opt : remove print [no ci]

ggml-ci

.devops/cpu.Dockerfile

@@ -14,9 +14,9 @@ WORKDIR /app
 COPY . .
 
 RUN if [ "$TARGETARCH" = "amd64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
     elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
     else \
         echo "Unsupported architecture"; \
         exit 1; \

.devops/cuda.Dockerfile

@@ -21,7 +21,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
     export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/intel.Dockerfile

@@ -17,7 +17,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
     echo "Building with dynamic libs" && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${OPT_SYCL_F16} && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${OPT_SYCL_F16} && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/llama-cli-cann.Dockerfile

@@ -22,7 +22,7 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
 
 RUN echo "Building with static libs" && \
     source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF -DLLAMA_BUILD_TESTS=OFF  && \
     cmake --build build --config Release --target llama-cli
 
 # TODO: use image with NNRT

.devops/musa.Dockerfile

@@ -35,7 +35,7 @@ COPY . .
 RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/rocm.Dockerfile

@@ -40,7 +40,7 @@ WORKDIR /app
 COPY . .
 
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
+    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
     && cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib \

.devops/vulkan.Dockerfile

@@ -16,7 +16,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1  -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.github/workflows/build-linux-cross.yml

@@ -4,18 +4,25 @@ on:
   workflow_call:
 
 jobs:
-  ubuntu-latest-riscv64-cpu-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-riscv64-cpu-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
       - name: Setup Riscv
         run: |
           sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   gcc-14-riscv64-linux-gnu \
@@ -40,21 +47,25 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  ubuntu-latest-riscv64-vulkan-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-riscv64-vulkan-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
       - name: Setup Riscv
         run: |
           sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   glslc \
@@ -82,21 +93,25 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  ubuntu-latest-arm64-vulkan-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-arm64-vulkan-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
       - name: Setup Arm64
         run: |
           sudo dpkg --add-architecture arm64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   glslc \

.github/workflows/build.yml

@@ -601,9 +601,8 @@ jobs:
             -DGGML_SYCL_F16=ON
           cmake --build build --config Release -j $(nproc)
 
-# Disabled for now due to sporadic issue syncing.
-#  build-linux-cross:
-#    uses: ./.github/workflows/build-linux-cross.yml
+  build-linux-cross:
+    uses: ./.github/workflows/build-linux-cross.yml
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest

README.md

@@ -17,7 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 ## Hot topics
 
 - **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` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated
+- 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
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim

cmake/build-info.cmake

@@ -41,14 +41,20 @@ endif()
 
 if(MSVC)
     set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    if (CMAKE_VS_PLATFORM_NAME)
+        set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    else()
+        set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
 else()
     execute_process(
-        COMMAND sh -c "\"$@\" --version | head -1" _ ${CMAKE_C_COMPILER}
+        COMMAND ${CMAKE_C_COMPILER} --version
         OUTPUT_VARIABLE OUT
         OUTPUT_STRIP_TRAILING_WHITESPACE
     )
+    string(REGEX REPLACE " *\n.*" "" OUT "${OUT}")
     set(BUILD_COMPILER ${OUT})
+
     execute_process(
         COMMAND ${CMAKE_C_COMPILER} -dumpmachine
         OUTPUT_VARIABLE OUT

common/CMakeLists.txt

@@ -39,7 +39,9 @@ add_custom_command(
     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} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+            -DCMAKE_C_COMPILER=${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

common/arg.cpp

@@ -43,6 +43,25 @@ std::initializer_list<enum llama_example> mmproj_examples = {
     // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
 };
 
+static std::string read_file(const std::string & fname) {
+    std::ifstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+    file.close();
+    return content;
+}
+
+static void write_file(const std::string & fname, const std::string & content) {
+    std::ofstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    file << content;
+    file.close();
+}
+
 common_arg & common_arg::set_examples(std::initializer_list<enum llama_example> examples) {
     this->examples = std::move(examples);
     return *this;
@@ -198,11 +217,11 @@ struct curl_slist_ptr {
 #define CURL_MAX_RETRY 3
 #define CURL_RETRY_DELAY_SECONDS 2
 
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds, const char * method_name) {
     int remaining_attempts = max_attempts;
 
     while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
 
         CURLcode res = curl_easy_perform(curl);
         if (res == CURLE_OK) {
@@ -213,6 +232,7 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
         LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
 
         remaining_attempts--;
+        if (remaining_attempts == 0) break;
         std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
     }
 
@@ -231,8 +251,6 @@ static bool common_download_file_single(const std::string & url, const std::stri
         return false;
     }
 
-    bool force_download = false;
-
     // Set the URL, allow to follow http redirection
     curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
     curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
@@ -256,7 +274,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
 
     // If the file exists, check its JSON metadata companion file.
     std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
+    nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead
     std::string etag;
     std::string last_modified;
 
@@ -266,14 +284,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         if (metadata_in.good()) {
             try {
                 metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
+                LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
                 if (metadata.contains("etag") && metadata.at("etag").is_string()) {
                     etag = metadata.at("etag");
                 }
@@ -281,10 +292,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
                     last_modified = metadata.at("lastModified");
                 }
             } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
+                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
             }
         }
+        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
     } else {
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
@@ -296,7 +307,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
     };
 
     common_load_model_from_url_headers headers;
+    bool head_request_ok = false;
+    bool should_download = !file_exists; // by default, we should download if the file does not exist
 
+    // get ETag to see if the remote file has changed
     {
         typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
         auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
@@ -325,23 +339,28 @@ static bool common_download_file_single(const std::string & url, const std::stri
         curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
         curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
 
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        // we only allow retrying once for HEAD requests
+        // this is for the use case of using running offline (no internet), retrying can be annoying
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
         if (!was_perform_successful) {
-            return false;
+            head_request_ok = false;
         }
 
         long http_code = 0;
         curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        if (http_code == 200) {
+            head_request_ok = true;
+        } else {
+            LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+            head_request_ok = false;
         }
     }
 
-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
+    // if head_request_ok is false, we don't have the etag or last-modified headers
+    // we leave should_download as-is, which is true if the file does not exist
+    if (head_request_ok) {
+        // check if ETag or Last-Modified headers are different
+        // if it is, we need to download the file again
         if (!etag.empty() && etag != headers.etag) {
             LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
             should_download = true;
@@ -350,6 +369,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
             should_download = true;
         }
     }
+
     if (should_download) {
         std::string path_temporary = path + ".downloadInProgress";
         if (file_exists) {
@@ -403,7 +423,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         // start the download
         LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
             llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS, "GET");
         if (!was_perform_successful) {
             return false;
         }
@@ -424,13 +444,15 @@ static bool common_download_file_single(const std::string & url, const std::stri
             {"etag", headers.etag},
             {"lastModified", headers.last_modified}
         });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+        write_file(metadata_path, metadata.dump(4));
+        LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
 
         if (rename(path_temporary.c_str(), path.c_str()) != 0) {
             LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
             return false;
         }
+    } else {
+        LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
     }
 
     return true;
@@ -605,16 +627,37 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
     // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
     // User-Agent header is already set in common_remote_get_content, no need to set it here
 
+    // we use "=" to avoid clashing with other component, while still being allowed on windows
+    std::string cached_response_fname = "manifest=" + hf_repo + "=" + tag + ".json";
+    string_replace_all(cached_response_fname, "/", "_");
+    std::string cached_response_path = fs_get_cache_file(cached_response_fname);
+
     // make the request
     common_remote_params params;
     params.headers = headers;
-    auto res = common_remote_get_content(url, params);
-    long res_code = res.first;
-    std::string res_str(res.second.data(), res.second.size());
+    long res_code = 0;
+    std::string res_str;
+    bool use_cache = false;
+    try {
+        auto res = common_remote_get_content(url, params);
+        res_code = res.first;
+        res_str = std::string(res.second.data(), res.second.size());
+    } catch (const std::exception & e) {
+        LOG_WRN("error: failed to get manifest: %s\n", e.what());
+        LOG_WRN("try reading from cache\n");
+        // try to read from cache
+        try {
+            res_str = read_file(cached_response_path);
+            res_code = 200;
+            use_cache = true;
+        } catch (const std::exception & e) {
+            throw std::runtime_error("error: failed to get manifest (check your internet connection)");
+        }
+    }
     std::string ggufFile;
     std::string mmprojFile;
 
-    if (res_code == 200) {
+    if (res_code == 200 || res_code == 304) {
         // extract ggufFile.rfilename in json, using regex
         {
             std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
@@ -631,6 +674,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
                 mmprojFile = match[1].str();
             }
         }
+        if (!use_cache) {
+            // if not using cached response, update the cache file
+            write_file(cached_response_path, res_str);
+        }
     } else if (res_code == 401) {
         throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
     } else {
@@ -1142,6 +1189,9 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
         fprintf(stderr, "%s\n", ex.what());
         ctx_arg.params = params_org;
         return false;
+    } catch (std::exception & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        exit(1); // for other exceptions, we exit with status code 1
     }
 
     return true;
@@ -1442,13 +1492,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-f", "--file"}, "FNAME",
         "a file containing the prompt (default: none)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
+            params.prompt = read_file(value);
             // store the external file name in params
             params.prompt_file = value;
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
             if (!params.prompt.empty() && params.prompt.back() == '\n') {
                 params.prompt.pop_back();
             }
@@ -1458,11 +1504,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-sysf", "--system-prompt-file"}, "FNAME",
         "a file containing the system prompt (default: none)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.system_prompt));
+            params.system_prompt = read_file(value);
             if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') {
                 params.system_prompt.pop_back();
             }
@@ -1887,15 +1929,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--grammar-file"}, "FNAME",
         "file to read grammar from",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.sampling.grammar)
-            );
+            params.sampling.grammar = read_file(value);
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1905,6 +1939,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sampling.grammar = json_schema_to_grammar(json::parse(value));
         }
     ).set_sparam());
+    add_opt(common_arg(
+        {"-jf", "--json-schema-file"}, "FILE",
+        "File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
+        [](common_params & params, const std::string & value) {
+            std::ifstream file(value);
+            if (!file) {
+                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            }
+            std::string schema;
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(schema)
+            );
+            params.sampling.grammar = json_schema_to_grammar(json::parse(schema));
+        }
+    ).set_sparam());
     add_opt(common_arg(
         {"--pooling"}, "{none,mean,cls,last,rank}",
         "pooling type for embeddings, use model default if unspecified",
@@ -2732,7 +2783,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP"));
     add_opt(common_arg(
         {"--cache-reuse"}, "N",
-        string_format("min chunk size to attempt reusing from the cache via KV shifting (default: %d)", params.n_cache_reuse),
+        string_format(
+            "min chunk size to attempt reusing from the cache via KV shifting (default: %d)\n"
+            "[(card)](https://ggml.ai/f0.png)", params.n_cache_reuse
+        ),
         [](common_params & params, int value) {
             params.n_cache_reuse = value;
         }
@@ -2815,14 +2869,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             "list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
         ),
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.chat_template));
+            params.chat_template = read_file(value);
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
     add_opt(common_arg(

common/common.cpp

@@ -1565,3 +1565,20 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
+    const int64_t ne_datapoint = llama_n_ctx(ctx);
+    const int64_t ndata        = (tokens.size() - ne_datapoint - 1) / stride;
+    ggml_opt_dataset_t result = ggml_opt_dataset_init(
+        GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
+
+    llama_token * data   = (llama_token *) ggml_opt_dataset_data(result)->data;
+    llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
+
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        memcpy(data   + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
+        memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
+    }
+
+    return result;
+}

common/common.h

@@ -664,3 +664,9 @@ const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 }
+
+//
+// training utils
+//
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);

convert_hf_to_gguf.py

@@ -16,6 +16,7 @@ from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
 from itertools import chain
+from transformers import AutoConfig
 
 import math
 import numpy as np
@@ -66,8 +67,6 @@ class ModelBase:
     part_names: list[str]
     is_safetensors: bool
     hparams: dict[str, Any]
-    block_count: int
-    tensor_map: gguf.TensorNameMap
     tensor_names: set[str] | None
     gguf_writer: gguf.GGUFWriter
     model_name: str | None
@@ -78,7 +77,11 @@ class ModelBase:
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
 
-    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
+    # subclasses should initialize this!
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
@@ -113,8 +116,6 @@ class ModelBase:
             if not self.is_safetensors:
                 self.part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
         self.hparams = ModelBase.load_hparams(self.dir_model) if hparams is None else hparams
-        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
         self.metadata_override = metadata_override
         self.model_name = model_name
@@ -417,15 +418,15 @@ class ModelBase:
 
     @staticmethod
     def load_hparams(dir_model: Path):
-        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-            hparams = json.load(f)
-            architectures = hparams.get("architectures")
-            if "text_config" in hparams:
-                hparams = {**hparams, **hparams["text_config"]}
-            if architectures is not None:
-                # preserve "architectures" from root level config
-                hparams["architectures"] = architectures
-            return hparams
+        try:
+            # for security reason, we don't allow loading remote code by default
+            # if a model need remote code, we will fallback to config.json
+            return AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
+        except Exception as e:
+            logger.warning(f"Failed to load model config from {dir_model}: {e}")
+            logger.warning("Trying to load config.json instead")
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                return json.load(f)
 
     @classmethod
     def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
@@ -454,6 +455,20 @@ class ModelBase:
 
 
 class TextModel(ModelBase):
+    model_type = ModelType.TEXT
+    hf_arch: str
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hf_arch = get_model_architecture(self.hparams, self.model_type)
+
+        if "text_config" in self.hparams:
+            # move the text_config to the root level
+            self.hparams = {**self.hparams, **self.hparams["text_config"]}
+
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
     @classmethod
     def __init_subclass__(cls):
         # can't use an abstract property, because overriding it without type errors
@@ -495,7 +510,7 @@ 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"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
@@ -1064,10 +1079,36 @@ class TextModel(ModelBase):
         if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_EOS)) is not None:
             self.gguf_writer.add_add_eos_token(field.parts[-1].tolist()[0])
 
+    def _try_set_pooling_type(self) -> None:
+        # get pooling path
+        pooling_path = None
+        module_path = self.dir_model / "modules.json"
+        if module_path.is_file():
+            with open(module_path, encoding="utf-8") as f:
+                modules = json.load(f)
+            for mod in modules:
+                if mod["type"] == "sentence_transformers.models.Pooling":
+                    pooling_path = mod["path"]
+                    break
+
+        # get pooling type
+        if pooling_path is not None:
+            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
+                pooling = json.load(f)
+            if pooling["pooling_mode_mean_tokens"]:
+                pooling_type = gguf.PoolingType.MEAN
+            elif pooling["pooling_mode_cls_token"]:
+                pooling_type = gguf.PoolingType.CLS
+            elif pooling["pooling_mode_lasttoken"]:
+                pooling_type = gguf.PoolingType.LAST
+            else:
+                raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
+            self.gguf_writer.add_pooling_type(pooling_type)
+
 
 class VisionModel(ModelBase):
+    model_type = ModelType.VISION
     model_arch = gguf.MODEL_ARCH.CLIP_VISION
-    n_text_embd = 0
     preprocessor_config: dict[str, Any]
     global_config: dict[str, Any]
 
@@ -1077,9 +1118,11 @@ class VisionModel(ModelBase):
         if self.model_arch != gguf.MODEL_ARCH.CLIP_VISION:
             raise TypeError("VisionModel must be subclassed with model_arch = gguf.MODEL_ARCH.CLIP_VISION")
 
-        # small hack to correct the number of layers
-        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, 128)
-        self.n_embd_text = self.find_hparam(["hidden_size", "n_embd"])
+        # get n_embd of the text model
+        if "text_config" not in self.hparams:
+            self.hparams["text_config"] = {}
+        text_config = {**self.hparams, **self.hparams["text_config"]}
+        self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
         assert self.n_embd_text > 0, "n_embd not found in hparams"
 
         if "vision_config" not in self.hparams:
@@ -1088,6 +1131,9 @@ class VisionModel(ModelBase):
         self.global_config = self.hparams
         self.hparams = self.hparams["vision_config"]
 
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth"])
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, self.block_count)
+
         # load preprocessor config
         with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
             self.preprocessor_config = json.load(f)
@@ -1105,12 +1151,12 @@ class VisionModel(ModelBase):
         self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"]))
         self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"]))
         self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"]))
-        self.gguf_writer.add_vision_block_count(self.find_hparam(["num_hidden_layers"]))
+        self.gguf_writer.add_vision_block_count(self.block_count)
         self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"]))
 
         # preprocessor config
         self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
-        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_mean"])
+        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
 
     def write_vocab(self):
         raise ValueError("VisionModel does not support vocab writing")
@@ -1726,23 +1772,12 @@ class StableLMModel(TextModel):
     "LlamaForCausalLM",
     "MistralForCausalLM",
     "MixtralForCausalLM",
-    "Idefics3ForConditionalGeneration",
-    "SmolVLMForConditionalGeneration",
+    "VLlama3ForCausalLM",
     "LlavaForConditionalGeneration")
 class LlamaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
     undo_permute = True
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        # fix for SmolVLM2, missing `num_attention_heads` in config.json
-        if self.hparams["architectures"][0] == "SmolVLMForConditionalGeneration":
-            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
-        # fix for Pixtral, missing `num_attention_heads` in config.json
-        if self.hparams["architectures"][0] == "LlavaForConditionalGeneration" \
-                and self.hparams.get("model_type") == "mistral":
-            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
-
     def set_vocab(self):
         try:
             self._set_vocab_sentencepiece()
@@ -1898,31 +1933,50 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("LlavaForConditionalGeneration")
+@ModelBase.register(
+    "LlavaForConditionalGeneration", # pixtral
+    "Mistral3ForConditionalGeneration", # mistral small 3.1
+)
 class LlavaVisionModel(VisionModel):
     img_break_tok_id = -1
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         if self.hparams["model_type"] == "pixtral":
-            # fix missing config.json values
-            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16)
-            self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 24)
-            self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 4096)
-            self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1024)
+            # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py
             self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
-            self.img_break_tok_id = 12 # see tokenizer_config.json
+            self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
+            logger.info(f"Image break token id: {self.img_break_tok_id}")
         else:
             raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
 
+    def get_token_id(self, token: str) -> int:
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+            added_tokens_decoder = json.load(f)['added_tokens_decoder']
+            for id_, token_data in added_tokens_decoder.items():
+                if token_data["content"] == token:
+                    return int(id_)
+        raise ValueError(f"Token '{token}' not found in tokenizer config.")
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
         if hparams["model_type"] == "pixtral":
             self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL)
-            # default values below are taken from HF tranformers code
             self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
-            self.gguf_writer.add_vision_use_silu(True)
+
+            # hidden_act
+            if hparams["hidden_act"] == "silu":
+                self.gguf_writer.add_vision_use_silu(True)
+            elif hparams["hidden_act"] == "gelu":
+                self.gguf_writer.add_vision_use_gelu(True)
+            else:
+                raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
+
+            # spatial_merge_size
+            if "spatial_merge_size" in self.global_config:
+                self.gguf_writer.add_vision_spatial_merge_size(self.global_config["spatial_merge_size"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -1951,13 +2005,12 @@ class LlavaVisionModel(VisionModel):
 class SmolVLMModel(VisionModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        # fix for SmolVLM2, missing some keys in config.json
-        # default values are taken from transformers code
         if self.hparams["model_type"] == "smolvlm_vision":
+            # fix for SmolVLM2, missing some keys in config.json
+            # default values are taken from transformers code
             self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1152)
             self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16)
             self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 3072)
-            self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 12)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -2519,7 +2572,7 @@ class QwenModel(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
 
-@ModelBase.register("Qwen2ForCausalLM")
+@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM")
 class Qwen2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.QWEN2
 
@@ -2531,12 +2584,18 @@ class Qwen2Model(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
+        self._try_set_pooling_type()
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
             if self.hparams["rope_scaling"].get("type") == "yarn":
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
                 self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
                 self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if self.hf_arch == "Qwen2Model":
+            name = f"model.{name}"  # map to Qwen2ForCausalLM tensors
+        yield from super().modify_tensors(data_torch, name, bid)
+
 
 @ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(TextModel):
@@ -2562,6 +2621,82 @@ class Qwen2VLModel(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
+class Qwen2VLVisionModel(VisionModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["image_size"] = self.hparams.get("image_size", 560)
+        # rename config.json values
+        self.hparams["num_attention_heads"] = self.hparams.get("num_heads")
+        self.hparams["num_hidden_layers"] = self.hparams.get("depth")
+        if "embed_dim" in self.hparams: # qwen2vl
+            self.hparams["intermediate_size"] = self.hparams.get("hidden_size")
+            self.hparams["hidden_size"] = self.hparams.get("embed_dim")
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        if self.global_config['model_type'] == 'qwen2_vl':
+            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN2VL)
+        elif self.global_config['model_type'] == 'qwen2_5_vl':
+            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN25VL)
+            self.gguf_writer.add_vision_use_silu(True)
+            # find n_wa_pattern (window attention pattern)
+            fullatt_block_indexes = hparams.get("fullatt_block_indexes")
+            assert fullatt_block_indexes is not None, "fullatt_block_indexes is required for qwen2_5_vl"
+            n_wa_pattern = fullatt_block_indexes[0] + 1
+            # validate n_wa_pattern
+            for i in range(1, len(fullatt_block_indexes)):
+                if fullatt_block_indexes[i] - fullatt_block_indexes[i - 1] != n_wa_pattern:
+                    raise ValueError(f"Invalid fullatt_block_indexes: {fullatt_block_indexes}")
+            self.gguf_writer.add_vision_n_wa_pattern(n_wa_pattern)
+        else:
+            raise ValueError(f"Unknown QwenVL model type: {self.global_config['model_type']}")
+        # default values below are taken from HF tranformers code
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6))
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, name, n_dims  # unused
+        if ".patch_embd." in new_name:
+            return gguf.GGMLQuantizationType.F16
+        if ".position_embd." in new_name:
+            return gguf.GGMLQuantizationType.F32
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        if name.startswith("visual."):
+            # process visual tensors
+            # split QKV tensors if needed
+            if ".qkv." in name:
+                if data_torch.ndim == 2: # weight
+                    c3, _ = data_torch.shape
+                else: # bias
+                    c3 = data_torch.shape[0]
+                assert c3 % 3 == 0
+                c = c3 // 3
+                wq = data_torch[:c]
+                wk = data_torch[c: c * 2]
+                wv = data_torch[c * 2:]
+                return [
+                    (self.map_tensor_name(name.replace("qkv", "q")), wq),
+                    (self.map_tensor_name(name.replace("qkv", "k")), wk),
+                    (self.map_tensor_name(name.replace("qkv", "v")), wv),
+                ]
+            elif 'patch_embed.proj.weight' in name:
+                # split Conv3D into Conv2Ds
+                c1, c2, kt, kh, kw = data_torch.shape
+                del c1, c2, kh, kw  # unused
+                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
+                return [
+                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight"  , data_torch[:, :, 0, ...]),
+                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...]),
+                ]
+            else:
+                return [(self.map_tensor_name(name), data_torch)]
+        return [] # skip other tensors
+
+
 @ModelBase.register("WavTokenizerDec")
 class WavTokenizerDecModel(TextModel):
     model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
@@ -3297,29 +3432,7 @@ class BertModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_causal_attention(False)
-
-        # get pooling path
-        pooling_path = None
-        module_path = self.dir_model / "modules.json"
-        if module_path.is_file():
-            with open(module_path, encoding="utf-8") as f:
-                modules = json.load(f)
-            for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
-                    pooling_path = mod["path"]
-                    break
-
-        # get pooling type
-        if pooling_path is not None:
-            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
-                pooling = json.load(f)
-            if pooling["pooling_mode_mean_tokens"]:
-                pooling_type = gguf.PoolingType.MEAN
-            elif pooling["pooling_mode_cls_token"]:
-                pooling_type = gguf.PoolingType.CLS
-            else:
-                raise NotImplementedError("Only MEAN and CLS pooling types supported")
-            self.gguf_writer.add_pooling_type(pooling_type)
+        self._try_set_pooling_type()
 
     def set_vocab(self):
         tokens, toktypes, tokpre = self.get_vocab_base()
@@ -3373,14 +3486,7 @@ class BertModel(TextModel):
 
         return [(self.map_tensor_name(name), data_torch)]
 
-
-@ModelBase.register("RobertaModel")
-class RobertaModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
+    def _xlmroberta_tokenizer_init(self) -> None:
         # we need the pad_token_id to know how to chop down position_embd matrix
         if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
             self._position_offset = 1 + pad_token_id
@@ -3389,82 +3495,7 @@ class RobertaModel(BertModel):
         else:
             self._position_offset = None
 
-    def set_vocab(self):
-        """Support BPE tokenizers for roberta models"""
-        bpe_tok_path = self.dir_model / "tokenizer.json"
-        if bpe_tok_path.exists():
-            self._set_vocab_gpt2()
-            self.gguf_writer.add_add_bos_token(True)
-            self.gguf_writer.add_add_eos_token(True)
-
-            # we need this to validate the size of the token_type embeddings
-            # though currently we are passing all zeros to the token_type embeddings
-            # "Sequence A" or "Sequence B"
-            self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
-
-        else:
-            return super().set_vocab()
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "roberta.", remove the prefix
-        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
-        if name.startswith("roberta."):
-            name = name[8:]
-
-        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
-        if name == "embeddings.position_embeddings.weight":
-            if self._position_offset is not None:
-                data_torch = data_torch[self._position_offset:,:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
-
-@ModelBase.register("NomicBertModel")
-class NomicBertModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.NOMIC_BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # the HF config claims n_ctx=8192, but it uses RoPE scaling
-        self.hparams["n_ctx"] = 2048
-
-        # SwigLU activation
-        assert self.hparams["activation_function"] == "swiglu"
-        # this doesn't do anything in the HF version
-        assert self.hparams["causal"] is False
-        # no bias tensors
-        assert self.hparams["qkv_proj_bias"] is False
-        assert self.hparams["mlp_fc1_bias"] is False
-        assert self.hparams["mlp_fc2_bias"] is False
-        # norm at end of layer
-        assert self.hparams["prenorm"] is False
-        # standard RoPE
-        assert self.hparams["rotary_emb_fraction"] == 1.0
-        assert self.hparams["rotary_emb_interleaved"] is False
-        assert self.hparams["rotary_emb_scale_base"] is None
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
-
-
-@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
-class XLMRobertaModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # we need the pad_token_id to know how to chop down position_embd matrix
-        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
-            self._position_offset = 1 + pad_token_id
-            if "max_position_embeddings" in self.hparams:
-                self.hparams["max_position_embeddings"] -= self._position_offset
-        else:
-            self._position_offset = None
-
-    def set_vocab(self):
+    def _xlmroberta_set_vocab(self) -> None:
         # to avoid TypeError: Descriptors cannot be created directly
         # exception when importing sentencepiece_model_pb2
         os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
@@ -3546,6 +3577,145 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
+
+@ModelBase.register("RobertaModel")
+class RobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def set_vocab(self):
+        """Support BPE tokenizers for roberta models"""
+        bpe_tok_path = self.dir_model / "tokenizer.json"
+        if bpe_tok_path.exists():
+            self._set_vocab_gpt2()
+            self.gguf_writer.add_add_bos_token(True)
+            self.gguf_writer.add_add_eos_token(True)
+
+            # we need this to validate the size of the token_type embeddings
+            # though currently we are passing all zeros to the token_type embeddings
+            # "Sequence A" or "Sequence B"
+            self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
+
+        else:
+            return super().set_vocab()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "roberta.", remove the prefix
+        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
+        if name.startswith("roberta."):
+            name = name[8:]
+
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("NomicBertModel")
+class NomicBertModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, **kwargs: Any):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model)
+
+        self.is_moe = bool(hparams.get("moe_every_n_layers"))
+        self.model_arch = gguf.MODEL_ARCH.NOMIC_BERT_MOE if self.is_moe else gguf.MODEL_ARCH.NOMIC_BERT
+
+        super().__init__(dir_model, ftype, fname_out, hparams=hparams, **kwargs)
+
+        self._tokenizer_is_xlmroberta = self._is_tokenizer_xlmroberta()
+        if self._tokenizer_is_xlmroberta:
+            self._xlmroberta_tokenizer_init()
+
+        npos, mtp = self.hparams["n_positions"], self.hparams.get("max_trained_positions", 2048)
+        if npos == 8192 and mtp == 2048:
+            self.hparams["n_positions"] = 2048  # nomic-embed-text v1 and v1.5 are trained for 2048 tokens.
+        elif npos == 2048 and mtp == 2048:
+            self.hparams["n_positions"] = 512   # nomic-embed-text-v2-moe is trained for 512 tokens.
+        else:
+            raise ValueError(f"unrecognized parameters: n_positions={npos}, max_trained_positions={mtp}")
+
+        assert self.hparams["activation_function"] == "gelu" if self.is_moe else "swiglu"
+
+        # this doesn't do anything in the HF version
+        assert self.hparams["causal"] is False
+        # no bias tensors unless MoE
+        assert self.hparams["qkv_proj_bias"] == self.is_moe
+        assert self.hparams["mlp_fc1_bias"]  == self.is_moe
+        assert self.hparams["mlp_fc2_bias"]  == self.is_moe
+
+        # norm at end of layer
+        assert self.hparams["prenorm"] is False
+        # standard RoPE
+        assert self.hparams["rotary_emb_fraction"] == 1.0
+        assert self.hparams["rotary_emb_interleaved"] is False
+        assert self.hparams["rotary_emb_scale_base"] is None
+
+    def set_vocab(self) -> None:
+        if self._tokenizer_is_xlmroberta:
+            return self._xlmroberta_set_vocab()
+        return super().set_vocab()
+
+    def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
+        # If the tensor is an experts bias tensor, skip it by returning an empty list.
+        if "mlp.experts.bias" in name:
+            return []  # Explicitly return an empty list.
+
+        if "mlp.experts.mlp.w1" in name:
+            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            name += ".weight"
+
+        if "mlp.experts.mlp.w2" in name:
+            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            data_torch = data_torch.transpose(1, 2)
+            name += ".weight"
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+        if self.is_moe:
+            self.gguf_writer.add_moe_every_n_layers(self.hparams["moe_every_n_layers"])
+            self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+            self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"])
+
+    def _is_tokenizer_xlmroberta(self) -> bool:
+        with open(self.dir_model / "tokenizer.json") as f:
+            tokenizer_json = json.load(f)
+        toktyp = tokenizer_json["model"]["type"]
+        if toktyp == "Unigram":
+            return True
+        if toktyp == "WordPiece":
+            return False
+        raise ValueError(f"unknown tokenizer: {toktyp}")
+
+
+@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
+class XLMRobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._xlmroberta_tokenizer_init()
+
+    def set_vocab(self):
+        self._xlmroberta_set_vocab()
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # if name starts with "roberta.", remove the prefix
         # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
@@ -5806,6 +5976,18 @@ def split_str_to_n_bytes(split_str: str) -> int:
     return n
 
 
+def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> str:
+    text_config = hparams.get("text_config", {})
+    vision_config = hparams.get("vision_config", {})
+    arch = hparams["architectures"][0]
+    # if "architectures" is found in the sub-config, use that instead
+    if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
+        arch = text_config["architectures"][0]
+    elif model_type == ModelType.VISION and vision_config.get("architectures") is not None:
+        arch = vision_config["architectures"][0]
+    return arch
+
+
 def main() -> None:
     args = parse_args()
 
@@ -5858,16 +6040,16 @@ def main() -> None:
 
     logger.info(f"Loading model: {dir_model.name}")
 
-    hparams = ModelBase.load_hparams(dir_model)
-
     if args.mmproj:
         if "mmproj" not in fname_out.name:
             fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-")
 
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
-        model_architecture = hparams["architectures"][0]
         model_type = ModelType.VISION if args.mmproj else ModelType.TEXT
+        hparams = ModelBase.load_hparams(dir_model)
+        model_architecture = get_model_architecture(hparams, model_type)
+        logger.info(f"Model architecture: {model_architecture}")
         try:
             model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type)
         except NotImplementedError:

examples/CMakeLists.txt

@@ -48,6 +48,7 @@ else()
     add_subdirectory(tokenize)
     add_subdirectory(tts)
     add_subdirectory(gen-docs)
+    add_subdirectory(training)
     if (NOT GGML_BACKEND_DL)
         # these examples use the backends directly and cannot be built with dynamic loading
         add_subdirectory(convert-llama2c-to-ggml)

examples/llama-bench/llama-bench.cpp

@@ -1133,8 +1133,6 @@ struct test {
             "build_commit", "build_number", "cpu_info",       "gpu_info",   "backends",     "model_filename",
             "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
             "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
-            "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "use_mmap",
-            "embeddings",   "n_prompt",     "n_gen",          "n_depth",    "test_time",    "avg_ns",
             "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
             "use_mmap",     "embeddings",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
             "avg_ns",       "stddev_ns",    "avg_ts",         "stddev_ts",

examples/llava/CMakeLists.txt

@@ -64,13 +64,7 @@ endif()
 add_executable(llama-llava-cli    deprecation-warning.cpp)
 add_executable(llama-gemma3-cli   deprecation-warning.cpp)
 add_executable(llama-minicpmv-cli deprecation-warning.cpp)
-
-set(TARGET llama-qwen2vl-cli)
-add_executable(${TARGET} qwen2vl-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-qwen2vl-cli)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+add_executable(llama-qwen2vl-cli  deprecation-warning.cpp)
 
 set(TARGET llama-mtmd-cli)
 add_executable(${TARGET} mtmd-cli.cpp)

examples/llava/README.md

@@ -34,6 +34,19 @@ llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
 
 # Pixtral 12B
 llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
+
+# Qwen 2 VL
+llama-mtmd-cli -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF
+
+# Qwen 2.5 VL
+llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF
+
+# Mistral Small 3.1 24B (IQ2_M quantization)
+llama-mtmd-cli -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF --chat-template mistral-v7
 ```
 
 ## How it works and what is `mmproj`?
@@ -57,7 +70,17 @@ Built upon `clip.cpp` (similar to `llava.cpp`), `libmtmd` offers several advanta
 
 ## How to obtain `mmproj`
 
-Multimodal projector (`mmproj`) files are specific to each model architecture. Please refer to the relevant guide for instructions on how to obtain or create them:
+Multimodal projector (`mmproj`) files are specific to each model architecture.
+
+For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file:
+- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support
+- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint
+- Qwen 2 VL and Qwen 2.5 VL (from [Qwen](https://huggingface.co/Qwen))
+- [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)
+
+For older models, please refer to the relevant guide for instructions on how to obtain or create them:
 
 - [LLaVA](../../docs/multimodal/llava.md)
 - [MobileVLM](../../docs/multimodal/MobileVLM.md)
@@ -67,9 +90,3 @@ Multimodal projector (`mmproj`) files are specific to each model architecture. P
 - [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
 - [IBM Granite Vision](../../docs/multimodal/granitevision.md)
 - [Google Gemma 3](../../docs/multimodal/gemma3.md)
-
-For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file:
-- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support
-- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
-- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
-- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint

examples/llava/clip-impl.h

@@ -2,8 +2,6 @@
 #include "gguf.h"
 #include "clip.h"
 
-#include "clip.h"
-
 #include <climits>
 #include <cstdarg>
 #include <string>
@@ -33,6 +31,7 @@
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
 #define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
 #define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
 
 #define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
 #define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
@@ -70,9 +69,11 @@
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_NORM     "mm.input_norm.weight"
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
 #define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
+#define TN_MM_PATCH_MERGER "mm.patch_merger.weight"     // mistral small 3.1
 #define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 
 // mimicpmv

examples/llava/clip.cpp

@@ -172,6 +172,7 @@ struct clip_hparams {
     std::unordered_set<int32_t> vision_feature_layer;
     int32_t attn_window_size = 0;
     int32_t n_wa_pattern = 0;
+    int32_t spatial_merge_size = 0;
 };
 
 struct clip_layer {
@@ -232,6 +233,7 @@ struct clip_vision_model {
     struct ggml_tensor * projection;
 
     // LLaVA projection
+    struct ggml_tensor * mm_input_norm_w = nullptr;
     struct ggml_tensor * mm_0_w = nullptr;
     struct ggml_tensor * mm_0_b = nullptr;
     struct ggml_tensor * mm_2_w = nullptr;
@@ -311,6 +313,7 @@ struct clip_vision_model {
 
     // pixtral
     struct ggml_tensor * token_embd_img_break = nullptr;
+    struct ggml_tensor * mm_patch_merger_w = nullptr;
 };
 
 struct clip_ctx {
@@ -637,6 +640,7 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
     const int d_head      = hidden_size / n_head;
     const int n_layer     = hparams.n_layer;
     const float eps       = hparams.eps;
+    const int n_merge     = hparams.spatial_merge_size;
 
     struct ggml_init_params params = {
         /*.mem_size   =*/ ctx->buf_compute_meta.size(),
@@ -721,7 +725,13 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         {
             ggml_tensor * gate_proj = ggml_mul_mat(ctx0, model.layers[il].ff_gate_w, cur);
             ggml_tensor * up_proj   = ggml_mul_mat(ctx0, model.layers[il].ff_up_w,   cur);
-            gate_proj = ggml_silu(ctx0, gate_proj); // pixtral uses silu
+            if (ctx->use_silu) {
+                gate_proj = ggml_silu(ctx0, gate_proj);
+            } else if (ctx->use_gelu) {
+                gate_proj = ggml_gelu(ctx0, gate_proj);
+            } else {
+                GGML_ABORT("Pixtral: Unsupported activation");
+            }
             cur = ggml_mul(ctx0, up_proj, gate_proj);
             cur = ggml_mul_mat(ctx0, model.layers[il].ff_down_w, cur);
         }
@@ -732,14 +742,42 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         embeddings = cur;
     }
 
-    // LlavaMultiModalProjector (with GELU activation)
+    // mistral small 3.1 patch merger
+    // ref: https://github.com/huggingface/transformers/blob/7a3e208892c06a5e278144eaf38c8599a42f53e7/src/transformers/models/mistral3/modeling_mistral3.py#L67
+    if (model.mm_patch_merger_w) {
+        GGML_ASSERT(hparams.spatial_merge_size > 0);
+
+        ggml_tensor * cur = embeddings;
+        cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.mm_input_norm_w);
+
+        // reshape image tokens to 2D grid
+        cur = ggml_reshape_3d(ctx0, cur, hidden_size, n_patches_x, n_patches_y);
+        cur = ggml_permute(ctx0, cur, 2, 0, 1, 3); // [x, y, hidden_size]
+        cur = ggml_cont(ctx0, cur);
+
+        // torch.nn.functional.unfold is just an im2col under the hood
+        // we just need a dummy kernel to make it work
+        ggml_tensor * kernel = ggml_view_3d(ctx0, cur, n_merge, n_merge, cur->ne[2], 0, 0, 0);
+        cur = ggml_im2col(ctx0, kernel, cur, n_merge, n_merge, 0, 0, 1, 1, true, inp->type);
+
+        // project to hidden_size
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
+        cur = ggml_mul_mat(ctx0, model.mm_patch_merger_w, cur);
+        embeddings = cur;
+    }
+
+    // LlavaMultiModalProjector (always using GELU activation)
     {
         embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings);
-        embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
+        if (model.mm_1_b) {
+            embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
+        }
 
         embeddings = ggml_gelu(ctx0, embeddings);
         embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
-        embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+        if (model.mm_2_b) {
+            embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+        }
     }
 
     // arrangement of the [IMG_BREAK] token
@@ -749,11 +787,14 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension
         // after the concatenation, we have a tensor with shape [hidden_size, n_patches_per_row + 1, n_rows]
 
+        const int p_y             = n_merge > 0 ? n_patches_y / n_merge : n_patches_y;
+        const int p_x             = n_merge > 0 ? n_patches_x / n_merge : n_patches_x;
+        const int p_total         = p_x * p_y;
         const int n_embd_text     = embeddings->ne[0];
-        const int n_tokens_output = num_patches + n_patches_y - 1; // one [IMG_BREAK] per row, except the last row
+        const int n_tokens_output = p_total + p_y - 1; // one [IMG_BREAK] per row, except the last row
 
-        ggml_tensor * cur = ggml_reshape_3d(ctx0, embeddings, n_embd_text, n_patches_x, n_patches_y);
-        ggml_tensor * tok = ggml_new_tensor_3d(ctx0, embeddings->type, n_embd_text, 1, n_patches_y);
+        ggml_tensor * cur = ggml_reshape_3d(ctx0, embeddings, n_embd_text, p_x, p_y);
+        ggml_tensor * tok = ggml_new_tensor_3d(ctx0, embeddings->type, n_embd_text, 1, p_y);
         tok = ggml_scale(ctx0, tok, 0.0); // clear the tensor
         tok = ggml_add(ctx0, tok, model.token_embd_img_break);
         cur = ggml_concat(ctx0, cur, tok, 1);
@@ -1574,7 +1615,7 @@ struct clip_model_loader {
     clip_ctx & ctx_clip;
     std::string fname;
 
-    size_t model_size; // in bytes
+    size_t model_size = 0; // in bytes
 
     // TODO @ngxson : we should not pass clip_ctx here, it should be clip_vision_model
     clip_model_loader(const char * fname, clip_ctx & ctx_clip) : ctx_clip(ctx_clip), fname(fname) {
@@ -1734,6 +1775,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_PIXTRAL:
                     {
                         hparams.rope_theta = 10000.0f;
+                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
                     } break;
                 case PROJECTOR_TYPE_QWEN25VL:
                     {
@@ -1748,6 +1790,8 @@ struct clip_model_loader {
             LOG_INF("%s: minicpmv_version:   %d\n", __func__, ctx_clip.minicpmv_version);
             LOG_INF("%s: proj_scale_factor:  %d\n", __func__, hparams.proj_scale_factor);
             LOG_INF("%s: n_wa_pattern:       %d\n", __func__, hparams.n_wa_pattern);
+            LOG_INF("%s: use_silu:           %d\n", __func__, ctx_clip.use_silu);
+            LOG_INF("%s: use_gelu:           %d\n", __func__, ctx_clip.use_gelu);
             LOG_INF("%s: model size:         %.2f MiB\n", __func__, model_size / 1024.0 / 1024.0);
             LOG_INF("%s: metadata size:      %.2f MiB\n", __func__, ggml_get_mem_size(ctx_meta.get()) / 1024.0 / 1024.0);
         }
@@ -1955,11 +1999,14 @@ struct clip_model_loader {
             case PROJECTOR_TYPE_PIXTRAL:
                 {
                     vision_model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
-                    vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
+                    vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"), false);
                     vision_model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
-                    vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                    vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"), false);
                     // [IMG_BREAK] token embedding
                     vision_model.token_embd_img_break = get_tensor(TN_TOK_IMG_BREAK);
+                    // for mistral small 3.1
+                    vision_model.mm_input_norm_w   = get_tensor(TN_MM_INP_NORM,     false);
+                    vision_model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false);
                 } break;
             default:
                 GGML_ASSERT(false && "unknown projector type");
@@ -2514,7 +2561,7 @@ struct llava_uhd {
 
         // no pinpoints, dynamically calculate the grid size (e.g. minicpmv)
 
-        auto best_size    = get_best_resize(original_size, slice_size, patch_size, has_slices);
+        auto best_size    = get_best_resize(original_size, slice_size, patch_size, !has_slices);
         res.overview_size = best_size;
 
         if (!has_slices) {
@@ -2823,15 +2870,18 @@ void clip_free(clip_ctx * ctx) {
     delete ctx;
 }
 
+// deprecated
 size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
-    return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
+    const int32_t nx = ctx->vision_model.hparams.image_size;
+    const int32_t ny = ctx->vision_model.hparams.image_size;
+    return clip_embd_nbytes_by_img(ctx, nx, ny);
 }
 
-size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) {
+size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h) {
     clip_image_f32 img;
     img.nx = img_w;
     img.ny = img_h;
-    return clip_n_patches_by_img(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float);
+    return clip_n_output_tokens(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float);
 }
 
 int32_t clip_get_image_size(const struct clip_ctx * ctx) {
@@ -2861,14 +2911,37 @@ size_t get_clip_image_grid_size(const struct clip_ctx * ctx) {
     return ctx->vision_model.hparams.image_grid_pinpoints.size();
 }
 
+// deprecated
 int clip_n_patches(const struct clip_ctx * ctx) {
     clip_image_f32 img;
     img.nx = ctx->vision_model.hparams.image_size;
     img.ny = ctx->vision_model.hparams.image_size;
-    return clip_n_patches_by_img(ctx, &img);
+    return clip_n_output_tokens(ctx, &img);
 }
 
+// deprecated
 int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
+    return clip_n_output_tokens(ctx, img);
+}
+
+int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
+    const auto & params = ctx->vision_model.hparams;
+    const int n_total = clip_n_output_tokens(ctx, img);
+    if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) {
+        return img->nx / (params.patch_size * 2) + (int)(img->nx % params.patch_size > 0);
+    }
+    return n_total;
+}
+
+int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
+    const auto & params = ctx->vision_model.hparams;
+    if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) {
+        return img->ny / (params.patch_size * 2) + (int)(img->ny % params.patch_size > 0);
+    }
+    return 1;
+}
+
+int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
     const auto & params = ctx->vision_model.hparams;
 
     int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
@@ -2898,8 +2971,9 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i
     } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) {
         n_patches /= ctx->vision_model.hparams.proj_scale_factor;
     } else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) {
-        int n_patches_x = img->nx / params.patch_size;
-        int n_patches_y = img->ny / params.patch_size;
+        int n_merge = ctx->vision_model.hparams.spatial_merge_size;
+        int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1);
+        int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1);
         n_patches = n_patches_y*n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
     }
 
@@ -3456,7 +3530,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->vision_model.mm_model_peg_0_b->ne[0];
         case PROJECTOR_TYPE_MLP:
         case PROJECTOR_TYPE_PIXTRAL:
-            return ctx->vision_model.mm_2_b->ne[0];
+            return ctx->vision_model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_MLP_NORM:
             return ctx->vision_model.mm_3_b->ne[0];
         case PROJECTOR_TYPE_MINICPMV:

examples/llava/clip.h

@@ -47,7 +47,7 @@ CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_par
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
+CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h);
 
 CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
 CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
@@ -59,9 +59,20 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
 CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
 
-CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
-CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
-CLIP_API int clip_n_mmproj_embd    (const struct clip_ctx * ctx);
+GGML_DEPRECATED(CLIP_API int clip_n_patches(const struct clip_ctx * ctx),
+    "use clip_n_output_tokens instead");
+GGML_DEPRECATED(CLIP_API int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img),
+    "use clip_n_output_tokens instead");
+
+CLIP_API int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// for M-RoPE, this will be the number of token positions in X and Y directions
+// for other models, X will be the total number of tokens and Y will be 1
+CLIP_API int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+CLIP_API int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// this should be equal to the embedding dimension of the text model
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
 CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);

examples/llava/llava.cpp

@@ -112,7 +112,7 @@ static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<
 }
 
 // Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
-static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out, clip_image_f32 * img_input) {
     struct {
         struct ggml_context * ctx;
     } model;
@@ -175,7 +175,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     model.ctx = ggml_init(params);
 
-    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_output_tokens(ctx_clip, img_input), num_images - 1); // example: 4096 x 576 x 4
     // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
     // fill it with the image embeddings, ignoring the base
     for (size_t i = 1; i < num_images; i++) {
@@ -214,8 +214,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
     // append without newline tokens (default behavior in llava_arch when not using unpad ):
-    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
-    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
+    memcpy(image_embd_out + clip_n_output_tokens(ctx_clip, img_input) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_output_tokens(ctx_clip, img_input));
 
     // Debug: Test single segments
     // Current findings: sending base image, sending a segment embedding all works similar to python
@@ -313,7 +313,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
                 image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
                 image_embd_v[i],
                 clip_embd_nbytes_by_img(ctx_clip, nx, ny));
-            n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
+            n_img_pos_out += clip_n_output_tokens(ctx_clip, img_res);
         }
         *n_img_pos = n_img_pos_out;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
@@ -342,8 +342,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     }
     else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
-        *n_img_pos = clip_n_patches(ctx_clip);
         clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        *n_img_pos = clip_n_output_tokens(ctx_clip, img_res);
         bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
         if (!encoded) {
             LOG_ERR("Unable to encode image\n");
@@ -381,7 +381,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
 
         int n_img_pos_out;
-        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
+        clip_image_f32 * img_input = clip_image_f32_get_img(img_res_v.get(), 0);
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out, img_input);
         *n_img_pos = n_img_pos_out;
 
         for (size_t i = 0; i < image_embd_v.size(); i++) {

examples/llava/mtmd-cli.cpp

@@ -72,6 +72,8 @@ struct mtmd_cli_context {
     llama_batch         batch;
     int                 n_batch;
 
+    std::vector<mtmd_bitmap> bitmaps;
+
     // note: we know that gemma3 template is "linear", meaning each turn is completely separated to another
     // so here we don't need to keep track of chat history
     common_chat_templates_ptr tmpls;
@@ -94,6 +96,7 @@ struct mtmd_cli_context {
             LOG_ERR("Model does not have chat template.\n");
             LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
             LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
+            LOG_ERR("  For Mistral Small 3.1, use '--chat-template mistral-v7'\n");
             exit(1);
         }
 
@@ -134,38 +137,14 @@ struct mtmd_cli_context {
             antiprompt_tokens.begin()
         );
     }
-};
 
-struct decode_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
+    bool load_image(const std::string & fname) {
+        mtmd_bitmap bitmap;
+        if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) {
+            return false;
         }
+        bitmaps.push_back(std::move(bitmap));
+        return true;
     }
 };
 
@@ -173,7 +152,7 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
     llama_tokens generated_tokens;
     for (int i = 0; i < n_predict; i++) {
         if (i > n_predict || !g_is_generating || g_is_interrupted) {
-            printf("\n");
+            LOG("\n");
             break;
         }
 
@@ -182,15 +161,15 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
         common_sampler_accept(smpl, token_id, true);
 
         if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
-            printf("\n");
+            LOG("\n");
             break; // end of generation
         }
 
-        printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
+        LOG("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
         fflush(stdout);
 
         if (g_is_interrupted) {
-            printf("\n");
+            LOG("\n");
             break;
         }
 
@@ -205,9 +184,7 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
     return 0;
 }
 
-static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
-    std::vector<mtmd_bitmap> bitmaps;
-
+static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, bool add_bos = false) {
     common_chat_templates_inputs tmpl_inputs;
     tmpl_inputs.messages = {msg};
     tmpl_inputs.add_generation_prompt = true;
@@ -215,15 +192,6 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vect
     auto formatted_chat = common_chat_templates_apply(ctx.tmpls.get(), tmpl_inputs);
     LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.prompt.c_str());
 
-    for (auto & fname : images_fname) {
-        mtmd_bitmap bitmap;
-        if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) {
-            LOG_ERR("Unable to load image %s\n", fname.c_str());
-            return 2; // image not found
-        }
-        bitmaps.push_back(std::move(bitmap));
-    }
-
     mtmd_input_text text;
     text.text          = formatted_chat.prompt;
     text.add_special   = add_bos;
@@ -232,18 +200,22 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vect
 
     if (g_is_interrupted) return 0;
 
-    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, bitmaps);
+    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, ctx.bitmaps);
     if (res != 0) {
         LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
         return 1;
     }
 
+    ctx.bitmaps.clear();
+
     if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks, ctx.n_past, 0, ctx.n_batch)) {
         LOG_ERR("Unable to eval prompt\n");
         return 1;
     }
 
-    ctx.n_past += mtmd_helper_get_n_tokens(chunks);
+    ctx.n_past += mtmd_helper_get_n_pos(chunks);
+
+    LOG("\n");
 
     return 0;
 }
@@ -267,7 +239,7 @@ int main(int argc, char ** argv) {
     }
 
     mtmd_cli_context ctx(params);
-    printf("%s: %s\n", __func__, params.model.path.c_str());
+    LOG("%s: loading model: %s\n", __func__, params.model.path.c_str());
 
     bool is_single_turn = !params.prompt.empty() && !params.image.empty();
 
@@ -300,7 +272,12 @@ int main(int argc, char ** argv) {
         common_chat_msg msg;
         msg.role = "user";
         msg.content = params.prompt;
-        if (eval_message(ctx, msg, params.image, true)) {
+        for (const auto & image : params.image) {
+            if (!ctx.load_image(image)) {
+                return 1; // error is already printed by libmtmd
+            }
+        }
+        if (eval_message(ctx, msg, true)) {
             return 1;
         }
         if (!g_is_interrupted && generate_response(ctx, smpl, n_predict)) {
@@ -315,7 +292,6 @@ int main(int argc, char ** argv) {
         LOG("\n");
 
         bool is_first_msg = true;
-        std::vector<std::string> images_fname;
         std::string content;
 
         while (!g_is_interrupted) {
@@ -340,10 +316,17 @@ int main(int argc, char ** argv) {
                 continue;
             }
             g_is_generating = true;
-            if (line.find("/image") == 0) {
+            if (line == "/image" || line.find("/image ") == 0) {
+                if (line.size() < 8) {
+                    LOG_ERR("ERR: Missing image filename\n");
+                    continue;
+                }
                 std::string image = line.substr(7);
-                images_fname.push_back(string_strip(image));
-                content += "<__image__>";
+                if (ctx.load_image(image)) {
+                    LOG("Image %s loaded\n", image.c_str());
+                    content += "<__image__>";
+                }
+                // else, error is already printed by libmtmd
                 continue;
             } else {
                 content += line;
@@ -351,26 +334,20 @@ int main(int argc, char ** argv) {
             common_chat_msg msg;
             msg.role = "user";
             msg.content = content;
-            int ret = eval_message(ctx, msg, images_fname, is_first_msg);
-            if (g_is_interrupted) break;
-            if (ret == 2) {
-                // non-fatal error
-                images_fname.clear();
-                content.clear();
-                continue;
-            }
+            int ret = eval_message(ctx, msg, is_first_msg);
             if (ret) {
                 return 1;
             }
+            if (g_is_interrupted) break;
             if (generate_response(ctx, smpl, n_predict)) {
                 return 1;
             }
-            images_fname.clear();
             content.clear();
             is_first_msg = false;
         }
     }
     if (g_is_interrupted) LOG("\nInterrupted by user\n");
+    LOG("\n\n");
     llama_perf_context_print(ctx.lctx);
     return g_is_interrupted ? 130 : 0;
 }

examples/llava/mtmd.cpp

@@ -40,11 +40,14 @@ struct mtmd_context {
     llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice
     llama_token tok_row_end       = LLAMA_TOKEN_NULL; // end of row
 
+    bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE
+
     // TODO @ngxson : add timings
 
     mtmd_context(const char * mmproj_fname,
                    const llama_model * text_model,
                    const mtmd_context_params & ctx_params) :
+        text_model   (text_model),
         print_timings(ctx_params.print_timings),
         n_threads    (ctx_params.n_threads),
         image_marker (ctx_params.image_marker)
@@ -56,9 +59,8 @@ struct mtmd_context {
         if (!ctx_clip) {
             throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
         }
-        this->text_model = text_model;
 
-        GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead");
+        use_mrope = clip_is_qwen2vl(ctx_clip);
 
         int minicpmv_version = clip_is_minicpmv(ctx_clip);
         if (minicpmv_version == 2) {
@@ -126,6 +128,7 @@ struct mtmd_image_tokens_data {
 struct mtmd_image_tokens {
     uint32_t nx; // number of tokens in x direction
     uint32_t ny; // number of tokens in y direction
+    bool use_mrope_pos = false; // use M-RoPE position counting (the whole image is 1 temporal position)
     uint32_t n_tokens() const { return nx * ny; }
     clip_image_f32_batch batch_f32; // preprocessed image patches
     std::string id; // optional user-defined ID, useful for KV cache tracking
@@ -202,6 +205,13 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
         string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
     }
 
+    else if (proj_type == PROJECTOR_TYPE_QWEN2VL || proj_type == PROJECTOR_TYPE_QWEN25VL) {
+        // <|vision_start|> ... (image embeddings) ... <|vision_end|>
+        marker_modified = "<|vision_start|>" + ctx->image_marker + "<|vision_end|>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
+    }
+
     // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
 
     std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
@@ -226,7 +236,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
 
         for (auto & entry : batch_f32.entries) {
             mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-            image_tokens->nx = clip_n_patches_by_img(ctx->ctx_clip, entry.get());
+            image_tokens->nx = clip_n_output_tokens(ctx->ctx_clip, entry.get());
             image_tokens->ny = 1;
             image_tokens->batch_f32.entries.push_back(std::move(entry));
             image_tokens->id = id;
@@ -322,12 +332,20 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
             } else {
                 size_t n_tokens = 0;
                 for (const auto & entry : batch_f32.entries) {
-                    n_tokens += clip_n_patches_by_img(ctx->ctx_clip, entry.get());
+                    n_tokens += clip_n_output_tokens(ctx->ctx_clip, entry.get());
                 }
 
                 mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-                image_tokens->nx = n_tokens;
-                image_tokens->ny = 1; // TODO
+                if (ctx->use_mrope) {
+                    // for Qwen2VL, we need this information for M-RoPE decoding positions
+                    image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_clip, batch_f32.entries[0].get());
+                    image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_clip, batch_f32.entries[0].get());
+                    image_tokens->use_mrope_pos = true;
+                } else {
+                    // other models, we only need the total number of tokens
+                    image_tokens->nx = n_tokens;
+                    image_tokens->ny = 1;
+                }
                 image_tokens->batch_f32 = std::move(batch_f32);
                 image_tokens->id = bitmaps[i_img].id; // optional
 
@@ -372,6 +390,13 @@ std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
     return image_tokens->id;
 }
 
+llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) {
+    if (image_tokens->use_mrope_pos) {
+        return 1; // for M-RoPE, the whole image is 1 in temporal dimension
+    }
+    return image_tokens->n_tokens();
+}
+
 int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
     int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
     ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
@@ -389,7 +414,7 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
         // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
         const auto & entries = image_tokens->batch_f32.entries;
         for (size_t i = 0; i < entries.size(); i++) {
-            int n_tokens_per_image = clip_n_patches_by_img(ctx->ctx_clip, entries[i].get());
+            int n_tokens_per_image = clip_n_output_tokens(ctx->ctx_clip, entries[i].get());
             ok = clip_image_encode(
                 ctx->ctx_clip,
                 ctx->n_threads,
@@ -417,7 +442,7 @@ size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
         if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
             n_tokens += chunk.tokens_text.size();
         } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            n_tokens += chunk.tokens_image->n_tokens();
+            n_tokens += mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
         } else {
             GGML_ASSERT(false && "chunk type not supported");
         }
@@ -425,22 +450,38 @@ size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
     return n_tokens;
 }
 
+llama_pos mtmd_helper_get_n_pos(mtmd_input_chunks & chunks) {
+    llama_pos n_pos = 0;
+    for (auto & chunk : chunks) {
+        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            n_pos += chunk.tokens_text.size();
+        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            n_pos += mtmd_image_tokens_get_n_pos(chunk.tokens_image.get());
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+    return n_pos;
+}
+
 // helper struct to make working with embd batch easier
 // note: this will be removed after llama_batch_ext refactoring
 struct decode_embd_batch {
+    int n_pos_per_embd;
+    int n_mmproj_embd;
     std::vector<llama_pos>      pos;
+    std::vector<llama_pos>      pos_view; // used by mrope
     std::vector<int32_t>        n_seq_id;
     std::vector<llama_seq_id>   seq_id_0;
     std::vector<llama_seq_id *> seq_ids;
     std::vector<int8_t>         logits;
     llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
+    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+        pos     .resize(n_tokens * n_pos_per_embd);
         n_seq_id.resize(n_tokens);
         seq_ids .resize(n_tokens + 1);
         logits  .resize(n_tokens);
         seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
         seq_ids [n_tokens] = nullptr;
         batch = {
             /*n_tokens       =*/ n_tokens,
@@ -451,13 +492,64 @@ struct decode_embd_batch {
             /*seq_id         =*/ seq_ids.data(),
             /*logits         =*/ logits.data(),
         };
-        for (int i = 0; i < n_tokens; i++) {
+    }
+
+    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.n_tokens; i++) {
             batch.pos     [i] = pos_0 + i;
             batch.n_seq_id[i] = 1;
             batch.seq_id  [i] = seq_id_0.data();
             batch.logits  [i] = false;
         }
     }
+
+    void set_position_mrope(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int y = 0; y < ny; y++) {
+            for (int x = 0; x < nx; x++) {
+                int i = y * nx + x;
+                pos[i                     ] = pos_0;
+                pos[i + batch.n_tokens    ] = pos_0 + y;
+                pos[i + batch.n_tokens * 2] = pos_0 + x;
+                pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
+            }
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    llama_batch get_view(int offset, int n_tokens) {
+        llama_pos * pos_ptr;
+        pos_view.clear();
+        pos_view.resize(n_tokens * n_pos_per_embd);
+        if (n_pos_per_embd > 1) {
+            // mrope
+            // for example, with layout of src: 1234...1234...1234...1234...
+            //       offset 2 will give us dst: 34...34...34...34...
+            for (int i = 0; i < n_pos_per_embd; i++) {
+                auto src = pos.begin() + i * batch.n_tokens + offset;
+                pos_view.insert(pos_view.end(), src, src + n_tokens);
+            }
+            pos_ptr = pos_view.data();
+        } else {
+            // normal
+            pos_ptr = pos.data() + offset;
+        }
+        return {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
+            /*pos            =*/ pos_ptr,
+            /*n_seq_id       =*/ batch.n_seq_id + offset,
+            /*seq_id         =*/ batch.seq_id   + offset,
+            /*logits         =*/ batch.logits   + offset,
+        };
+    }
 };
 
 int32_t mtmd_helper_eval(mtmd_context * ctx,
@@ -470,6 +562,7 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
     llama_pos n_past = pos0;
     llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
     int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
+    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
 
     for (auto & chunk : chunks) {
         bool is_last = &chunk == &chunks.back();
@@ -497,7 +590,7 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
             }
 
         } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            GGML_ASSERT(!is_last && "logits for last image chunk is not yet support");
+            GGML_ASSERT(!is_last && "logits for last image chunk is not yet supported");
             GGML_ASSERT(chunk.tokens_image != nullptr);
             int64_t t0 = ggml_time_ms();
             if (ctx->print_timings) {
@@ -517,6 +610,16 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
             int32_t i_batch = 0;
             int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
             float * embd = mtmd_get_output_embd(ctx);
+            decode_embd_batch batch_embd(embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
+
+            const int nx = mtmd_image_tokens_get_nx(chunk.tokens_image.get());
+            const int ny = mtmd_image_tokens_get_ny(chunk.tokens_image.get());
+
+            if (mtmd_decode_use_mrope(ctx)) {
+                batch_embd.set_position_mrope(n_past, nx, ny, seq_id);
+            } else {
+                batch_embd.set_position_normal(n_past, seq_id);
+            }
 
             if (mtmd_decode_use_non_causal(ctx)) {
                 llama_set_causal_attn(lctx, false);
@@ -524,15 +627,14 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
             }
 
             while (i_batch < n_img_batches) { // split into batches
-                int32_t pos_offset = i_batch*n_batch;
-                int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
-                float * embd_batch = embd + pos_offset*n_mmproj_embd;
-                decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0);
+                int pos_offset = i_batch*n_batch;
+                int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
+                llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);
 
-                printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
+                LOG_INF("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
 
                 int64_t t1 = ggml_time_ms();
-                ret = llama_decode(lctx, batch_img.batch);
+                ret = llama_decode(lctx, batch_embd_view);
                 if (ret != 0) {
                     LOG_ERR("failed to decode image\n");
                     llama_set_causal_attn(lctx, true); // restore causal attn
@@ -545,9 +647,11 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
                 }
 
                 i_batch++;
-                n_past += n_tokens_batch;
             }
 
+            // for mrope, one image is one single **temporal** position
+            n_past += mtmd_decode_use_mrope(ctx) ? 1 : n_tokens;
+
             if (mtmd_decode_use_non_causal(ctx)) {
                 llama_set_causal_attn(lctx, true);
             }
@@ -595,6 +699,10 @@ bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
     return false;
 }
 
+bool mtmd_decode_use_mrope(mtmd_context * ctx) {
+    return ctx->use_mrope;
+}
+
 void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
     mtmd_image_tokens_free(val);
 }

examples/llava/mtmd.h

@@ -102,6 +102,7 @@ MTMD_API size_t      mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * im
 MTMD_API size_t      mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens);
 MTMD_API size_t      mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens);
 MTMD_API std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens);
+MTMD_API llama_pos   mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens); // number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
 MTMD_API void        mtmd_image_tokens_free(mtmd_image_tokens * image_tokens);
 
 // returns 0 on success
@@ -114,15 +115,21 @@ MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 // whether we need to set non-causal mask before llama_decode
 MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
 
+// whether the current model use M-RoPE for llama_decode
+MTMD_API bool mtmd_decode_use_mrope(mtmd_context * ctx);
+
 
 
 //
 // helper functions (can be implemented based on other functions)
 //
 
-// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past
+// helper to count the total number of tokens from a list of chunks, useful to keep track of KV cache
 MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks);
 
+// helper to count the total position of tokens from a list of chunks, useful to keep track of n_past
+MTMD_API llama_pos mtmd_helper_get_n_pos(mtmd_input_chunks & chunks);
+
 // helper function that automatically:
 // 1. run llama_decode() on text chunks
 // 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()

examples/llava/qwen2_vl_surgery.py

@@ -1,217 +0,0 @@
-import argparse
-from typing import Dict, List, Optional
-
-import torch
-import numpy as np
-from gguf import *
-from transformers import (
-    AutoProcessor,
-    Qwen2VLConfig,
-    Qwen2VLProcessor,
-    Qwen2VLForConditionalGeneration,
-    Qwen2_5_VLConfig, # type: ignore[reportAttributeAccessIssue]
-    Qwen2_5_VLForConditionalGeneration, # type: ignore[reportAttributeAccessIssue]
-)
-
-
-VISION = "clip.vision"
-
-
-def k(raw_key: str, arch: str) -> str:
-    return raw_key.format(arch=arch)
-
-
-def get_n_wa_pattern(fullatt_block_indexes: Optional[List[int]]):
-    if fullatt_block_indexes is None:
-        return 0
-    n_wa = fullatt_block_indexes[0]
-    for a, b in zip(fullatt_block_indexes, fullatt_block_indexes[1:]):
-        if b - a - 1 != n_wa:
-            raise ValueError(
-                f"window/full attention layer should have fix pattern of "
-                f"for each full-attention layer followed by {n_wa} window-attention layers"
-            )
-    return n_wa + 1
-
-
-class VL2:
-
-    @staticmethod
-    def to_gguf_name(name: str) -> str:
-        og = name
-        name = name.replace("text_model", "t").replace("vision_model", "v")
-        name = name.replace("blocks", "blk").replace("embeddings.", "")
-        name = name.replace("attn.", "attn_")
-        name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
-        # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
-        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-        name = name.replace("merger.mlp", 'mm')
-        print(f"[to_gguf_name] {og} --> {name}")
-        return name
-
-    @classmethod
-    def find_vision_tensors(cls, qwen2vl, dtype) -> Dict[str, np.ndarray]:
-        vision_model = qwen2vl.visual
-        tensor_map = {}
-        for name, ten in vision_model.state_dict().items():
-            ten = ten.numpy()
-            if 'qkv' in name:
-                if ten.ndim == 2: # weight
-                    c3, _ = ten.shape
-                else:             # bias
-                    c3 = ten.shape[0]
-                assert c3 % 3 == 0
-                c = c3 // 3
-                wq = ten[:c]
-                wk = ten[c: c * 2]
-                wv = ten[c * 2:]
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
-            elif 'merger' in name:
-                if name.endswith("ln_q.weight"):
-                    tensor_map['v.post_ln.weight'] = ten
-                elif name.endswith("ln_q.bias"):
-                    tensor_map['v.post_ln.bias'] = ten
-                else:
-                    # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
-                    tensor_map[cls.to_gguf_name(name)] = ten
-            elif 'patch_embed.proj.weight' in name:
-                # NOTE: split Conv3D into Conv2Ds
-                c1, c2, kt, kh, kw = ten.shape
-                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
-                tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
-                tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
-            else:
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}")] = ten
-
-        for new_name, ten in tensor_map.items():
-            if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
-                tensor_map[new_name] = ten.astype(np.float32)
-            else:
-                tensor_map[new_name] = ten.astype(dtype)
-        tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
-        return tensor_map
-
-
-class VL25(VL2):
-
-    @staticmethod
-    def to_gguf_name(name: str) -> str:
-        og = name
-        name = name.replace("text_model", "t").replace("vision_model", "v")
-        name = name.replace("blocks", "blk").replace("embeddings.", "")
-        name = name.replace("attn.", "attn_")
-        name = name.replace("mlp.down_proj", "ffn_down").replace("mlp.up_proj", "ffn_up")
-        name = name.replace("mlp.gate_proj", "ffn_gate").replace("proj.", "out.")
-        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-        name = name.replace("merger.mlp", 'mm')
-        print(f"[vl25][to_gguf_name] {og} --> {name}")
-        return name
-
-
-def main(args):
-    if args.data_type == 'fp32':
-        dtype = torch.float32
-        np_dtype = np.float32
-        ftype = 0
-    elif args.data_type == 'fp16':
-        dtype = torch.float16
-        np_dtype = np.float16
-        ftype = 1
-    else:
-        raise ValueError()
-
-    local_model = False
-    model_path = ""
-    model_name = args.model_name
-    print("model_name: ", model_name)
-    if args.model_type == "qwen2vl":
-        qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
-            model_name, torch_dtype=dtype, device_map="cpu"
-        )
-        cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-        vcfg = cfg.vision_config
-    else:
-        qwen2vl = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-            model_name, torch_dtype=dtype, device_map="cpu"
-        )
-        cfg: Qwen2_5_VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-        vcfg = cfg.vision_config
-
-    if os.path.isdir(model_name):
-        local_model = True
-        if model_name.endswith(os.sep):
-            model_name = model_name[:-1]
-        model_path = model_name
-        model_name = os.path.basename(model_name)
-    fname_out = f"{model_name.replace('/', '-').lower()}-vision.gguf"
-
-    fout = GGUFWriter(path=fname_out, arch="clip")
-    fout.add_description("image encoder for Qwen2VL")
-
-    fout.add_file_type(ftype)
-    fout.add_bool("clip.has_text_encoder", False)
-    fout.add_bool("clip.has_vision_encoder", True)
-    fout.add_bool("clip.has_qwen2vl_merger", True)
-
-    print(cfg.vision_config)
-    if 'silu' in cfg.vision_config.hidden_act.lower():
-        fout.add_bool("clip.use_silu", True)
-        fout.add_bool("clip.use_gelu", False)
-    elif 'gelu' in cfg.vision_config.hidden_act.lower():
-        fout.add_bool("clip.use_silu", False)
-        fout.add_bool("clip.use_gelu", 'quick' not in cfg.vision_config.hidden_act.lower())
-    else:
-        raise ValueError()
-
-    if args.model_type == "qwen2.5vl":
-        fout.add_uint32("clip.vision.n_wa_pattern", get_n_wa_pattern(vcfg.fullatt_block_indexes))
-        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.hidden_size)
-        fout.add_uint32("clip.vision.projection_dim", vcfg.out_hidden_size)
-        fout.add_string("clip.projector_type", "qwen2.5vl_merger")
-    else:
-        fout.add_string("clip.projector_type", "qwen2vl_merger")
-        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
-        fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
-
-    if args.model_type == "qwen2.5vl":
-        tensor_map = VL25.find_vision_tensors(qwen2vl, np_dtype)
-    else:
-        tensor_map = VL2.find_vision_tensors(qwen2vl, np_dtype)
-    for name, data in tensor_map.items():
-        fout.add_tensor(name, data)
-
-    fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
-    fout.add_uint32("clip.vision.image_size", 14 * 40)  # some reasonable size that is divable by (14*2)
-    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
-    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
-    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
-    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 0)  # not sure what this does, put 0 here as a placeholder
-    fout.add_name(model_name)
-    """
-    HACK: Since vision rope related parameter aren't stored in the `Qwen2VLConfig,
-            it will be hardcoded in the `clip_image_build_graph` from `clip.cpp`.
-    """
-
-    if local_model:
-        processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_path)
-    else:
-        processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name)
-    fout.add_array("clip.vision.image_mean", processor.image_processor.image_mean) # type: ignore[reportAttributeAccessIssue]
-    fout.add_array("clip.vision.image_std", processor.image_processor.image_std) # type: ignore[reportAttributeAccessIssue]
-
-    fout.write_header_to_file()
-    fout.write_kv_data_to_file()
-    fout.write_tensors_to_file()
-    fout.close()
-    print("save model as: ", fname_out)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
-    parser.add_argument("--model_type", nargs='?', choices=['qwen2vl', 'qwen2.5vl'], default="qwen2vl")
-    parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
-    args = parser.parse_args()
-    main(args)

examples/llava/qwen2vl-test.cpp

@@ -27,6 +27,8 @@
 #include <cassert>
 #include <cmath>
 
+// THIS FILE IS ONLY USED FOR TESTING THE QWEN2VL MODEL
+// IT IS NOT A PRODUCTION CODE
 
 static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
                                      int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) {

examples/llava/tests.sh

@@ -36,12 +36,6 @@ add_test() {
     arr_tmpl+=("$tmpl")
 }
 
-add_test_big() {
-    if [ "$RUN_BIG_TESTS" = true ]; then
-        add_test "$@"
-    fi
-}
-
 add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-500M-Instruct-GGUF:Q8_0"
 add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-2.2B-Instruct-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0"
@@ -54,11 +48,20 @@ add_test "llama-mtmd-cli"  "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K"  # model from openbmb is corrupted
 add_test "llama-mtmd-cli"  "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
 add_test "llama-mtmd-cli"  "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
-add_test "llama-qwen2vl-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
-add_test "llama-qwen2vl-cli"  "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
 
 # to test the big models, run: ./tests.sh big
-add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M"
+if [ "$RUN_BIG_TESTS" = true ]; then
+    add_test "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli" "ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF" "mistral-v7"
+    add_test "llama-mtmd-cli" "ggml-org/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli" "ggml-org/Qwen2-VL-7B-Instruct-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
+    add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-7B-Instruct-GGUF:Q4_K_M"
+    # add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-32B-Instruct-GGUF:Q4_K_M" # does not work on my mac M3 Ultra
+    # add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-72B-Instruct-GGUF:Q4_K_M" # too big
+fi
 
 # these models always give the wrong answer, not sure why
 # add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M"

examples/rpc/rpc-server.cpp

@@ -304,8 +304,9 @@ int main(int argc, char * argv[]) {
         get_backend_memory(&free_mem, &total_mem);
     }
     const char * cache_dir = nullptr;
-    std::string cache_dir_str = fs_get_cache_directory() + "rpc/";
+    std::string cache_dir_str;
     if (params.use_cache) {
+        cache_dir_str = fs_get_cache_directory() + "rpc/";
         if (!fs_create_directory_with_parents(cache_dir_str)) {
             fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str());
             return 1;

examples/server/README.md

@@ -154,7 +154,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
-| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
+| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
 | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
 | `--slots` | enable slots monitoring endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
 | `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |

examples/server/utils.hpp

@@ -642,9 +642,31 @@ static json oaicompat_completion_params_parse(
         throw std::runtime_error("Cannot use custom grammar constraints with tools.");
     }
 
+    // if the assistant message appears at the end of list, we do not add end-of-turn token
+    // for ex. this can be useful to modify the reasoning process in reasoning models
+    bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant";
+    common_chat_msg last_message;
+    if (prefill_assistant_message) {
+        last_message = inputs.messages.back();
+        inputs.messages.pop_back();
+
+        /* sanity check, max one assistant message at the end of the list */
+        if (!inputs.messages.empty() && inputs.messages.back().role == "assistant"){
+            throw std::runtime_error("Cannot have 2 or more assistant messages at the end of the list.");
+        }
+
+        inputs.extract_reasoning = false;
+        inputs.add_generation_prompt = true;
+    }
+
     // Apply chat template to the list of messages
     auto chat_params = common_chat_templates_apply(tmpls, inputs);
 
+    /* Append assistant prefilled message */
+    if (prefill_assistant_message) {
+         chat_params.prompt += last_message.content;
+    }
+
     llama_params["chat_format"]      = static_cast<int>(chat_params.format);
     llama_params["prompt"]           = chat_params.prompt;
     if (!chat_params.grammar.empty()) {

examples/training/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-finetune)
+add_executable(${TARGET} finetune.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/training/README.md

@@ -0,0 +1,17 @@
+# llama.cpp/examples/training
+
+This directory contains examples related to language model training using llama.cpp/GGML.
+So far finetuning is technically functional (for FP32 models and limited hardware setups) but the code is very much WIP.
+Finetuning of Stories 260K and LLaMA 3.2 1b seems to work with 24 GB of memory.
+**For CPU training, compile llama.cpp without any additional backends such as CUDA.**
+**For CUDA training, use the maximum number of GPU layers.**
+
+Proof of concept:
+
+``` sh
+export model_name=llama_3.2-1b && export quantization=f32
+./build/bin/finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
+./build/bin/perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
+```
+
+The perplexity value of the finetuned model should be lower after training on the test set for 2 epochs.

examples/training/finetune.cpp

@@ -0,0 +1,97 @@
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    params.logits_all = true;
+    params.escape = false;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
+        return 1;
+    }
+
+    if (params.use_mmap) {
+        LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n", __func__);
+        params.use_mmap = false;
+    }
+    if (params.cache_type_k == GGML_TYPE_F16) {
+        LOG_INF("%s: force changing k cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
+        params.cache_type_k = GGML_TYPE_F32;
+    }
+    if (params.cache_type_v == GGML_TYPE_F16) {
+        LOG_INF("%s: force changing v cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
+        params.cache_type_v = GGML_TYPE_F32;
+    }
+
+    common_init();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model and apply lora adapter, if any
+    common_init_result llama_init = common_init_from_params(params);
+    llama_model_ptr   & model = llama_init.model;
+    llama_context_ptr & ctx   = llama_init.context;
+
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n", __func__);
+        return 1;
+    }
+
+    // print system information
+    {
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+    }
+
+    constexpr float val_split = 0.05f;
+
+    std::vector<llama_token> tokens = common_tokenize(ctx.get(), params.prompt, true);
+    ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get())/2);
+
+    struct ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr);
+    optimizer_params.adamw.alpha = 1e-7f; // learning rate
+
+    struct llama_opt_params lopt_params {
+        /*n_ctx_train     =*/ 0,
+        /*param_filter    =*/ llama_opt_param_filter_all,
+        /*param_filter_ud =*/ nullptr,
+        /*get_opt_pars    =*/ ggml_opt_get_constant_optimizer_params,
+        /*get_opt_pars_ud =*/ &optimizer_params,
+    };
+    llama_opt_init(ctx.get(), model.get(), lopt_params);
+
+    const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split);
+
+    ggml_opt_result_t result_train = ggml_opt_result_init();
+    ggml_opt_result_t result_eval  = ggml_opt_result_init();
+
+    for (int epoch = 0; epoch < 2; ++epoch) {
+        llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
+            ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
+        fprintf(stderr, "\n");
+
+        ggml_opt_result_reset(result_train);
+        ggml_opt_result_reset(result_eval);
+    }
+    ggml_opt_result_free(result_train);
+    ggml_opt_result_free(result_eval);
+
+    llama_model_save_to_file(model.get(), "finetuned-model.gguf");
+
+    llama_backend_free();
+
+    return 0;
+}

ggml/CMakeLists.txt

@@ -360,3 +360,27 @@ write_basic_package_version_file(
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
+
+if (MSVC)
+    set(MSVC_WARNING_FLAGS
+        /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
+    )
+    function(disable_msvc_warnings target_name)
+        if(TARGET ${target_name})
+            target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
+        endif()
+    endfunction()
+
+    disable_msvc_warnings(ggml-base)
+    disable_msvc_warnings(ggml)
+    disable_msvc_warnings(ggml-cpu)
+    disable_msvc_warnings(ggml-cpu-x64)
+    disable_msvc_warnings(ggml-cpu-sse42)
+    disable_msvc_warnings(ggml-cpu-sandybridge)
+    disable_msvc_warnings(ggml-cpu-haswell)
+    disable_msvc_warnings(ggml-cpu-skylakex)
+    disable_msvc_warnings(ggml-cpu-icelake)
+    disable_msvc_warnings(ggml-cpu-alderlake)
+endif()

ggml/include/ggml-cpp.h

@@ -24,7 +24,7 @@ typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;
 
 struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };
 
-typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
+typedef std::unique_ptr<ggml_gallocr, ggml_gallocr_deleter> ggml_gallocr_ptr;
 
 // ggml-backend
 

ggml/include/ggml-opt.h

@@ -37,13 +37,16 @@ extern "C" {
     // ====== Dataset ======
 
     GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
-            int64_t ne_datapoint, // number of elements per datapoint
-            int64_t ne_label,     // number of elements per label
-            int64_t ndata,        // total number of datapoints/labels
-            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+            enum ggml_type type_data,    // the type for the internal data tensor
+            enum ggml_type type_label,   // the type for the internal labels tensor
+            int64_t        ne_datapoint, // number of elements per datapoint
+            int64_t        ne_label,     // number of elements per label
+            int64_t        ndata,        // total number of datapoints/labels
+            int64_t        ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
     GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
 
     // get underlying tensors that store the data
+    GGML_API int64_t              ggml_opt_dataset_ndata (ggml_opt_dataset_t dataset);
     GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
     GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
 
@@ -56,13 +59,19 @@ extern "C" {
             struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
             struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
             int64_t              ibatch);
+    GGML_API void ggml_opt_dataset_get_batch_host(
+            ggml_opt_dataset_t   dataset,
+            void               * data_batch,
+            size_t               nb_data_batch,
+            void               * labels_batch,
+            int64_t              ibatch);
 
     // ====== Model / Context ======
 
     enum ggml_opt_build_type {
-        GGML_OPT_BUILD_TYPE_FORWARD,
-        GGML_OPT_BUILD_TYPE_GRAD,
-        GGML_OPT_BUILD_TYPE_OPT,
+        GGML_OPT_BUILD_TYPE_FORWARD = 10,
+        GGML_OPT_BUILD_TYPE_GRAD    = 20,
+        GGML_OPT_BUILD_TYPE_OPT     = 30,
     };
 
     // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
@@ -81,20 +90,22 @@ extern "C" {
     // userdata can be used to pass arbitrary data
     typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
 
-    // returns the default optimizer params (constant)
+    // returns the default optimizer params (constant, hard-coded values)
     // userdata is not used
     GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
 
+    // casts userdata to ggml_opt_optimizer_params and returns it
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata);
+
     // parameters for initializing a new optimization context
     struct ggml_opt_params {
         ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
 
-        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
-
-        // the forward graph is defined by inputs and outputs
-        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
-        struct ggml_tensor * inputs;
-        struct ggml_tensor * outputs;
+        // by default the forward graph needs to be reconstructed for each eval
+        // if ctx_compute, inputs, and outputs are set the graphs are instead allocated statically
+        struct ggml_context * ctx_compute;
+        struct ggml_tensor  * inputs;
+        struct ggml_tensor  * outputs;
 
         enum ggml_opt_loss_type  loss_type;
         enum ggml_opt_build_type build_type;
@@ -107,12 +118,9 @@ extern "C" {
 
     // get parameters for an optimization context with defaults set where possible
     // parameters for which no sensible defaults exist are supplied as arguments to this function
-    GGML_API ggml_opt_params ggml_opt_default_params(
-            ggml_backend_sched_t      backend_sched,
-            struct ggml_context     * ctx_compute,
-            struct ggml_tensor      * inputs,
-            struct ggml_tensor      * outputs,
-            enum ggml_opt_loss_type   loss_type);
+    GGML_API struct ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t    backend_sched,
+            enum ggml_opt_loss_type loss_type);
 
     GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
     GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
@@ -121,6 +129,7 @@ extern "C" {
     GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
 
     // get underlying tensors that store data
+    // if not using static graphs these pointers become invalid with the next call to ggml_opt_alloc
     GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
     GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
     GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
@@ -128,11 +137,12 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
     GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
 
+    // get the gradient accumulator for a node from the forward graph
     GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
 
     // ====== Optimization Result ======
 
-    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API ggml_opt_result_t ggml_opt_result_init(void);
     GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
     GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
 
@@ -144,11 +154,20 @@ extern "C" {
 
     // ====== Computation ======
 
-    // do forward pass, increment result if not NULL
-    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+    // if not using static graphs, this function must be called prior to ggml_opt_alloc
+    GGML_API void ggml_opt_prepare_alloc(
+        ggml_opt_context_t    opt_ctx,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * gf,
+        struct ggml_tensor  * inputs,
+        struct ggml_tensor  * outputs);
 
-    // do forward pass, increment result if not NULL, do backward pass
-    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+    // allocate the next graph for evaluation, either forward or forward + backward
+    // must be called exactly once prior to calling ggml_opt_eval
+    GGML_API void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward);
+
+    // do forward pass, increment result if not NULL, do backward pass if allocated
+    GGML_API void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
 
     // ############################################################################
     // ## The high-level functions start here. They do not depend on any private ##
@@ -200,9 +219,9 @@ extern "C" {
     // fit model defined by inputs and outputs to dataset
     GGML_API void ggml_opt_fit(
             ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
-            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
-            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
-            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            struct ggml_context           * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            struct ggml_tensor            * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            struct ggml_tensor            * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
             ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
             enum ggml_opt_loss_type         loss_type,      // loss to minimize
             ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)

ggml/include/ggml.h

@@ -764,7 +764,7 @@ extern "C" {
     // Tensor flags
     GGML_API void ggml_set_input(struct ggml_tensor * tensor);
     GGML_API void ggml_set_output(struct ggml_tensor * tensor);
-    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API void ggml_set_param(struct ggml_tensor * tensor);
     GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
 
     //
@@ -934,7 +934,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_repeat_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b); // sum up values that are adjacent in dims > 0 instead of repeated with same stride
 
     // concat a and b along dim
     // used in stable-diffusion
@@ -2045,15 +2045,14 @@ extern "C" {
 
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(
-        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
-        struct ggml_context * ctx_compute, // context for gradient computation
-        struct ggml_cgraph  * cgraph,
-        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
+        struct ggml_context *  ctx,        // context for gradient computation
+        struct ggml_cgraph  *  cgraph,
+        struct ggml_tensor  ** grad_accs);
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
     GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads);
-    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads);
     GGML_API void                 ggml_graph_cpy       (struct ggml_cgraph * src, struct ggml_cgraph * dst);
     GGML_API void                 ggml_graph_reset     (struct ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1
     GGML_API void                 ggml_graph_clear     (struct ggml_cgraph * cgraph);

ggml/src/ggml-alloc.c

@@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
     size_t node_size = 0;
     if (!node->data && !node->view_src) {
-        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        // If we previously had data but don't now then reallocate
+        if (talloc->buffer_id < 0) {
+            return false;
+        }
         node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
     }
     return talloc->size_max >= node_size;

ggml/src/ggml-backend.cpp

@@ -1109,7 +1109,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             const int node_backend_id = tensor_backend_id(node);
 
-            assert(node_backend_id != -1); // all nodes should be assigned by now
+            assert(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -352,10 +352,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
             message(STATUS "z15 target")
-            list(APPEND ARCH_FLAGS -march=z15 -mtune=z15)
+            list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
             message(STATUS "z16 target")
-            list(APPEND ARCH_FLAGS -march=z16 -mtune=z16)
+            list(APPEND ARCH_FLAGS -march=z16)
+        elseif (${S390X_M} MATCHES "9175|9176")
+            # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+            message(STATUS "z17 target")
+            list(APPEND ARCH_FLAGS -march=z17)
         else()
             message(STATUS "Unknown target")
             message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")

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

@@ -1054,6 +1054,493 @@ class tinyBLAS_Q0_AVX {
       } \
    } \
 
+template <typename TA, typename TB, typename TC>
+class tinyBLAS_BF16_PPC {
+  public:
+    tinyBLAS_BF16_PPC(int64_t k,
+                const TA *A, int64_t lda,
+                const TB *B, int64_t ldb,
+                TC *C, int64_t ldc,
+                int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    void vector_permute_store(vec_t *c, int numVec, unsigned char *vecOffset) {
+        vec_t t[8], s[8];
+        vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+
+        if (numVec == 2) {
+            t[0] = vec_perm(c[0], c[1], swiz1);
+            t[1] = vec_perm(c[2], c[3], swiz1);
+            s[0] = vec_perm(t[0], t[1], swiz3);
+            s[1] = vec_perm(t[0], t[1], swiz4);
+            vec_xst(s[0], 0, (vec_t*)vecOffset);
+            vec_xst(s[1], 0, (vec_t*)(vecOffset + 16));
+        } else if (numVec == 4) {
+            t[0] = vec_perm(c[0], c[1], swiz1);
+            t[1] = vec_perm(c[0], c[1], swiz2);
+            t[2] = vec_perm(c[2], c[3], swiz1);
+            t[3] = vec_perm(c[2], c[3], swiz2);
+            s[0] = vec_perm(t[0], t[2], swiz3);
+            s[1] = vec_perm(t[0], t[2], swiz4);
+            s[2] = vec_perm(t[1], t[3], swiz3);
+            s[3] = vec_perm(t[1], t[3], swiz4);
+            for (int i = 0; i < 4; ++i)
+                vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16));
+        } else if (numVec == 8) {
+            for (int i = 0; i < 4; i += 2) {
+                t[i+0] = vec_perm(c[i+0], c[i+1], swiz1);
+                t[i+1] = vec_perm(c[i+0], c[i+1], swiz2);
+            }
+            for (int i = 4; i < 8; i += 2) {
+                t[i+0] = vec_perm(c[i+0], c[i+1], swiz1);
+                t[i+1] = vec_perm(c[i+0], c[i+1], swiz2);
+            }
+            s[0] = vec_perm(t[0], t[2], swiz3);
+            s[1] = vec_perm(t[0], t[2], swiz4);
+            s[2] = vec_perm(t[1], t[3], swiz3);
+            s[3] = vec_perm(t[1], t[3], swiz4);
+            s[4] = vec_perm(t[4], t[6], swiz3);
+            s[5] = vec_perm(t[4], t[6], swiz4);
+            s[6] = vec_perm(t[5], t[7], swiz3);
+            s[7] = vec_perm(t[5], t[7], swiz4);
+            for (int i = 0; i < 8; ++i)
+                vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16));
+        }
+    }
+
+    void packNormal(const TA* a, int64_t lda, int rows, int cols, unsigned char* vec) {
+        int64_t i, j;
+        TA *aoffset = NULL;
+        unsigned char *vecOffset = NULL;
+        TA * aoffsets[8];
+        vector unsigned char c_arr[8];
+        aoffset = const_cast<TA*>(a);
+        vecOffset = vec;
+        j = (rows >> 3);
+        if (j > 0) {
+            do {
+                if (cols == 4) {
+                    aoffsets[0] = aoffset;
+                    for (int it = 1; it < 4; ++it)
+                        aoffsets[it] = aoffsets[it-1] + lda;
+                    aoffset += 4 * lda;
+                    for (int i = 0; i < 4; ++i)
+                        c_arr[i] = vec_xl(0, (vector unsigned char*)aoffsets[i]);
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int i = 0; i<4; i++)
+                        aoffsets[i] = aoffsets[i]+lda;
+                    vecOffset +=64;
+                }
+                i = (cols >> 3);
+                if (i > 0) {
+                    aoffsets[0] = aoffset;
+                    for (int it = 1; it < 8; ++it) {
+                        aoffsets[it] = aoffsets[it-1] + lda;
+                    }
+                    aoffset += 8 * lda;
+                    do {
+                        for (int it = 0; it < 8; ++it)
+                            c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                        vector_permute_store(c_arr, 8, vecOffset);
+                        for (int it = 0; it < 8; ++it)
+                            aoffsets[it] = aoffsets[it] + 8*lda;
+                        vecOffset += 128;
+                        i--;
+                    } while(i > 0);
+                }
+                j--;
+            } while(j > 0);
+        }
+        if (rows & 4) {
+            aoffsets[0] = aoffset;
+            for (int it = 1; it < 4; ++it)
+                aoffsets[it] = aoffsets[it-1] + lda;
+            aoffset += 4 * lda;
+            if (cols == 4) {
+                for (int it = 0; it < 4; ++it)
+                    c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                vector_permute_store(c_arr, 2, vecOffset);
+                for (int it = 0; it< 4; it++)
+                    aoffsets[it] = aoffsets[it] + lda;
+                vecOffset += 32;
+            }
+            i = (cols >> 3);
+            if (i > 0) {
+                do {
+                    for (int it = 0; it < 4; ++it)
+                        c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int it = 0; it< 4; it++)
+                        aoffsets[it] = aoffsets[it] + 8*lda;
+                    vecOffset += 64;
+                    i--;
+                } while(i > 0);
+            }
+        }
+        if (rows & 3) {
+            aoffsets[0] = aoffset;
+            for (int it = 1; it < 4; ++it)
+                aoffsets[it] = aoffsets[it-1] + lda;
+            if (cols == 4) {
+                switch(rows) {
+                    case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]);
+                    case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]);
+                    case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]);
+                        break;
+                }
+                vector_permute_store(c_arr, 2, vecOffset);
+                for (int it = 0; it< 4; it++)
+                     aoffsets[it] = aoffsets[it] + lda;
+                vecOffset += 32;
+            }
+            i = (cols >> 3);
+            if (i > 0) {
+                do {
+                    switch(rows) {
+                        case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]);
+                        case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]);
+                        case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]);
+                            break;
+                    }
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int it = 0; it <4; it++)
+                         aoffsets[it] = aoffsets[it] + 8* lda;
+                    vecOffset += 64;
+                    i--;
+                } while(i > 0);
+            }
+        }
+    }
+
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        int m_rem = MIN(m - m0, 8);
+        int n_rem = MIN(n - n0, 8);
+
+        if (m_rem >= 8 && n_rem >= 8) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if (m_rem >= 4 && n_rem >= 8) {
+            mc = 4;
+            nc = 8;
+            gemm<4,8>(m0, m, n0, n);
+        } else if (m_rem >=8 && n_rem >=4){
+                mc = 8;
+                nc = 4;
+                gemm<8,4>(m0, m, n0, n);
+        } else if ((m_rem < 4) && (n_rem >= 8)) {
+            nc = 8;
+            switch(m_rem) {
+                case 1:
+                    mc = 1;
+                    gemm_Mx8<1>(m0, m, n0, n);
+                    break;
+                case 2:
+                    mc = 2;
+                    gemm_Mx8<2>(m0, m, n0, n);
+                    break;
+                case 3:
+                    mc = 3;
+                    gemm_Mx8<3>(m0, m, n0, n);
+                    break;
+                default:
+                    return;
+            }
+        } else if (m_rem >= 4 && n_rem >= 4) {
+            mc = 4;
+            nc = 4;
+            gemm_small<4, 4>(m0, m, n0, n);
+        } else if ((m_rem > 4) && (n_rem < 4)) {
+            mc = 4;
+            switch(n_rem) {
+                case 1:
+                    nc = 1;
+                    gemm_small<4, 1>(m0, m, n0, n);
+                    break;
+                case 2:
+                    nc = 2;
+                    gemm_small<4, 2>(m0, m, n0, n);
+                    break;
+                case 3:
+                    nc = 3;
+                    gemm_small<4, 3>(m0, m, n0, n);
+                    break;
+
+                default:
+                    return;
+            }
+        } else {
+            switch((m_rem << 4) | n_rem) {
+                case 0x43:
+                    mc = 4;
+                    nc = 3;
+                    gemm_small<4, 3>(m0, m, n0, n);
+                    break;
+                case 0x42:
+                    mc = 4;
+                    nc = 2;
+                    gemm_small<4, 2>(m0, m, n0, n);
+                    break;
+                case 0x41:
+                    mc = 4;
+                    nc = 1;
+                    gemm_small<4, 1>(m0, m, n0, n);
+                    break;
+                case 0x34:
+                    mc = 3;
+                    nc = 4;
+                    gemm_small<3, 4>(m0, m, n0, n);
+                    break;
+                case 0x33:
+                    mc = 3;
+                    nc = 3;
+                    gemm_small<3, 3>(m0, m, n0, n);
+                    break;
+                case 0x32:
+                    mc = 3;
+                    nc = 2;
+                    gemm_small<3, 2>(m0, m, n0, n);
+                    break;
+                case 0x31:
+                    mc = 3;
+                    nc = 1;
+                    gemm_small<3, 1>(m0, m, n0, n);
+                    break;
+                case 0x24:
+                    mc = 2;
+                    nc = 4;
+                    gemm_small<2,4>(m0, m, n0, n);
+                    break;
+                case 0x23:
+                    mc = 2;
+                    nc = 3;
+                    gemm_small<2, 3>(m0, m, n0, n);
+                    break;
+                case 0x22:
+                    mc = 2;
+                    nc = 2;
+                    gemm_small<2, 2>(m0, m, n0, n);
+                    break;
+                case 0x21:
+                    mc = 2;
+                    nc = 1;
+                    gemm_small<2, 1>(m0, m, n0, n);
+                    break;
+                case 0x14:
+                    mc = 1;
+                    nc = 4;
+                    gemm_small<1, 4>(m0, m, n0, n);
+                    break;
+                case 0x13:
+                    mc = 1;
+                    nc = 3;
+                    gemm_small<1, 3>(m0, m, n0, n);
+                    break;
+                case 0x12:
+                    mc = 1;
+                    nc = 2;
+                    gemm_small<1, 2>(m0, m, n0, n);
+                    break;
+                case 0x11:
+                    mc = 1;
+                    nc = 1;
+                    gemm_small<1, 1>(m0, m, n0, n);
+                    break;
+                default:
+                    return;
+            }
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    void KERNEL_4x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[4], vec_B[8] , vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int l = 0; l < k; l+=8) {
+            packNormal((A+(ii*lda)+l), lda, 4, 8, (uint8_t*)vec_A);
+            packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+            }
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+    }
+
+    void KERNEL_8x4(int64_t ii, int64_t jj) {
+        vec_t vec_A[8], vec_B[4] , vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int l = 0; l < k; l+=8) {
+            packNormal((A+(ii*lda)+l), lda, 8, 8, (uint8_t*)vec_A);
+            packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x+4], vec_B[x]);
+            }
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii+4, jj);
+    }
+
+
+    void KERNEL_8x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[8], vec_B[8], vec_C[4];
+        acc_t acc_0, acc_1, acc_2, acc_3;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        __builtin_mma_xxsetaccz(&acc_2);
+        __builtin_mma_xxsetaccz(&acc_3);
+        for (int l = 0; l < k; l+=8) {
+            packNormal(A+(ii*lda)+l, lda, 8, 8, (uint8_t*)vec_A);
+            packNormal(B+(jj*ldb)+l, ldb, 8, 8, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, (vec_t)vec_A[x], (vec_t)vec_B[x+4]);
+                __builtin_mma_xvbf16ger2pp(&acc_2, (vec_t)vec_A[x+4], (vec_t)vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_3, (vec_t)vec_A[x+4], (vec_t)vec_B[x+4]);
+            }
+        }
+
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+        SAVE_ACC(&acc_2, ii+4, jj);
+        SAVE_ACC(&acc_3, ii+4, jj+4);
+    }
+
+    template<int RM, int RN>
+    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            vec_t vec_C[4];
+            acc_t acc_0;
+            __builtin_mma_xxsetaccz(&acc_0);
+            vec_t vec_A[2], vec_B[2];
+            for (int l=0; l<k; l+=4) {
+                packNormal(A+(ii*lda)+l, lda, RM, 4, (uint8_t*)vec_A);
+                packNormal(B+(jj*ldb)+l, ldb, RN, 4, (uint8_t*)vec_B);
+                for (int x = 0; x<2; x++) {
+                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_0);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < RN; J++) {
+                    *((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+        }
+    }
+
+    template<int RM>
+    void gemm_Mx8(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int RN = 8;
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            vec_t vec_C[4];
+            acc_t acc_0, acc_1;
+            __builtin_mma_xxsetaccz(&acc_0);
+            __builtin_mma_xxsetaccz(&acc_1);
+            vec_t vec_A[4], vec_B[8];
+            for (int l=0; l<k; l+=8) {
+                packNormal(A+(ii*lda)+l, lda, RM, 8, (uint8_t*)vec_A);
+                packNormal(B+(jj*ldb)+l, ldb, RN, 8, (uint8_t*)vec_B);
+                for (int x = 0; x<4; x++) {
+                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                    __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_0);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_1);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((TC*)(C+ii+((jj+4+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+        }
+    }
+
+    template<int RM, int RN>
+    inline void kernel(int64_t ii, int64_t jj) {
+       if constexpr(RM == 4 && RN == 8) {
+          KERNEL_4x8(ii,jj);
+       } else if constexpr(RM == 8 && RN == 8) {
+          KERNEL_8x8(ii,jj);
+       } else if constexpr(RM == 8 && RN == 4) {
+          KERNEL_8x4(ii,jj);
+       } else {
+          static_assert(false, "RN/RM values not supported");
+       }
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            kernel<RM, RN>(ii, jj);
+        }
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+
 template <typename TA, typename TB, typename TC>
 class tinyBLAS_Q0_PPC {
   public:
@@ -2202,6 +2689,7 @@ class tinyBLAS_PPC {
         boffset = vec;
         j = (rows >> 3);
         if (j > 0) {
+
             do {
                 aoffset1 = aoffset;
                 aoffset2 = aoffset1 + lda;
@@ -2875,9 +3363,22 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                 (float *)C, ldc};
             return tb.matmul(m, n);
         }
+#elif defined(__MMA__)
+        if ((k % 8))
+                return false;
+        if(Btype == GGML_TYPE_BF16) {
+           tinyBLAS_BF16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
+            (const ggml_bf16_t *)A, lda,
+            (const ggml_bf16_t *)B, ldb,
+            (float *)C, ldc,
+            params->ith, params->nth};
+        tb.matmul(m, n);
+        return true;
+        }
 #endif
         return false;
     }
+
     case GGML_TYPE_F16: {
 #if defined(__AVX512F__)
         if (Btype == GGML_TYPE_F16) {

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

@@ -341,7 +341,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F32_EPR  4
 
 #define GGML_F32x4              vector float
-#define GGML_F32x4_ZERO         0.0f
+#define GGML_F32x4_ZERO         {0.0f}
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)
 #define GGML_F32x4_STORE(p, r)  vec_xst(r, 0, p)

ggml/src/ggml-cuda/CMakeLists.txt

@@ -133,6 +133,7 @@ if (CUDAToolkit_FOUND)
                 COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
                 OUTPUT_VARIABLE CUDA_CCVER
                 ERROR_QUIET
+                OUTPUT_STRIP_TRAILING_WHITESPACE
             )
         else()
             if (CUDA_CCFULLVER MATCHES Apple)
@@ -143,7 +144,7 @@ if (CUDAToolkit_FOUND)
             string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
         endif()
 
-        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+        message(STATUS "CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER})
         list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later

ggml/src/ggml-cuda/convert.cu

@@ -1,6 +1,8 @@
 #include "convert.cuh"
 #include "dequantize.cuh"
 
+#include <cstdint>
+
 #define CUDA_Q8_0_NE_ALIGN 2048
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -570,30 +572,46 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t
 }
 
 template <typename src_t, typename dst_t>
-static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
-    const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+static __global__ void convert_unary(
+        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t s01, const int64_t s02, const int64_t s03) {
+    const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
-    if (i >= k) {
+    if (i00 >= ne00) {
         return;
     }
 
+    const int64_t i01 = blockIdx.y;
+    const int64_t i02 = blockIdx.z % ne02;
+    const int64_t i03 = blockIdx.z / ne02;
+
     const src_t * x = (const src_t *) vx;
 
-    y[i] = float(x[i]);
+    const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
+    const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
+    y[iy] = float(x[ix]);
 }
 
 template <typename src_t, typename dst_t>
-static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+static void convert_unary_cuda(const void * vx, dst_t * y,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
+    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
+    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
+        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary_cont_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    convert_unary_cuda<src_t>(vx, y, k, 1, 1, 1, k, k, k, stream);
 }
 
 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
+            return convert_unary_cont_cuda<float>;
         case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
+            return convert_unary_cont_cuda<half>;
         default:
             return nullptr;
     }
@@ -643,9 +661,9 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
+            return convert_unary_cont_cuda<float>;
         case GGML_TYPE_BF16:
-            return convert_unary_cuda<nv_bfloat16>;
+            return convert_unary_cont_cuda<nv_bfloat16>;
         default:
             return nullptr;
     }
@@ -692,7 +710,18 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
+            return convert_unary_cont_cuda<half>;
+        case GGML_TYPE_BF16:
+            return convert_unary_cont_cuda<nv_bfloat16>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float>;
         case GGML_TYPE_BF16:
             return convert_unary_cuda<nv_bfloat16>;
         default:

ggml/src/ggml-cuda/convert.cuh

@@ -3,7 +3,7 @@
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
 
 template<typename T>
-using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
+using to_t_cuda_t = void (*)(const void * x, T * y, int64_t k, cudaStream_t stream);
 
 typedef to_t_cuda_t<float> to_fp32_cuda_t;
 typedef to_t_cuda_t<half> to_fp16_cuda_t;
@@ -14,3 +14,13 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type);
 
 to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);
+
+// TODO more general support for non-contiguous inputs
+
+template<typename T>
+using to_t_nc_cuda_t = void (*)(const void * x, T * y,
+    int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
+    int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
+
+typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
+to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);

ggml/src/ggml-cuda/cpy.cu

@@ -592,6 +592,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
         graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
     }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
 #endif
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));

ggml/src/ggml-cuda/getrows.cu

@@ -33,8 +33,8 @@ static __global__ void k_get_rows(
     dfloat2 v;
     dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst_row[iybs + iqs + 0]        = v.x;
-    dst_row[iybs + iqs + y_offset] = v.y;
+    dst_row[iybs + iqs + 0]        = float(v.x);
+    dst_row[iybs + iqs + y_offset] = float(v.y);
 }
 
 template<typename src0_t, typename dst_t>
@@ -60,7 +60,7 @@ static __global__ void k_get_rows_float(
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
     const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
 
-    dst_row[i00] = src0_row[i00];
+    dst_row[i00] = float(src0_row[i00]);
 }
 
 template<typename grad_t, typename dst_t>
@@ -86,120 +86,159 @@ static __global__ void k_get_rows_back_float(
     dst[dst_row*ncols + col] = sum;
 }
 
-template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
+template<int qk, int qr, dequantize_kernel_t dq, typename dst_t>
+static void get_rows_cuda_q(
+        const void * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
 
     // 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 s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
 
-    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);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
 
     GGML_ASSERT(ne00 % 2 == 0);
 
     k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10, ne11,*/ ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
-
-    GGML_UNUSED(dst);
 }
 
-template<typename src0_t>
+template<typename src0_t, typename dst_t>
 static void get_rows_cuda_float(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    GGML_ASSERT(ne13 == 1);
-
+        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
 
     // 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 s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
 
-    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);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
 
     k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10, ne11,*/ ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
+}
 
-    GGML_UNUSED(dst);
+template <typename dst_t>
+static void ggml_cuda_get_rows_switch_src0_type(
+        const void * src0_d, const ggml_type src0_type, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    switch (src0_type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float((const half *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda_q<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda_q<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda_q<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda_q<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type));
+            break;
+    }
+}
+
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream) {
+    switch (dst_type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (nv_bfloat16 *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            GGML_ABORT("%s: unsupported dst type: %s\n", __func__, ggml_type_name(dst_type));
+            break;
+    }
 }
 
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    const void    * src0_d = (const void    *) src0->data;
-    const int32_t * src1_d = (const int32_t *) src1->data;
-    float         * dst_d  = (float         *) dst->data;
-
     cudaStream_t stream = ctx.stream();
 
+    GGML_TENSOR_BINARY_OP_LOCALS
+
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ne13 == 1);
 
     GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
     GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
     GGML_ASSERT(dst->nb[0]  == ggml_type_size(dst->type));
 
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            get_rows_cuda_float(src0, src1, dst, (const half *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_F32:
-            get_rows_cuda_float(src0, src1, dst, (const float *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        default:
-            // TODO: k-quants
-            GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            break;
-    }
+    get_rows_cuda(src0->data, src0->type, (const int32_t *) src1->data, dst->data, dst->type,
+        ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
 }
 
 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/getrows.cuh

@@ -3,6 +3,13 @@
 #define CUDA_GET_ROWS_BLOCK_SIZE 256
 #define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256
 
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream);
+
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -1551,7 +1551,7 @@ static void ggml_cuda_op_mul_mat(
 
             if (src1_on_device && src1_is_contiguous) {
                 quantize_src1(
-                    dev[id].src1_ddf, dev[id].src1_ddq, src0->type, ne10,
+                    dev[id].src1_ddf, nullptr, dev[id].src1_ddq, src0->type, ne10,
                     nb11/sizeof(float), nb12/sizeof(float), nb13/sizeof(float),
                     src1_padded_col_size, ne11, ne12, ne13, stream);
                 CUDA_CHECK(cudaGetLastError());
@@ -1649,7 +1649,7 @@ static void ggml_cuda_op_mul_mat(
 
                 if (quantize_src1 && !src1_is_contiguous) {
                     quantize_src1(
-                        src1_ddf_i, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
+                        src1_ddf_i, nullptr, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
                         src1_padded_col_size, src1_ncols, 1, 1, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
@@ -1720,15 +1720,15 @@ static __global__ void k_compute_batched_ptrs(
         size_t  nb12, size_t  nb13,
         size_t  nbd2, size_t  nbd3,
         int64_t r2,   int64_t r3) {
-    int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
-    int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
+    const int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
+    const int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
 
     if (i13 >= ne13 || i12 >= ne12) {
         return;
     }
 
-    int64_t i03 = i13 / r3;
-    int64_t i02 = i12 / r2;
+    const int64_t i03 = i13 / r3;
+    const int64_t i02 = i12 / r2;
 
     ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
     ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
@@ -1742,6 +1742,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
+    // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
+    // As long as dst is contiguous this does not matter though.
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t ne_dst = ggml_nelements(dst);
@@ -1750,21 +1754,31 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 
     CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
 
-    void * src0_ddq = src0->data;
-    half * src0_f16 = (half *) src0_ddq;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
+    const half * src0_f16 = (const half *) src0->data;
+    float * dst_ddf = (float *) dst->data;
+
+    const half * src1_f16 = (const half *) src1->data;
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    GGML_ASSERT(nb10 == ts_src1);
+    int64_t s11 = nb11 / ts_src1;
+    int64_t s12 = nb12 / ts_src1;
+    int64_t s13 = nb13 / ts_src1;
+    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
 
     // convert src1 to fp16
-    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
     if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+        const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
         const int64_t ne_src1 = ggml_nelements(src1);
         src1_f16_alloc.alloc(ne_src1);
         GGML_ASSERT(to_fp16_cuda != nullptr);
-        to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
+
+        to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
+
+        src1_f16 = src1_f16_alloc.get();
+        s11 = ne10;
+        s12 = ne11*s11;
+        s13 = ne12*s12;
     }
-    half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
 
     ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
     char * dst_t;
@@ -1824,13 +1838,13 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                 int i02 = i12 / r2;
 
                 CUBLAS_CHECK(
-                        cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F,   nb01/sizeof(half),
-                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F,   nb11/sizeof(float),
-                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
-                            cu_compute_type,
-                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+                cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+                    ne01, ne11, ne10,
+                    alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F,   nb01/sizeof(half),
+                                          src1_f16 + i13*s13  + i12*s12,  CUDA_R_16F,   s11,
+                    beta,  (      char *)    dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
+                    cu_compute_type,
+                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
             }
         }
     }
@@ -1841,15 +1855,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, (const char *) src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00,  // strideA
-                       (const char *) src1_f16, CUDA_R_16F,   nb11/nb10, nb12/nb10,  // strideB
-                beta,  (      char *)    dst_t, cu_data_type, ne01,       nb2/nb0,   // strideC
+                alpha, src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00, // strideA
+                       src1_f16, CUDA_R_16F,   s11,       s12,       // strideB
+                beta,     dst_t, cu_data_type, ne0,       ne1*ne0,   // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     } else {
         // use cublasGemmBatchedEx
-        const int ne23 = ne12*ne13;
+        const int64_t ne23 = ne12*ne13;
 
         ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
         ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
@@ -1861,8 +1875,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                 ne12, ne13,
                 ne23,
                 nb02, nb03,
-                src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
-                src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
+                src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
+                src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
                 nbd2, nbd3,
                 r2, r3);
         CUDA_CHECK(cudaGetLastError());
@@ -1871,8 +1885,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
                 alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/nb10,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
+                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   s11,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
                 ne23,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1935,8 +1949,10 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
     } else if (!split && use_mul_mat_vec_q) {
         ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
+    } else if (!split && use_mul_mat_q) {
+        ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
-            dst->op_params[0] == GGML_PREC_DEFAULT && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+            !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_mul_mat_vec) {
@@ -1950,183 +1966,145 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     }
 }
 
-struct mmid_row_mapping {
-    int32_t i1;
-    int32_t i2;
-};
-
-static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
-                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
-                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
-                                                 int64_t ne11, int64_t ne10,
-                                                 size_t nb11, size_t nb12) {
-    int32_t iid1 = blockIdx.x;
-    int32_t id = blockIdx.y;
-
-    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
-
-    if (row_id_i != i02) {
-        return;
-    }
-
-    const int64_t i11 = id % ne11;
-    const int64_t i12 = iid1;
-
-    __shared__ int src1_row;
-    if (threadIdx.x == 0) {
-        src1_row = atomicAdd(cur_src1_row, 1);
-        row_mapping[src1_row] = {id, iid1};
-    }
-    __syncthreads();
-
-    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
-    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
-
-    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
-        src1_row_contiguous[i] = src1_row_original[i];
-    }
-}
-
-static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
-                                                  const mmid_row_mapping * __restrict__ row_mapping,
-                                                  int64_t ne0,
-                                                  size_t nb1, size_t nb2) {
-    int32_t i = blockIdx.x;
-
-    const int32_t i1 = row_mapping[i].i1;
-    const int32_t i2 = row_mapping[i].i2;
-
-    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
-    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
-
-    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
-        dst_row_original[j] = dst_row_contiguous[j];
-    }
-}
-
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * ids  = dst->src[2];
 
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && ne2 == 1) {
-        if (ggml_is_quantized(src0->type)) {
-            ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
-        } else {
-            ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst);
-        }
-        return;
-    }
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
 
-    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        if (ne2 == 1) {
+            if (ggml_is_quantized(src0->type)) {
+                ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+            } else {
+                ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst);
+            }
+            return;
+        }
+
+        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12)) {
+            ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst);
+            return;
+        }
+    }
 
     cudaStream_t stream = ctx.stream();
 
-    const int64_t n_as = ne02;
-    const int64_t n_ids = ids->ne[0];
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const ggml_type type_src1_sorted = (src0->type == GGML_TYPE_F16 && !fast_fp16_hardware_available(cc))
+        || ggml_is_quantized(src0->type) ? GGML_TYPE_F32 : src0->type;
+    const ggml_type type_dst_sorted  = GGML_TYPE_F32;
+    const size_t ts_src1_sorted = ggml_type_size(type_src1_sorted);
+    const size_t ts_dst_sorted  = ggml_type_size(type_dst_sorted);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<int32_t> ids_to_sorted_host;
+    ids_to_sorted_host.reserve(2*ne_get_rows);
+    std::vector<int32_t> ids_from_sorted_host(ne_get_rows);
+
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool(), 2*ne_get_rows);
+
+    std::vector<int32_t> tokens_per_expert(ne02);
+
+    ggml_cuda_pool_alloc<char> src1_sorted(ctx.pool(), ne12*n_expert_used*ne10*ts_src1_sorted);
+    ggml_cuda_pool_alloc<char>  dst_sorted(ctx.pool(), ne2 *n_expert_used* ne0*ts_dst_sorted);
 
     std::vector<char> ids_host(ggml_nbytes(ids));
-    const char * ids_dev = (const char *) ids->data;
-    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
     CUDA_CHECK(cudaStreamSynchronize(stream));
 
-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row  = *dst;
-
-    char * src0_original = (char *) src0->data;
-    char * src1_original = (char *) src1->data;
-    char * dst_original  = (char *)  dst->data;
-
-    src0_row.ne[2] = 1;
-    src0_row.ne[3] = 1;
-    src0_row.nb[3] = nb02;
-
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
-
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
-
-    ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
-    ggml_cuda_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
-
-    src1_row.data = src1_contiguous.get();
-    dst_row.data  =  dst_contiguous.get();
-
-    for (int64_t i02 = 0; i02 < n_as; i02++) {
-        int64_t num_src1_rows = 0;
-
-        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-            for (int64_t id = 0; id < n_ids; id++) {
-                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-
-                GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
-
-                if (row_id_i != i02) {
-                    continue;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_from_sorted_host[i12*n_expert_used + iex] = ids_to_sorted_host.size();
+                    ids_to_sorted_host.push_back(i12*ne11 + iex % ne11);
+                    tokens_per_expert[i02]++;
+                    break;
                 }
-
-                num_src1_rows++;
             }
         }
+    }
+    GGML_ASSERT(ids_to_sorted_host.size() == size_t(ne_get_rows));
 
-        if (num_src1_rows == 0) {
+    ids_to_sorted_host.insert(ids_to_sorted_host.end(), ids_from_sorted_host.begin(), ids_from_sorted_host.end());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_to_sorted_host.data(), 2*ne_get_rows*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_to_sorted   = ids_buf_dev.ptr + 0*ne_get_rows;
+    const int32_t * ids_from_sorted = ids_buf_dev.ptr + 1*ne_get_rows;
+
+    get_rows_cuda(src1->data, src1->type, ids_to_sorted, src1_sorted.ptr, type_src1_sorted,
+        ne10, nb11, nb12, nb13,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    char * src1_data_cur = (char *) src1_sorted.ptr;
+    char *  dst_data_cur = (char *)  dst_sorted.ptr;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) {
+        if (tokens_per_expert[i02] == 0) {
             continue;
         }
 
-        ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
-        ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
-        CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
+        ggml_tensor src0_slice = *src0;
+        src0_slice.ne[2] = 1;
+        src0_slice.nb[3] = src0_slice.nb[2];
+        src0_slice.data  = (char *) src0->data + i02*nb02;
 
-        {
-            dim3 block_dims(std::min((unsigned int)ne10, 768u));
-            dim3 grid_dims(ids->ne[1], n_ids);
-            k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                    src1_original, src1_contiguous.get(),
-                    dev_cur_src1_row.get(), dev_row_mapping.get(),
-                    ids_dev, i02, ids->nb[1], ids->nb[0],
-                    ne11, ne10,
-                    nb11, nb12);
-            CUDA_CHECK(cudaGetLastError());
-        }
+        ggml_tensor src1_slice;
+        memset(&src1_slice, 0, sizeof(src1_slice));
+        src1_slice.buffer = src1->buffer;
+        src1_slice.type   = type_src1_sorted;
+        src1_slice.ne[0]  = ne10;
+        src1_slice.ne[1]  = tokens_per_expert[i02];
+        src1_slice.ne[2]  = 1;
+        src1_slice.ne[3]  = 1;
+        src1_slice.nb[0]  = ts_src1_sorted;
+        src1_slice.nb[1]  = src1_slice.ne[0] * src1_slice.nb[0];
+        src1_slice.nb[2]  = src1_slice.ne[1] * src1_slice.nb[1];
+        src1_slice.nb[3]  = src1_slice.ne[2] * src1_slice.nb[2];
+        src1_slice.data   = src1_data_cur;
 
-        src0_row.data = src0_original + i02*nb02;
+        ggml_tensor dst_slice;
+        memset(&dst_slice, 0, sizeof(dst_slice));
+        dst_slice.buffer = dst->buffer;
+        dst_slice.type   = type_dst_sorted;
+        dst_slice.ne[0]  = ne0;
+        dst_slice.ne[1]  = tokens_per_expert[i02];
+        dst_slice.ne[2]  = 1;
+        dst_slice.ne[3]  = 1;
+        dst_slice.nb[0]  = ts_dst_sorted;
+        dst_slice.nb[1]  = dst_slice.ne[0] * dst_slice.nb[0];
+        dst_slice.nb[2]  = dst_slice.ne[1] * dst_slice.nb[1];
+        dst_slice.nb[3]  = dst_slice.ne[2] * dst_slice.nb[2];
+        dst_slice.data   = dst_data_cur;
 
-        GGML_ASSERT(nb11 == sizeof(float)*ne10);
-        GGML_ASSERT(nb1 == sizeof(float)*ne0);
+        ggml_cuda_mul_mat(ctx, &src0_slice, &src1_slice, &dst_slice);
+        CUDA_CHECK(cudaGetLastError());
 
-        src1_row.ne[1] = num_src1_rows;
-        src1_row.nb[1] = nb11;
-        src1_row.nb[2] = num_src1_rows*nb11;
-        src1_row.nb[3] = num_src1_rows*nb11;
-
-        dst_row.ne[1] = num_src1_rows;
-        dst_row.nb[1] = nb1;
-        dst_row.nb[2] = num_src1_rows*nb1;
-        dst_row.nb[3] = num_src1_rows*nb1;
-
-        ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
-
-        {
-            dim3 block_dims(std::min((unsigned int)ne0, 768u));
-            dim3 grid_dims(num_src1_rows);
-            k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                    dst_original, dst_contiguous.get(),
-                    dev_row_mapping.get(),
-                    ne0,
-                    nb1, nb2);
-            CUDA_CHECK(cudaGetLastError());
-        }
+        src1_data_cur += src1_slice.nb[2];
+        dst_data_cur  +=  dst_slice.nb[2];
     }
+
+    get_rows_cuda(dst_sorted.ptr, type_dst_sorted, ids_from_sorted, dst->data, dst->type,
+        ne0, ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        nb1, nb2, nb3, stream);
 }
 
 static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {

ggml/src/ggml-cuda/mmq.cu

@@ -1,37 +1,10 @@
 #include "mmq.cuh"
+#include "quantize.cuh"
 
-void ggml_cuda_op_mul_mat_q(
-    ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
+#include <vector>
 
-    const int64_t ne00 = src0->ne[0];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    GGML_ASSERT(ne10 % QK8_1 == 0);
-
-    const int64_t ne0 = dst->ne[0];
-
-    const int64_t row_diff = row_high - row_low;
-    const int64_t stride00 = ne00 / ggml_blck_size(src0->type);
-
-    int id = ggml_cuda_get_device();
-    const int cc = ggml_cuda_info().devices[id].cc;
-
-    // the main device has a larger memory buffer to hold the results from all GPUs
-    // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
-
-    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
-    // Also its fixup needs to allocate a temporary buffer in the memory pool.
-    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
-        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
-    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
-
-    switch (src0->type) {
+static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    switch (args.type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
             break;
@@ -90,10 +63,195 @@ void ggml_cuda_op_mul_mat_q(
             GGML_ABORT("fatal error");
             break;
     }
+}
+
+void ggml_cuda_mul_mat_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    cudaStream_t stream = ctx.stream();
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);
+
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(        nb0        == ts_dst);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+
+    const char  * src0_d = (const char  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s3  =  dst->nb[3] / ts_dst;
+
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
+
+    if (!ids) {
+        const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+            get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+        ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+        {
+            const int64_t s11 = src1->nb[1] / ts_src1;
+            const int64_t s12 = src1->nb[2] / ts_src1;
+            const int64_t s13 = src1->nb[3] / ts_src1;
+            quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type,
+                ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+        }
+
+        const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+        const int64_t s13 = ne12*s12;
+
+        const mmq_args args = {
+            src0_d, src0->type, (const int *) src1_q8_1.ptr, nullptr, nullptr, dst_d,
+            ne00, ne01, ne1, s01, s1,
+            ne02, ne12, s02, s12, s2,
+            ne03, ne13, s03, s13, s3,
+            use_stream_k};
+        ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+        return;
+    }
+
+    GGML_ASSERT(ne13 == 1);
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    std::vector<int32_t> ids_src1_host;
+    ids_src1_host.reserve(ne_get_rows);
+    std::vector<int32_t> ids_dst_host;
+    ids_dst_host.reserve(ne_get_rows);
+    std::vector<int32_t> tokens_per_expert_host(ne02);
+    std::vector<int32_t> expert_bounds_host(ne02 + 1);
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_src1_host.push_back(i12*(nb12/nb11) + iex % ne11);
+                    ids_dst_host.push_back(i12*ne1 + iex);
+                    tokens_per_expert_host[i02]++;
+                    break;
+                }
+            }
+        }
+    }
+
+    int32_t cumsum = 0;
+    for (int64_t i = 0; i < ne02; ++i) {
+        expert_bounds_host[i] = cumsum;
+        cumsum += tokens_per_expert_host[i];
+    }
+    expert_bounds_host[ne02] = cumsum;
+
+    std::vector<int32_t> ids_buf_host;
+    ids_buf_host.reserve(ids_src1_host.size() + ids_dst_host.size() + expert_bounds_host.size());
+    ids_buf_host.insert(ids_buf_host.end(), ids_src1_host.begin(), ids_src1_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), ids_dst_host.begin(), ids_dst_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), expert_bounds_host.begin(), expert_bounds_host.end());
+    ids_buf_dev.alloc(ids_buf_host.size() + get_mmq_x_max_host(cc)); // Expert bounds are padded on device.
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_buf_host.data(), ids_buf_host.size()*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_src1_dev      = ids_buf_dev.ptr;
+    const int32_t * ids_dst_dev       = ids_src1_dev + ids_src1_host.size();
+    const int32_t * expert_bounds_dev = ids_dst_dev + ids_dst_host.size();
+
+    const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+        get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+    ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+    const int64_t ne11_flat = ne12*n_expert_used;
+    const int64_t ne12_flat = 1;
+    const int64_t ne13_flat = 1;
+
+    {
+        const int64_t s11 = src1->nb[1] / ts_src1;
+        const int64_t s12 = src1->nb[2] / ts_src1;
+        const int64_t s13 = src1->nb[2] / ts_src1;
+        quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
+            ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
+    }
+
+    const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+    const int64_t s13 = ne12*s12;
+
+    // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
+    const mmq_args args = {
+        src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d,
+        ne00, ne01, ne_get_rows, s01, s1,
+        ne02, ne02, s02, s12, s2,
+        ne03, ne13, s03, s13, s3,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+}
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+    const int64_t stride01 = ne00 / ggml_blck_size(src0->type);
+
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
+        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
+    const mmq_args args = {
+        src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
+        ne00, row_diff, src1_ncols, stride01, nrows_dst,
+        1, 1, 0, 0, 0,
+        1, 1, 0, 0, 0,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 
     GGML_UNUSED(src1);
     GGML_UNUSED(dst);
     GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(src1_padded_row_size);
 }
 
 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {

ggml/src/ggml-cuda/mmq.cuh

@@ -13,9 +13,10 @@ using namespace ggml_cuda_mma;
 #define MMQ_ITER_K 256
 #define MMQ_NWARPS 8
 
-typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int & kbx0, const int & i_max, const int & stride);
-typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00);
-typedef void (*mmq_write_back_t)(const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max);
+typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride);
+typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00);
+typedef void (*mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted,
+    float * __restrict__ dst, const int stride, const int i_max, const int j_max);
 
 enum mmq_q8_1_ds_layout {
     MMQ_Q8_1_DS_LAYOUT_D4,
@@ -233,7 +234,7 @@ static constexpr __device__ int mmq_get_granularity_device(const int /* mmq_x */
 // ------------------------------------------------------------
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -289,7 +290,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -328,7 +329,7 @@ static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -384,7 +385,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -423,7 +424,7 @@ static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -495,7 +496,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -565,7 +566,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -621,7 +622,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -651,7 +652,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps, mmq_q8_1_ds_layout ds_layout>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     typedef tile<16, 8, int> tile_A;
     typedef tile< 8, 8, int> tile_B;
@@ -732,7 +733,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -762,7 +763,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     typedef tile<16, 8, int> tile_A;
     typedef tile< 8, 8, int> tile_B;
@@ -839,7 +840,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
     const int   * x_qs = (const int   *) x;
@@ -871,7 +872,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -955,7 +956,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1011,7 +1012,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1074,7 +1075,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -1201,7 +1202,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1298,7 +1299,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1340,7 +1341,7 @@ static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, co
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1437,7 +1438,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1469,7 +1470,7 @@ static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1578,7 +1579,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1610,7 +1611,7 @@ static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1693,7 +1694,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1726,7 +1727,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -1835,7 +1836,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_nl(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1893,7 +1894,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xxs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1951,7 +1952,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2007,7 +2008,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2070,7 +2071,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_xxs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2126,7 +2127,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2189,7 +2190,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq1_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2245,7 +2246,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_xs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2306,8 +2307,8 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template<int mmq_x, int mmq_y, int nwarps, bool need_check>
 static __device__ __forceinline__ void mmq_write_back_dp4a(
-    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
-
+        const float * __restrict__ sum, const int32_t * __restrict__ ids_dst, float * __restrict__ dst,
+        const int stride, const int i_max, const int j_max) {
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
         const int j = j0 + threadIdx.y;
@@ -2324,15 +2325,15 @@ static __device__ __forceinline__ void mmq_write_back_dp4a(
                 continue;
             }
 
-            dst[j*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            dst[ids_dst[j]*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
         }
     }
 }
 
 template<int mmq_x, int mmq_y, int nwarps, bool need_check>
 static __device__ __forceinline__ void mmq_write_back_mma(
-    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
-
+        const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst,
+        const int stride, const int i_max, const int j_max) {
     typedef tile<16, 8, int> tile_C;
 
     constexpr int granularity = mmq_get_granularity_device(mmq_x);
@@ -2362,7 +2363,7 @@ static __device__ __forceinline__ void mmq_write_back_mma(
                     continue;
                 }
 
-                dst[j*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l];
+                dst[ids_dst[j]*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l];
             }
         }
     }
@@ -2518,17 +2519,18 @@ struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_XS> {
 };
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check, bool fixup>
-static __device__ void mul_mat_q_process_tile(
-    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-    const int & ne00, const int & ne01, const int & stride01, const int & ne10, const int & ne11, const int & stride11, const int & ne0,
-    const int & it, const int & jt, const int & kb0_start, const int & kb0_stop) {
+static __device__ __forceinline__ void mul_mat_q_process_tile(
+        const char * __restrict__ x, const int offset_x, const int * __restrict__ y,
+        const int * __restrict__ ids_dst, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+        const int nrows_x, const int ncols_y, const int stride_row_x, const int stride_col_dst,
+        const int tile_x_max_i, const int tile_y_max_j, const int kb0_start, const int kb0_stop) {
 
     constexpr int              qk         = ggml_cuda_type_traits<type>::qk;
     constexpr int              mmq_y      = get_mmq_y_device();
     constexpr load_tiles_mmq_t load_tiles = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::load_tiles;
 
-    extern __shared__ char data_mul_mat_q[];
-    int * tile_y = (int *) data_mul_mat_q;
+    extern __shared__ int data_mul_mat_q[];
+    int * tile_y = data_mul_mat_q + mmq_x;
     int * tile_x = tile_y + GGML_PAD(mmq_x*(WARP_SIZE + WARP_SIZE/QI8_1), nwarps*WARP_SIZE);
 
 #ifdef NEW_MMA_AVAILABLE
@@ -2543,16 +2545,11 @@ static __device__ void mul_mat_q_process_tile(
 
     float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
 
-    const int tile_x_max_i = ne01 - it*mmq_y - 1;
-    const int tile_y_max_j = ne11 - jt*mmq_x - 1;
-
-    const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int));
-
     for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) {
-        load_tiles(x, tile_x, stride01*it*mmq_y + kb0, tile_x_max_i, stride01);
+        load_tiles(x, tile_x, offset_x + kb0, tile_x_max_i, stride_row_x);
 
         {
-            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
+            const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
 #pragma unroll
             for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
                 int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
@@ -2568,7 +2565,7 @@ static __device__ void mul_mat_q_process_tile(
         __syncthreads();
 
         {
-            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
+            const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
 #pragma unroll
             for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
                 int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
@@ -2585,12 +2582,10 @@ static __device__ void mul_mat_q_process_tile(
     }
 
     if (fixup) {
-        write_back(sum, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
+        write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
     } else {
-        write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j);
+        write_back(sum, ids_dst, dst, stride_col_dst, tile_x_max_i, tile_y_max_j);
     }
-
-    GGML_UNUSED(ne00); GGML_UNUSED(ne10);
 }
 
 
@@ -2609,8 +2604,11 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 static __global__ void mul_mat_q(
-    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-    const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
+        const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
+        const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+        const int ncols_x, const int nrows_x, const int ncols_y, const int stride_row_x, const int stride_col_dst,
+        const int channel_ratio, const int nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 
     // Skip unused template specializations for faster compilation:
     if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
@@ -2621,26 +2619,85 @@ static __global__ void mul_mat_q(
     constexpr int qk    = ggml_cuda_type_traits<type>::qk;
     constexpr int mmq_y = get_mmq_y_device();
 
+    const int ntx = (ncols_y + mmq_x - 1) / mmq_x; // Number of tiles x
+    const int nty = (nrows_x + mmq_y - 1) / mmq_y; // Number of tiles y
+
+    // Initialize the ids for writing back data with just the index.
+    // For regular matrix multiplications this is never changed.
+    // For MoE the correct indices are loaded from ids_dst.
+    extern __shared__ int ids_dst_shared[]; // Stored at beginning of shared memory.
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+        const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+        if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+            break;
+        }
+
+        ids_dst_shared[j] = j;
+    }
+
     // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
 #if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
     {
+        const int wt = blockIdx.z / nchannels_y;
+        const int zt = blockIdx.z - wt*nchannels_y;
+        const int jt = blockIdx.y;
+        const int it = blockIdx.x;
+
+        // Defaults for regular matrix multiplication:
+        int col_low    = 0;
+        int col_high   = ncols_y;
+        int col_diff   = ncols_y;
+        int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+        int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+        if (ids_dst) {
+            col_low  = expert_bounds[zt + 0];
+            col_high = expert_bounds[zt + 1];
+            col_diff = col_high - col_low;
+
+            offset_y   = 0;
+            offset_dst = 0;
+
+            if (jt*mmq_x >= col_diff) {
+                return;
+            }
+
+#pragma unroll
+            for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+                const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                    break;
+                }
+
+                ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
+            }
+        }
+
+        offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+        offset_dst += it*mmq_y;
+
+        const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+        const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+        const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
+
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-                blockIdx.x, blockIdx.y, 0, ne00/qk);
+            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+             tile_x_max_i, tile_y_max_j, 0, ncols_x/qk);
         return;
     }
 #endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
 
-    const     int64_t blocks_per_ne00 = ne00 / qk;
+    const     int64_t blocks_per_ne00 = ncols_x / qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
 
-    const int ntx = (ne11 + mmq_x - 1) / mmq_x; // Number of tiles x
-    const int nty = (ne01 + mmq_y - 1) / mmq_y; // Number of tiles y
-
     // kbc == k block continuous, current index in continuous ijk space.
-    int64_t kbc      = (int64_t) blockIdx.x     *blocks_per_ne00*ntx*nty / gridDim.x;
-    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x;
+    int64_t kbc      = (int64_t) blockIdx.x     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
 
     kbc      -= (kbc      % blocks_per_ne00) % blocks_per_iter;
     kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_iter;
@@ -2649,13 +2706,64 @@ static __global__ void mul_mat_q(
     int kb0_start = kbc % blocks_per_ne00;
     int kb0_stop  = min(blocks_per_ne00, kb0_start + kbc_stop - kbc);
     while (kbc < kbc_stop && kb0_stop == blocks_per_ne00) {
-        const int jt =  kbc /    (blocks_per_ne00*nty);                    // j index of current tile.
-        const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; // i index of current tile.
+        int tmp = kbc;
+        const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+        tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+        const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+        tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+        const int zt = tmp / (ntx*blocks_per_ne00);
+        tmp -= zt * (ntx*blocks_per_ne00);
+        const int jt = tmp / blocks_per_ne00;
+
+        // Defaults for regular matrix multiplication:
+        int col_low    = 0;
+        int col_high   = ncols_y;
+        int col_diff   = ncols_y;
+        int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+        int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+        if (ids_dst) {
+            col_low  = expert_bounds[zt + 0];
+            col_high = expert_bounds[zt + 1];
+            col_diff = col_high - col_low;
+
+            offset_y   = 0;
+            offset_dst = 0;
+
+            if (jt*mmq_x >= col_diff) {
+                kbc += blocks_per_ne00;
+                kbc -= kbc % blocks_per_ne00;
+
+                kb0_start = 0;
+                kb0_stop  = min(blocks_per_ne00, kbc_stop - kbc);
+
+                continue;
+            }
+
+#pragma unroll
+            for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+                const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                    break;
+                }
+
+                ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
+            }
+        }
+
+        offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+        offset_dst += it*mmq_y;
+
+        const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+        const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+        const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
         constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-             it, jt, kb0_start, kb0_stop);
+            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+             tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 
         kbc += blocks_per_ne00;
         kbc -= kbc % blocks_per_ne00;
@@ -2668,55 +2776,106 @@ static __global__ void mul_mat_q(
         return;
     }
 
-    const int jt =  kbc /    (blocks_per_ne00*nty);
-    const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
+    int tmp = kbc;
+    const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+    tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+    const int zt = tmp / (ntx*blocks_per_ne00);
+    tmp -= zt * (ntx*blocks_per_ne00);
+    const int jt = tmp / blocks_per_ne00;
+
+    // Defaults for regular matrix multiplication:
+    int col_low    = 0;
+    int col_high   = ncols_y;
+    int col_diff   = ncols_y;
+    int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+    int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+    if (ids_dst) {
+        col_low  = expert_bounds[zt + 0];
+        col_high = expert_bounds[zt + 1];
+        col_diff = col_high - col_low;
+
+        offset_y   = 0;
+        offset_dst = 0;
+
+        if (jt*mmq_x >= col_diff) {
+            return;
+        }
+
+        // The memory layout for the fixup buffer is always contiguous, therefore reset ids:
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+            const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+            if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                break;
+            }
+
+            ids_dst_shared[j] = j;
+        }
+    }
+
+    offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+    offset_dst += it*mmq_y;
+
+    const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+    const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+    const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
     constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-        (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-            it, jt, kb0_start, kb0_stop);
+        (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+         tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 }
 
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 static __global__ void mul_mat_q_stream_k_fixup(
-    float * __restrict__ dst, const float * __restrict__ tmp_last_tile, const int ne00, const int ne01, const int ne11, const int ne0, const int block_num_mmq) {
-
+        const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
+        const int ncols_x, const int nrows_x, const int ncols_y, const int stride_col_dst,
+        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst) {
     constexpr int     mmq_y           = get_mmq_y_device();
     constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
-    const     int64_t blocks_per_ne00 = ne00 / qk;
+    const     int64_t blocks_per_ne00 = ncols_x / qk;
 
     float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
 
-    const int ntx = (ne11 + mmq_x - 1) / mmq_x;
-    const int nty = (ne01 + mmq_y - 1) / mmq_y;
+    const int ntx  = (ncols_y + mmq_x - 1) / mmq_x;
+    const int nty  = (nrows_x + mmq_y - 1) / mmq_y;
+
+    const int bidx0 = blockIdx.x;
+
+    // kbc == k block continuous, current index in continuous ijk space.
+    int64_t kbc0      = (int64_t) bidx0     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+    int64_t kbc0_stop = (int64_t)(bidx0 + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+
+    kbc0      -= (kbc0      % blocks_per_ne00) % blocks_per_iter;
+    kbc0_stop -= (kbc0_stop % blocks_per_ne00) % blocks_per_iter;
+
+    const bool did_not_have_any_data   = kbc0 == kbc0_stop;
+    const bool wrote_beginning_of_tile = kbc0 % blocks_per_ne00 == 0;
+    const bool did_not_write_last      = kbc0/blocks_per_ne00 == kbc0_stop/blocks_per_ne00 && kbc0_stop % blocks_per_ne00 != 0;
+    if (did_not_have_any_data || wrote_beginning_of_tile || did_not_write_last) {
+        return;
+    }
 
     bool any_fixup = false;
 
-    const int bidx_start = ((blockIdx.y*nty + blockIdx.x)     * block_num_mmq)                           / (gridDim.y*gridDim.x);
-    const int bidx_stop  = ((blockIdx.y*nty + blockIdx.x + 1) * block_num_mmq + gridDim.y*gridDim.x - 1) / (gridDim.y*gridDim.x);
+    // Iterate over previous blocks and sum up partial sums written to fixup buffer.
+    // All CUDA blocks that get here must have a previous block that needs a fixup.
+    int64_t bidx = bidx0 - 1;
+    int64_t kbc_stop = kbc0;
+    while(true) {
+        int64_t kbc = bidx*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+        kbc -= (kbc % blocks_per_ne00) % blocks_per_iter;
 
-    int64_t kbc_0;
-    int64_t kbc_stop_0 = (int64_t) bidx_start*blocks_per_ne00*ntx*nty / block_num_mmq;
-
-    for (int bidx = bidx_start; bidx < bidx_stop; ++bidx) {
-        kbc_0 = kbc_stop_0;
-        kbc_stop_0 = (int64_t) (bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq;
-
-        const int64_t kbc      = kbc_0      - (kbc_0      % blocks_per_ne00) % blocks_per_iter;
-        const int64_t kbc_stop = kbc_stop_0 - (kbc_stop_0 % blocks_per_ne00) % blocks_per_iter;
-
-        // Skip fixup tile if the MMQ CUDA block never wrote anything to it:
-        if (kbc == kbc_stop || kbc_stop % blocks_per_ne00 == 0) {
-            continue;
-        }
-
-        const int jt =  kbc_stop /    (blocks_per_ne00*nty);
-        const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
-
-        // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block:
-        if ((unsigned)it != blockIdx.x || (unsigned)jt != blockIdx.y) {
+        if (kbc == kbc_stop) { // Did not have any data.
+            bidx--;
+            kbc_stop = kbc;
             continue;
         }
 
@@ -2733,16 +2892,71 @@ static __global__ void mul_mat_q_stream_k_fixup(
                 sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i];
             }
         }
+
+        // If this block started in a previous tile we are done and don't need to combine additional partial results.
+        if (kbc % blocks_per_ne00 == 0 || kbc/blocks_per_ne00 < kbc0/blocks_per_ne00) {
+            break;
+        }
+        bidx--;
+        kbc_stop = kbc;
     }
 
     if (!any_fixup) {
         return;
     }
 
-    dst += blockIdx.y*mmq_x*ne0 + blockIdx.x*mmq_y;
+    int tmp = kbc0;
+    const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+    tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+    const int zt = tmp / (ntx*blocks_per_ne00);
+    tmp -= zt * (ntx*blocks_per_ne00);
+    const int jt = tmp / blocks_per_ne00;
 
-    const int i_max = ne01 - blockIdx.x*mmq_y - 1;
-    const int j_max = ne11 - blockIdx.y*mmq_x - 1;
+    if (!ids_dst) {
+        const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst + it*mmq_y;
+        dst += offset_dst;
+
+        const int i_max = nrows_x - it*mmq_y - 1;
+        const int j_max = ncols_y - jt*mmq_x - 1;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j > j_max) {
+                return;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                if (need_check && i > i_max) {
+                    continue;
+                }
+
+                dst[j*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            }
+        }
+        return;
+    }
+
+    __shared__ int ids_dst_shared[mmq_x];
+    const int col_low  = expert_bounds[zt + 0];
+    const int col_high = expert_bounds[zt + 1];
+    const int col_diff = col_high - col_low;
+
+    for (int j = threadIdx.y*WARP_SIZE + threadIdx.x; j < mmq_x; j += nwarps*WARP_SIZE) {
+        ids_dst_shared[j] = ids_dst[col_low + j];
+    }
+
+    const int offset_dst = it*mmq_y;
+    dst += offset_dst;
+
+    const int i_max = nrows_x  - it*mmq_y - 1;
+    const int j_max = col_diff - jt*mmq_x - 1;
 
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
@@ -2760,26 +2974,27 @@ static __global__ void mul_mat_q_stream_k_fixup(
                 continue;
             }
 
-            dst[j*ne0 + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            dst[ids_dst_shared[j]*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
         }
     }
 }
 
 struct mmq_args {
-    const char * x; const char * y; float * dst;
-    int64_t ne00; int64_t ne01; int64_t stride01;
-    int64_t ne10; int64_t ne11; int64_t stride11;
-    int64_t ne0;
+    const char * x; ggml_type type_x; const int * y; const int32_t * ids_dst; const int32_t * expert_bounds; float * dst;
+    int64_t ncols_x; int64_t nrows_x; int64_t ncols_y; int64_t stride_row_x; int64_t nrows_dst;
+    int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst;
+    int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
     bool use_stream_k;
 };
 
 template<ggml_type type>
-static int mmq_get_shmem(const int mmq_x, const int mmq_y, const int cc) {
+static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int cc) {
     const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
     const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
-    const int shmem_x = new_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
-    const int shmem_y = mmq_x*sizeof(block_q8_1_mmq);
-    return shmem_x + GGML_PAD(shmem_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
+    const size_t nbs_ids = mmq_x*sizeof(int);
+    const size_t nbs_x = new_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
+    const size_t nbs_y = mmq_x*sizeof(block_q8_1_mmq);
+    return nbs_ids + nbs_x + GGML_PAD(nbs_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
 }
 
 template <ggml_type type, int mmq_x>
@@ -2791,86 +3006,114 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const dim3 block_dims(WARP_SIZE, MMQ_NWARPS, 1);
 
-    const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
+    const int nbytes_shared = mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc);
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-    static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-    if (!shmem_limit_raised[id]) {
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
-        shmem_limit_raised[id] = true;
+    static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+    if (!shared_memory_limit_raised[id]) {
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
+        shared_memory_limit_raised[id] = true;
     }
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
 
-    const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
-    const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
-    const dim3 block_nums_xy_tiling(nty, ntx, 1);
+    const int nty  = (args.nrows_x + mmq_y - 1) / mmq_y;
+    const int ntx  = (args.ncols_y + mmq_x - 1) / mmq_x;
+    const int ntzw = args.nchannels_y * args.nsamples_y;
+    const dim3 block_nums_xy_tiling(nty, ntx, ntzw);
+
+    GGML_ASSERT(args.nchannels_y % args.nchannels_x == 0);
+    GGML_ASSERT(args.nsamples_y  % args.nsamples_x  == 0);
+    const int channel_ratio = args.nchannels_y / args.nchannels_x;
+    const int sample_ratio  = args.nsamples_y  / args.nsamples_x;
 
     if (!args.use_stream_k) {
-        if (args.ne01 % mmq_y == 0) {
+        if (args.nrows_x % mmq_y == 0) {
             constexpr bool need_check = false;
-            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
-                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
+                (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
+                 args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+                 channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
         } else {
             constexpr bool need_check = true;
-            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
-                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
+                (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
+                 args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+                 channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
         }
         return;
     }
 
-    const dim3 block_nums_mmq(nsm, 1, 1);
+    const dim3 block_nums_stream_k(nsm, 1, 1);
+    const bool fixup_needed = ntx*nty*ntzw % nsm != 0;
 
     ggml_cuda_pool & pool = ctx.pool(id);
-    ggml_cuda_pool_alloc<float> tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y);
+    ggml_cuda_pool_alloc<float> tmp_fixup(pool);
+    if (fixup_needed) {
+        tmp_fixup.alloc(block_nums_stream_k.x * mmq_x*mmq_y);
+    }
 
-    if (args.ne01 % mmq_y == 0) {
+    if (args.nrows_x % mmq_y == 0) {
         constexpr bool need_check = false;
 
-        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
-            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
+            (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
+             args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+             channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
 
-        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
-            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+        if (!fixup_needed) {
+            return;
+        }
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_y,
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
     } else {
         constexpr bool need_check = true;
 
-        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
-            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
+            (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
+             args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+             channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
 
-        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
-            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+        if (!fixup_needed) {
+            return;
+        }
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_y,
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
     }
 }
 
 template <ggml_type type>
 void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
-    const int id    = ggml_cuda_get_device();
-    const int cc    = ggml_cuda_info().devices[id].cc;
-    const int smpbo = ggml_cuda_info().devices[id].smpbo;
+    const int    id    = ggml_cuda_get_device();
+    const int    cc    = ggml_cuda_info().devices[id].cc;
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
     const int mmq_x_max = get_mmq_x_max_host(cc);
     const int mmq_y = get_mmq_y_host(cc);
-    const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
 
     int mmq_x_best  = 0;
-    int nparts_best = INT_MAX;
+    int ntiles_x_best = INT_MAX;
 
-    for (int mmq_x = 8; mmq_x <= mmq_x_max && nparts_best > 1; mmq_x += 8) {
+    for (int mmq_x = 8; mmq_x <= mmq_x_max && ntiles_x_best > 1; mmq_x += 8) {
         const int granularity = mmq_get_granularity_host(mmq_x, cc);
 
-        if (mmq_x % granularity != 0 || mmq_get_shmem<type>(mmq_x, mmq_y, cc) > smpbo) {
+        if (mmq_x % granularity != 0 || mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc) > smpbo) {
             continue;
         }
 
-        const int ntiles_x = (args.ne11 + mmq_x - 1) / mmq_x;
-        const int nwaves_xy_tiling = ntiles_x*block_num_y;
-        const int nparts = use_stream_k ? ntiles_x : nwaves_xy_tiling;
+        const int ntiles_x = (args.ncols_y + mmq_x - 1) / mmq_x;
 
-        if (nparts < nparts_best) {
-            mmq_x_best  = mmq_x;
-            nparts_best = nparts;
+        if (ntiles_x < ntiles_x_best) {
+            mmq_x_best = mmq_x;
+            ntiles_x_best = ntiles_x;
         }
     }
 
@@ -2954,6 +3197,9 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS);
 
 // -------------------------------------------------------------------------------------------------------------------------
 
+void ggml_cuda_mul_mat_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+
 void ggml_cuda_op_mul_mat_q(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,

ggml/src/ggml-cuda/mmvq.cu

@@ -158,7 +158,7 @@ static __global__ void mul_mat_vec_q(
     const     int blocks_per_row_x = ncols_x / qk;
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-    // The MUL_MAT_ID code path with ids != nullptr is only implemetned for ncols_dst == 1.
+    // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
     const int channel_dst = blockIdx.y;
     const int channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]          : channel_dst / channel_ratio;
     const int channel_y   = ncols_dst == 1 && ids ? channel_dst % nchannels_y : channel_dst;
@@ -507,7 +507,7 @@ void ggml_cuda_mul_mat_vec_q(
     GGML_ASSERT(        nb0        == ts_dst);
     GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
 
-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
 
     const float   * src1_d =       (const float   *) src1->data;
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
@@ -519,7 +519,7 @@ void ggml_cuda_mul_mat_vec_q(
         const int64_t s11 = src1->nb[1] / ts_src1;
         const int64_t s12 = src1->nb[2] / ts_src1;
         const int64_t s13 = src1->nb[3] / ts_src1;
-        quantize_row_q8_1_cuda(src1_d, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+        quantize_row_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
     }
 
     const int64_t s01 = src0->nb[1] / ts_src0;

ggml/src/ggml-cuda/quantize.cu

@@ -49,29 +49,38 @@ static __global__ void quantize_q8_1(
 
 template <mmq_q8_1_ds_layout ds_layout>
 static __global__ void quantize_mmq_q8_1(
-    const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
+        const float * __restrict__ x, const int32_t * __restrict__ ids, void * __restrict__ vy,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int ne1, const int ne2) {
 
     constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
     constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
 
-    const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
 
-    if (ix0 >= kx0_padded) {
+    if (i0 >= ne0) {
         return;
     }
 
-    const float4 * x4 = (const float4 *) x;
+    const int64_t i1 = blockIdx.y;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;
 
-    const int64_t ix1 = kx1*blockIdx.z + blockIdx.y;
+    const int64_t i00 = i0;
+    const int64_t i01 = ids ? ids[i1] : i1;
+    const int64_t i02 = i2;
+    const int64_t i03 = i3;
+
+    const float4 * x4 = (const float4 *) x;
 
     block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
 
     const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
-    const int64_t ib  = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y;                   // block index in channel
-    const int64_t iqs = ix0 % (4*QK8_1);                                            // quant index in block
+    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y;                    // block index in channel
+    const int64_t iqs = i0 % (4*QK8_1);                                             // quant index in block
 
     // Load 4 floats per thread and calculate max. abs. value between them:
-    const float4 xi = ix0 < kx0 ? x4[(ix1*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+    const float4 xi = i0 < ne00 ? x4[(i03*s03 + i02*s02 + i01*s01 + i00)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
     float amax = fabsf(xi.x);
     amax = fmaxf(amax, fabsf(xi.y));
     amax = fmaxf(amax, fabsf(xi.z));
@@ -87,7 +96,7 @@ static __global__ void quantize_mmq_q8_1(
     if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) {
         sum = xi.x + xi.y + xi.z + xi.w;
 
-        // Exchange calculate sum across vals_per_sum/4 threads.
+        // Calculate sums across vals_per_sum/4 threads.
 #pragma unroll
         for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
             sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
@@ -137,9 +146,10 @@ static __global__ void quantize_mmq_q8_1(
 }
 
 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
-
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
+    GGML_ASSERT(!ids);
     GGML_ASSERT(ne0 % QK8_1 == 0);
 
     const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
@@ -150,9 +160,9 @@ void quantize_row_q8_1_cuda(
 }
 
 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
-
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
     GGML_ASSERT(ne0 % (4*QK8_1) == 0);
 
     const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
@@ -161,21 +171,18 @@ void quantize_mmq_q8_1_cuda(
     switch (mmq_get_q8_1_ds_layout(type_src0)) {
         case MMQ_Q8_1_DS_LAYOUT_D4:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         case MMQ_Q8_1_DS_LAYOUT_DS4:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_DS4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         case MMQ_Q8_1_DS_LAYOUT_D2S6:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D2S6>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         default:
             GGML_ABORT("fatal error");
             break;
     }
-    GGML_UNUSED(s01);
-    GGML_UNUSED(s02);
-    GGML_UNUSED(s03);
 }

ggml/src/ggml-cuda/quantize.cuh

@@ -12,13 +12,16 @@ static_assert(MATRIX_ROW_PADDING %    CUDA_QUANTIZE_BLOCK_SIZE      == 0, "Risk
 static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access.");
 
 typedef void (*quantize_cuda_t)(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);

ggml/src/ggml-opt.cpp

@@ -28,16 +28,19 @@ struct ggml_opt_dataset {
 };
 
 struct ggml_opt_context {
-    ggml_backend_sched_t    backend_sched        = nullptr;
-    ggml_cgraph           * allocated_graph      = nullptr;
-    ggml_cgraph           * allocated_graph_copy = nullptr;
-    struct ggml_context   * ctx_static           = nullptr;
-    struct ggml_context   * ctx_static_cpu       = nullptr;
-    struct ggml_context   * ctx_compute          = nullptr;
-    struct ggml_context   * ctx_copy             = nullptr;
-    ggml_backend_buffer_t   buf_static           = nullptr;
-    ggml_backend_buffer_t   buf_static_cpu       = nullptr;
-    std::mt19937            rng;
+    ggml_backend_sched_t       backend_sched        = nullptr;
+    ggml_cgraph              * allocated_graph      = nullptr;
+    ggml_cgraph              * allocated_graph_copy = nullptr;
+    struct ggml_context      * ctx_static           = nullptr;
+    struct ggml_context      * ctx_cpu              = nullptr;
+    struct ggml_context      * ctx_compute          = nullptr;
+    struct ggml_context      * ctx_copy             = nullptr;
+    ggml_backend_buffer_t      buf_static           = nullptr;
+    ggml_backend_buffer_t      buf_cpu              = nullptr;
+    std::mt19937               rng;
+    enum ggml_opt_loss_type    loss_type;
+    enum ggml_opt_build_type   build_type;
+    enum ggml_opt_build_type   build_type_alloc;
 
     struct ggml_tensor * inputs  = nullptr;
     struct ggml_tensor * outputs = nullptr;
@@ -50,6 +53,11 @@ struct ggml_opt_context {
     struct ggml_cgraph * gf      = nullptr;
     struct ggml_cgraph * gb_grad = nullptr;
     struct ggml_cgraph * gb_opt  = nullptr;
+    bool static_graphs           = false;
+    bool eval_ready              = false;
+    std::vector<struct ggml_tensor *> grad_accs;
+    std::vector<struct ggml_tensor *> grad_m;
+    std::vector<struct ggml_tensor *> grad_v;
 
     int64_t iter               = 1;
     int32_t opt_period         = 1;
@@ -73,7 +81,13 @@ struct ggml_opt_result {
 
 // ====== Dataset ======
 
-ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) {
+ggml_opt_dataset_t ggml_opt_dataset_init(
+        enum ggml_type type_data,
+        enum ggml_type type_label,
+        int64_t        ne_datapoint,
+        int64_t        ne_label,
+        int64_t        ndata,
+        int64_t        ndata_shard) {
     GGML_ASSERT(ne_datapoint >  0);
     GGML_ASSERT(ne_label     >= 0);
     GGML_ASSERT(ndata        >  0);
@@ -92,11 +106,11 @@ ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label,
         result->ctx = ggml_init(params);
     }
 
-    result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata);
+    result->data = ggml_new_tensor_2d(result->ctx, type_data, ne_datapoint, ndata);
     result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata;
 
     if (ne_label > 0) {
-        result->labels = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_label, ndata);
+        result->labels = ggml_new_tensor_2d(result->ctx, type_label, ne_label, ndata);
         result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata;
     } else {
         result->labels = nullptr;
@@ -119,6 +133,10 @@ void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) {
     delete dataset;
 }
 
+int64_t ggml_opt_dataset_ndata(ggml_opt_dataset_t dataset) {
+    return dataset->ndata;
+}
+
 struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) {
     return dataset->data;
 }
@@ -144,6 +162,8 @@ void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor *
     GGML_ASSERT(   data_batch && ggml_is_contiguous(data_batch));
     GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch));
     GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+    GGML_ASSERT(                   data_batch->type == dataset->data->type);
+    GGML_ASSERT(!labels_batch || labels_batch->type == dataset->labels->type);
 
     const size_t nb_data_batch = ggml_nbytes(data_batch);
     GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
@@ -171,6 +191,31 @@ void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor *
     }
 }
 
+void ggml_opt_dataset_get_batch_host(ggml_opt_dataset_t dataset, void * data_batch, size_t nb_data_batch, void * labels_batch, int64_t ibatch) {
+    GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+    GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
+
+    const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
+
+    GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
+
+    for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
+        const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
+
+        const char * ptr_data       = (const char *) dataset->data->data + ishard      *dataset->nbs_data;
+        char       * ptr_data_batch = (char       *) data_batch          + ishard_batch*dataset->nbs_data;
+        memcpy(ptr_data_batch, ptr_data, dataset->nbs_data);
+
+        if (!labels_batch) {
+            continue;
+        }
+
+        const char * ptr_labels       = (const char *) dataset->labels->data + ishard      *dataset->nbs_labels;
+        char       * ptr_labels_batch = (char       *) labels_batch          + ishard_batch*dataset->nbs_labels;
+        memcpy(ptr_labels_batch, ptr_labels, dataset->nbs_labels);
+    }
+}
+
 // ====== Model / Context ======
 
 struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) {
@@ -187,17 +232,18 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
     return result;
 }
 
+struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata) {
+    return *((struct ggml_opt_optimizer_params *) userdata);
+}
+
 struct ggml_opt_params ggml_opt_default_params(
         ggml_backend_sched_t      backend_sched,
-        struct ggml_context     * ctx_compute,
-        struct ggml_tensor      * inputs,
-        struct ggml_tensor      * outputs,
         enum ggml_opt_loss_type   loss_type) {
     return {
         /*backend_sched   =*/ backend_sched,
-        /*ctx_compute     =*/ ctx_compute,
-        /*inputs          =*/ inputs,
-        /*logits          =*/ outputs,
+        /*ctx_compute     =*/ nullptr,
+        /*inputs          =*/ nullptr,
+        /*logits          =*/ nullptr,
         /*loss_type       =*/ loss_type,
         /*build_type      =*/ GGML_OPT_BUILD_TYPE_OPT,
         /*opt_period      =*/ 1,
@@ -266,195 +312,246 @@ static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) {
     return dst;
 }
 
-static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
-    GGML_ASSERT(graph);
-    if (opt_ctx->allocated_graph == graph) {
-        return;
-    }
+static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
+    GGML_ASSERT(opt_ctx->ctx_compute && "no compute context set, either use static graphs or set one with ggml_opt_prepare_alloc");
+    GGML_ASSERT((!opt_ctx->static_graphs || opt_ctx->inputs->data) && "when using static graphs the inputs must be allocated statically");
 
-    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+    const bool accumulate = opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD &&
+        !(opt_ctx->static_graphs && opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period == 1);
 
-    {
-        ggml_init_params params = {
-            /*.mem_size   =*/ ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE,
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        ggml_free(opt_ctx->ctx_copy);
-        opt_ctx->ctx_copy = ggml_init(params);
-    }
-
-    opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
-
-    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
-    opt_ctx->allocated_graph = graph;
-}
-
-ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
-    ggml_opt_context_t result = new struct ggml_opt_context;
-    result->backend_sched   = params.backend_sched;
-    result->ctx_compute     = params.ctx_compute;
-    result->inputs          = params.inputs;
-    result->outputs         = params.outputs;
-    result->opt_period      = params.opt_period;
-    result->get_opt_pars    = params.get_opt_pars;
-    result->get_opt_pars_ud = params.get_opt_pars_ud;
-
-    GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
-    GGML_ASSERT(result->opt_period >= 1);
-
-    const bool accumulate = params.build_type == GGML_OPT_BUILD_TYPE_GRAD ||
-        (params.build_type == GGML_OPT_BUILD_TYPE_OPT && result->opt_period > 1);
-
-    ggml_set_input(result->inputs);
-    ggml_set_output(result->outputs);
-
-    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
-    ggml_build_forward_expand(result->gf, result->outputs);
+    ggml_set_input(opt_ctx->inputs);
+    ggml_set_output(opt_ctx->outputs);
 
     int n_param = 0;
-    for (int i = 0; i < result->gf->n_nodes; ++i) {
-        if (result->gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
+    for (int i = 0; i < opt_ctx->gf->n_nodes; ++i) {
+        const struct ggml_tensor * node = opt_ctx->gf->nodes[i];
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             n_param++;
         }
+        GGML_ASSERT(!(node->flags & GGML_TENSOR_FLAG_LOSS) && "support for extra loss terms not implemented");
     }
 
-    {
+    if (!opt_ctx->ctx_static) {
         // The static context is used for:
-        //   - gradients (1 tensor per param if using gradient accumulation)
+        //   - gradients (1 per loss, 1 tensor per param if using gradient accumulation)
         //   - optimizer momenta (2 tensors per param)
-        //   - labels
-        //   - loss + its gradient (up to 5 tensors)
-        //   - pred
-        //   - ncorrect (2 tensors).
-        const size_t tensors_per_param = (accumulate ? 1 : 0) + (params.build_type == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
-        const size_t size_meta = (tensors_per_param*n_param + 9) * ggml_tensor_overhead();
+        //   - labels (if using static graphs)
+        //   - loss (if using static graphs, up to 5 tensors)
+        //   - pred (if using static graphs)
+        //   - ncorrect (if using static graphs, 2 tensors).
+        constexpr size_t n_loss = 1;
+        const size_t tensors_per_param = (accumulate ? 1 : 0) +
+            (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t tensors_const = opt_ctx->static_graphs ? 9 : 0;
+        const size_t size_meta = (n_loss + tensors_per_param*n_param + tensors_const) * ggml_tensor_overhead();
         struct ggml_init_params params = {
             /*.mem_size   =*/ size_meta,
             /*.mem_buffer =*/ nullptr,
             /*.no_alloc   =*/ true,
         };
-        result->ctx_static = ggml_init(params);
+        opt_ctx->ctx_static = ggml_init(params);
     }
+    GGML_ASSERT(opt_ctx->build_type <= opt_ctx->build_type_alloc);
+
     {
-        // The static cpu context is used for:
-        //   - optimizer parameters (1 for the entire context)
+        // The cpu context is allocated statically if using static graphs, dynamically otherwise.
+        // It is used for:
+        //   - optimizer parameters (1 shared for all optimizer invocations)
         const size_t size_meta = 1 * ggml_tensor_overhead();
         struct ggml_init_params params = {
             /*.mem_size   =*/ size_meta,
             /*.mem_buffer =*/ nullptr,
             /*.no_alloc   =*/ true,
         };
-        result->ctx_static_cpu = ggml_init(params);
+        ggml_free(opt_ctx->ctx_cpu);
+        opt_ctx->ctx_cpu = ggml_init(params);
+
+        ggml_backend_buffer_free(opt_ctx->buf_cpu);
+        opt_ctx->buf_cpu = nullptr;
     }
 
+    struct ggml_context * ctx_results = opt_ctx->static_graphs ? opt_ctx->ctx_static : opt_ctx->ctx_compute;
 
-    switch (params.loss_type) {
+    switch (opt_ctx->loss_type) {
         case GGML_OPT_LOSS_TYPE_MEAN: {
-            result->loss = ggml_sum(result->ctx_static, result->outputs);
-            ggml_set_name(result->loss, "loss_sum");
-            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
-            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
-            ggml_set_name(result->loss, "loss_mean");
-            result->loss_per_datapoint = true;
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs);
+            ggml_set_name(opt_ctx->loss, "loss_sum");
+            const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs));
+            opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale);
+            ggml_set_name(opt_ctx->loss, "loss_mean");
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
         case GGML_OPT_LOSS_TYPE_SUM: {
-            result->loss = ggml_sum(result->ctx_static, result->outputs);
-            ggml_set_name(result->loss, "loss_sum");
-            result->loss_per_datapoint = false;
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs);
+            ggml_set_name(opt_ctx->loss, "loss_sum");
+            opt_ctx->loss_per_datapoint = false;
             break;
         }
         case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: {
-            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
-            ggml_set_input(result->labels);
-            ggml_set_name(result->labels, "labels");
-            result->loss = ggml_cross_entropy_loss(result->ctx_static, result->outputs, result->labels);
-            ggml_set_name(result->loss, "loss_cross_entropy");
-            if (result->opt_period > 1) {
-                result->loss = ggml_scale(result->ctx_static, result->loss, 1.0f / result->opt_period);
-                ggml_set_name(result->loss, "loss_cross_entropy_scaled");
+            opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs);
+            ggml_set_input(opt_ctx->labels);
+            ggml_set_name(opt_ctx->labels, "labels");
+            opt_ctx->loss = ggml_cross_entropy_loss(ctx_results, opt_ctx->outputs, opt_ctx->labels);
+            ggml_set_name(opt_ctx->loss, "loss_cross_entropy");
+            if (opt_ctx->opt_period > 1) {
+                opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, 1.0f / opt_ctx->opt_period);
+                ggml_set_name(opt_ctx->loss, "loss_cross_entropy_scaled");
             }
-            result->loss_per_datapoint = true;
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
         case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: {
-            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
-            ggml_set_input(result->labels);
-            ggml_set_name(result->labels, "labels");
-            result->loss = ggml_sub(result->ctx_static, result->outputs, result->labels);
-            ggml_set_name(result->loss, "loss_error");
-            result->loss = ggml_sqr(result->ctx_static, result->loss);
-            ggml_set_name(result->loss, "loss_squared_error");
-            result->loss = ggml_sum(result->ctx_static, result->loss);
-            ggml_set_name(result->loss, "loss_sum_squared_error");
-            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
-            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
-            ggml_set_name(result->loss, "loss_mean_squared_error");
-            result->loss_per_datapoint = true;
+            opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs);
+            ggml_set_input(opt_ctx->labels);
+            ggml_set_name(opt_ctx->labels, "labels");
+            opt_ctx->loss = ggml_sub(ctx_results, opt_ctx->outputs, opt_ctx->labels);
+            ggml_set_name(opt_ctx->loss, "loss_error");
+            opt_ctx->loss = ggml_sqr(ctx_results, opt_ctx->loss);
+            ggml_set_name(opt_ctx->loss, "loss_squared_error");
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->loss);
+            ggml_set_name(opt_ctx->loss, "loss_sum_squared_error");
+            const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs));
+            opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale);
+            ggml_set_name(opt_ctx->loss, "loss_mean_squared_error");
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
     }
-    ggml_set_output(result->loss);
-    ggml_set_loss(result->loss);
-    ggml_build_forward_expand(result->gf, result->loss);
+    ggml_set_output(opt_ctx->loss);
+    ggml_set_loss(opt_ctx->loss);
+    ggml_build_forward_expand(opt_ctx->gf, opt_ctx->loss);
 
-    result->pred = ggml_argmax(result->ctx_static, result->outputs);
-    ggml_set_name(result->pred, "pred");
-    ggml_set_output(result->pred);
-    ggml_build_forward_expand(result->gf, result->pred);
+    if (opt_ctx->loss_type == GGML_OPT_LOSS_TYPE_CROSS_ENTROPY) {
+        opt_ctx->pred = ggml_argmax(ctx_results, opt_ctx->outputs);
+        ggml_set_name(opt_ctx->pred, "pred");
+        ggml_set_output(opt_ctx->pred);
+        ggml_build_forward_expand(opt_ctx->gf, opt_ctx->pred);
 
-    if (result->labels) {
-        result->ncorrect = ggml_count_equal(result->ctx_static, result->pred, ggml_argmax(result->ctx_static, result->labels));
-        ggml_set_name(result->ncorrect, "ncorrect");
-        ggml_set_output(result->ncorrect);
-        ggml_build_forward_expand(result->gf, result->ncorrect);
-    } else {
-        result->ncorrect = nullptr;
+        opt_ctx->ncorrect = ggml_count_equal(ctx_results, opt_ctx->pred, ggml_argmax(ctx_results, opt_ctx->labels));
+        ggml_set_name(opt_ctx->ncorrect, "ncorrect");
+        ggml_set_output(opt_ctx->ncorrect);
+        ggml_build_forward_expand(opt_ctx->gf, opt_ctx->ncorrect);
     }
 
-    if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
-        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        return result;
+    if (opt_ctx->buf_static) {
+        if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
+            return;
+        }
+    } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_FORWARD) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(
+            opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        return;
     }
 
-    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
-    result->gb_grad = ggml_graph_dup(result->ctx_compute, result->gf);
-    ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
+    if (opt_ctx->grad_accs.empty()) {
+        GGML_ASSERT(opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD);
 
-    if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
-        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        ggml_graph_reset(result->gb_grad);
-        return result;
-    }
+        const int n_nodes = opt_ctx->gf->n_nodes;
+        opt_ctx->grad_accs.resize(n_nodes);
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor * node = opt_ctx->gf->nodes[i];
+            if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
+                opt_ctx->grad_accs[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+            } else {
+                opt_ctx->grad_accs[i] = nullptr;
+            }
+        }
 
-    GGML_ASSERT(params.build_type == GGML_OPT_BUILD_TYPE_OPT);
-
-    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
-    result->gb_opt = ggml_graph_dup(result->ctx_compute, result->gb_grad);
-
-    result->adamw_params = ggml_new_tensor_1d(result->ctx_static_cpu, GGML_TYPE_F32, 7);
-    ggml_set_input(result->adamw_params);
-    ggml_set_name(result->adamw_params, "adamw_params");
-
-    for (int i = result->gf->n_nodes-1; i >= 0; --i) {
-        struct ggml_tensor * node = result->gb_opt->nodes[i];
-        struct ggml_tensor * grad = ggml_graph_get_grad(result->gb_opt, node);
-
-        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-            struct ggml_tensor * m        = ggml_dup_tensor(result->ctx_static, node);
-            struct ggml_tensor * v        = ggml_dup_tensor(result->ctx_static, node);
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(result->ctx_compute, node, grad, m, v, result->adamw_params);
-            ggml_build_forward_expand(result->gb_opt, opt_step);
+        if (opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
+            opt_ctx->grad_m.resize(n_nodes);
+            opt_ctx->grad_v.resize(n_nodes);
+            for (int i = 0; i < n_nodes; ++i) {
+                ggml_tensor * node = opt_ctx->gf->nodes[i];
+                if (node->flags & GGML_TENSOR_FLAG_PARAM) {
+                    opt_ctx->grad_m[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+                    opt_ctx->grad_v[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+                } else {
+                    opt_ctx->grad_m[i] = nullptr;
+                    opt_ctx->grad_v[i] = nullptr;
+                }
+            }
         }
     }
 
-    result->buf_static = ggml_backend_alloc_ctx_tensors(
-        result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
+    opt_ctx->gb_grad = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gf, /*force_grads =*/ true);
+    ggml_build_backward_expand(opt_ctx->ctx_compute, opt_ctx->gb_grad, opt_ctx->grad_accs.data());
 
-    result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
+    if (opt_ctx->buf_static) {
+        if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_GRAD) {
+            return;
+        }
+    } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_GRAD) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
 
-    ggml_graph_reset(result->gb_opt);
+    GGML_ASSERT(opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT);
+
+    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
+    opt_ctx->gb_opt = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gb_grad, /*force_grads =*/ true);
+
+    opt_ctx->adamw_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, 7);
+    ggml_set_input(opt_ctx->adamw_params);
+    ggml_set_name(opt_ctx->adamw_params, "adamw_params");
+
+    for (int i = opt_ctx->gf->n_nodes-1; i >= 0; --i) {
+        struct ggml_tensor * node = opt_ctx->gb_opt->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(opt_ctx->gb_opt, node);
+
+        if (grad && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
+            struct ggml_tensor * m        = opt_ctx->grad_m[i];
+            struct ggml_tensor * v        = opt_ctx->grad_v[i];
+            struct ggml_tensor * opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, opt_ctx->adamw_params);
+
+            ggml_set_name(m,        (std::string("AdamW m for ")    + std::string(node->name)).c_str());
+            ggml_set_name(v,        (std::string("AdamW v for ")    + std::string(node->name)).c_str());
+            ggml_set_name(opt_step, (std::string("AdamW step for ") + std::string(node->name)).c_str());
+
+            ggml_build_forward_expand(opt_ctx->gb_opt, opt_step);
+        }
+    }
+
+    if (!opt_ctx->buf_static) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(
+            opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        ggml_graph_reset(opt_ctx->gb_opt);
+    }
+
+    opt_ctx->buf_cpu = ggml_backend_alloc_ctx_tensors_from_buft(opt_ctx->ctx_cpu, ggml_backend_cpu_buffer_type());
+}
+
+ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
+    ggml_opt_context_t result = new struct ggml_opt_context;
+    result->backend_sched    = params.backend_sched;
+    result->ctx_compute      = params.ctx_compute;
+    result->loss_type        = params.loss_type;
+    result->build_type       = params.build_type;
+    result->build_type_alloc = params.build_type;
+    result->inputs           = params.inputs;
+    result->outputs          = params.outputs;
+    result->opt_period       = params.opt_period;
+    result->get_opt_pars     = params.get_opt_pars;
+    result->get_opt_pars_ud  = params.get_opt_pars_ud;
+
+    GGML_ASSERT(result->opt_period >= 1);
+
+    result->static_graphs = result->ctx_compute;
+
+    if (!result->static_graphs) {
+        GGML_ASSERT(!result->inputs);
+        GGML_ASSERT(!result->outputs);
+        return result;
+    }
+
+    GGML_ASSERT(result->inputs);
+    GGML_ASSERT(result->outputs);
+
+    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
+    ggml_build_forward_expand(result->gf, result->outputs);
+
+    ggml_opt_build(result);
 
     return result;
 }
@@ -464,9 +561,9 @@ void ggml_opt_free(ggml_opt_context_t opt_ctx) {
         return;
     }
     ggml_backend_buffer_free(opt_ctx->buf_static);
-    ggml_backend_buffer_free(opt_ctx->buf_static_cpu);
+    ggml_backend_buffer_free(opt_ctx->buf_cpu);
     ggml_free(opt_ctx->ctx_static);
-    ggml_free(opt_ctx->ctx_static_cpu);
+    ggml_free(opt_ctx->ctx_cpu);
     delete opt_ctx;
 }
 
@@ -582,8 +679,80 @@ void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, doubl
 
 // ====== Computation ======
 
-static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_opt_result * result) {
-    if (graph != opt_ctx->gf) {
+void ggml_opt_prepare_alloc(
+        ggml_opt_context_t    opt_ctx,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * gf,
+        struct ggml_tensor  * inputs,
+        struct ggml_tensor  * outputs) {
+    GGML_ASSERT(!opt_ctx->static_graphs);
+    opt_ctx->ctx_compute = ctx_compute;
+    opt_ctx->gf          = gf;
+    opt_ctx->inputs      = inputs;
+    opt_ctx->outputs     = outputs;
+}
+
+void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward) {
+    GGML_ASSERT(!opt_ctx->eval_ready);
+    if (backward) {
+        const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+        opt_ctx->build_type = opt_i_next == 0 ? GGML_OPT_BUILD_TYPE_OPT : GGML_OPT_BUILD_TYPE_GRAD;
+    } else {
+        opt_ctx->build_type = GGML_OPT_BUILD_TYPE_FORWARD;
+    }
+
+    if (!opt_ctx->static_graphs) {
+        ggml_opt_build(opt_ctx);
+    }
+
+    struct ggml_cgraph * graph = nullptr;
+    switch (opt_ctx->build_type) {
+        case GGML_OPT_BUILD_TYPE_FORWARD: {
+            graph = opt_ctx->gf;
+        } break;
+        case GGML_OPT_BUILD_TYPE_GRAD: {
+            graph = opt_ctx->gb_grad;
+        } break;
+        case GGML_OPT_BUILD_TYPE_OPT: {
+            graph = opt_ctx->gb_opt;
+        } break;
+    }
+    GGML_ASSERT(graph);
+
+    if (opt_ctx->allocated_graph == graph) {
+        opt_ctx->eval_ready = true;
+        return;
+    }
+
+    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+
+    if (opt_ctx->static_graphs) {
+        ggml_init_params params = {
+            /*.mem_size   =*/ graph->size*ggml_tensor_overhead() + ggml_graph_overhead_custom(graph->size, graph->grads),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ggml_free(opt_ctx->ctx_copy);
+        opt_ctx->ctx_copy = ggml_init(params);
+
+        opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
+    } else {
+        opt_ctx->allocated_graph_copy = graph;
+    }
+
+    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->allocated_graph = graph;
+
+    if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period > 1 && opt_ctx->opt_i == 0) {
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
+
+    opt_ctx->eval_ready = true;
+}
+
+void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result) {
+    GGML_ASSERT(opt_ctx->eval_ready);
+    if (opt_ctx->allocated_graph == opt_ctx->gb_opt) {
         struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
 
         GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
@@ -609,9 +778,19 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
         adamw_par_data[6] = beta2h;
     }
 
-    ggml_opt_alloc_graph(opt_ctx, graph);
     ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
     opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt;
+    opt_ctx->opt_i = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+
+    if (!opt_ctx->static_graphs) {
+        opt_ctx->gf                   = nullptr;
+        opt_ctx->gb_grad              = nullptr;
+        opt_ctx->gb_opt               = nullptr;
+        opt_ctx->allocated_graph      = nullptr;
+        opt_ctx->allocated_graph_copy = nullptr;
+    }
+
+    opt_ctx->eval_ready = false;
 
     if (!result) {
         return;
@@ -635,12 +814,14 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
     ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss));
     result->loss.push_back(loss);
 
-    GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
-    std::vector<int32_t> pred(ndata);
-    ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
-    result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+    if (opt_ctx->pred) {
+        GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
+        std::vector<int32_t> pred(ndata);
+        ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
+        result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+    }
 
-    if (!opt_ctx->labels || result->ncorrect < 0) {
+    if (!opt_ctx->ncorrect || result->ncorrect < 0) {
         result->ncorrect = -1;
         return;
     }
@@ -652,26 +833,6 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
     result->ncorrect += ncorrect;
 }
 
-void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
-    ggml_opt_eval_graph(opt_ctx, opt_ctx->gf, result);
-}
-
-void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
-    if (opt_ctx->opt_period == 1) {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
-        return;
-    }
-
-    const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
-    if (opt_i_next == 0) {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
-        ggml_opt_reset(opt_ctx, /*optimizer =*/ false);
-    } else {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_grad, result);
-    }
-    opt_ctx->opt_i = opt_i_next;
-}
-
 // ====== High-Level Functions ======
 
 void ggml_opt_epoch(
@@ -700,16 +861,18 @@ void ggml_opt_epoch(
     int64_t ibatch = 0;
     int64_t t_loop_start = ggml_time_us();
     for (; ibatch < ibatch_split; ++ibatch) {
+        ggml_opt_alloc(opt_ctx, /*backward =*/ true);
         ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
-        ggml_opt_forward_backward(opt_ctx, result_train);
+        ggml_opt_eval(opt_ctx, result_train);
         if (callback_train) {
             callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start);
         }
     }
     t_loop_start = ggml_time_us();
     for (; ibatch < nbatches; ++ibatch) {
+        ggml_opt_alloc(opt_ctx, /*backward =*/ false);
         ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
-        ggml_opt_forward(opt_ctx, result_eval);
+        ggml_opt_eval(opt_ctx, result_eval);
         if (callback_eval) {
             callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start);
         }
@@ -726,13 +889,26 @@ void ggml_opt_epoch_callback_progress_bar(
         int64_t            t_start_us) {
     fprintf(stderr, "%s[", train ? "train: " : "val:   ");
 
-    constexpr int64_t bar_length = 25;
+    // The progress bar consists of partially filled blocks, unicode has 8 separate fill levels.
+    constexpr int64_t bar_length = 8;
+    const int64_t ibatch8 = 8 * ibatch;
     for (int64_t j = 0; j < bar_length; ++j) {
-        const int64_t ibatch_j = ibatch_max * j/bar_length;
-        if (ibatch_j < ibatch) {
-            fprintf(stderr, "=");
-        } else if (ibatch_max * (j - 1)/bar_length < ibatch) {
-            fprintf(stderr, ">");
+        if        (ibatch_max * (8*j + 8) / bar_length < ibatch8) {
+            fprintf(stderr, "\u2588"); // full block
+        } else if (ibatch_max * (8*j + 7) / bar_length < ibatch8) {
+            fprintf(stderr, "\u2589"); // 7/8 filled
+        } else if (ibatch_max * (8*j + 6) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258A"); // 6/8 filled
+        } else if (ibatch_max * (8*j + 5) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258B"); // 5/8 filled
+        } else if (ibatch_max * (8*j + 4) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258C"); // 4/8 filled
+        } else if (ibatch_max * (8*j + 3) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258D"); // 3/8 filled
+        } else if (ibatch_max * (8*j + 2) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258E"); // 2/8 filled
+        } else if (ibatch_max * (8*j + 1) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258F"); // 1/8 filled
         } else {
             fprintf(stderr, " ");
         }
@@ -764,8 +940,8 @@ void ggml_opt_epoch_callback_progress_bar(
     const int64_t t_eta_m = t_eta_s / 60;
     t_eta_s -= t_eta_m * 60;
 
-    fprintf(stderr, "| data=%06" PRId64 "/%06" PRId64 ", loss=%.6lf+-%.6lf, accuracy=%.2lf+-%.2lf%%, "
-            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 ", ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 "]\r",
+    fprintf(stderr, "] data=%07" PRId64 "/%07" PRId64 " loss=%.5lf±%.5lf acc=%.2lf±%.2lf%% "
+            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " \r",
             idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc,
             t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s);
     if (ibatch == ibatch_max) {
@@ -806,7 +982,10 @@ void ggml_opt_fit(
 
     int64_t epoch = 1;
 
-    ggml_opt_params params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
+    ggml_opt_params params = ggml_opt_default_params(backend_sched, loss_type);
+    params.ctx_compute     = ctx_compute;
+    params.inputs          = inputs;
+    params.outputs         = outputs;
     params.opt_period      = opt_period;
     params.get_opt_pars    = get_opt_pars;
     params.get_opt_pars_ud = &epoch;

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -518,6 +518,11 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
     result.view_src = reinterpret_cast<uint64_t>(tensor->view_src);
     result.view_offs = tensor->view_offs;
     result.data = reinterpret_cast<uint64_t>(tensor->data);
+
+    // Avoid sending uninitialized data over the wire
+    memset(result.name, 0, sizeof(result.name));
+    memset(result.padding, 0, sizeof(result.padding));
+
     snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
     return result;
 }

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -71,6 +71,22 @@ if (Vulkan_FOUND)
         add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
     endif()
 
+    # Compile a test shader to determine whether GL_EXT_bfloat16 is supported.
+    # If it's not, there will be an error to stderr.
+    # If it's supported, set a define to indicate that we should compile those shaders
+    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp"
+                    OUTPUT_VARIABLE glslc_output
+                    ERROR_VARIABLE glslc_error)
+
+    if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_bfloat16.*")
+        message(STATUS "GL_EXT_bfloat16 not supported by glslc")
+        set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT OFF)
+    else()
+        message(STATUS "GL_EXT_bfloat16 supported by glslc")
+        set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT ON)
+        add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+    endif()
+
     target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
     target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
 
@@ -142,6 +158,7 @@ if (Vulkan_FOUND)
                     -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT_GLSLC_SUPPORT}
                     -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT}
                     -DGGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT=${GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT}
+                    -DGGML_VULKAN_BFLOAT16_GLSLC_SUPPORT=${GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT}
             BUILD_COMMAND ${CMAKE_COMMAND} --build .
             INSTALL_COMMAND ${CMAKE_COMMAND} --install .
             INSTALL_DIR ${CMAKE_BINARY_DIR}

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

@@ -51,6 +51,24 @@
 
 #include "ggml-vulkan-shaders.hpp"
 
+// remove this once it's more widely available in the SDK
+#if !defined(VK_KHR_shader_bfloat16)
+
+#define VK_KHR_shader_bfloat16 1
+#define VK_KHR_SHADER_BFLOAT16_SPEC_VERSION                          1
+#define VK_KHR_SHADER_BFLOAT16_EXTENSION_NAME                        "VK_KHR_shader_bfloat16"
+#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR ((VkStructureType)1000141000)
+#define VK_COMPONENT_TYPE_BFLOAT16_KHR                               ((VkComponentTypeKHR)1000141000)
+
+typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR {
+    VkStructureType                       sType;
+    void*                                 pNext;
+    VkBool32                              shaderBFloat16Type;
+    VkBool32                              shaderBFloat16DotProduct;
+    VkBool32                              shaderBFloat16CooperativeMatrix;
+} VkPhysicalDeviceShaderBfloat16FeaturesKHR;
+#endif
+
 #define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1))
 #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
@@ -266,8 +284,9 @@ struct vk_device_struct {
     bool subgroup_require_full_support;
 
     bool coopmat_support;
-    bool coopmat_acc_f32_support;
-    bool coopmat_acc_f16_support;
+    bool coopmat_acc_f32_support {};
+    bool coopmat_acc_f16_support {};
+    bool coopmat_bf16_support {};
     uint32_t coopmat_m;
     uint32_t coopmat_n;
     uint32_t coopmat_k;
@@ -293,6 +312,7 @@ struct vk_device_struct {
 
     vk_matmul_pipeline pipeline_matmul_f32 {};
     vk_matmul_pipeline pipeline_matmul_f32_f16 {};
+    vk_matmul_pipeline pipeline_matmul_bf16 {};
     vk_matmul_pipeline2 pipeline_matmul_f16;
     vk_matmul_pipeline2 pipeline_matmul_f16_f32;
 
@@ -301,6 +321,7 @@ struct vk_device_struct {
     vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_COUNT];
 
     vk_matmul_pipeline pipeline_matmul_id_f32 {};
+    vk_matmul_pipeline pipeline_matmul_id_bf16 {};
     vk_matmul_pipeline2 pipeline_matmul_id_f16;
     vk_matmul_pipeline2 pipeline_matmul_id_f16_f32;
 
@@ -333,8 +354,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_clamp_f32;
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32;
-    vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
-    vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16;
+    vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16, pipeline_cpy_f32_bf16;
+    vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f32_bf16;
     vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_norm_f32;
@@ -368,6 +389,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
+    vk_pipeline pipeline_conv2d_dw_whcn_f32;
+    vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
     // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
     vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
@@ -680,6 +703,24 @@ struct vk_op_rwkv_wkv7_push_constants {
     uint32_t H;
 };
 
+struct vk_op_conv2d_dw_push_constants {
+    uint32_t ne;
+    uint32_t batches;
+    uint32_t channels;
+    uint32_t dst_w;
+    uint32_t dst_h;
+    uint32_t src_w;
+    uint32_t src_h;
+    uint32_t knl_w;
+    uint32_t knl_h;
+    int32_t stride_x;
+    int32_t stride_y;
+    int32_t pad_x;
+    int32_t pad_y;
+    int32_t dilation_x;
+    int32_t dilation_y;
+};
+
 struct vk_op_upscale_push_constants {
     uint32_t ne; uint32_t a_offset; uint32_t d_offset;
     uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
@@ -1791,6 +1832,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     if (!device->pipeline_matmul_id_f32) {
         device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     }
+    if (!device->pipeline_matmul_bf16) {
+        device->pipeline_matmul_bf16 = std::make_shared<vk_matmul_pipeline_struct>();
+    }
+    if (!device->pipeline_matmul_id_bf16) {
+        device->pipeline_matmul_id_bf16 = std::make_shared<vk_matmul_pipeline_struct>();
+    }
 
     std::vector<std::future<void>> compiles;
     auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint,
@@ -1900,6 +1947,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT)   \
 
         CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+        }
+#endif
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
@@ -1921,6 +1973,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f16,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
 
         CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+        }
+#endif
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
@@ -1974,6 +2031,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, )
+        }
+#endif
 
         if (device->coopmat_acc_f16_support) {
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2022,6 +2084,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+        }
+#endif
 
         if (device->coopmat_acc_f16_support) {
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2104,6 +2171,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2139,6 +2208,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2191,6 +2262,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2226,6 +2299,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2246,8 +2321,26 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc,  matmul_id_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-#undef CREATE_MM
     }
+    // reusing CREATE_MM from the fp32 path
+    if ((device->coopmat2 || device->coopmat_support)
+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+        && !device->coopmat_bf16_support
+#endif
+        ) {
+        // use scalar tile sizes
+        l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
+        m_warptile = { 128,  64,  64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 };
+        s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, 1, subgroup_size_8 };
+
+        l_wg_denoms = {128, 128, 1 };
+        m_wg_denoms = { 64,  64, 1 };
+        s_wg_denoms = { 32,  32, 1 };
+
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+    }
+#undef CREATE_MM
 
     // mul mat vec
 
@@ -2266,6 +2359,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f32_f32_len, mul_mat_vec_bf16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
@@ -2288,6 +2382,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f16_f32_len, mul_mat_vec_bf16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
@@ -2311,6 +2406,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
@@ -2356,6 +2452,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     // get_rows
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32",  get_rows_f32_len,  get_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16",  get_rows_f16_len,  get_rows_f16_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_BF16], "get_rows_bf16", get_rows_bf16_len, get_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -2373,6 +2470,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32",  get_rows_f32_f32_len,  get_rows_f32_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32",  get_rows_f16_f32_len,  get_rows_f16_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_BF16], "get_rows_bf16_f32", get_rows_bf16_f32_len, get_rows_bf16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -2399,7 +2497,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
         }
     }
-    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
@@ -2410,10 +2508,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_bf16,"cpy_f32_bf16",cpy_f32_bf16_len,cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
     if (device->float_controls_rte_fp16) {
         ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
@@ -2529,6 +2630,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f32, "conv2d_dw_cwhn_f32", conv2d_dw_cwhn_f32_len, conv2d_dw_cwhn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
+
     for (auto &c : compiles) {
         c.wait();
     }
@@ -2578,6 +2682,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         bool coopmat2_support = false;
         device->coopmat_support = false;
         device->integer_dot_product = false;
+        bool bfloat16_support = false;
 
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
@@ -2608,6 +2713,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
                        !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) {
                 device->integer_dot_product = true;
 #endif
+            } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 &&
+                       !getenv("GGML_VK_DISABLE_BFLOAT16")) {
+                bfloat16_support = true;
             }
         }
 
@@ -2794,6 +2902,17 @@ static vk_device ggml_vk_get_device(size_t idx) {
         }
 #endif
 
+#if defined(VK_KHR_shader_bfloat16)
+        VkPhysicalDeviceShaderBfloat16FeaturesKHR bfloat16_features {};
+        bfloat16_features.pNext = nullptr;
+        bfloat16_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR;
+        if (bfloat16_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&bfloat16_features;
+            last_struct = (VkBaseOutStructure *)&bfloat16_features;
+            device_extensions.push_back("VK_KHR_shader_bfloat16");
+        }
+#endif
+
         VkPhysicalDeviceMaintenance4Features maint4_features {};
         maint4_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_FEATURES;
         if (maintenance4_support) {
@@ -2991,6 +3110,25 @@ static vk_device ggml_vk_get_device(size_t idx) {
                     device->coopmat_int_n = prop.NSize;
                     device->coopmat_int_k = prop.KSize;
                 }
+#if defined(VK_KHR_shader_bfloat16) && defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+                if (prop.AType == VK_COMPONENT_TYPE_BFLOAT16_KHR &&
+                    prop.BType == VK_COMPONENT_TYPE_BFLOAT16_KHR &&
+                    prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR &&
+                    prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR &&
+                    (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup
+                ) {
+                    // coopmat sizes not set yet
+                    if (device->coopmat_m == 0) {
+                        device->coopmat_bf16_support = true;
+                        device->coopmat_m = prop.MSize;
+                        device->coopmat_n = prop.NSize;
+                        device->coopmat_k = prop.KSize;
+                    } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) {
+                        // Only enable if shape is identical
+                        device->coopmat_bf16_support = true;
+                    }
+                }
+#endif
             }
 
             if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) {
@@ -2998,11 +3136,19 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n");
                 device->coopmat_support = false;
             }
+            if (getenv("GGML_VK_DISABLE_BFLOAT16")) {
+                device->coopmat_bf16_support = false;
+            }
         }
 
         if (device->coopmat_support) {
             device_extensions.push_back("VK_KHR_cooperative_matrix");
         }
+#if defined(VK_KHR_shader_bfloat16)
+        if (device->coopmat_bf16_support) {
+            device_extensions.push_back("VK_KHR_shader_bfloat16");
+        }
+#endif
 #endif
         device->name = GGML_VK_NAME + std::to_string(idx);
 
@@ -3459,6 +3605,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
         return ctx->device->pipeline_matmul_f32_f16;
     }
+    if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) {
+        return ctx->device->pipeline_matmul_bf16;
+    }
     if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_f16_f32.f16acc;
@@ -3530,6 +3679,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
     switch (a_type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -3562,6 +3712,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_id_f32;
     }
+    if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) {
+        return ctx->device->pipeline_matmul_id_bf16;
+    }
     if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_id_f16_f32.f16acc;
@@ -3615,6 +3768,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context
     switch (a_type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4350,6 +4504,13 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_cpy_f16_f16;
         }
     }
+    if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_BF16) {
+        if (contig) {
+            return ctx->device->pipeline_contig_cpy_f32_bf16;
+        } else {
+            return ctx->device->pipeline_cpy_f32_bf16;
+        }
+    }
     if (src->type == GGML_TYPE_F32) {
         switch (to) {
         case GGML_TYPE_Q4_0:
@@ -4477,8 +4638,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) ||
                               !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
+                              (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) ||
                               !ggml_vk_dim01_contiguous(src1);
 
+    // If src0 is BF16, try to use a BF16 x BF16 multiply
+    ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
     bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0;
@@ -4488,25 +4653,25 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
 
     if (mmp == nullptr) {
         // Fall back to f16 dequant mul mat
-        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]);
         quantize_y = false;
     }
 
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
-    const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig);
+    const bool qy_needs_dequant = !quantize_y && ((src1->type != f16_type && !y_f32_kernel) || y_non_contig);
 
     if (qx_needs_dequant) {
         // Fall back to dequant + f16 mulmat
-        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]);
     }
 
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
 
-    const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)));
+    const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)));
     const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && ne11 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type));
+    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type));
 
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11;
@@ -4527,12 +4692,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     vk_pipeline to_q8_1 = nullptr;
 
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -4949,6 +5114,8 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t nb01 = src0->nb[1];
     const uint64_t nb02 = src0->nb[2];
 
+    const uint64_t nb12 = src1->nb[2];
+
     // const uint64_t ne10 = src1->ne[0];
     const uint64_t ne11 = src1->ne[1];
     const uint64_t ne12 = src1->ne[2];
@@ -4974,6 +5141,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
+    const uint32_t channel_stride_y = nb12 / sizeof(float);
 
     const uint64_t qx_sz = ggml_nbytes(src0);
     const uint64_t qy_sz = ggml_nbytes(src1);
@@ -5004,7 +5172,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
 
     // compute
-    const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+    const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
         { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
@@ -5029,7 +5197,7 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
     // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four)
     // when ne12 and ne13 are one.
     } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) &&
-               (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+               (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
     } else {
         ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun);
@@ -5097,27 +5265,31 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) ||
                               !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
+                              (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) ||
                               !ggml_vk_dim01_contiguous(src1);
 
+    // If src0 is BF16, try to use a BF16 x BF16 multiply
+    ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
+    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]);
 
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
-    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
+    const bool qy_needs_dequant = (src1->type != f16_type && !y_f32_kernel) || y_non_contig;
 
     if (qx_needs_dequant) {
         // Fall back to dequant + f16 mulmat
-        mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]);
     }
 
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
 
-    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? GGML_TYPE_F16 : src0->type));
+    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? f16_type : src0->type));
     const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type);
+    vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type);
 
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
@@ -5136,12 +5308,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     vk_pipeline to_fp16_vk_1 = nullptr;
 
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -5988,6 +6160,15 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_leaky_relu_f32;
         }
         return nullptr;
+    case GGML_OP_CONV_2D_DW:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            if (ggml_is_contiguous(src1)) {
+                return ctx->device->pipeline_conv2d_dw_whcn_f32;
+            } else if (ggml_is_contiguous_channels(src1)) {
+                return ctx->device->pipeline_conv2d_dw_cwhn_f32;
+            }
+        }
+        return nullptr;
     default:
         return nullptr;
     }
@@ -6014,6 +6195,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_ROPE:
     case GGML_OP_RMS_NORM:
+    case GGML_OP_CONV_2D_DW:
         return true;
     default:
         return false;
@@ -6310,6 +6492,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
     case GGML_OP_UNARY:
+    case GGML_OP_CONV_2D_DW:
         {
             const uint32_t ne = ggml_nelements(dst);
             if (ne > 262144) {
@@ -7096,6 +7279,30 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c
     }, dryrun);
 }
 
+static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    vk_op_conv2d_dw_push_constants p{};
+    p.ne = ggml_nelements(dst);
+    p.channels = dst->ne[2];
+    p.batches = dst->ne[3];
+    p.dst_w = dst->ne[0];
+    p.dst_h = dst->ne[1];
+    p.src_w = src1->ne[0];
+    p.src_h = src1->ne[1];
+    p.knl_w = src0->ne[0];
+    p.knl_h = src0->ne[1];
+    p.stride_x = dst->op_params[0];
+    p.stride_y = dst->op_params[1];
+    p.pad_x = dst->op_params[2];
+    p.pad_y = dst->op_params[3];
+    p.dilation_x = dst->op_params[4];
+    p.dilation_y = dst->op_params[5];
+
+    GGML_ASSERT(src0->ne[3] == p.channels);
+    GGML_ASSERT(src1->ne[3] == p.batches);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun);
+}
+
 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const float * op_params = (const float *)dst->op_params;
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
@@ -8116,6 +8323,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_POOL_2D:
+    case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
     case GGML_OP_RWKV_WKV7:
     case GGML_OP_LEAKY_RELU:
@@ -8179,6 +8387,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_POOL_2D:
+        case GGML_OP_CONV_2D_DW:
         case GGML_OP_LEAKY_RELU:
             {
                 // These operations all go through ggml_vk_op_f32, so short-circuit and
@@ -8352,6 +8561,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_POOL_2D:
         ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_CONV_2D_DW:
+        ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_LEAKY_RELU:
         ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun);
@@ -8473,6 +8686,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_POOL_2D:
+    case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
     case GGML_OP_RWKV_WKV7:
     case GGML_OP_LEAKY_RELU:
@@ -9227,6 +9441,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 switch (src0_type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -9262,10 +9477,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (a->ne[3] != b->ne[3]) {
                     return false;
                 }
-                if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) ||
+                if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16) ||
                     !(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) {
                     return false;
                 }
+                if (op->src[0]->type == GGML_TYPE_BF16 && op->src[1]->type == GGML_TYPE_F16) {
+                    // We currently don't have a bf16 x f16 shader, or an fp16->bf16 copy shader.
+                    // So don't support this combination for now.
+                    return false;
+                }
 
                 return true;
             } break;
@@ -9338,6 +9558,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 switch (op->src[0]->type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -9368,6 +9589,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     switch (src1_type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -9442,6 +9664,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_CONV_2D_DW:
         case GGML_OP_POOL_2D:
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:

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

@@ -12,6 +12,9 @@ endif()
 if (GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
 endif()
+if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+    add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+endif()
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)

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

@@ -18,7 +18,11 @@ void main() {
     // fast path for when all four iterations are in-bounds
     if (idx + (num_iter-1)*num_threads < p.ne) {
         [[unroll]] for (uint i = 0; i < num_iter; ++i) {
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+
+#if defined(DATA_D_BF16)
+            float f = float(data_a[get_aoffset() + idx]);
+            data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
             data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
             data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
@@ -31,7 +35,10 @@ void main() {
                 continue;
             }
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+#if defined(DATA_D_BF16)
+            float f = float(data_a[get_aoffset() + idx]);
+            data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
             data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
             data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];

ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp

@@ -0,0 +1,105 @@
+#version 450
+
+#include "types.comp"
+
+layout (push_constant) uniform parameter
+{
+    uint ne;
+    uint batches;
+    uint channels;
+    uint dst_w;
+    uint dst_h;
+    uint src_w;
+    uint src_h;
+    uint knl_w;
+    uint knl_h;
+    int stride_x;
+    int stride_y;
+    int pad_x;
+    int pad_y;
+    int dilation_x;
+    int dilation_y;
+} p;
+
+layout (binding = 0) readonly buffer A {A_TYPE knl_data[];};
+layout (binding = 1) readonly buffer B {B_TYPE src_data[];};
+layout (binding = 2) writeonly buffer D {D_TYPE dst_data[];};
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE conv_2d_dw_whcn(uint idx) {
+    uint i0 = idx / p.dst_w;
+    uint dst_x = idx - i0 * p.dst_w;
+    uint i1 = i0 / p.dst_h;
+    uint dst_y = i0 - i1 * p.dst_h;
+    uint n = i1 / p.channels;
+    uint c = i1 - n * p.channels;
+
+    uint src_i = n * p.channels * p.src_h * p.src_w + c * p.src_h * p.src_w;
+    uint knl_i = c * p.knl_h * p.knl_w;
+
+    FLOAT_TYPE sum = 0.0;
+    for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+        uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
+        if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int
+            continue;
+        }
+        for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+            uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
+            if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int
+                continue;
+            }
+            FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * p.src_w + src_x]);
+            FLOAT_TYPE k = FLOAT_TYPE(knl_data[knl_i + knl_y * p.knl_w + knl_x]);
+            sum = fma(v, k, sum);
+        }
+    }
+    return sum;
+}
+
+FLOAT_TYPE conv_2d_dw_cwhn(uint idx) {
+    uint i0 = idx / p.channels;
+    uint c = idx - i0 * p.channels;
+    uint i1 = i0 / p.dst_w;
+    uint dst_x = i0 - i1 * p.dst_w;
+    uint n = i1 / p.dst_h;
+    uint dst_y = i1 - n * p.dst_h;
+
+    uint src_i = n * p.channels * p.src_h * p.src_w;
+    uint src_row = p.src_w * p.channels;
+    uint knl_row = p.knl_w * p.channels;
+
+    FLOAT_TYPE sum = 0.0;
+    for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+        uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
+        if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int
+            continue;
+        }
+        for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+            uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
+            if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int
+                continue;
+            }
+            FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * src_row + src_x * p.channels + c]);
+            FLOAT_TYPE k = FLOAT_TYPE(knl_data[        knl_y * knl_row + knl_x * p.channels + c]);
+            sum = fma(v, k, sum);
+        }
+    }
+    return sum;
+}
+
+void main() {
+    uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+    if (idx >= p.ne) {
+        return;
+    }
+
+    FLOAT_TYPE result =
+#ifdef WHCN
+        conv_2d_dw_whcn(idx);
+#else
+        conv_2d_dw_cwhn(idx);
+#endif
+    dst_data[idx] = D_TYPE(result);
+}
+

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

@@ -12,7 +12,10 @@ void main() {
         return;
     }
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+#if defined(DATA_D_BF16)
+    float f = float(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
     data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
 #else
     data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)];

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

@@ -23,6 +23,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_BF16)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(bf16_to_fp32(data_a[a_offset + ib]), bf16_to_fp32(data_a[a_offset + ib + 1]));
+}
+#endif
+
 #if defined(DATA_A_Q4_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
@@ -428,7 +434,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
-#if defined(DATA_A_F32) || defined(DATA_A_F16)
+#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(0, 0);
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -482,7 +482,7 @@ float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCo
     const uint ib8 = (idx & 0x18) >> 3;  // 0..3
     const uint iqs = 8 * ib32 + ib8;
 
-    const uint8_t qs = bl.block.qs[iqs];
+    const uint qs = bl.block.qs[iqs];
     const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3]));
 
     const float dscale = float(bl.block.d) * 0.25 * (0.5 + float(signscale >> 28));

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

@@ -20,9 +20,14 @@ void main() {
     const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
     const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);
+#if defined(DATA_A_BF16)
+    FLOAT_TYPE v = FLOAT_TYPE(bf16_to_fp32(data_a[a_offset + i00]));
 #else
-    data_d[d_offset + i00] = data_a[a_offset + i00];
+    FLOAT_TYPE v = FLOAT_TYPE(data_a[a_offset + i00]);
+#endif
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[d_offset + i00] = D_TYPE(v);
+#else
+    data_d[d_offset + i00] = D_TYPE(v);
 #endif
 }

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

@@ -6,7 +6,7 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16)
 #define K_PER_ITER 8
 #else
 #define K_PER_ITER 2

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

@@ -21,7 +21,9 @@ layout (push_constant) uniform parameter
     uint nrows_x;
     uint row_stride_x;
     uint channel_stride_x;
+    uint channel_stride_y;
     uint channel_x_divisor;
+    uint ne12;
     uint b_offset;
     uint d_offset;
 } p;
@@ -33,6 +35,7 @@ void main() {
     const uint row_x     = gl_GlobalInvocationID.y;
     const uint channel   = gl_GlobalInvocationID.z;
     const uint channel_x = channel / p.channel_x_divisor;
+    const uint channel_y = channel % p.ne12;
 
     const uint nrows_y   = p.ncols_x;
     const uint nrows_dst = p.nrows_x;
@@ -56,7 +59,7 @@ void main() {
                 const uint row_y = col_x;
 
                 const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-                const uint iy = channel*nrows_y + row_y;
+                const uint iy = channel_y*p.channel_stride_y + row_y;
 
                 const vec4 av4 = vec4(data_a_v4[ix / 4]);
                 const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -72,7 +75,7 @@ void main() {
             const uint row_y = col_x;
 
             const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel*nrows_y + row_y;
+            const uint iy = channel_y*p.channel_stride_y + row_y;
 
             const vec4 av4 = vec4(data_a_v4[ix / 4]);
             const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -89,7 +92,7 @@ void main() {
             const uint row_y = col_x;
 
             const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel*nrows_y + row_y;
+            const uint iy = channel_y*p.channel_stride_y + row_y;
 
             const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
 

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

@@ -10,6 +10,10 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
 
+#if defined(DATA_A_BF16) && defined(COOPMAT)
+#extension GL_EXT_bfloat16 : enable
+#endif
+
 #ifdef COOPMAT
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_KHR_memory_scope_semantics : enable
@@ -29,6 +33,10 @@
 #define LOAD_VEC_B 1
 #endif
 
+#if !defined(TO_FLOAT_TYPE)
+#define TO_FLOAT_TYPE FLOAT_TYPE
+#endif
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -202,8 +210,8 @@ void main() {
 #endif
 
 #ifdef COOPMAT
-    coopmat<float16_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
-    coopmat<float16_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
+    coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
+    coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
     coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
 
     [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
@@ -248,6 +256,21 @@ void main() {
                 buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(0.0f);
             }
 #endif
+#elif defined(DATA_A_BF16)
+#if LOAD_VEC_A == 4
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+            buf_a[buf_idx    ] = TO_FLOAT_TYPE(data_a[idx].x);
+            buf_a[buf_idx + 1] = TO_FLOAT_TYPE(data_a[idx].y);
+            buf_a[buf_idx + 2] = TO_FLOAT_TYPE(data_a[idx].z);
+            buf_a[buf_idx + 3] = TO_FLOAT_TYPE(data_a[idx].w);
+#else
+            if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) {
+                buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
+            } else {
+                buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(uint16_t(0));
+            }
+#endif
 #elif defined(DATA_A_Q4_0)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
             const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a;
@@ -695,13 +718,13 @@ void main() {
             const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
 #endif
             const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B;
-            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x);
-            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y);
-            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z);
-            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w);
+            buf_b[buf_idx + 0] = TO_FLOAT_TYPE(data_b[idx].x);
+            buf_b[buf_idx + 1] = TO_FLOAT_TYPE(data_b[idx].y);
+            buf_b[buf_idx + 2] = TO_FLOAT_TYPE(data_b[idx].z);
+            buf_b[buf_idx + 3] = TO_FLOAT_TYPE(data_b[idx].w);
 #elif !MUL_MAT_ID
             if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) {
-                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
             } else {
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
             }
@@ -709,7 +732,7 @@ void main() {
             const uint row_i = ic * BN + loadc_b + l;
             if (row_i < _ne1) {
                 const u16vec2 row_idx = row_ids[row_i];
-                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
             } else {
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
             }

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

@@ -14,6 +14,9 @@
 #extension GL_EXT_buffer_reference : enable
 #extension GL_KHR_shader_subgroup_ballot : enable
 #extension GL_KHR_shader_subgroup_vote : enable
+#ifdef DATA_A_BF16
+#extension GL_EXT_bfloat16 : enable
+#endif
 
 #include "types.comp"
 
@@ -80,6 +83,12 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #define store_scales(a)
 #endif
 
+#if defined(DATA_A_BF16)
+#define MAT_TYPE bfloat16_t
+#else
+#define MAT_TYPE FLOAT_TYPE
+#endif
+
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
 
@@ -271,8 +280,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose);
@@ -286,8 +295,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose);
@@ -310,8 +319,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose);
@@ -325,8 +334,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose);
@@ -350,8 +359,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose);
@@ -365,8 +374,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose);
@@ -405,8 +414,8 @@ void main() {
                 fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
             }
 
-            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
             coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID

ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp

@@ -0,0 +1,7 @@
+#version 460
+
+#extension GL_EXT_bfloat16 : require
+
+void main()
+{
+}

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

@@ -33,6 +33,19 @@
 #endif
 #endif
 
+#if defined(DATA_A_BF16)
+#define QUANT_K 1
+#define QUANT_R 1
+
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
+#define A_TYPE uint16_t
+#elif LOAD_VEC_A == 4
+#define A_TYPE u16vec4
+#elif LOAD_VEC_A == 8
+#error unsupported
+#endif
+#endif
+
 #define QUANT_K_Q4_0 32
 #define QUANT_R_Q4_0 2
 
@@ -1343,4 +1356,18 @@ void init_iq_shmem(uvec3 wgsize)
 }
 #endif
 
+// returns the bfloat value in the low 16b.
+// See ggml_compute_fp32_to_bf16
+uint32_t fp32_to_bf16(float f)
+{
+    uint32_t u = floatBitsToUint(f);
+    u = (u + (0x7fff + ((u >> 16) & 1))) >> 16;
+    return u;
+}
+
+float bf16_to_fp32(uint32_t u)
+{
+    return uintBitsToFloat(u << 16);
+}
+
 #endif // !defined(GGML_TYPES_COMP)

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

@@ -63,7 +63,8 @@ const std::vector<std::string> type_names = {
     "iq3_xxs",
     "iq3_s",
     "iq4_xs",
-    "iq4_nl"
+    "iq4_nl",
+    "bf16",
 };
 
 namespace {
@@ -296,7 +297,6 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
     std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
 
     std::map<std::string, std::string> base_dict = {
-        {"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"},
         {"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"},
     };
     std::string shader_name = "matmul";
@@ -318,12 +318,45 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
 
     const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
 
-    // Shaders with f16 B_TYPE
-    string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    auto const &FLOAT_TYPE = [&](const std::string &t) -> std::string {
+        if (t == "bf16") {
+            // scalar path promotes to float
+            if (!coopmat && !coopmat2) {
+                return "float";
+            }
+            return "bfloat16_t";
+        }
+        if (coopmat2 || fp16) {
+            return "float16_t";
+        }
+        return "float";
+    };
 
-    string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+    // Shaders with f16 B_TYPE
+    string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+
+    string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+
+    // bf16
+    {
+        std::string load_vec_a_unaligned = "1";
+        // For aligned matmul loads
+        std::string load_vec_a = coopmat2 ? "1" : "4";
+
+        // scalar path promotes to float
+        std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32";
+
+        // If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader
+#if !defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (!(coopmat || coopmat2))
+#endif
+        {
+            string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},   {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_bf16",         source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                      {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"},                          {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
+        }
+    }
 
     for (const auto& tname : type_names) {
         std::string load_vec_quant = "2";
@@ -332,26 +365,30 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
         else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl"))
             load_vec_quant = "4";
 
+        if (tname == "bf16") {
+            continue;
+        }
+
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
         // For unaligned, load one at a time for f32/f16, or two at a time for quants
-        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : load_vec_quant;
+        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant;
         // For aligned matmul loads
-        std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : load_vec_quant;
+        std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
 
         // don't generate f32 variants for coopmat2
         if (!coopmat2) {
-            string_to_spv(shader_name + "_" + tname + "_f32",         source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32",         source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
         if (tname != "f16" && tname != "f32") {
-            string_to_spv(shader_name + "_" + tname + "_f16",         source_name,  merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name,  merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16",         source_name,  merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name,  merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
         if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
-            string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
         }
 #endif
     }
@@ -393,6 +430,7 @@ void process_shaders() {
             if (tname == "f32") {
                 continue;
             }
+            if (tname == "bf16") continue;
 
             if (tname == "f16") {
                 string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
@@ -417,12 +455,12 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
-        if (tname != "f16") {
+        if (tname != "f16" && tname != "bf16") {
             string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
         }
 
         if (!string_ends_with(tname, "_k")) {
-            shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp";
+            shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp";
 
             if (tname == "f16") {
                 string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}));
@@ -447,9 +485,11 @@ void process_shaders() {
     string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
     string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
     string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
     string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
 
     for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
         string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
@@ -544,6 +584,9 @@ void process_shaders() {
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
+    string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
+    string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
+
     for (auto &c : compiles) {
         c.wait();
     }

ggml/src/ggml.c

@@ -5776,10 +5776,9 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
 }
 
 void ggml_build_backward_expand(
-        struct ggml_context * ctx_static,
-        struct ggml_context * ctx_compute,
-        struct ggml_cgraph  * cgraph,
-        bool                  accumulate) {
+        struct ggml_context *  ctx,
+        struct ggml_cgraph  *  cgraph,
+        struct ggml_tensor  ** grad_accs) {
     GGML_ASSERT(cgraph->n_nodes > 0);
     GGML_ASSERT(cgraph->grads);
     GGML_ASSERT(cgraph->grad_accs);
@@ -5852,21 +5851,24 @@ void ggml_build_backward_expand(
         GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW ||
             node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
 
-        const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
-        GGML_ASSERT(igrad != GGML_HASHSET_FULL);
-        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad));
-        if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
-            cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node);
-            cgraph->grads[igrad]     = cgraph->grad_accs[igrad];
-            ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name);
+        const size_t ihash = ggml_hash_find(&cgraph->visited_hash_set, node);
+        GGML_ASSERT(ihash != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, ihash));
+        if (grad_accs && grad_accs[i]) {
+            cgraph->grad_accs[ihash] = grad_accs[i];
+            cgraph->grads[ihash]     = cgraph->grad_accs[ihash];
+        } else if (node->flags & GGML_TENSOR_FLAG_LOSS) {
+            // loss tensors always need a gradient accumulator
+            cgraph->grad_accs[ihash] = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+            cgraph->grads[ihash]     = cgraph->grad_accs[ihash];
         }
-        grads_needed[igrad] = true;
+        grads_needed[ihash] = true;
     }
 
     for (int i = n_nodes_f - 1; i >= 0; --i) {
         // inplace operations to add gradients are not created by ggml_compute_backward except for gradient accumulation
         // use allocator to automatically make inplace operations
-        ggml_compute_backward(ctx_compute, cgraph, i, grads_needed);
+        ggml_compute_backward(ctx, cgraph, i, grads_needed);
     }
 
     free(grads_needed);
@@ -6012,8 +6014,8 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
     }
 }
 
-struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) {
-    struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads != NULL);
+struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads) {
+    struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads || force_grads);
     ggml_graph_cpy(cgraph, result);
     return result;
 }
@@ -6341,8 +6343,8 @@ void ggml_set_output(struct ggml_tensor * tensor) {
     tensor->flags |= GGML_TENSOR_FLAG_OUTPUT;
 }
 
-void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor) {
-    GGML_UNUSED(ctx); // TODO: remove this parameter
+void ggml_set_param(struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->op == GGML_OP_NONE);
     tensor->flags |= GGML_TENSOR_FLAG_PARAM;
 }
 

gguf-py/gguf/constants.py

@@ -104,6 +104,7 @@ class Keys:
         EXPERT_WEIGHTS_SCALE              = "{arch}.expert_weights_scale"
         EXPERT_WEIGHTS_NORM               = "{arch}.expert_weights_norm"
         EXPERT_GATING_FUNC                = "{arch}.expert_gating_func"
+        MOE_EVERY_N_LAYERS                = "{arch}.moe_every_n_layers"
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
@@ -230,8 +231,10 @@ class Keys:
         BLOCK_COUNT         = "clip.vision.block_count"
         IMAGE_MEAN          = "clip.vision.image_mean"
         IMAGE_STD           = "clip.vision.image_std"
+        SPATIAL_MERGE_SIZE  = "clip.vision.spatial_merge_size"
         USE_GELU            = "clip.use_gelu"
         USE_SILU            = "clip.use_silu"
+        N_WA_PATTERN        = "clip.vision.n_wa_pattern" # used by qwen2.5vl
 
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
@@ -267,6 +270,7 @@ class MODEL_ARCH(IntEnum):
     REFACT           = auto()
     BERT             = auto()
     NOMIC_BERT       = auto()
+    NOMIC_BERT_MOE   = auto()
     JINA_BERT_V2     = auto()
     BLOOM            = auto()
     STABLELM         = auto()
@@ -489,6 +493,7 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_FFN_DOWN       = auto()
     V_PRE_NORM           = auto()
     V_POST_NORM          = auto()
+    V_MM_INP_NORM        = auto()
     V_MM_INP_PROJ        = auto() # gemma3
     V_MM_SOFT_EMB_NORM   = auto() # gemma3
     V_RESMPL_POS_EMBD_K  = auto() # minicpmv
@@ -503,6 +508,7 @@ class MODEL_TENSOR(IntEnum):
     V_RESMPL_PROJ        = auto() # minicpmv
     V_RESMPL_QUERY       = auto() # minicpmv
     V_TOK_EMBD_IMG_BREAK = auto() # pixtral
+    V_MM_PATCH_MERGER    = auto() # mistral small 3.1
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -521,6 +527,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.REFACT:           "refact",
     MODEL_ARCH.BERT:             "bert",
     MODEL_ARCH.NOMIC_BERT:       "nomic-bert",
+    MODEL_ARCH.NOMIC_BERT_MOE:   "nomic-bert-moe",
     MODEL_ARCH.JINA_BERT_V2:     "jina-bert-v2",
     MODEL_ARCH.BLOOM:            "bloom",
     MODEL_ARCH.STABLELM:         "stablelm",
@@ -744,6 +751,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
     MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
     MODEL_TENSOR.V_MM_INP_PROJ:             "mm.input_projection",
+    MODEL_TENSOR.V_MM_INP_NORM:             "mm.input_norm",
     MODEL_TENSOR.V_MM_SOFT_EMB_NORM:        "mm.soft_emb_norm",
     MODEL_TENSOR.V_RESMPL_POS_EMBD_K:       "resampler.pos_embd_k",
     MODEL_TENSOR.V_RESMPL_ATTN_Q:           "resampler.attn.q",
@@ -757,6 +765,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_RESMPL_PROJ:             "resampler.proj",
     MODEL_TENSOR.V_RESMPL_QUERY:            "resampler.query",
     MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK:      "v.token_embd.img_break", # pixtral
+    MODEL_TENSOR.V_MM_PATCH_MERGER:         "mm.patch_merger", # mistral small 3.1
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -780,6 +789,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_PRE_NORM,
         MODEL_TENSOR.V_POST_NORM,
         MODEL_TENSOR.V_MM_INP_PROJ,
+        MODEL_TENSOR.V_MM_INP_NORM,
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM,
         MODEL_TENSOR.V_RESMPL_POS_EMBD_K,
         MODEL_TENSOR.V_RESMPL_ATTN_Q,
@@ -793,6 +803,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_RESMPL_PROJ,
         MODEL_TENSOR.V_RESMPL_QUERY,
         MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
+        MODEL_TENSOR.V_MM_PATCH_MERGER,
     ],
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -960,6 +971,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.NOMIC_BERT_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
     MODEL_ARCH.JINA_BERT_V2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
@@ -2006,6 +2033,8 @@ class PoolingType(IntEnum):
     NONE = 0
     MEAN = 1
     CLS  = 2
+    LAST = 3
+    RANK = 4
 
 
 class GGMLQuantizationType(IntEnum):
@@ -2136,6 +2165,8 @@ class VisionProjectorType:
     GEMMA3 = "gemma3"
     IDEFICS3 = "idefics3"
     PIXTRAL = "pixtral"
+    QWEN2VL = "qwen2vl_merger"
+    QWEN25VL = "qwen2.5vl_merger"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/gguf_writer.py

@@ -728,6 +728,9 @@ class GGUFWriter:
     def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None:
         self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value)
 
+    def add_moe_every_n_layers(self, value: int) -> None:
+        self.add_uint32(Keys.LLM.MOE_EVERY_N_LAYERS.format(arch=self.arch), value)
+
     def add_swin_norm(self, value: bool) -> None:
         self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
 
@@ -969,6 +972,9 @@ class GGUFWriter:
     def add_vision_image_std(self, values: Sequence[float]) -> None:
         self.add_array(Keys.ClipVision.IMAGE_STD, values)
 
+    def add_vision_spatial_merge_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.SPATIAL_MERGE_SIZE, value)
+
     def add_vision_use_gelu(self, value: bool) -> None:
         self.add_bool(Keys.ClipVision.USE_GELU, value)
 
@@ -978,6 +984,9 @@ class GGUFWriter:
     def add_vision_projector_scale_factor(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value)
 
+    def add_vision_n_wa_pattern(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)
+
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:

gguf-py/gguf/tensor_mapping.py

@@ -290,6 +290,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
             "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
             "language_model.model.layers.{bid}.feed_forward.router", # llama4
+            "encoder.layers.{bid}.mlp.router.layer",            # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
@@ -322,6 +323,7 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
             "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
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
@@ -337,6 +339,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.experts.up_proj",         # qwen2moe olmoe (merged)
             "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
             "language_model.model.layers.{bid}.feed_forward.experts.up_proj", # llama4
+            "encoder.layers.{bid}.mlp.experts.mlp.w1",        # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_UP_SHEXP: (
@@ -418,6 +421,7 @@ class TensorNameMap:
             "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
             "model.layers.{bid}.block_sparse_moe.experts.w2",    # phimoe (merged)
             "language_model.model.layers.{bid}.feed_forward.experts.down_proj", # llama4
+            "encoder.layers.{bid}.mlp.experts.mlp.w2",           # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_DOWN_SHEXP: (
@@ -892,6 +896,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MMPROJ: (
             "multi_modal_projector.linear_{bid}",
+            "visual.merger.mlp.{bid}", # qwen2vl
         ),
 
         MODEL_TENSOR.V_MMPROJ_FC: (
@@ -915,6 +920,7 @@ class TensorNameMap:
             "vpm.embeddings.patch_embedding",
             "model.vision_model.embeddings.patch_embedding", # SmolVLM
             "vision_tower.patch_conv", # pixtral
+            "visual.patch_embed.proj", # qwen2vl
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_POS: (
@@ -928,6 +934,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.q_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral
+            "visual.blocks.{bid}.attn.q", # qwen2vl, generated
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_K: (
@@ -935,6 +942,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.k_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral
+            "visual.blocks.{bid}.attn.k", # qwen2vl, generated
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_V: (
@@ -942,6 +950,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.v_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral
+            "visual.blocks.{bid}.attn.v", # qwen2vl, generated
         ),
 
         MODEL_TENSOR.V_ENC_INPUT_NORM: (
@@ -949,6 +958,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.layer_norm1",
             "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral
+            "visual.blocks.{bid}.norm1", # qwen2vl
         ),
 
         MODEL_TENSOR.V_ENC_OUTPUT: (
@@ -956,6 +966,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.out_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral
+            "visual.blocks.{bid}.attn.proj", # qwen2vl
         ),
 
         MODEL_TENSOR.V_ENC_OUTPUT_NORM: (
@@ -963,17 +974,24 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.layer_norm2",
             "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM
             "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral
+            "visual.blocks.{bid}.norm2", # qwen2vl
         ),
 
+        # some namings are messed up because the original llava code swapped fc1 and fc2
+        # we have no better way to fix it, just be careful
+        # new models like pixtral use the correct naming
         MODEL_TENSOR.V_ENC_FFN_UP: (
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1",
             "vpm.encoder.layers.{bid}.mlp.fc1",
             "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped)
             "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
+            "visual.blocks.{bid}.mlp.fc2", # qwen2vl
+            "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl
         ),
 
         MODEL_TENSOR.V_ENC_FFN_GATE: (
             "vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral
+            "visual.blocks.{bid}.mlp.gate_proj", # qwen2.5vl
         ),
 
         MODEL_TENSOR.V_ENC_FFN_DOWN: (
@@ -981,6 +999,8 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.mlp.fc2",
             "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped)
             "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
+            "visual.blocks.{bid}.mlp.fc1", # qwen2vl
+            "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
         ),
 
         MODEL_TENSOR.V_PRE_NORM: (
@@ -991,12 +1011,17 @@ class TensorNameMap:
         MODEL_TENSOR.V_POST_NORM: (
             "vision_tower.vision_model.post_layernorm",
             "model.vision_model.post_layernorm", # SmolVLM
+            "visual.merger.ln_q", # qwen2vl
         ),
 
         MODEL_TENSOR.V_MM_INP_PROJ: (
             "multi_modal_projector.mm_input_projection",
         ),
 
+        MODEL_TENSOR.V_MM_INP_NORM: (
+            "multi_modal_projector.norm",
+        ),
+
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
             "multi_modal_projector.mm_soft_emb_norm",
         ),
@@ -1048,6 +1073,10 @@ class TensorNameMap:
         MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: (
             "v.token_embd.img_break", # for pixtral, this is a generated vector
         ),
+
+        MODEL_TENSOR.V_MM_PATCH_MERGER: (
+            "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1
+        ),
     }
 
     # architecture-specific block mappings

include/llama.h

@@ -4,6 +4,7 @@
 #include "ggml.h"
 #include "ggml-cpu.h"
 #include "ggml-backend.h"
+#include "ggml-opt.h"
 
 #include <stddef.h>
 #include <stdint.h>
@@ -445,6 +446,10 @@ extern "C" {
                                  size_t    n_paths,
               struct llama_model_params    params);
 
+    LLAMA_API void llama_model_save_to_file(
+            const struct llama_model * model,
+                        const char * path_model);
+
     DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model),
             "use llama_model_free instead");
 
@@ -924,7 +929,7 @@ extern "C" {
     // Frees a batch of tokens allocated with llama_batch_init()
     LLAMA_API void llama_batch_free(struct llama_batch batch);
 
-    // Processes a batch of tokens with the ecoder part of the encoder-decoder model.
+    // Processes a batch of tokens with the encoder part of the encoder-decoder model.
     // Stores the encoder output internally for later use by the decoder cross-attention layers.
     //   0 - success
     // < 0 - error. the KV cache state is restored to the state before this call
@@ -932,7 +937,7 @@ extern "C" {
             struct llama_context * ctx,
               struct llama_batch   batch);
 
-    // Positive return values does not mean a fatal error, but rather a warning.
+    // A positive return value does not mean a fatal error, but rather a warning.
     //   0 - success
     //   1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
     // < 0 - error. the KV cache state is restored to the state before this call
@@ -1232,6 +1237,7 @@ extern "C" {
         "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    /// Setting k <= 0 makes this a noop
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
@@ -1427,6 +1433,37 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
+    //
+    // training
+    //
+
+    // function that returns whether or not a given tensor is a trainable parameter
+    typedef bool (*llama_opt_param_filter)(const struct ggml_tensor * tensor, void * userdata);
+
+    // always returns true
+    LLAMA_API bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata);
+
+    struct llama_opt_params {
+        uint32_t n_ctx_train; // assumed context size post training, use context size specified in llama_context if 0
+
+        llama_opt_param_filter param_filter; // callback for determining which tensors are trainable parameters
+        void * param_filter_ud;              // userdata for determining which tensors are trainable parameters
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);
+
+    LLAMA_API void llama_opt_epoch(
+            struct llama_context    * lctx,
+            ggml_opt_dataset_t        dataset,
+            ggml_opt_result_t         result_train,
+            ggml_opt_result_t         result_eval,
+            int64_t                   idata_split,
+            ggml_opt_epoch_callback   callback_train,
+            ggml_opt_epoch_callback   callback_eval);
+
 #ifdef __cplusplus
 }
 #endif

scripts/compare-llama-bench.py

@@ -19,9 +19,9 @@ logger = logging.getLogger("compare-llama-bench")
 
 # Properties by which to differentiate results per commit:
 KEY_PROPERTIES = [
-    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "model_filename", "model_type", "n_batch", "n_ubatch",
-    "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v", "use_mmap", "no_kv_offload",
-    "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen"
+    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "tensor_buft_overrides", "model_filename", "model_type",
+    "n_batch", "n_ubatch", "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v",
+    "use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen", "n_depth"
 ]
 
 # Properties that are boolean and are converted to Yes/No for the table:
@@ -30,11 +30,11 @@ BOOL_PROPERTIES = ["embeddings", "cpu_strict", "use_mmap", "no_kv_offload", "fla
 # Header names for the table:
 PRETTY_NAMES = {
     "cpu_info": "CPU", "gpu_info": "GPU", "backends": "Backends", "n_gpu_layers": "GPU layers",
-    "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]",
-    "model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size",
-    "embeddings": "Embeddings", "cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll",
-    "n_threads": "Threads", "type_k": "K type", "type_v": "V type", "split_mode": "Split mode", "main_gpu": "Main GPU",
-    "no_kv_offload": "NKVO", "flash_attn": "FlashAttention", "tensor_split": "Tensor split", "use_mmap": "Use mmap",
+    "tensor_buft_overrides": "Tensor overrides", "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]",
+    "model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size", "embeddings": "Embeddings",
+    "cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll", "n_threads": "Threads", "type_k": "K type", "type_v": "V type",
+    "use_mmap": "Use mmap", "no_kv_offload": "NKVO", "split_mode": "Split mode", "main_gpu": "Main GPU", "tensor_split": "Tensor split",
+    "flash_attn": "FlashAttention",
 }
 
 DEFAULT_SHOW = ["model_type"]  # Always show these properties by default.
@@ -281,12 +281,12 @@ def get_rows(properties):
     The returned rows are unique in terms of property combinations.
     """
     select_string = ", ".join(
-        [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
+        [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "tb.n_depth", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
     equal_string = " AND ".join(
         [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [
             f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"]
     )
-    group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt"])
+    group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt", "tb.n_depth"])
     query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} "
              f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
     return cursor.execute(query).fetchall()
@@ -309,7 +309,7 @@ else:
     rows_full = get_rows(KEY_PROPERTIES)
     properties_different = []
     for i, kp_i in enumerate(KEY_PROPERTIES):
-        if kp_i in DEFAULT_SHOW or kp_i == "n_prompt" or kp_i == "n_gen":
+        if kp_i in DEFAULT_SHOW or kp_i in ["n_prompt", "n_gen", "n_depth"]:
             continue
         for row_full in rows_full:
             if row_full[i] != rows_full[0][i]:
@@ -340,17 +340,20 @@ else:
 
 table = []
 for row in rows_show:
-    n_prompt = int(row[-4])
-    n_gen    = int(row[-3])
+    n_prompt = int(row[-5])
+    n_gen    = int(row[-4])
+    n_depth  = int(row[-3])
     if n_prompt != 0 and n_gen == 0:
         test_name = f"pp{n_prompt}"
     elif n_prompt == 0 and n_gen != 0:
         test_name = f"tg{n_gen}"
     else:
         test_name = f"pp{n_prompt}+tg{n_gen}"
+    if n_depth != 0:
+        test_name = f"{test_name}@d{n_depth}"
     #           Regular columns    test name    avg t/s values              Speedup
     #            VVVVVVVVVVVVV     VVVVVVVVV    VVVVVVVVVVVVVV              VVVVVVV
-    table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
+    table.append(list(row[:-5]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
 
 # Some a-posteriori fixes to make the table contents prettier:
 for bool_property in BOOL_PROPERTIES:
@@ -376,7 +379,7 @@ if "gpu_info" in show:
         for gns in GPU_NAME_STRIP:
             row_table[ip] = row_table[ip].replace(gns, "")
 
-        gpu_names = row_table[ip].split("/")
+        gpu_names = row_table[ip].split(", ")
         num_gpus = len(gpu_names)
         all_names_the_same = len(set(gpu_names)) == 1
         if len(gpu_names) >= 2 and all_names_the_same:

scripts/sync-ggml.last

@@ -1 +1 @@
-13bcf9ce50651a8b4238ec6d136f46f2c1b23b6f
+0482de9c63b9134eb462c7732888c0ee0dbc2755

src/CMakeLists.txt

@@ -23,6 +23,7 @@ add_library(llama
             llama-memory.cpp
             llama-mmap.cpp
             llama-model-loader.cpp
+            llama-model-saver.cpp
             llama-model.cpp
             llama-quant.cpp
             llama-sampling.cpp

src/llama-arch.cpp

@@ -19,6 +19,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_REFACT,           "refact"           },
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
+    { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
@@ -106,6 +107,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
+    { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -472,6 +474,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_NOMIC_BERT_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {

src/llama-arch.h

@@ -23,6 +23,7 @@ enum llm_arch {
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -110,6 +111,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,

src/llama-batch.cpp

@@ -189,7 +189,7 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
     return ubatch;
 }
 
-void llama_sbatch::from_batch(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, bool logits_all) {
     GGML_ASSERT(batch.n_tokens >= 0);
     this->batch = &batch;
     this->n_embd = n_embd;
@@ -203,6 +203,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
     for (size_t i = 0; i < n_tokens; ++i) {
         ids[i] = i;
     }
+
     if (simple_split) {
         seq.resize(1);
         llama_sbatch_seq & s = seq[0];
@@ -212,6 +213,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         s.length = n_tokens;
         return;
     }
+
     std::sort(ids.begin(), ids.end(),
             [&batch](size_t a, size_t b) {
                 int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
@@ -239,6 +241,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
                 return n_seq_a > n_seq_b;
             }
     );
+
     // init seq
     llama_sbatch_seq * last_seq = nullptr;
 
@@ -262,6 +265,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         seq.push_back(new_seq);
         last_seq = &seq.back();
     }
+
     // keep shared prompts first at the end, then sort by length descending.
     std::sort(seq.begin(), seq.end(),
             [](llama_sbatch_seq & a, llama_sbatch_seq & b) {

src/llama-batch.h

@@ -70,7 +70,8 @@ struct llama_sbatch {
     // sequence-wise split
     llama_ubatch split_seq(size_t n_ubatch);
 
-    void from_batch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+    llama_sbatch() = default;
+    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
 };
 
 // temporary allocate memory for the input batch if needed

src/llama-chat.cpp

@@ -447,8 +447,16 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4 || tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
         ss << "[gMASK]" << "<sop>";
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n" << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
         for (auto message : chat) {
             std::string role(message->role);
             ss << "<|" << role << "|>" << "\n" << message->content;

src/llama-context.h

@@ -7,6 +7,7 @@
 #include "llama-adapter.h"
 
 #include "ggml-cpp.h"
+#include "ggml-opt.h"
 
 #include <map>
 #include <vector>
@@ -27,7 +28,12 @@ struct llama_context {
 
     void synchronize();
 
-    const llama_model & get_model() const;
+    const llama_model   & get_model()   const;
+    const llama_cparams & get_cparams() const;
+
+    ggml_backend_sched_t get_sched() const;
+
+    ggml_context * get_ctx_compute() const;
 
     uint32_t n_ctx()         const;
     uint32_t n_ctx_per_seq() const;
@@ -128,6 +134,32 @@ struct llama_context {
     llama_perf_context_data perf_get_data() const;
     void perf_reset();
 
+    //
+    // training
+    //
+
+    void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
+
+    void opt_epoch(
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,
+            ggml_opt_result_t       result_eval,
+            int64_t                 idata_split,
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    void opt_epoch_iter(
+            ggml_opt_dataset_t               dataset,
+            ggml_opt_result_t                result,
+            const std::vector<llama_token> & tokens,
+            const std::vector<llama_token> & labels_sparse,
+            llama_batch                    & batch,
+            ggml_opt_epoch_callback          callback,
+            bool                             train,
+            int64_t                          idata_in_loop,
+            int64_t                          ndata_in_loop,
+            int64_t                          t_loop_start);
+
 private:
     //
     // output
@@ -137,49 +169,30 @@ private:
     // Returns max number of outputs for which space was reserved.
     int32_t output_reserve(int32_t n_outputs);
 
-    // make the outputs have the same order they had in the user-provided batch
-    // TODO: maybe remove this
-    void output_reorder();
-
     //
     // graph
     //
 
+public:
     int32_t graph_max_nodes() const;
 
     // zero-out inputs and create the ctx_compute for the compute graph
     ggml_cgraph * graph_init();
 
-    llm_graph_result_ptr graph_build(
-            ggml_context * ctx,
-             ggml_cgraph * gf,
-      const llama_ubatch & ubatch,
-          llm_graph_type   gtype);
-
     // returns the result of ggml_backend_sched_graph_compute_async execution
     ggml_status graph_compute(
             ggml_cgraph * gf,
                    bool   batched);
 
+private:
+    llm_graph_result_ptr graph_build(
+            ggml_context * ctx,
+             ggml_cgraph * gf,
+      const llama_ubatch & ubatch,
+          llm_graph_type   gtype);
+
     llm_graph_cb graph_get_cb() const;
 
-    // used by kv_self_update()
-    ggml_tensor * build_rope_shift(
-        ggml_context * ctx0,
-        ggml_tensor * cur,
-        ggml_tensor * shift,
-        ggml_tensor * factors,
-              float   freq_base,
-              float   freq_scale) const;
-
-    llm_graph_result_ptr build_kv_self_shift(
-            ggml_context * ctx0,
-            ggml_cgraph * gf) const;
-
-    llm_graph_result_ptr build_kv_self_defrag(
-            ggml_context * ctx0,
-            ggml_cgraph * gf) const;
-
     // TODO: read/write lora adapters and cvec
     size_t state_write_data(llama_io_write_i & io);
     size_t state_read_data (llama_io_read_i  & io);
@@ -196,11 +209,10 @@ private:
     llama_cparams       cparams;
     llama_adapter_cvec  cvec;
     llama_adapter_loras loras;
-    llama_sbatch        sbatch;
 
     llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
 
-    std::unique_ptr<llama_kv_cache_unified> kv_self;
+    std::unique_ptr<llama_memory_i> memory;
 
     // TODO: remove
     bool logits_all = false;
@@ -230,6 +242,9 @@ private:
 
     ggml_context_ptr ctx_compute;
 
+    // training
+    ggml_opt_context_t opt_ctx = nullptr;
+
     ggml_threadpool_t threadpool       = nullptr;
     ggml_threadpool_t threadpool_batch = nullptr;
 

src/llama-graph.cpp

@@ -55,13 +55,16 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && pos) {
         const int64_t n_tokens = ubatch->n_tokens;
 
-        if (ubatch->token && n_pos_per_embd > 1) {
+        if (ubatch->token && n_pos_per_embd == 4) {
             // in case we're using M-RoPE with text tokens, convert the 1D positions to 4D
-            // the other dimensions are all 0, they are unused for text tokens
-            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd, 0);
+            // the 3 first dims are the same, and 4th dim is all 0
+            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd);
             // copy the first dimension
             for (int i = 0; i < n_tokens; ++i) {
-                pos_data[i] = ubatch->pos[i];
+                pos_data[               i] = ubatch->pos[i];
+                pos_data[    n_tokens + i] = ubatch->pos[i];
+                pos_data[2 * n_tokens + i] = ubatch->pos[i];
+                pos_data[3 * n_tokens + i] = 0; // 4th dim is 0
             }
             ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
         } else {
@@ -281,24 +284,7 @@ void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
 
         // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
         for (uint32_t i = 0; i < n_kv; ++i) {
-            const uint32_t  cell_id = i + kv_self->head;
-
-            //////////////////////////////////////////////
-            // TODO: this should not mutate the KV cache !
-            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
-
-            // prevent out-of-bound sources
-            if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) {
-                kv_cell.src = cell_id;
-            }
-
-            data[i] = kv_cell.src;
-
-            // TODO: do not mutate the KV cache
-            // ensure copy only happens once
-            if (kv_cell.src != (int32_t) cell_id) {
-                kv_cell.src = cell_id;
-            }
+            data[i] = kv_self->s_copy(i);
         }
     }
 }
@@ -314,18 +300,7 @@ void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
 
         // clear unused states
         for (int i = 0; i < n_kv; ++i) {
-            const uint32_t  cell_id = i + kv_self->head;
-
-            //////////////////////////////////////////////
-            // TODO: this should not mutate the KV cache !
-            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
-
-            data[i] = (float) (kv_cell.src >= 0);
-
-            // only clear once
-            if (kv_cell.src < 0) {
-                kv_cell.src = cell_id;
-            }
+            data[i] = kv_self->s_mask(i);
         }
     }
 }
@@ -925,28 +900,35 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
     cb(up, "ffn_moe_up", il);
 
-    ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(gate, "ffn_moe_gate", il);
+    ggml_tensor * experts = nullptr;
+    if (gate_exps) {
+        cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+        cb(cur, "ffn_moe_gate", il);
+    } else {
+        cur = up;
+    }
 
     switch (type_op) {
         case LLM_FFN_SILU:
             {
-                gate = ggml_silu(ctx0, gate);
-                cb(gate, "ffn_moe_silu", il);
+                cur = ggml_silu(ctx0, cur);
+                cb(cur, "ffn_moe_silu", il);
             } break;
         case LLM_FFN_GELU:
             {
-                gate = ggml_gelu(ctx0, gate);
-                cb(gate, "ffn_moe_gelu", il);
+                cur = ggml_gelu(ctx0, cur);
+                cb(cur, "ffn_moe_gelu", il);
             } break;
         default:
             GGML_ABORT("fatal error");
     }
 
-    ggml_tensor * par = ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens]
-    cb(par, "ffn_moe_gate_par", il);
+    if (gate_exps) {
+        cur = ggml_mul(ctx0, cur, up); // [n_ff, n_expert_used, n_tokens]
+        cb(cur, "ffn_moe_gate_par", il);
+    }
 
-    ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens]
+    experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
     if (!weight_before_ffn) {
@@ -989,6 +971,7 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
         //cb(inp->tokens, "inp_tokens", -1);
         ggml_set_input(inp->tokens);
+        res->t_tokens = inp->tokens;
 
         cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
 
@@ -1095,7 +1078,7 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_s_copy() const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     auto inp = std::make_unique<llm_graph_input_s_copy>(kv_self);
 
@@ -1112,7 +1095,7 @@ ggml_tensor * llm_graph_context::build_inp_s_copy() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_s_mask() const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     auto inp = std::make_unique<llm_graph_input_s_mask>(kv_self);
 
@@ -1426,8 +1409,6 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // store to KV cache
     {
-        GGML_ASSERT(!kv_self->recurrent);
-
         const auto kv_head = kv_self->head;
 
         GGML_ASSERT(kv_self->size == n_ctx);
@@ -1577,7 +1558,7 @@ ggml_tensor * llm_graph_context::build_copy_mask_state(
          ggml_tensor * state_mask,
              int32_t   n_state,
              int32_t   n_seqs) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto n_kv    = kv_self->n;
     const auto kv_head = kv_self->head;
@@ -1609,7 +1590,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
          ggml_tensor * state_mask,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto token_shift_count = hparams.token_shift_count;
 
@@ -1630,7 +1611,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
          ggml_tensor * token_shift,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto token_shift_count = hparams.token_shift_count;
     const auto n_embd = hparams.n_embd;

src/llama-graph.h

@@ -19,6 +19,7 @@ struct llama_cparams;
 
 class llama_memory_i;
 class llama_kv_cache_unified;
+class llama_kv_cache_recurrent;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -186,26 +187,26 @@ public:
 
 class llm_graph_input_s_copy : public llm_graph_input_i {
 public:
-    llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    llm_graph_input_s_copy(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {}
     virtual ~llm_graph_input_s_copy() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_recurrent * kv_self;
 };
 
 class llm_graph_input_s_mask : public llm_graph_input_i {
 public:
-    llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    llm_graph_input_s_mask(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {}
     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_unified * kv_self;
+    const llama_kv_cache_recurrent * kv_self;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -297,6 +298,7 @@ class llm_graph_result_i {
 public:
     virtual ~llm_graph_result_i() = default;
 
+    virtual ggml_tensor * get_tokens()      = 0;
     virtual ggml_tensor * get_logits()      = 0;
     virtual ggml_tensor * get_embd()        = 0;
     virtual ggml_tensor * get_embd_pooled() = 0;
@@ -311,6 +313,7 @@ class llm_graph_result : public llm_graph_result_i {
 public:
     virtual ~llm_graph_result() = default;
 
+    ggml_tensor * get_tokens()      override { return t_tokens; }
     ggml_tensor * get_logits()      override { return t_logits; }
     ggml_tensor * get_embd()        override { return t_embd; }
     ggml_tensor * get_embd_pooled() override { return t_embd_pooled; }
@@ -327,6 +330,7 @@ public:
     }
 
     // important graph nodes
+    ggml_tensor * t_tokens      = nullptr;
     ggml_tensor * t_logits      = nullptr;
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;
@@ -350,8 +354,8 @@ struct llm_graph_params {
     const llama_cparams & cparams;
     const llama_ubatch  & ubatch;
 
-    ggml_backend_sched * sched;
-    ggml_backend * backend_cpu;
+    ggml_backend_sched_t sched;
+    ggml_backend_t backend_cpu;
 
     const llama_adapter_cvec  * cvec;
     const llama_adapter_loras * loras;
@@ -402,9 +406,9 @@ struct llm_graph_context {
 
     ggml_context * ctx0 = nullptr;
 
-    ggml_backend_sched * sched;
+    ggml_backend_sched_t sched;
 
-    ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
+    ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
 
     const llama_adapter_cvec  * cvec;
     const llama_adapter_loras * loras;

src/llama-hparams.h

@@ -66,6 +66,7 @@ struct llama_hparams {
     float    expert_weights_scale = 0.0;
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
+    uint32_t moe_every_n_layers   = 0;
 
     float f_norm_eps;
     float f_norm_rms_eps;

src/llama-kv-cache.h

@@ -2,32 +2,72 @@
 
 #include "llama.h"
 #include "llama-io.h"
+#include "llama-graph.h"
 #include "llama-memory.h"
 
 #include "ggml-cpp.h"
 
-#include <functional>
 #include <set>
 #include <vector>
 
 struct llama_cparams;
 struct llama_hparams;
 struct llama_ubatch;
+struct llama_sbatch;
+struct llama_model;
+struct llama_context;
 
 struct llama_kv_cache : public llama_memory_i {
-    using llama_memory_i::llama_memory_i;
+    virtual ~llama_kv_cache() = default;
 
-    virtual void restore() = 0; // call if batch processing fails - restores the cache state
-    virtual void commit() = 0;  // call after successful batch processing - clears any pending state
+    // call if batch processing fails - restores the cache state
+    virtual void restore() = 0;
 
-    virtual int32_t get_n_tokens()   const = 0;
-    virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+    // call after successful batch processing - clears any pending state
+    virtual void commit()  = 0;
 
-    virtual bool get_can_shift() const = 0;
+    // process any pending defrag/shift/etc. operations
+    // optionally call once before processing a new batch
+    virtual bool update(llama_context & lctx) = 0;
+
+    // schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing
+    virtual void defrag_sched(float thold) = 0;
+
+    // simulate full cache, used for allocating worst-case compute buffers
+    virtual void set_full() = 0;
+
+    //
+    // batch processing
+    //
+
+    virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0;
+
+    // different KV caches require different batch splitting strategies
+    virtual llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const = 0;
+
+    // find an empty slot of size "n_tokens" in the cache
+    virtual bool find_slot(const llama_ubatch & batch) = 0;
+
+    // getters
+    virtual int32_t   get_n_tokens()   const = 0;
+    virtual int32_t   get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+    virtual llama_pos get_pos_max()    const = 0;
+    virtual bool      get_can_shift()  const = 0;
 
     bool get_can_edit() const override { return get_can_shift(); }
+
+    //
+    // state write/read
+    //
+
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
 };
 
+//
+// llama_kv_cache_guard
+//
+
 struct llama_kv_cache_guard {
     llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {}
 
@@ -43,65 +83,50 @@ private:
     llama_kv_cache * kv;
 };
 
-struct llama_kv_cell {
-    llama_pos pos   = -1;
-    llama_pos delta =  0;
-    int32_t   src   = -1; // used by recurrent state models to copy states
-    int32_t   tail  = -1;
+//
+// llama_kv_cache_unified
+//
 
-    std::set<llama_seq_id> seq_id;
-
-    bool has_seq_id(const llama_seq_id & id) const {
-        return seq_id.find(id) != seq_id.end();
-    }
-
-    bool is_empty() const {
-        return seq_id.empty();
-    }
-
-    bool is_same_seq(const llama_kv_cell & other) const {
-        return seq_id == other.seq_id;
-    }
-};
-
-// ring-buffer of cached KV data
-// TODO: pimpl
 // TODO: add notion of max sequences
 class llama_kv_cache_unified : public llama_kv_cache {
 public:
-    // can be used to query data from the model if needed
-    struct callbacks {
-        std::function<ggml_tensor * (uint32_t n_ctx_per_seq, int il)> get_rope_factors;
+    struct kv_cell {
+        llama_pos pos   = -1;
+        llama_pos delta =  0;
+
+        std::set<llama_seq_id> seq_id;
+
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
+
+        bool is_empty() const {
+            return seq_id.empty();
+        }
+
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
     };
 
+    static uint32_t get_padding(const llama_cparams & cparams);
+
     llama_kv_cache_unified(
-            const llama_hparams & hparams,
-            callbacks             cbs);
-
-    virtual ~llama_kv_cache_unified() = default;
-
-    // TODO: become constructor
-    bool init(
-            const llama_model & model,   // TODO: do not reference the model
-          const llama_cparams & cparams,
+            const llama_model & model,
                     ggml_type   type_k,
                     ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
                      uint32_t   kv_size,
-                         bool   offload);
+                     uint32_t   padding);
 
-    int32_t get_n_tokens()   const override;
-    int32_t get_used_cells() const override;
+    ~llama_kv_cache_unified() = default;
 
-    size_t total_size() const;
-
-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos pos_max() const;
+    //
+    // llama_memory_i
+    //
 
     void clear() override;
-    void defrag() override;
-
-    virtual void restore() override;
-    virtual void commit() 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;
@@ -111,63 +136,40 @@ public:
 
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    bool get_can_shift() const override;
+    //
+    // llama_kv_cache
+    //
+
+    void restore() override;
+    void commit()  override;
+
+    bool update(llama_context & ctx) override;
+
+    void defrag_sched(float thold) override;
+
+    void set_full() override;
+
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
+
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
-    // find an empty slot of size "n_tokens" in the cache
     // updates the cache head
     // Note: On success, it's important that cache.head points
     // to the first cell of the slot.
-    bool find_slot(const llama_ubatch & batch);
+    bool find_slot(const llama_ubatch & batch) override;
 
-    // TODO: maybe not needed
-    uint32_t get_padding(const llama_cparams & cparams) const;
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;
 
-    // find how many cells are currently in use
-    uint32_t cell_max() const;
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos get_pos_max() const override;
 
-    size_t size_k_bytes() const;
-    size_t size_v_bytes() const;
-
-    // defrag
-
-    struct {
-        std::vector<uint32_t> ids;
-    } defrag_info;
-
-    // return true if cells have been moved
-    bool defrag_prepare(int32_t n_max_nodes);
-
-    // commit/restore cache
-
-    struct slot_range {
-        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
-        uint32_t c1 = 0;
-    };
-
-    // pending cell updates that are not yet committed
-    struct {
-        std::vector<slot_range> ranges;
-    } pending;
+    bool get_can_shift() const override;
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1);
-
-    // members
-
-    const llama_hparams & hparams;
-
-    callbacks cbs;
-
-    bool has_shift = false;
-    bool do_defrag = false;
-
-    // TODO: remove this and implement llama_kv_cache_recurrent instead
-    bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token
-
-    bool v_trans   = true;  // the value tensor is transposed
-    bool can_shift = false;
+    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;
 
     // Note: The value of head isn't only used to optimize searching
     // for a free KV slot. llama_decode_impl also uses it, so it
@@ -179,18 +181,213 @@ public:
     // computed before each graph build
     uint32_t n = 0;
 
-    std::vector<llama_kv_cell> cells;
+    std::vector<kv_cell> cells;
 
     std::vector<ggml_tensor *> k_l; // per layer
     std::vector<ggml_tensor *> v_l;
 
 private:
+    const llama_model & model;
+    const llama_hparams & hparams;
+
+    bool has_shift = false;
+    bool do_defrag = false;
+
+    bool v_trans   = true;  // the value tensor is transposed
+    bool can_shift = false;
+
+    // required padding
+    uint32_t padding = 1;
+
     ggml_type type_k = GGML_TYPE_F16;
     ggml_type type_v = GGML_TYPE_F16;
 
     std::vector<ggml_context_ptr>        ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
+    // defrag
+    struct {
+        std::vector<uint32_t> ids;
+    } defrag_info;
+
+    // return true if cells have been moved
+    bool defrag_prepare(int32_t n_max_nodes);
+
+    // commit/restore cache
+    struct slot_range {
+        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
+        uint32_t c1 = 0;
+    };
+
+    // pending cell updates that are not yet committed
+    struct {
+        std::vector<slot_range> ranges;
+    } pending;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
+    ggml_tensor * build_rope_shift(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_tensor * cur,
+                    ggml_tensor * shift,
+                    ggml_tensor * factors,
+                          float   freq_base,
+                          float   freq_scale) const;
+
+    llm_graph_result_ptr build_graph_shift(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_graph_defrag(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_cgraph * gf) 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;
+
+    bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
+};
+
+//
+// llama_kv_cache_recurrent
+//
+
+class llama_kv_cache_recurrent : public llama_kv_cache {
+public:
+    struct kv_cell {
+        llama_pos pos  = -1;
+        int32_t   src  = -1; // used to copy states
+        int32_t   tail = -1;
+
+        std::set<llama_seq_id> seq_id;
+
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
+
+        bool is_empty() const {
+            return seq_id.empty();
+        }
+
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
+    };
+
+    llama_kv_cache_recurrent(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   offload,
+                     uint32_t   kv_size);
+
+    ~llama_kv_cache_recurrent() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    void clear() 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 delta) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    //
+    // llama_kv_cache
+    //
+
+    void restore() override;
+    void commit()  override;
+
+    bool update(llama_context & lctx) override;
+
+    void defrag_sched(float thold) override;
+
+    void set_full() override;
+
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
+
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
+
+    bool find_slot(const llama_ubatch & batch) override;
+
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;
+
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos get_pos_max() const override;
+
+    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;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
+
+    // Note: The value of head isn't only used to optimize searching
+    // for a free KV slot. llama_decode_impl also uses it, so it
+    // cannot be freely changed after a slot has been allocated.
+    uint32_t head = 0;
+    uint32_t size = 0;
+    uint32_t used = 0; // used cells (i.e. at least one seq_id)
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    std::vector<kv_cell> cells;
+
+    std::vector<ggml_tensor *> k_l; // per layer
+    std::vector<ggml_tensor *> v_l;
+
+private:
+    //const llama_model & model;
+    const llama_hparams & hparams;
+
+    // commit/restore cache
+    // TODO: rework for recurrent cache
+    struct slot_range {
+        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
+        uint32_t c1 = 0;
+    };
+
+    // pending cell updates that are not yet committed
+    struct {
+        std::vector<slot_range> ranges;
+    } pending;
+
+    ggml_type type_k = GGML_TYPE_F16;
+    ggml_type type_v = GGML_TYPE_F16;
+
+    std::vector<ggml_context_ptr>        ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_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;
 
@@ -198,11 +395,6 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
 
-// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified
-//class llama_kv_cache_recurrent : public llama_kv_cache_unified {
-//public:
-//    using llama_kv_cache_unified::llama_kv_cache_unified;
-//};
 
 //
 // kv cache view

src/llama-memory.h

@@ -2,12 +2,22 @@
 
 #include "llama.h"
 
+struct llama_memory_params {
+    // kv cache
+    ggml_type type_k;
+    ggml_type type_v;
+
+    // parameters for other types of memory
+    // ...
+};
+
 // general concept of LLM memory
 // the KV cache is a type of LLM memory, but there can be other types
 class llama_memory_i {
 public:
+    virtual ~llama_memory_i() = default;
+
     virtual void clear() = 0;
-    virtual void defrag() = 0;
 
     virtual bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) = 0;
     virtual void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0;

src/llama-model-loader.cpp

@@ -301,12 +301,12 @@ namespace GGUFMeta {
             GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
 
         switch (arr_info.gt) {
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
-            case GGUF_TYPE_INT32:   GGML_ASSERT(
-                                            (std::is_same<T,  int32_t>::value) ||
-                                            (std::is_same<T, uint32_t>::value));  break;
+            case GGUF_TYPE_UINT32:
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
+                                                (std::is_same<T, uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
         }
 
         result.resize(arr_info.length);
@@ -330,12 +330,12 @@ namespace GGUFMeta {
             GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
 
         switch (arr_info.gt) {
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
-            case GGUF_TYPE_INT32:   GGML_ASSERT(
-                                            (std::is_same<T,  int32_t>::value) ||
-                                            (std::is_same<T, uint32_t>::value));  break;
+            case GGUF_TYPE_UINT32:
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
+                                                (std::is_same<T, uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
         }
 
         if (arr_info.length > N_MAX) {

src/llama-model-saver.cpp

@@ -0,0 +1,281 @@
+#include "llama-model-saver.h"
+
+#include "gguf.h"
+
+#include "llama.h"
+#include "llama-hparams.h"
+#include "llama-model.h"
+#include "llama-vocab.h"
+
+#include <string>
+
+llama_model_saver::llama_model_saver(const struct llama_model & model) : model(model), llm_kv(model.arch) {
+    gguf_ctx = gguf_init_empty();
+}
+
+llama_model_saver::~llama_model_saver() {
+    gguf_free(gguf_ctx);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) {
+    gguf_set_val_u32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const int32_t value) {
+    gguf_set_val_i32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const float value) {
+    gguf_set_val_f32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const bool value) {
+    gguf_set_val_bool(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const char * value) {
+    gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+[[noreturn]]
+void llama_model_saver::add_kv(const enum llm_kv key, const char value) {
+    GGML_UNUSED(key);
+    GGML_UNUSED(value);
+    GGML_ABORT("fatal error"); // this should never be called, only needed to make the template below compile
+}
+
+template <typename Container>
+void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, const bool per_layer) {
+    const size_t n_values = per_layer ? size_t(model.hparams.n_layer) : value.size();
+    GGML_ASSERT(n_values <= value.size());
+
+    if (n_values == 0) {
+        return;
+    }
+
+    if (per_layer) {
+        bool all_values_the_same = true;
+        for (size_t i = 1; i < n_values; ++i) {
+            if (value[i] != value[0]) {
+                all_values_the_same = false;
+                break;
+            }
+        }
+        if (all_values_the_same) {
+            add_kv(key, value[0]);
+            return;
+        }
+    }
+
+    if (std::is_same<typename Container::value_type, uint8_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT8, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, int8_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT8, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, uint32_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT32, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, int32_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT32, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, float>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_FLOAT32, value.data(), n_values);
+    } else if (std::is_same<Container, std::string>::value) {
+        gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), reinterpret_cast<const char *>(value.data()));
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const std::vector<std::string> & value) {
+    std::vector<const char *> tmp(value.size());
+    for (size_t i = 0; i < value.size(); ++i) {
+        tmp[i] = value[i].c_str();
+    }
+    gguf_set_arr_str(gguf_ctx, llm_kv(key).c_str(), tmp.data(), tmp.size());
+}
+
+void llama_model_saver::add_tensor(const struct ggml_tensor * tensor) {
+    if (!tensor) {
+        return;
+    }
+    if (gguf_find_tensor(gguf_ctx, tensor->name) >= 0) {
+        GGML_ASSERT(std::string(tensor->name) == "rope_freqs.weight"); // FIXME
+        return;
+    }
+    gguf_add_tensor(gguf_ctx, tensor);
+}
+
+void llama_model_saver::add_kv_from_model() {
+    const llama_hparams & hparams = model.hparams;
+    const llama_vocab   & vocab   = model.vocab;
+
+    const int32_t n_vocab = vocab.n_tokens();
+    std::vector<std::string> tokens(n_vocab);
+    std::vector<float>       scores(n_vocab);
+    std::vector<int32_t>     token_types(n_vocab);
+
+    for (int32_t id = 0; id < n_vocab; ++id) {
+        const llama_vocab::token_data & token_data = vocab.get_token_data(id);
+
+        tokens[id] = token_data.text;
+        scores[id] = token_data.score;
+
+        switch(token_data.attr) {
+            case LLAMA_TOKEN_ATTR_UNKNOWN:      token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN;      break;
+            case LLAMA_TOKEN_ATTR_UNUSED:       token_types[id] = LLAMA_TOKEN_TYPE_UNUSED;       break;
+            case LLAMA_TOKEN_ATTR_NORMAL:       token_types[id] = LLAMA_TOKEN_TYPE_NORMAL;       break;
+            case LLAMA_TOKEN_ATTR_CONTROL:      token_types[id] = LLAMA_TOKEN_TYPE_CONTROL;      break;
+            case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;
+            case LLAMA_TOKEN_ATTR_BYTE:         token_types[id] = LLAMA_TOKEN_TYPE_BYTE;         break;
+            case LLAMA_TOKEN_ATTR_UNDEFINED:
+            default:                            token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED;    break;
+        }
+    }
+
+    // add_kv(LLM_KV_GENERAL_TYPE,                      ???);
+    add_kv(LLM_KV_GENERAL_ARCHITECTURE,              model.arch_name());
+    // add_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION,      ???);
+    // add_kv(LLM_KV_GENERAL_ALIGNMENT,                 ???);
+    add_kv(LLM_KV_GENERAL_NAME,                      model.name);
+    // add_kv(LLM_KV_GENERAL_AUTHOR,                    ???);
+    // add_kv(LLM_KV_GENERAL_VERSION,                   ???);
+    // add_kv(LLM_KV_GENERAL_URL,                       ???);
+    // add_kv(LLM_KV_GENERAL_DESCRIPTION,               ???);
+    // add_kv(LLM_KV_GENERAL_LICENSE,                   ???);
+    // add_kv(LLM_KV_GENERAL_SOURCE_URL,                ???);
+    // add_kv(LLM_KV_GENERAL_SOURCE_HF_REPO,            ???);
+
+    add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());
+    add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);
+    add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);
+    add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer);
+    add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
+    add_kv(LLM_KV_FEED_FORWARD_LENGTH,               hparams.n_ff_arr, true);
+    add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    add_kv(LLM_KV_USE_PARALLEL_RESIDUAL,             hparams.use_par_res);
+    // add_kv(LLM_KV_TENSOR_DATA_LAYOUT,                ???);
+    add_kv(LLM_KV_EXPERT_COUNT,                      hparams.n_expert);
+    add_kv(LLM_KV_EXPERT_USED_COUNT,                 hparams.n_expert_used);
+    add_kv(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
+    add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale);
+    add_kv(LLM_KV_POOLING_TYPE,                      uint32_t(hparams.pooling_type));
+    add_kv(LLM_KV_LOGIT_SCALE,                       hparams.f_logit_scale);
+    add_kv(LLM_KV_DECODER_START_TOKEN_ID,            hparams.dec_start_token_id);
+    add_kv(LLM_KV_ATTN_LOGIT_SOFTCAPPING,            hparams.f_attn_logit_softcapping);
+    add_kv(LLM_KV_FINAL_LOGIT_SOFTCAPPING,           hparams.f_final_logit_softcapping);
+    add_kv(LLM_KV_SWIN_NORM,                         hparams.swin_norm);
+    add_kv(LLM_KV_RESCALE_EVERY_N_LAYERS,            hparams.rescale_every_n_layers);
+    add_kv(LLM_KV_TIME_MIX_EXTRA_DIM,                hparams.time_mix_extra_dim);
+    add_kv(LLM_KV_TIME_DECAY_EXTRA_DIM,              hparams.time_decay_extra_dim);
+    add_kv(LLM_KV_RESIDUAL_SCALE,                    hparams.f_residual_scale);
+    add_kv(LLM_KV_EMBEDDING_SCALE,                   hparams.f_embedding_scale);
+
+    add_kv(LLM_KV_ATTENTION_HEAD_COUNT,              hparams.n_head_arr, true);
+    add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);
+    add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS,          hparams.f_max_alibi_bias);
+    add_kv(LLM_KV_ATTENTION_CLAMP_KQV,               hparams.f_clamp_kqv);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v);
+    add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,           hparams.f_norm_eps);
+    add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+    add_kv(LLM_KV_ATTENTION_CAUSAL,                  hparams.causal_attn);
+    add_kv(LLM_KV_ATTENTION_Q_LORA_RANK,             hparams.n_lora_q);
+    add_kv(LLM_KV_ATTENTION_KV_LORA_RANK,            hparams.n_lora_kv);
+    add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,  hparams.n_rel_attn_bkts);
+    add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);
+    add_kv(LLM_KV_ATTENTION_SCALE,                   hparams.f_attention_scale);
+
+    const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
+
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot);
+    add_kv(LLM_KV_ROPE_FREQ_BASE,                    hparams.rope_freq_base_train);
+    // add_kv(LLM_KV_ROPE_SCALE_LINEAR,                 rope_scaling_factor); // old name
+    add_kv(LLM_KV_ROPE_SCALING_TYPE,                 llama_rope_scaling_type_name(hparams.rope_scaling_type_train));
+    add_kv(LLM_KV_ROPE_SCALING_FACTOR,               rope_scaling_factor);
+    add_kv(LLM_KV_ROPE_SCALING_ATTN_FACTOR,          hparams.rope_attn_factor);
+    add_kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,         hparams.n_ctx_orig_yarn);
+    add_kv(LLM_KV_ROPE_SCALING_FINETUNED,            hparams.rope_finetuned);
+    add_kv(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,         hparams.rope_yarn_log_mul);
+
+    // TODO: implement split file support
+    // add_kv(LLM_KV_SPLIT_NO,                          ???);
+    // add_kv(LLM_KV_SPLIT_COUNT,                       ???);
+    // add_kv(LLM_KV_SPLIT_TENSORS_COUNT,               ???);
+
+    add_kv(LLM_KV_SSM_INNER_SIZE,                    hparams.ssm_d_inner);
+    add_kv(LLM_KV_SSM_CONV_KERNEL,                   hparams.ssm_d_conv);
+    add_kv(LLM_KV_SSM_STATE_SIZE,                    hparams.ssm_d_state);
+    add_kv(LLM_KV_SSM_TIME_STEP_RANK,                hparams.ssm_dt_rank);
+    add_kv(LLM_KV_SSM_DT_B_C_RMS,                    hparams.ssm_dt_b_c_rms);
+
+    add_kv(LLM_KV_WKV_HEAD_SIZE,                     hparams.wkv_head_size);
+
+    add_kv(LLM_KV_TOKENIZER_MODEL,                   vocab.get_tokenizer_model());
+    add_kv(LLM_KV_TOKENIZER_PRE,                     vocab.get_tokenizer_pre());
+    add_kv(LLM_KV_TOKENIZER_LIST,                    tokens);
+    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE,              token_types);
+    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,        vocab.n_token_types());
+    add_kv(LLM_KV_TOKENIZER_SCORES,                  scores);
+    add_kv(LLM_KV_TOKENIZER_MERGES,                  vocab.get_bpe_merges());
+    // FIXME llama_token is type i32 but when reading in a GGUF file u32 is expected, not an issue for writing though
+    add_kv(LLM_KV_TOKENIZER_BOS_ID,                  uint32_t(vocab.token_bos()));
+    add_kv(LLM_KV_TOKENIZER_EOS_ID,                  uint32_t(vocab.token_eos()));
+    add_kv(LLM_KV_TOKENIZER_EOT_ID,                  uint32_t(vocab.token_eot()));
+    add_kv(LLM_KV_TOKENIZER_EOM_ID,                  uint32_t(vocab.token_eom()));
+    add_kv(LLM_KV_TOKENIZER_UNK_ID,                  uint32_t(vocab.token_unk()));
+    add_kv(LLM_KV_TOKENIZER_SEP_ID,                  uint32_t(vocab.token_sep()));
+    add_kv(LLM_KV_TOKENIZER_PAD_ID,                  uint32_t(vocab.token_pad()));
+    // add_kv(LLM_KV_TOKENIZER_CLS_ID,                  uint32_t(vocab.token_bos())); // deprecated
+    // add_kv(LLM_KV_TOKENIZER_MASK_ID,                 ???);
+    add_kv(LLM_KV_TOKENIZER_ADD_BOS,                 vocab.get_add_bos());
+    add_kv(LLM_KV_TOKENIZER_ADD_EOS,                 vocab.get_add_eos());
+    add_kv(LLM_KV_TOKENIZER_ADD_PREFIX,              vocab.get_add_space_prefix());
+    add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,         vocab.get_remove_extra_whitespaces());
+    add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,    vocab.get_precompiled_charsmap());
+    // add_kv(LLM_KV_TOKENIZER_HF_JSON,                 ???);
+    // add_kv(LLM_KV_TOKENIZER_RWKV,                    ???);
+    add_kv(LLM_KV_TOKENIZER_FIM_PRE_ID,              uint32_t(vocab.token_fim_pre()));
+    add_kv(LLM_KV_TOKENIZER_FIM_SUF_ID,              uint32_t(vocab.token_fim_suf()));
+    add_kv(LLM_KV_TOKENIZER_FIM_MID_ID,              uint32_t(vocab.token_fim_mid()));
+    add_kv(LLM_KV_TOKENIZER_FIM_PAD_ID,              uint32_t(vocab.token_fim_pad()));
+    add_kv(LLM_KV_TOKENIZER_FIM_REP_ID,              uint32_t(vocab.token_fim_rep()));
+    add_kv(LLM_KV_TOKENIZER_FIM_SEP_ID,              uint32_t(vocab.token_fim_sep()));
+
+    // TODO: implement LoRA support
+    // add_kv(LLM_KV_ADAPTER_TYPE,                      ???);
+    // add_kv(LLM_KV_ADAPTER_LORA_ALPHA,                ???);
+
+    // deprecated
+    // add_kv(LLM_KV_TOKENIZER_PREFIX_ID,               ???);
+    // add_kv(LLM_KV_TOKENIZER_SUFFIX_ID,               ???);
+    // add_kv(LLM_KV_TOKENIZER_MIDDLE_ID,               ???);
+}
+
+void llama_model_saver::add_tensors_from_model() {
+    if (std::string(model.output->name) != std::string(model.tok_embd->name)) {
+        add_tensor(model.tok_embd); // some models use the same tensor for tok_embd and output
+    }
+    add_tensor(model.type_embd);
+    add_tensor(model.pos_embd);
+    add_tensor(model.tok_norm);
+    add_tensor(model.tok_norm_b);
+    add_tensor(model.output_norm);
+    add_tensor(model.output_norm_b);
+    add_tensor(model.output);
+    add_tensor(model.output_b);
+    add_tensor(model.output_norm_enc);
+    add_tensor(model.cls);
+    add_tensor(model.cls_b);
+    add_tensor(model.cls_out);
+    add_tensor(model.cls_out_b);
+
+    for (const struct llama_layer & layer : model.layers) {
+        for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
+            add_tensor(reinterpret_cast<const struct ggml_tensor * const *>(&layer)[i]);
+        }
+    }
+}
+
+void llama_model_saver::save(const std::string & path_model) {
+    gguf_write_to_file(gguf_ctx, path_model.c_str(), false);
+}
+

src/llama-model-saver.h

@@ -0,0 +1,37 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-arch.h"
+
+#include <vector>
+
+struct llama_model_saver {
+    struct gguf_context * gguf_ctx = nullptr;
+    const struct llama_model & model;
+    const struct LLM_KV llm_kv;
+
+    llama_model_saver(const struct llama_model & model);
+    ~llama_model_saver();
+
+    void add_kv(enum llm_kv key, uint32_t     value);
+    void add_kv(enum llm_kv key, int32_t      value);
+    void add_kv(enum llm_kv key, float        value);
+    void add_kv(enum llm_kv key, bool         value);
+    void add_kv(enum llm_kv key, const char * value);
+
+    [[noreturn]]
+    void add_kv(enum llm_kv key, char value); // needed to make the template below compile
+
+    template <typename Container>
+    void add_kv(enum llm_kv key, const Container & value, bool per_layer = false);
+
+    void add_kv(enum llm_kv key, const std::vector<std::string> & value);
+
+    void add_tensor(const struct ggml_tensor * tensor);
+
+    void add_kv_from_model();
+
+    void add_tensors_from_model();
+
+    void save(const std::string & path_model);
+};

src/llama-model.cpp

@@ -40,14 +40,17 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_335M:          return "335M";
         case LLM_TYPE_410M:          return "410M";
         case LLM_TYPE_450M:          return "450M";
+        case LLM_TYPE_475M:          return "475M";
         case LLM_TYPE_770M:          return "770M";
         case LLM_TYPE_780M:          return "780M";
         case LLM_TYPE_0_5B:          return "0.5B";
+        case LLM_TYPE_0_6B:          return "0.6B";
         case LLM_TYPE_1B:            return "1B";
         case LLM_TYPE_1_3B:          return "1.3B";
         case LLM_TYPE_1_4B:          return "1.4B";
         case LLM_TYPE_1_5B:          return "1.5B";
         case LLM_TYPE_1_6B:          return "1.6B";
+        case LLM_TYPE_1_7B:          return "1.7B";
         case LLM_TYPE_1_8B:          return "1.8B";
         case LLM_TYPE_2B:            return "2B";
         case LLM_TYPE_2_8B:          return "2.8B";
@@ -66,6 +69,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_15B:           return "15B";
         case LLM_TYPE_16B:           return "16B";
         case LLM_TYPE_20B:           return "20B";
+        case LLM_TYPE_27B:           return "27B";
         case LLM_TYPE_30B:           return "30B";
         case LLM_TYPE_32B:           return "32B";
         case LLM_TYPE_34B:           return "34B";
@@ -74,6 +78,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_65B:           return "65B";
         case LLM_TYPE_70B:           return "70B";
         case LLM_TYPE_236B:          return "236B";
+        case LLM_TYPE_290B:          return "290B";
         case LLM_TYPE_314B:          return "314B";
         case LLM_TYPE_671B:          return "671B";
         case LLM_TYPE_SMALL:         return "0.1B";
@@ -88,10 +93,10 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_16x3_8B:       return "16x3.8B";
         case LLM_TYPE_10B_128x3_66B: return "10B+128x3.66B";
         case LLM_TYPE_57B_A14B:      return "57B.A14B";
-        case LLM_TYPE_27B:           return "27B";
-        case LLM_TYPE_290B:          return "290B";
         case LLM_TYPE_17B_16E:       return "17Bx16E (Scout)";
         case LLM_TYPE_17B_128E:      return "17Bx128E (Maverick)";
+        case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_235B_A22B:     return "235B.A22B";
         default:                     return "?B";
     }
 }
@@ -111,6 +116,10 @@ static const std::map<llama_rope_scaling_type, const char *> LLAMA_ROPE_SCALING_
     { LLAMA_ROPE_SCALING_TYPE_LONGROPE,   "longrope"   },
 };
 
+std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type) {
+    return LLAMA_ROPE_SCALING_TYPES.at(rope_scaling_type);
+}
+
 static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) {
     for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) {
         if (kv.second == name) {
@@ -695,13 +704,19 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
             } break;
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
                 ml.get_key(LLM_KV_ATTENTION_CAUSAL,           hparams.causal_attn);
                 ml.get_key(LLM_KV_POOLING_TYPE,               hparams.pooling_type);
+                ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS,         hparams.moe_every_n_layers, 0);
 
                 if (hparams.n_layer == 12 && hparams.n_embd == 768) {
-                    type = LLM_TYPE_137M;
+                    if (arch == LLM_ARCH_NOMIC_BERT) {
+                        type = LLM_TYPE_137M;
+                    } else if (arch == LLM_ARCH_NOMIC_BERT_MOE && hparams.moe_every_n_layers == 2) {
+                        type = LLM_TYPE_475M;
+                    }
                 }
             } break;
         case LLM_ARCH_BLOOM:
@@ -762,6 +777,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             // fall through
         case LLM_ARCH_QWEN2:
             {
+                ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 24: type = hparams.n_embd == 1024 ? LLM_TYPE_0_5B : LLM_TYPE_1B; break;
@@ -791,6 +807,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
+                    case 28: type = hparams.n_embd == 1024 ? LLM_TYPE_0_6B : LLM_TYPE_1_7B; break;
+                    case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break;
+                    case 40: type = LLM_TYPE_14B; break;
+                    case 64: type = LLM_TYPE_32B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -800,6 +820,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
+                    case 48: type = LLM_TYPE_30B_A3B; break;
+                    case 94: type = LLM_TYPE_235B_A22B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -2057,6 +2079,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_BERT:
             case LLM_ARCH_NOMIC_BERT:
+            case LLM_ARCH_NOMIC_BERT_MOE:
                 {
                     tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
                     type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0);
@@ -2090,20 +2113,31 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
                         }
 
+                        if (arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                            layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
+                        }
+
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {n_embd, n_embd}, 0);
 
                         layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
                         layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);
 
-                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,        "weight", i), {n_embd, n_ff}, 0);
-                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN,      "weight", i), {n_ff, n_embd}, 0);
-
-                        if (arch == LLM_ARCH_BERT) {
+                        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
                             layer.bo         = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
-                            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, 0);
-                            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
                         } else {
-                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,        "weight", i), {n_embd, n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN,      "weight", i), {n_ff, n_embd}, 0);
+
+                            if (arch == LLM_ARCH_BERT || arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                                layer.bo         = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+                                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, 0);
+                                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
+                            } else {
+                                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+                            }
                         }
 
                         layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
@@ -4208,7 +4242,7 @@ uint64_t llama_model::n_elements() const {
 }
 
 void llama_model::print_info() const {
-    const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train);
+    const std::string rope_scaling_type = llama_rope_scaling_type_name(hparams.rope_scaling_type_train);
 
     auto print_f = [](const std::function<uint32_t(uint32_t)> & f, uint32_t n) {
         bool is_var = false;
@@ -4269,7 +4303,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: causal attn      = %d\n",     __func__, hparams.causal_attn);
         LLAMA_LOG_INFO("%s: pooling type     = %d\n",     __func__, hparams.pooling_type);
         LLAMA_LOG_INFO("%s: rope type        = %d\n",     __func__, hparams.rope_type);
-        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type);
+        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type.c_str());
         LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
@@ -4416,6 +4450,19 @@ const ggml_tensor * llama_model::get_tensor(const char * name) const {
     return it->second;
 }
 
+ggml_tensor * llama_model::get_rope_factors(uint32_t n_ctx_per_seq, int il) const {
+    // choose long/short freq factors based on the context size
+    if (layers[il].rope_freqs != nullptr) {
+        return layers[il].rope_freqs;
+    }
+
+    if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
+        return layers[il].rope_long;
+    }
+
+    return layers[il].rope_short;
+}
+
 struct llm_build_llama : public llm_graph_context {
     llm_build_llama(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -4456,7 +4503,7 @@ struct llm_build_llama : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -4681,7 +4728,7 @@ struct llm_build_deci : public llm_graph_context {
             } else if (n_head > 0) {
                 // self-attention
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -5730,6 +5777,11 @@ struct llm_build_bert : public llm_graph_context {
                 cur = build_lora_mm(model.layers[il].wqkv, cur);
                 cb(cur, "wqkv", il);
 
+                if (model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
                 Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
                 Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
                 Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
@@ -5782,13 +5834,29 @@ struct llm_build_bert : public llm_graph_context {
             cb(ffn_inp, "ffn_inp", il);
 
             // feed-forward network
-            if (model.arch == LLM_ARCH_BERT) {
+            if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
+                // MoE branch
+                cur = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        nullptr,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        hparams.n_expert,
+                        hparams.n_expert_used,
+                        LLM_FFN_GELU,
+                        false, false,
+                        0.0f,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+                cb(cur, "ffn_moe_out", il);
+            } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                         NULL,                      NULL,                        NULL,
                         model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                         NULL,
                         LLM_FFN_GELU, LLM_FFN_SEQ, il);
+                cb(cur, "ffn_out", il);
             } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   NULL,                        NULL,
@@ -5796,6 +5864,7 @@ struct llm_build_bert : public llm_graph_context {
                         model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                         NULL,
                         LLM_FFN_GELU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
             } else {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   NULL, NULL,
@@ -5803,8 +5872,8 @@ struct llm_build_bert : public llm_graph_context {
                         model.layers[il].ffn_down, NULL, NULL,
                         NULL,
                         LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
             }
-            cb(cur, "ffn_out", il);
 
             // attentions bypass the intermediate layer
             cur = ggml_add(ctx0, cur, ffn_inp);
@@ -7141,7 +7210,7 @@ struct llm_build_phi3 : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for 128k context
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 ggml_tensor* attn_norm_output = build_norm(inpL,
                         model.layers[il].attn_norm,
@@ -7893,7 +7962,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
-            ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+            ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
             // norm
             cur = build_norm(inpL,
@@ -8660,7 +8729,7 @@ struct llm_build_mamba : public llm_graph_context {
              ggml_tensor * state_mask,
       const llama_ubatch & ubatch,
                      int   il) const {
-        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto kv_head = kv_self->head;
 
@@ -8961,7 +9030,7 @@ struct llm_build_cohere2 : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for 128k context
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -9899,7 +9968,7 @@ struct llm_build_deepseek : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -10149,7 +10218,6 @@ struct llm_build_deepseek2 : public llm_graph_context {
 
                     // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
                     ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
-                    ggml_mul_mat_set_prec(q_nope_absorbed, GGML_PREC_F32);
                     cb(q_nope_absorbed, "q_nope_absorbed", il);
 
                     // {kv_lora_rank, n_head, n_tokens}
@@ -11264,7 +11332,7 @@ struct llm_build_exaone : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -11409,7 +11477,7 @@ struct llm_build_rwkv6_base : public llm_graph_context {
             ggml_tensor * state_mask,
             const llama_ubatch & ubatch,
             int   il) const {
-        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto n_tokens = ubatch.n_tokens;
         const auto n_seqs = ubatch.n_seqs;
@@ -11805,7 +11873,7 @@ struct llm_build_rwkv7_base : public llm_graph_context {
             ggml_tensor *& first_layer_value,
             const llama_ubatch & ubatch,
             int   il) const {
-        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto n_tokens = ubatch.n_tokens;
         const auto n_seqs = ubatch.n_seqs;
@@ -12645,7 +12713,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -12765,7 +12833,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
     }
 };
 
-llama_memory_i * llama_model::create_memory() const {
+llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
     llama_memory_i * res;
 
     switch (arch) {
@@ -12775,26 +12843,29 @@ llama_memory_i * llama_model::create_memory() const {
         case LLM_ARCH_RWKV7:
         case LLM_ARCH_ARWKV7:
             {
-                res = new llama_kv_cache_unified(hparams, {
-                    /*.get_rope_factors =*/ nullptr
-                });
+                res = new llama_kv_cache_recurrent(
+                        *this,
+                        GGML_TYPE_F32,
+                        GGML_TYPE_F32,
+                        cparams.offload_kqv,
+                        std::max((uint32_t) 1, cparams.n_seq_max));
             } break;
         default:
             {
-                res = new llama_kv_cache_unified(hparams, {
-                    /*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) {
-                        // choose long/short freq factors based on the context size
-                        if (layers[il].rope_freqs != nullptr) {
-                            return layers[il].rope_freqs;
-                        }
+                const auto padding = llama_kv_cache_unified::get_padding(cparams);
 
-                        if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
-                            return layers[il].rope_long;
-                        }
+                cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
 
-                        return layers[il].rope_short;
-                    }
-                });
+                LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
+
+                res = new llama_kv_cache_unified(
+                        *this,
+                        params.type_k,
+                        params.type_v,
+                        !cparams.flash_attn,
+                        cparams.offload_kqv,
+                        cparams.n_ctx,
+                        padding);
             }
     }
 
@@ -12843,6 +12914,7 @@ llm_graph_result_ptr llama_model::build_graph(
         case LLM_ARCH_BERT:
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
             {
                 llm = std::make_unique<llm_build_bert>(*this, params, gf);
             } break;
@@ -13175,8 +13247,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DECI:
         case LLM_ARCH_BAICHUAN:
         case LLM_ARCH_STARCODER:
-        case LLM_ARCH_PLAMO:
-        case LLM_ARCH_ORION:
         case LLM_ARCH_INTERNLM2:
         case LLM_ARCH_MINICPM:
         case LLM_ARCH_XVERSE:
@@ -13201,6 +13271,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DBRX:
         case LLM_ARCH_BERT:
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_STABLELM:
         case LLM_ARCH_BITNET:
         case LLM_ARCH_QWEN:
@@ -13213,6 +13284,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:
         case LLM_ARCH_PHIMOE:
+        case LLM_ARCH_PLAMO:
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_GEMMA2:
         case LLM_ARCH_GEMMA3:
@@ -13220,6 +13292,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_OPENELM:
         case LLM_ARCH_GPTNEOX:
         case LLM_ARCH_CODESHELL:
+        case LLM_ARCH_ORION:
         case LLM_ARCH_NEMOTRON:
         case LLM_ARCH_EXAONE:
         case LLM_ARCH_MINICPM3: