_copy(int i) const;
-    float   s_mask(int i) const;
-
     // state write/load
 
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
@@ -73,10 +73,14 @@ public:
     // computed before each graph build
     uint32_t n = 0;
 
+    // first zero-ed state
+    int32_t rs_z = -1;
+
     // TODO: optimize for recurrent state needs
-    struct kv_cell {
+    struct mem_cell {
         llama_pos pos  = -1;
-        int32_t   src  = -1; // used to copy states
+        int32_t   src  = -1; // used to know where states should be copied from
+        int32_t   src0 = -1; // like src, but only used when setting the inputs (allowing to copy once)
         int32_t   tail = -1;
 
         std::set<llama_seq_id> seq_id;
@@ -89,15 +93,16 @@ public:
             return seq_id.empty();
         }
 
-        bool is_same_seq(const kv_cell & other) const {
+        bool is_same_seq(const mem_cell & other) const {
             return seq_id == other.seq_id;
         }
     };
 
-    std::vector<kv_cell> cells;
+    std::vector<mem_cell> cells;
 
-    std::vector<ggml_tensor *> k_l; // per layer
-    std::vector<ggml_tensor *> v_l;
+    // per layer
+    std::vector<ggml_tensor *> r_l;
+    std::vector<ggml_tensor *> s_l;
 
 private:
     //const llama_model & model;
@@ -110,8 +115,8 @@ private:
 
     size_t total_size() const;
 
-    size_t size_k_bytes() const;
-    size_t size_v_bytes() const;
+    size_t size_r_bytes() const;
+    size_t size_s_bytes() const;
 
     void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
@@ -120,24 +125,22 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
 
-class llama_kv_cache_recurrent_state : public llama_memory_state_i {
+class llama_memory_recurrent_state : public llama_memory_state_i {
 public:
     // used for errors
-    llama_kv_cache_recurrent_state(llama_memory_status status);
+    llama_memory_recurrent_state(llama_memory_status status);
 
     // used to create a full-cache state
-    llama_kv_cache_recurrent_state(
-            llama_memory_status status,
-            llama_kv_cache_recurrent * kv);
+    llama_memory_recurrent_state(
+            llama_memory_recurrent * mem);
 
     // used to create a state from a batch
-    llama_kv_cache_recurrent_state(
-            llama_memory_status status,
-            llama_kv_cache_recurrent * kv,
+    llama_memory_recurrent_state(
+            llama_memory_recurrent * mem,
             llama_sbatch sbatch,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_recurrent_state();
+    virtual ~llama_memory_recurrent_state();
 
     //
     // llama_memory_state_i
@@ -152,23 +155,23 @@ public:
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_recurrent_state specific API
+    // llama_memory_recurrent_state specific API
     //
 
-    uint32_t get_n_kv() const;
+    uint32_t get_n_rs() const;
     uint32_t get_head() const;
+    int32_t  get_rs_z() const;
     uint32_t get_size() const;
 
-    ggml_tensor * get_k_l(int32_t il) const;
-    ggml_tensor * get_v_l(int32_t il) const;
+    ggml_tensor * get_r_l(int32_t il) const;
+    ggml_tensor * get_s_l(int32_t il) const;
 
     int32_t s_copy(int i) const;
-    float   s_mask(int i) const;
 
 private:
     const llama_memory_status status;
 
-    llama_kv_cache_recurrent * kv;
+    llama_memory_recurrent * mem;
 
     llama_sbatch sbatch;
 

src/llama-memory.h

@@ -73,8 +73,7 @@ struct llama_memory_i {
     virtual llama_memory_state_ptr init_batch(
             const llama_batch & batch,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) = 0;
+            bool embd_all) = 0;
 
     // simulate full cache, used for allocating worst-case compute buffers
     virtual llama_memory_state_ptr init_full() = 0;

src/llama-model.h

@@ -73,6 +73,7 @@ enum llm_type {
     LLM_TYPE_40B,
     LLM_TYPE_65B,
     LLM_TYPE_70B,
+    LLM_TYPE_142B,
     LLM_TYPE_236B,
     LLM_TYPE_290B,
     LLM_TYPE_314B,

src/llama-quant.cpp

@@ -585,7 +585,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) {
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
-                gguf_set_val_i32(ctx_out.get(), o.key, o.val_i64);
+                // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {

src/llama-vocab.cpp

@@ -9,16 +9,16 @@
 
 #include <algorithm>
 #include <cassert>
+#include <cctype>
 #include <cfloat>
-#include <climits>
 #include <cstdarg>
 #include <cstring>
 #include <forward_list>
+#include <limits>
 #include <map>
 #include <queue>
 #include <set>
 #include <unordered_map>
-#include <cctype>
 
 //
 // helpers
@@ -1987,6 +1987,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<|eom_id|>"
                     || t.first == "<EOT>"
                     || t.first == "_<EOT>"
+                    || t.first == "<|end_of_text|>"
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2059,9 +2060,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
     //NOTE: Per token attributes are missing from the GGUF file.
     //TODO: Extract attributes from GGUF file.
     {
-        auto _contains_any = [] (const std::string & str, const std::vector<std::string> & substrs) -> bool {
+        auto _contains_any = [] (const std::string & str, const std::vector<std::string_view> & substrs) -> bool {
             for (const auto & substr : substrs) {
-                if (str.find(substr) < std::string::npos) {
+                if (str.find(substr) != std::string::npos) {
                     return true;
                 }
             }
@@ -2572,6 +2573,10 @@ int32_t llama_vocab::impl::token_to_piece(llama_token token, char * buf, int32_t
     // copy piece chars to output text buffer
     // skip up to 'lstrip' leading spaces before copying
     auto _try_copy = [=] (const char * token, size_t size) -> int32_t {
+        if (size >= static_cast<size_t>(std::numeric_limits<int32_t>::max())) {
+            GGML_ABORT("invalid token size: %zu exceeds int32_t limit", size);
+        }
+
         for (int32_t i = 0; i < lstrip && size && *token == ' '; ++i) {
             token++;
             size--;
@@ -2768,26 +2773,26 @@ void llama_vocab::impl::print_info() const {
     LLAMA_LOG_INFO("%s: n_merges         = %u\n",     __func__, (uint32_t) bpe_ranks.size());
 
     // special tokens
-    if (special_bos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: BOS token        = %d '%s'\n", __func__, special_bos_id,     id_to_token[special_bos_id].text.c_str() );  }
-    if (special_eos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOS token        = %d '%s'\n", __func__, special_eos_id,     id_to_token[special_eos_id].text.c_str() );  }
-    if (special_eot_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOT token        = %d '%s'\n", __func__, special_eot_id,     id_to_token[special_eot_id].text.c_str() );  }
-    if (special_eom_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOM token        = %d '%s'\n", __func__, special_eom_id,     id_to_token[special_eom_id].text.c_str() );  }
-    if (special_unk_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: UNK token        = %d '%s'\n", __func__, special_unk_id,     id_to_token[special_unk_id].text.c_str() );  }
-    if (special_sep_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: SEP token        = %d '%s'\n", __func__, special_sep_id,     id_to_token[special_sep_id].text.c_str() );  }
-    if (special_pad_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: PAD token        = %d '%s'\n", __func__, special_pad_id,     id_to_token[special_pad_id].text.c_str() );  }
-    if (special_mask_id != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: MASK token       = %d '%s'\n", __func__, special_mask_id,    id_to_token[special_mask_id].text.c_str() ); }
+    if (special_bos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: BOS token        = %d '%s'\n", __func__, special_bos_id,     id_to_token.at(special_bos_id).text.c_str() );  }
+    if (special_eos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOS token        = %d '%s'\n", __func__, special_eos_id,     id_to_token.at(special_eos_id).text.c_str() );  }
+    if (special_eot_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOT token        = %d '%s'\n", __func__, special_eot_id,     id_to_token.at(special_eot_id).text.c_str() );  }
+    if (special_eom_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOM token        = %d '%s'\n", __func__, special_eom_id,     id_to_token.at(special_eom_id).text.c_str() );  }
+    if (special_unk_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: UNK token        = %d '%s'\n", __func__, special_unk_id,     id_to_token.at(special_unk_id).text.c_str() );  }
+    if (special_sep_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: SEP token        = %d '%s'\n", __func__, special_sep_id,     id_to_token.at(special_sep_id).text.c_str() );  }
+    if (special_pad_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: PAD token        = %d '%s'\n", __func__, special_pad_id,     id_to_token.at(special_pad_id).text.c_str() );  }
+    if (special_mask_id != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: MASK token       = %d '%s'\n", __func__, special_mask_id,    id_to_token.at(special_mask_id).text.c_str() ); }
 
-    if (linefeed_id != LLAMA_TOKEN_NULL)        { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, linefeed_id,        id_to_token[linefeed_id].text.c_str() ); }
+    if (linefeed_id != LLAMA_TOKEN_NULL)        { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, linefeed_id,        id_to_token.at(linefeed_id).text.c_str() ); }
 
-    if (special_fim_pre_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PRE token    = %d '%s'\n", __func__, special_fim_pre_id, id_to_token[special_fim_pre_id].text.c_str() ); }
-    if (special_fim_suf_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SUF token    = %d '%s'\n", __func__, special_fim_suf_id, id_to_token[special_fim_suf_id].text.c_str() ); }
-    if (special_fim_mid_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM MID token    = %d '%s'\n", __func__, special_fim_mid_id, id_to_token[special_fim_mid_id].text.c_str() ); }
-    if (special_fim_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PAD token    = %d '%s'\n", __func__, special_fim_pad_id, id_to_token[special_fim_pad_id].text.c_str() ); }
-    if (special_fim_rep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM REP token    = %d '%s'\n", __func__, special_fim_rep_id, id_to_token[special_fim_rep_id].text.c_str() ); }
-    if (special_fim_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SEP token    = %d '%s'\n", __func__, special_fim_sep_id, id_to_token[special_fim_sep_id].text.c_str() ); }
+    if (special_fim_pre_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PRE token    = %d '%s'\n", __func__, special_fim_pre_id, id_to_token.at(special_fim_pre_id).text.c_str() ); }
+    if (special_fim_suf_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SUF token    = %d '%s'\n", __func__, special_fim_suf_id, id_to_token.at(special_fim_suf_id).text.c_str() ); }
+    if (special_fim_mid_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM MID token    = %d '%s'\n", __func__, special_fim_mid_id, id_to_token.at(special_fim_mid_id).text.c_str() ); }
+    if (special_fim_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PAD token    = %d '%s'\n", __func__, special_fim_pad_id, id_to_token.at(special_fim_pad_id).text.c_str() ); }
+    if (special_fim_rep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM REP token    = %d '%s'\n", __func__, special_fim_rep_id, id_to_token.at(special_fim_rep_id).text.c_str() ); }
+    if (special_fim_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SEP token    = %d '%s'\n", __func__, special_fim_sep_id, id_to_token.at(special_fim_sep_id).text.c_str() ); }
 
     for (const auto & id : special_eog_ids) {
-        LLAMA_LOG_INFO( "%s: EOG token        = %d '%s'\n", __func__, id, id_to_token[id].text.c_str() );
+        LLAMA_LOG_INFO( "%s: EOG token        = %d '%s'\n", __func__, id, id_to_token.at(id).text.c_str() );
     }
 
     LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, max_token_len);

src/llama.cpp

@@ -198,14 +198,18 @@ static struct llama_model * llama_model_load_from_file_impl(
 
     // if using single GPU mode, remove all except the main GPU
     if (params.split_mode == LLAMA_SPLIT_MODE_NONE) {
-        if (params.main_gpu < 0 || params.main_gpu >= (int)model->devices.size()) {
-            LLAMA_LOG_ERROR("%s: invalid value for main_gpu: %d (available devices: %d)\n", __func__, params.main_gpu, (int)model->devices.size());
-            llama_model_free(model);
-            return nullptr;
+        if (params.main_gpu < 0) {
+            model->devices.clear();
+        } else {
+            if (params.main_gpu >= (int)model->devices.size()) {
+                LLAMA_LOG_ERROR("%s: invalid value for main_gpu: %d (available devices: %zu)\n", __func__, params.main_gpu, model->devices.size());
+                llama_model_free(model);
+                return nullptr;
+            }
+            ggml_backend_dev_t main_gpu = model->devices[params.main_gpu];
+            model->devices.clear();
+            model->devices.push_back(main_gpu);
         }
-        ggml_backend_dev_t main_gpu = model->devices[params.main_gpu];
-        model->devices.clear();
-        model->devices.push_back(main_gpu);
     }
 
     for (auto * dev : model->devices) {

src/unicode.cpp

@@ -204,12 +204,17 @@ static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
     // 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<wchar_t>> conv;
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
 #endif
 
     return conv.from_bytes(s);

tests/CMakeLists.txt

@@ -42,6 +42,34 @@ function(llama_test target)
     set_property(TEST ${TEST_NAME} PROPERTY LABELS ${LLAMA_TEST_LABEL})
 endfunction()
 
+function(llama_test_cmd target)
+    include(CMakeParseArguments)
+    set(options)
+    set(oneValueArgs NAME LABEL WORKING_DIRECTORY)
+    set(multiValueArgs ARGS)
+    cmake_parse_arguments(LLAMA_TEST "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+    if (NOT DEFINED LLAMA_TEST_LABEL)
+        set(LLAMA_TEST_LABEL "main")
+    endif()
+    if (NOT DEFINED LLAMA_TEST_WORKING_DIRECTORY)
+        set(LLAMA_TEST_WORKING_DIRECTORY .)
+    endif()
+    if (DEFINED LLAMA_TEST_NAME)
+        set(TEST_NAME ${LLAMA_TEST_NAME})
+    else()
+        set(TEST_NAME ${target})
+    endif()
+
+    add_test(
+        NAME ${TEST_NAME}
+        WORKING_DIRECTORY ${LLAMA_TEST_WORKING_DIRECTORY}
+        COMMAND ${target}
+        ${LLAMA_TEST_ARGS})
+
+    set_property(TEST ${TEST_NAME} PROPERTY LABELS ${LLAMA_TEST_LABEL})
+endfunction()
+
 # Builds and runs a test source file.
 # Optional args:
 # - NAME: name of the executable & test target (defaults to the source file name without extension)
@@ -83,25 +111,31 @@ endfunction()
 # build test-tokenizer-0 target once and add many tests
 llama_build(test-tokenizer-0.cpp)
 
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-bert-bge          ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bert-bge.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-command-r         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-command-r.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-coder.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm      ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-llm.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-falcon            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-gpt-2             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-bpe         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-spm         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-spm.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-mpt               ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-phi-3             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-phi-3.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-qwen2             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-qwen2.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-bert-bge          ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-bert-bge.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-command-r         ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-command-r.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-deepseek-coder.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm      ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-deepseek-llm.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-falcon            ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-falcon.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-gpt-2             ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-gpt-2.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-bpe         ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-bpe.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-spm         ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-spm.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-mpt               ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-mpt.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-phi-3             ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-phi-3.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-qwen2             ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-qwen2.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact            ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-refact.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder         ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-starcoder.gguf)
 
-# TODO: missing HF tokenizer for this model in convert_hf_to_gguf_update.py, see https://github.com/ggml-org/llama.cpp/pull/13847
-# llama_test(test-tokenizer-0 NAME test-tokenizer-0-nomic-bert-moe    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-nomic-bert-moe.gguf)
+if (NOT WIN32)
+    llama_test_cmd(
+        ${CMAKE_CURRENT_SOURCE_DIR}/test-tokenizers-repo.sh
+        NAME test-tokenizers-ggml-vocabs
+        WORKING_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}
+        ARGS https://huggingface.co/ggml-org/vocabs ${PROJECT_SOURCE_DIR}/models/ggml-vocabs
+    )
+endif()
 
 if (LLAMA_LLGUIDANCE)
-    llama_build_and_test(test-grammar-llguidance.cpp ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
+    llama_build_and_test(test-grammar-llguidance.cpp ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-bpe.gguf)
 endif ()
 
 if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
@@ -113,8 +147,8 @@ if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
     llama_build_and_test(test-chat.cpp)
     # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
     if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
-        llama_build_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
-        target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../tools/server)
+        llama_build_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})
+        target_include_directories(test-json-schema-to-grammar PRIVATE ${PROJECT_SOURCE_DIR}/tools/server)
     endif()
 
     if (NOT GGML_BACKEND_DL)
@@ -127,20 +161,20 @@ if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
     llama_build(test-tokenizer-1-bpe.cpp)
 
     # TODO: disabled due to slowness
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-aquila    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-falcon    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-2     ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-neox  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-llama-bpe ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf --ignore-merges)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-mpt       ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-refact    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-starcoder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-aquila    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-aquila.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-falcon    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-falcon.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-2     ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-gpt-2.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-neox  ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-gpt-neox.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-llama-bpe ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-bpe.gguf --ignore-merges)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-mpt       ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-mpt.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-refact    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-refact.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-starcoder ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-starcoder.gguf)
 
     # build test-tokenizer-1-spm target once and add many tests
     llama_build(test-tokenizer-1-spm.cpp)
 
-    llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-spm.gguf)
-    #llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-baichuan  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
+    llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-spm.gguf)
+    #llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-baichuan  ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-baichuan.gguf)
 
     # llama_build_and_test(test-double-float.cpp) # SLOW
 endif()
@@ -151,6 +185,8 @@ llama_build_and_test(test-json-partial.cpp)
 llama_build_and_test(test-log.cpp)
 llama_build_and_test(test-regex-partial.cpp)
 
+llama_build_and_test(test-thread-safety.cpp ARGS -hf ggml-org/models -hff tinyllamas/stories15M-q4_0.gguf -ngl 99 -p "The meaning of life is" -n 128 -c 256 -ub 32 -np 4)
+
 # this fails on windows (github hosted runner) due to curl DLL not found (exit code 0xc0000135)
 if (NOT WIN32)
     llama_build_and_test(test-arg-parser.cpp)

tests/test-thread-safety.cpp

@@ -0,0 +1,152 @@
+// thread safety test
+// - Loads a copy of the same model on each GPU, plus a copy on the CPU
+// - Creates n_parallel (--parallel) contexts per model
+// - Runs inference in parallel on each context
+
+#include <thread>
+#include <vector>
+#include <atomic>
+#include "llama.h"
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "sampling.h"
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+        return 1;
+    }
+
+    common_init();
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+
+    //llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
+    //    if (level == GGML_LOG_LEVEL_ERROR) {
+    //        common_log_add(common_log_main(), level, "%s", text);
+    //    }
+    //}, NULL);
+
+    auto cparams = common_context_params_to_llama(params);
+
+    int dev_count = ggml_backend_dev_count();
+    int gpu_dev_count = 0;
+    for (int i = 0; i < dev_count; ++i) {
+        auto * dev = ggml_backend_dev_get(i);
+        if (dev && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            gpu_dev_count++;
+        }
+    }
+    const int num_models = gpu_dev_count + 1 + 1; // GPUs + 1 CPU model + 1 layer split
+    //const int num_models = std::max(1, gpu_dev_count);
+    const int num_contexts = std::max(1, params.n_parallel);
+
+    std::vector<llama_model_ptr> models;
+    std::vector<std::thread> threads;
+    std::atomic<bool> failed = false;
+
+    for (int m = 0; m < num_models; ++m) {
+        auto mparams = common_model_params_to_llama(params);
+
+        if (m < gpu_dev_count) {
+            mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
+            mparams.main_gpu = m;
+        } else if (m == gpu_dev_count) {
+            mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
+            mparams.main_gpu = -1; // CPU model
+        } else {
+            mparams.split_mode = LLAMA_SPLIT_MODE_LAYER;;
+        }
+
+        llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
+        if (model == NULL) {
+            LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+            return 1;
+        }
+
+        models.emplace_back(model);
+    }
+
+    for  (int m = 0; m < num_models; ++m) {
+        auto * model = models[m].get();
+        for (int c = 0; c < num_contexts; ++c) {
+            threads.emplace_back([&, m, c, model]() {
+                LOG_INF("Creating context %d/%d for model %d/%d\n", c + 1, num_contexts, m + 1, num_models);
+
+                llama_context_ptr ctx { llama_init_from_model(model, cparams) };
+                if (ctx == NULL) {
+                    LOG_ERR("failed to create context\n");
+                    failed.store(true);
+                    return;
+                }
+
+                std::unique_ptr<common_sampler, decltype(&common_sampler_free)> sampler { common_sampler_init(model, params.sampling), common_sampler_free };
+                if (sampler == NULL) {
+                    LOG_ERR("failed to create sampler\n");
+                    failed.store(true);
+                    return;
+                }
+
+                llama_batch batch = {};
+                {
+                    auto prompt = common_tokenize(ctx.get(), params.prompt, true);
+                    if (prompt.empty()) {
+                        LOG_ERR("failed to tokenize prompt\n");
+                        failed.store(true);
+                        return;
+                    }
+                    batch = llama_batch_get_one(prompt.data(), prompt.size());
+                    if (llama_decode(ctx.get(), batch)) {
+                        LOG_ERR("failed to decode prompt\n");
+                        failed.store(true);
+                        return;
+                    }
+                }
+
+                const auto * vocab = llama_model_get_vocab(model);
+                std::string result = params.prompt;
+
+                for (int i = 0; i < params.n_predict; i++) {
+                    llama_token token;
+                    if (batch.n_tokens > 0) {
+                        token = common_sampler_sample(sampler.get(), ctx.get(), batch.n_tokens - 1);
+                    } else {
+                        token = llama_vocab_bos(vocab);
+                    }
+
+                    result += common_token_to_piece(ctx.get(), token);
+
+                    if (llama_vocab_is_eog(vocab, token)) {
+                        break;
+                    }
+
+                    batch = llama_batch_get_one(&token, 1);
+                    if (llama_decode(ctx.get(), batch)) {
+                        LOG_ERR("Model %d/%d, Context %d/%d: failed to decode\n", m + 1, num_models, c + 1, num_contexts);
+                        failed.store(true);
+                        return;
+                    }
+                }
+
+                LOG_INF("Model %d/%d, Context %d/%d: %s\n\n", m + 1, num_models, c + 1, num_contexts, result.c_str());
+            });
+        }
+    }
+
+    for (auto & thread : threads) {
+        thread.join();
+    }
+
+    if (failed) {
+        LOG_ERR("One or more threads failed.\n");
+        return 1;
+    }
+
+    LOG_INF("All threads finished without errors.\n");
+    return 0;
+}

tests/test-tokenizers-repo.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+
+if [ $# -lt 2 ]; then
+    printf "Usage: $0 <git-repo> <target-folder> [<test-exe>]\n"
+    exit 1
+fi
+
+if [ $# -eq 3 ]; then
+    toktest=$3
+else
+    toktest="./test-tokenizer-0"
+fi
+
+if [ ! -x $toktest ]; then
+    printf "Test executable \"$toktest\" not found!\n"
+    exit 1
+fi
+
+repo=$1
+folder=$2
+
+if [ -d $folder ] && [ -d $folder/.git ]; then
+    (cd $folder; git pull)
+else
+    git clone $repo $folder
+fi
+
+shopt -s globstar
+for gguf in $folder/**/*.gguf; do
+    if [ -f $gguf.inp ] && [ -f $gguf.out ]; then
+        $toktest $gguf
+    else
+        printf "Found \"$gguf\" without matching inp/out files, ignoring...\n"
+    fi
+done
+

tools/llama-bench/llama-bench.cpp

@@ -267,6 +267,7 @@ struct cmd_params {
     int                              delay;
     bool                             verbose;
     bool                             progress;
+    bool                             no_warmup;
     output_formats                   output_format;
     output_formats                   output_format_stderr;
 };
@@ -303,6 +304,7 @@ static const cmd_params cmd_params_defaults = {
     /* delay                */ 0,
     /* verbose              */ false,
     /* progress             */ false,
+    /* no_warmup            */ false,
     /* output_format        */ MARKDOWN,
     /* output_format_stderr */ NONE,
 };
@@ -325,6 +327,7 @@ static void print_usage(int /* argc */, char ** argv) {
            output_format_str(cmd_params_defaults.output_format_stderr));
     printf("  -v, --verbose                             verbose output\n");
     printf("  --progress                                print test progress indicators\n");
+    printf("  --no-warmup                               skip warmup runs before benchmarking\n");
     printf("\n");
     printf("test parameters:\n");
     printf("  -m, --model <filename>                    (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
@@ -425,6 +428,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     params.prio                 = cmd_params_defaults.prio;
     params.delay                = cmd_params_defaults.delay;
     params.progress             = cmd_params_defaults.progress;
+    params.no_warmup            = cmd_params_defaults.no_warmup;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -798,6 +802,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 params.verbose = true;
             } else if (arg == "--progress") {
                 params.progress = true;
+            } else if (arg == "--no-warmup") {
+                params.no_warmup = true;
             } else {
                 invalid_param = true;
                 break;
@@ -1925,25 +1931,27 @@ int main(int argc, char ** argv) {
         llama_attach_threadpool(ctx, threadpool, NULL);
 
         // warmup run
-        if (t.n_prompt > 0) {
-            if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count);
+        if (!params.no_warmup) {
+            if (t.n_prompt > 0) {
+                if (params.progress) {
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count);
+                }
+                //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
+                bool res = test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
+                if (!res) {
+                    fprintf(stderr, "%s: error: failed to run prompt warmup\n", __func__);
+                    exit(1);
+                }
             }
-            //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
-            bool res = test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
-            if (!res) {
-                fprintf(stderr, "%s: error: failed to run prompt warmup\n", __func__);
-                exit(1);
-            }
-        }
-        if (t.n_gen > 0) {
-            if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count);
-            }
-            bool res = test_gen(ctx, 1, t.n_threads);
-            if (!res) {
-                fprintf(stderr, "%s: error: failed to run gen warmup\n", __func__);
-                exit(1);
+            if (t.n_gen > 0) {
+                if (params.progress) {
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count);
+                }
+                bool res = test_gen(ctx, 1, t.n_threads);
+                if (!res) {
+                    fprintf(stderr, "%s: error: failed to run gen warmup\n", __func__);
+                    exit(1);
+                }
             }
         }