fix(hexagon): use padded stride for ssm-conv weights (#24470)

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

.devops/cann.Dockerfile

@@ -13,6 +13,20 @@ ARG APP_REVISION=N/A
 # BUILD STAGE
 # Compile all binary files and libraries
 # ==============================================================================
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 FROM ${CANN_BASE_IMAGE} AS build
 
 # -- Install build dependencies --
@@ -26,6 +40,8 @@ WORKDIR /app
 # -- Copy project files --
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 # -- Set CANN environment variables (required for compilation) --
 # Using ENV instead of `source` allows environment variables to persist across the entire image layer
 ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest

.devops/cpu.Dockerfile

@@ -3,6 +3,20 @@ ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
 
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 FROM docker.io/ubuntu:$UBUNTU_VERSION AS build
 
 ARG TARGETARCH
@@ -16,6 +30,8 @@ WORKDIR /app
 
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 RUN if [ "$TARGETARCH" = "amd64" ] || [ "$TARGETARCH" = "arm64" ]; then \
         cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
     else \

.devops/cuda.Dockerfile

@@ -11,6 +11,20 @@ ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
 
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
 ARG GCC_VERSION
@@ -26,6 +40,8 @@ WORKDIR /app
 
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
     export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \

.devops/intel.Dockerfile

@@ -5,6 +5,20 @@ ARG APP_REVISION=N/A
 
 ## Build Image
 
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 FROM docker.io/intel/deep-learning-essentials:$ONEAPI_VERSION AS build
 
 ARG GGML_SYCL_F16=ON
@@ -22,6 +36,8 @@ WORKDIR /app
 
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         echo "GGML_SYCL_F16 is set" \
         && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON" \

.devops/musa.Dockerfile

@@ -10,6 +10,20 @@ ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
 
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
 # MUSA architecture to build for (defaults to all supported archs)
@@ -29,6 +43,8 @@ WORKDIR /app
 
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
     fi && \

.devops/openvino.Dockerfile

@@ -22,6 +22,20 @@ ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
 
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 ## Build Image
 FROM docker.io/ubuntu:${UBUNTU_VERSION} AS build
 
@@ -69,6 +83,8 @@ WORKDIR /app
 
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 # Build Stage
 RUN bash -c "source ${OpenVINO_DIR}/setupvars.sh && \
     cmake -B build/ReleaseOV -G Ninja \

.devops/rocm.Dockerfile

@@ -11,6 +11,20 @@ ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
 
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 ### Build image
 FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 
@@ -38,6 +52,8 @@ WORKDIR /app
 
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
     cmake -S . -B build \
         -DGGML_HIP=ON \

.devops/vulkan.Dockerfile

@@ -3,6 +3,20 @@ ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
 
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 FROM docker.io/ubuntu:$UBUNTU_VERSION AS build
 
 # Install build tools
@@ -17,6 +31,8 @@ WORKDIR /app
 
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
     cmake --build build --config Release -j$(nproc)
 

.devops/zendnn.Dockerfile

@@ -3,6 +3,20 @@ ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
 
+ARG NODE_VERSION=24
+
+FROM docker.io/node:$NODE_VERSION AS web
+
+ARG APP_VERSION
+
+WORKDIR /app/tools/ui
+
+COPY tools/ui/package.json tools/ui/package-lock.json ./
+RUN npm ci
+
+COPY tools/ui/ ./
+RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
+
 FROM docker.io/ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
@@ -14,6 +28,8 @@ WORKDIR /app
 
 COPY . .
 
+COPY --from=web /app/tools/ui/dist tools/ui/dist
+
 RUN cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_ZENDNN=ON && \
     cmake --build build -j $(nproc)
 

.dockerignore

@@ -10,6 +10,8 @@
 
 build*/
 
+tools/ui/node_modules/
+
 models/*
 
 /llama-cli

.github/workflows/docker.yml

@@ -58,6 +58,13 @@ jobs:
           git tag ${{ steps.srctag.outputs.name }} || exit 0
           git push origin ${{ steps.srctag.outputs.name }} || exit 0
 
+  build_ui:
+    name: Build UI
+    needs: create_tag
+    uses: ./.github/workflows/ui-build.yml
+    with:
+      hf_ui_version: ${{ needs.create_tag.outputs.source_tag }}
+
   prepare_matrices:
     name: Prepare Docker matrices
     runs-on: ubuntu-24.04
@@ -79,7 +86,7 @@ jobs:
           [
             { "tag": "cpu", "dockerfile": ".devops/cpu.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" },
             { "tag": "cpu", "dockerfile": ".devops/cpu.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-arm" },
-            { "tag": "cpu", "dockerfile": ".devops/s390x.Dockerfile", "platforms": "linux/s390x", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-s390x" },
+            { "tag": "cpu", "dockerfile": ".devops/s390x.Dockerfile", "platforms": "linux/s390x", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-s390x", "prebuilt_ui": true },
             { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "12.8.1", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
             { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "12.8.1", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
             { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.3.0", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
@@ -135,7 +142,7 @@ jobs:
 
   push_to_registry:
     name: Push Docker image to Docker Registry
-    needs: [prepare_matrices, create_tag]
+    needs: [prepare_matrices, create_tag, build_ui]
 
     runs-on: ${{ matrix.config.runs_on }}
     strategy:
@@ -150,6 +157,13 @@ jobs:
           fetch-depth: 0
           ref: ${{ needs.create_tag.outputs.source_tag }}
 
+      - name: Download prebuilt UI
+        if: ${{ matrix.config.prebuilt_ui == true }}
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
+        with:
+          name: ui-build
+          path: tools/ui/dist
+
       - name: Set up QEMU
         if: ${{ contains(matrix.config.platforms, 'linux/amd64') }}
         uses: docker/setup-qemu-action@ce360397dd3f832beb865e1373c09c0e9f86d70a # v4

.github/workflows/release.yml

@@ -1627,6 +1627,7 @@ jobs:
             **Windows:**
             - [Windows x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-x64.zip)
             - [Windows arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-arm64.zip)
+            - [Windows arm64 (OpenCL Adreno)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-opencl-adreno-arm64.zip)
             - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-12.4-x64.zip) - [CUDA 12.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-12.4-x64.zip)
             - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.3-x64.zip) - [CUDA 13.3 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.3-x64.zip)
             - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)

.pi/gg/SYSTEM.md

@@ -25,13 +25,3 @@ Commits:
 - Do not explicitly set the git author in commits - rely on the default git config
 - Always use `--no-gpg-sign` when committing
 - Never `git push` without explicit confirmation from the user
-
-Resources (read on demand):
-- [CONTRIBUTING.md](CONTRIBUTING.md)
-- [Build documentation](docs/build.md)
-- [Server usage documentation](tools/server/README.md)
-- [Server development documentation](tools/server/README-dev.md)
-- [PEG parser](docs/development/parsing.md)
-- [Auto parser](docs/autoparser.md)
-- [Jinja engine](common/jinja/README.md)
-- [PR template](.github/pull_request_template.md)

common/arg.cpp

@@ -17,6 +17,7 @@
 #   define NOMINMAX
 #endif
 #include <windows.h>
+#include <shellapi.h>
 #endif
 
 #define JSON_ASSERT GGML_ASSERT
@@ -285,58 +286,15 @@ static std::string clean_file_name(const std::string & fname) {
     return clean_fname;
 }
 
-static bool common_params_handle_remote_preset(common_params & params, llama_example ex) {
-    GGML_ASSERT(!params.model.hf_repo.empty());
-
-    // the returned hf_repo is without tag
-    auto [hf_repo, hf_tag] = common_download_split_repo_tag(params.model.hf_repo);
-
-    // "latest" tag (default if not specified) is translated to "default" preset
-    if (hf_tag == "latest") {
-        hf_tag = "default";
-    }
-
-    std::string model_endpoint = common_get_model_endpoint();
-    auto preset_url = model_endpoint + hf_repo + "/resolve/main/preset.ini";
-
-    // prepare local path for caching
-    auto preset_fname = clean_file_name(hf_repo + "_preset.ini");
-    auto preset_path = fs_get_cache_file(preset_fname);
-    common_download_opts opts;
-    opts.bearer_token = params.hf_token;
-    opts.offline = params.offline;
-
-    LOG_TRC("%s: looking for remote preset at %s\n", __func__, preset_url.c_str());
-    const int status = common_download_file_single(preset_url, preset_path, opts);
-    const bool has_preset = status >= 200 && status < 400;
-
-    // remote preset is optional, so we don't error out if not found
-    if (has_preset) {
-        LOG_TRC("%s: applying remote preset from %s\n", __func__, preset_url.c_str());
-        common_preset_context ctx(ex, /* only_remote_allowed */ true);
-        common_preset global;
-        auto remote_presets = ctx.load_from_ini(preset_path, global);
-        remote_presets = ctx.cascade(global, remote_presets);
-        if (remote_presets.find(hf_tag) != remote_presets.end()) {
-            common_preset preset = remote_presets.at(hf_tag);
-            LOG_INF("\n%s", preset.to_ini().c_str()); // to_ini already added trailing newline
-            preset.apply_to_params(params);
-        } else {
-            throw std::runtime_error("Remote preset.ini does not contain [" + std::string(hf_tag) + "] section");
-        }
-    } else {
-        LOG_TRC("%s: no remote preset found, skipping\n", __func__);
-    }
-
-    return has_preset;
-}
-
 struct handle_model_result {
     bool found_mmproj = false;
     common_params_model mmproj;
 
     bool found_mtp = false;
     common_params_model mtp;
+
+    bool found_preset = false;
+    std::string preset_path;
 };
 
 static handle_model_result common_params_handle_model(struct common_params_model & model,
@@ -345,7 +303,6 @@ static handle_model_result common_params_handle_model(struct common_params_model
 
     if (!model.docker_repo.empty()) {
         model.path = common_docker_resolve_model(model.docker_repo);
-        model.name = model.docker_repo;
     } else if (!model.hf_repo.empty()) {
         // If -m was used with -hf, treat the model "path" as the hf_file to download
         if (model.hf_file.empty() && !model.path.empty()) {
@@ -355,11 +312,16 @@ static handle_model_result common_params_handle_model(struct common_params_model
         common_download_opts hf_opts = opts;
         auto download_result = common_download_model(model, hf_opts);
 
+        if (!download_result.preset_path.empty()) {
+            result.found_preset = true;
+            result.preset_path = download_result.preset_path;
+            return result; // skip everything else if preset.ini is used
+        }
+
         if (download_result.model_path.empty()) {
             throw std::runtime_error("failed to download model from Hugging Face");
         }
 
-        model.name = model.hf_repo;
         model.path = download_result.model_path;
 
         if (!download_result.mmproj_path.empty()) {
@@ -454,6 +416,17 @@ bool common_params_handle_models(common_params & params, llama_example curr_ex)
 
     try {
         auto res = common_params_handle_model(params.model, opts);
+        if (res.found_preset) {
+            if (!params.models_preset.empty()) {
+                throw std::invalid_argument("cannot use both --models-preset and -hf with a preset.ini file");
+            }
+            // if HF repo is a preset repo, we simply run server in router mode with the preset.ini file
+            params.models_preset_hf = params.model.hf_repo; // only for showing a warning
+            params.models_preset    = res.preset_path;
+            params.model = common_params_model{}; // make sure to clear model, so server starts in router mode
+            return true;
+        }
+
         if (params.no_mmproj) {
             params.mmproj = {};
         } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
@@ -601,30 +574,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
     // parse the first time to get -hf option (used for remote preset)
     parse_cli_args();
 
-    // export_graph_ops loads only metadata
-    const bool skip_model_download = ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;
-
-    // maybe handle remote preset
-    if (!params.model.hf_repo.empty() && !skip_model_download) {
-        std::string cli_hf_repo = params.model.hf_repo;
-        bool has_preset = common_params_handle_remote_preset(params, ctx_arg.ex);
-
-        // special case: if hf_repo explicitly set by preset, we need to preserve it (ignore CLI value)
-        // this is useful when we have one HF repo pointing to other HF repos (one model - multiple GGUFs)
-        std::string preset_hf_repo = params.model.hf_repo;
-        bool preset_has_hf_repo = preset_hf_repo != cli_hf_repo;
-
-        if (has_preset) {
-            // re-parse CLI args to override preset values
-            parse_cli_args();
-        }
-
-        // preserve hf_repo from preset if needed
-        if (preset_has_hf_repo) {
-            params.model.hf_repo = preset_hf_repo;
-        }
-    }
-
     postprocess_cpu_params(params.cpuparams,       nullptr);
     postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
 
@@ -635,15 +584,21 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
-    // handle model and download
-    if (!skip_model_download) {
-        common_params_handle_models(params, ctx_arg.ex);
-    }
+    // export_graph_ops loads only metadata
+    const bool skip_model_download = ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;
 
-    // model is required (except for server)
-    // TODO @ngxson : maybe show a list of available models in CLI in this case
-    if (params.model.path.empty() && ctx_arg.ex != LLAMA_EXAMPLE_SERVER && !skip_model_download && !params.usage && !params.completion) {
-        throw std::invalid_argument("error: --model is required\n");
+    if (!skip_model_download) {
+        // handle model and download
+        common_params_handle_models(params, ctx_arg.ex);
+
+        // model is required (except for server)
+        // TODO @ngxson : maybe show a list of available models in CLI in this case
+        if (params.model.path.empty()
+                && ctx_arg.ex != LLAMA_EXAMPLE_SERVER
+                && !params.usage
+                && !params.completion) {
+            throw std::invalid_argument("error: --model is required\n");
+        }
     }
 
     if (params.escape) {
@@ -937,7 +892,44 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
     return true;
 }
 
+#ifdef _WIN32
+struct utf8_argv {
+    std::vector<std::string> buf;
+    std::vector<char*> ptrs;
+};
+
+static utf8_argv make_utf8_argv() {
+    utf8_argv out;
+    int wargc = 0;
+    LPWSTR* wargv = CommandLineToArgvW(GetCommandLineW(), &wargc);
+    if (!wargv) return out;
+
+    out.buf.reserve(wargc);
+    for (int i = 0; i < wargc; ++i) {
+        int n = WideCharToMultiByte(CP_UTF8, WC_ERR_INVALID_CHARS, wargv[i], -1, nullptr, 0, nullptr, nullptr);
+        if (n <= 0) { out.buf.emplace_back(); continue; }
+        auto& s = out.buf.emplace_back();
+        s.resize(static_cast<size_t>(n - 1));
+        (void)WideCharToMultiByte(CP_UTF8, 0, wargv[i], -1, s.data(), n, nullptr, nullptr);
+    }
+    LocalFree(wargv);
+
+    out.ptrs.reserve(out.buf.size() + 1);
+    for (auto& s : out.buf) out.ptrs.push_back(s.data());
+    out.ptrs.push_back(nullptr);
+    return out;
+}
+#endif
+
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
+#ifdef _WIN32
+    auto utf8 = make_utf8_argv();
+    // repair argv only when it matches the process command line
+    if (static_cast<int>(utf8.buf.size()) == argc) {
+        argv = utf8.ptrs.data();
+    }
+#endif
+
     auto ctx_arg = common_params_parser_init(params, ex, print_usage);
     const common_params params_org = ctx_arg.params; // the example can modify the default params
 
@@ -2874,62 +2866,26 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.api_prefix = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_PREFIX"));
-    // Deprecated: use --ui-config instead (kept for backward compat)
     add_opt(common_arg(
-        {"--webui-config"}, "JSON",
-        "[DEPRECATED: use --ui-config] JSON that provides default WebUI settings (overrides WebUI defaults)",
-        [](common_params & params, const std::string & value) {
-            params.ui_config_json = value;
-            params.webui_config_json = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_CONFIG"));
-
-    add_opt(common_arg(
-        {"--ui-config"}, "JSON",
+        {"--ui-config", "--webui-config"}, "JSON",
         "JSON that provides default UI settings (overrides UI defaults)",
         [](common_params & params, const std::string & value) {
             params.ui_config_json = value;
-            params.webui_config_json = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_UI_CONFIG"));
-
-    // Deprecated: use --ui-config-file instead (kept for backward compat)
     add_opt(common_arg(
-        {"--webui-config-file"}, "PATH",
-        "[DEPRECATED: use --ui-config-file] JSON file that provides default WebUI settings (overrides WebUI defaults)",
-        [](common_params & params, const std::string & value) {
-            params.ui_config_json = read_file(value);
-            params.webui_config_json = params.ui_config_json;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_CONFIG_FILE"));
-
-    add_opt(common_arg(
-        {"--ui-config-file"}, "PATH",
+        {"--ui-config-file", "--webui-config-file"}, "PATH",
         "JSON file that provides default UI settings (overrides UI defaults)",
         [](common_params & params, const std::string & value) {
             params.ui_config_json = read_file(value);
-            params.webui_config_json = params.ui_config_json;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_UI_CONFIG_FILE"));
-
-    // Deprecated: use --ui-mcp-proxy instead (kept for backward compat)
     add_opt(common_arg(
-        {"--webui-mcp-proxy"},
-        {"--no-webui-mcp-proxy"},
-        "[DEPRECATED: use --ui-mcp-proxy/--no-ui-mcp-proxy] experimental: whether to enable MCP CORS proxy",
-        [](common_params & params, bool value) {
-            params.ui_mcp_proxy = value;
-            params.webui_mcp_proxy = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI_MCP_PROXY"));
-
-    add_opt(common_arg(
-        {"--ui-mcp-proxy"},
-        {"--no-ui-mcp-proxy"},
+        {"--ui-mcp-proxy", "--webui-mcp-proxy"},
+        {"--no-ui-mcp-proxy", "--no-webui-mcp-proxy"},
         "experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: disabled)",
         [](common_params & params, bool value) {
             params.ui_mcp_proxy = value;
-            params.webui_mcp_proxy = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_UI_MCP_PROXY"));
     add_opt(common_arg(
@@ -2941,24 +2897,26 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.server_tools = parse_csv_row(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TOOLS"));
-    // Deprecated: use --ui/--no-ui instead (kept for backward compat)
     add_opt(common_arg(
-        {"--webui"},
-        {"--no-webui"},
-        "[DEPRECATED: use --ui/--no-ui] whether to enable the Web UI",
+        {"-ag", "--agent"},
+        {"-no-ag", "--no-agent"},
+        "whether to enable CORS proxy and all built-in tools - do not enable in untrusted environments (default: disabled)",
         [](common_params & params, bool value) {
-            params.ui = value;
-            params.webui = value;
+            if (value) {
+                params.server_tools = {"all"};
+                params.ui_mcp_proxy = true;
+            } else {
+                params.server_tools.clear();
+                params.ui_mcp_proxy = false;
+            }
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_WEBUI"));
-
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_AGENT"));
     add_opt(common_arg(
-        {"--ui"},
-        {"--no-ui"},
+        {"--ui", "--webui"},
+        {"--no-ui", "--no-webui"},
         string_format("whether to enable the Web UI (default: %s)", params.ui ? "enabled" : "disabled"),
         [](common_params & params, bool value) {
             params.ui = value;
-            params.webui = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_UI"));
     add_opt(common_arg(
@@ -2989,7 +2947,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_API_KEY"));
     add_opt(common_arg(
         {"--api-key-file"}, "FNAME",
-        "path to file containing API keys (default: none)",
+        "path to file containing API keys, one per line; lines starting with a hash are treated as comments (default: none)",
         [](common_params & params, const std::string & value) {
             std::ifstream key_file(value);
             if (!key_file) {
@@ -2997,7 +2955,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
             std::string key;
             while (std::getline(key_file, key)) {
-                if (!key.empty()) {
+                if (!key.empty() && key[0] != '#') {
                     params.api_keys.push_back(key);
                 }
             }

common/common.cpp

@@ -1074,6 +1074,18 @@ std::vector<common_file_info> fs_list(const std::string & path, bool include_dir
     return files;
 }
 
+std::ifstream fs_open_ifstream(const std::string & fname, std::ios_base::openmode mode) {
+#ifdef _WIN32
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, fname.c_str(), -1, NULL, 0);
+    if (!wlen) { return std::ifstream(); }
+    std::vector<wchar_t> wfname(wlen);
+    (void)MultiByteToWideChar(CP_UTF8, 0, fname.c_str(), -1, wfname.data(), wlen);
+    return std::ifstream(wfname.data(), mode);
+#else
+    return std::ifstream(fname, mode);
+#endif
+}
+
 //
 // TTY utils
 //
@@ -2034,7 +2046,7 @@ bool common_prompt_batch_decode(
 }
 
 size_t common_prompt_checkpoint::size() const {
-    return data_tgt.size() + data_dft.size();
+    return data_tgt.size() + data_dft.size() + data_spec.size();
 }
 
 bool common_prompt_checkpoint::empty() const {
@@ -2049,6 +2061,7 @@ void common_prompt_checkpoint::clear() {
 
     data_tgt.clear();
     data_dft.clear();
+    data_spec.clear();
 }
 
 void common_prompt_checkpoint::update_pos(
@@ -2138,4 +2151,5 @@ void common_prompt_checkpoint::clear_tgt() {
 
 void common_prompt_checkpoint::clear_dft() {
     data_dft.clear();
+    data_spec.clear();
 }

common/common.h

@@ -295,7 +295,16 @@ struct common_params_model {
     std::string hf_repo     = ""; // HF repo                                                // NOLINT
     std::string hf_file     = ""; // HF file                                                // NOLINT
     std::string docker_repo = ""; // Docker repo                                            // NOLINT
-    std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
+
+    std::string get_name() {
+        if (!hf_repo.empty()) {
+            return hf_repo;
+        }
+        if (!docker_repo.empty()) {
+            return docker_repo;
+        }
+        return path;
+    }
 };
 
 // draft-model-based speculative decoding parameters
@@ -363,7 +372,7 @@ struct common_params_speculative {
 
     uint32_t need_n_rs_seq() const {
         bool needs_rs_seq = std::any_of(types.begin(), types.end(), [&](auto t) {
-            return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP;
+            return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3;
         });
 
         return needs_rs_seq ? draft.n_max : 0u;
@@ -624,12 +633,6 @@ struct common_params {
 
     // UI configs
     bool ui = true;
-
-    // Deprecated: use ui, ui_mcp_proxy, ui_config_json instead
-    bool webui = ui;
-    bool webui_mcp_proxy = false;
-    std::string webui_config_json;
-
     bool ui_mcp_proxy = false;
     std::string ui_config_json;
 
@@ -642,10 +645,11 @@ struct common_params {
     std::vector<std::string> server_tools;
 
     // router server configs
-    std::string models_dir    = ""; // directory containing models for the router server
-    std::string models_preset = ""; // directory containing model presets for the router server
-    int models_max = 4;             // maximum number of models to load simultaneously
-    bool models_autoload = true;    // automatically load models when requested via the router server
+    std::string models_dir    = "";     // directory containing models for the router server
+    std::string models_preset = "";     // directory containing model presets for the router server
+    int models_max = 4;                 // maximum number of models to load simultaneously
+    bool models_autoload = true;        // automatically load models when requested via the router server
+    std::string models_preset_hf = "";  // show a warning about remote presets on router loaded (if not empty)
 
     bool log_json = false;
 
@@ -847,6 +851,9 @@ struct common_file_info {
 };
 std::vector<common_file_info> fs_list(const std::string & path, bool include_directories);
 
+// fs open, also handle UTF8 on Windows
+std::ifstream fs_open_ifstream(const std::string & fname, std::ios_base::openmode mode);
+
 //
 // TTY utils
 //
@@ -1064,6 +1071,10 @@ struct common_prompt_checkpoint {
     std::vector<uint8_t> data_tgt;
     std::vector<uint8_t> data_dft;
 
+    // (optional) speculative-decoding implementation state stashed with the checkpoint
+    // (e.g. eagle3's deferred-boundary g_embd row)
+    std::vector<uint8_t> data_spec;
+
     size_t size() const;
 
     bool empty() const;

common/download.cpp

@@ -696,6 +696,7 @@ struct hf_plan {
     hf_cache::hf_files model_files;
     hf_cache::hf_file mmproj;
     hf_cache::hf_file mtp;
+    hf_cache::hf_file preset; // if set, only this file is downloaded
 };
 
 static hf_plan get_hf_plan(const common_params_model  & model,
@@ -717,6 +718,14 @@ static hf_plan get_hf_plan(const common_params_model  & model,
         return plan;
     }
 
+    // if preset.ini exists in the repo root, download only that file
+    for (const auto & f : all) {
+        if (f.path == "preset.ini") {
+            plan.preset = f;
+            return plan;
+        }
+    }
+
     hf_cache::hf_file primary;
 
     if (!model.hf_file.empty()) {
@@ -794,14 +803,19 @@ common_download_model_result common_download_model(const common_params_model  &
 
     if (is_hf) {
         hf = get_hf_plan(model, opts, download_mmproj, download_mtp);
-        for (const auto & f : hf.model_files) {
-            tasks.push_back({f.url, f.local_path});
-        }
-        if (!hf.mmproj.path.empty()) {
-            tasks.push_back({hf.mmproj.url, hf.mmproj.local_path});
-        }
-        if (!hf.mtp.path.empty()) {
-            tasks.push_back({hf.mtp.url, hf.mtp.local_path});
+        if (!hf.preset.path.empty()) {
+            // if preset.ini exists, only download that file alone
+            tasks.push_back({hf.preset.url, hf.preset.local_path});
+        } else {
+            for (const auto & f : hf.model_files) {
+                tasks.push_back({f.url, f.local_path});
+            }
+            if (!hf.mmproj.path.empty()) {
+                tasks.push_back({hf.mmproj.url, hf.mmproj.local_path});
+            }
+            if (!hf.mtp.path.empty()) {
+                tasks.push_back({hf.mtp.url, hf.mtp.local_path});
+            }
         }
     } else if (!model.url.empty()) {
         tasks = get_url_tasks(model);
@@ -835,17 +849,22 @@ common_download_model_result common_download_model(const common_params_model  &
     }
 
     if (is_hf) {
-        for (const auto & f : hf.model_files) {
-            hf_cache::finalize_file(f);
-        }
-        result.model_path = hf.primary.final_path;
+        if (!hf.preset.path.empty()) {
+            // if preset.ini is used, do not set other paths
+            result.preset_path = hf_cache::finalize_file(hf.preset);
+        } else {
+            for (const auto & f : hf.model_files) {
+                hf_cache::finalize_file(f);
+            }
+            result.model_path = hf.primary.final_path;
 
-        if (!hf.mmproj.path.empty()) {
-            result.mmproj_path = hf_cache::finalize_file(hf.mmproj);
-        }
+            if (!hf.mmproj.path.empty()) {
+                result.mmproj_path = hf_cache::finalize_file(hf.mmproj);
+            }
 
-        if (!hf.mtp.path.empty()) {
-            result.mtp_path = hf_cache::finalize_file(hf.mtp);
+            if (!hf.mtp.path.empty()) {
+                result.mtp_path = hf_cache::finalize_file(hf.mtp);
+            }
         }
     } else {
         result.model_path = model.path;

common/download.h

@@ -63,6 +63,7 @@ struct common_download_model_result {
     std::string model_path;
     std::string mmproj_path;
     std::string mtp_path;
+    std::string preset_path;
 };
 
 // throw if the file is missing or invalid (e.g. ETag check failed)

common/preset.cpp

@@ -16,48 +16,6 @@ static std::string rm_leading_dashes(const std::string & str) {
     return str.substr(pos);
 }
 
-// only allow a subset of args for remote presets for security reasons
-// do not add more args unless absolutely necessary
-// args that output to files are strictly prohibited
-static std::set<std::string> get_remote_preset_whitelist(const std::map<std::string, common_arg> & key_to_opt) {
-    static const std::set<std::string> allowed_options = {
-        "model-url",
-        "hf-repo",
-        "hf-repo-draft",
-        "hf-repo-v", // vocoder
-        "hf-file-v", // vocoder
-        "mmproj-url",
-        "pooling",
-        "jinja",
-        "batch-size",
-        "ubatch-size",
-        "cache-reuse",
-        "chat-template-kwargs",
-        "mmap",
-        // note: sampling params are automatically allowed by default
-        // negated args will be added automatically if the positive arg is specified above
-    };
-
-    std::set<std::string> allowed_keys;
-
-    for (const auto & it : key_to_opt) {
-        const std::string & key = it.first;
-        const common_arg & opt = it.second;
-        if (allowed_options.find(key) != allowed_options.end() || opt.is_sampling) {
-            allowed_keys.insert(key);
-            // also add variant keys (args without leading dashes and env vars)
-            for (const auto & arg : opt.get_args()) {
-                allowed_keys.insert(rm_leading_dashes(arg));
-            }
-            for (const auto & env : opt.get_env()) {
-                allowed_keys.insert(env);
-            }
-        }
-    }
-
-    return allowed_keys;
-}
-
 std::vector<std::string> common_preset::to_args(const std::string & bin_path) const {
     std::vector<std::string> args;
 
@@ -300,16 +258,10 @@ static std::string parse_bool_arg(const common_arg & arg, const std::string & ke
     return value;
 }
 
-common_preset_context::common_preset_context(llama_example ex, bool only_remote_allowed)
+common_preset_context::common_preset_context(llama_example ex)
         : ctx_params(common_params_parser_init(default_params, ex)) {
     common_params_add_preset_options(ctx_params.options);
     key_to_opt = get_map_key_opt(ctx_params);
-
-    // setup allowed keys if only_remote_allowed is true
-    if (only_remote_allowed) {
-        filter_allowed_keys = true;
-        allowed_keys = get_remote_preset_whitelist(key_to_opt);
-    }
 }
 
 common_presets common_preset_context::load_from_ini(const std::string & path, common_preset & global) const {

common/preset.h

@@ -60,7 +60,7 @@ struct common_preset_context {
     std::set<std::string> allowed_keys;
 
     // if only_remote_allowed is true, only accept whitelisted keys
-    common_preset_context(llama_example ex, bool only_remote_allowed = false);
+    common_preset_context(llama_example ex);
 
     // load presets from INI file
     common_presets load_from_ini(const std::string & path, common_preset & global) const;

common/sampling.cpp

@@ -259,6 +259,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
              }
         }
     }
+    if (!grmr && !grammar_str.empty()) {
+        throw std::runtime_error("failed to parse grammar");
+    }
 
     // Compute prefill tokens from the generation prompt
     std::vector<llama_token> prefill_tokens;

common/speculative.cpp

@@ -161,6 +161,10 @@ struct common_speculative_impl {
 
     virtual void accept(llama_seq_id seq_id, uint16_t n_accepted, bool is_other) = 0;
 
+    // (optional) serialize/restore per-seq internal state (e.g. eagle3's deferred boundary).
+    virtual bool get_state(llama_seq_id /*seq_id*/, std::vector<uint8_t> & /*data*/) const { return false; }
+    virtual void set_state(llama_seq_id /*seq_id*/, const std::vector<uint8_t> & /*data*/) {}
+
     // true if this implementation requires the target context to extract post-norm embeddings
     virtual bool need_embd() const = 0;
 
@@ -841,6 +845,49 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
                     (size_t) n_embd_dec * sizeof(float));
     }
 
+    // we only need to stash the deferred boundary's g_embd row for recurrent/hybrid targets:
+    // their single-position checkpoints drop it on restore
+    bool need_boundary_stash() const {
+        const llama_model * model_tgt = llama_get_model(params.ctx_tgt);
+        return llama_model_is_recurrent(model_tgt) || llama_model_is_hybrid(model_tgt);
+    }
+
+    bool get_state(llama_seq_id seq_id, std::vector<uint8_t> & data) const override {
+        if (!need_boundary_stash()) {
+            return false;
+        }
+        if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq || pending_pos_last[seq_id] < 0) {
+            return false;
+        }
+
+        const llama_pos          pos = pending_pos_last[seq_id];
+        const std::vector<float> & g = pending_g_last[seq_id];
+
+        data.resize(sizeof(llama_pos) + g.size() * sizeof(float));
+        std::memcpy(data.data(),                     &pos,     sizeof(llama_pos));
+        std::memcpy(data.data() + sizeof(llama_pos), g.data(), g.size() * sizeof(float));
+        return true;
+    }
+
+    void set_state(llama_seq_id seq_id, const std::vector<uint8_t> & data) override {
+        if (!need_boundary_stash()) {
+            return;
+        }
+        if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
+            return;
+        }
+        if (data.size() != sizeof(llama_pos) + (size_t) n_embd_dec * sizeof(float)) {
+            return;
+        }
+
+        llama_pos pos = -1;
+        std::memcpy(&pos, data.data(), sizeof(llama_pos));
+
+        pending_pos_last[seq_id] = pos;
+        pending_g_last[seq_id].resize(n_embd_dec);
+        std::memcpy(pending_g_last[seq_id].data(), data.data() + sizeof(llama_pos), (size_t) n_embd_dec * sizeof(float));
+    }
+
     bool need_embd() const override {
         return false;
     }
@@ -2118,6 +2165,31 @@ void common_speculative_accept(common_speculative * spec, llama_seq_id seq_id, u
     }
 }
 
+// TODO: support the case of more than one speculative implementations having a state
+bool common_speculative_get_state(common_speculative * spec, llama_seq_id seq_id, std::vector<uint8_t> & data) {
+    if (spec == nullptr) {
+        return false;
+    }
+
+    for (auto & impl : spec->impls) {
+        if (impl->get_state(seq_id, data)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+void common_speculative_set_state(common_speculative * spec, llama_seq_id seq_id, const std::vector<uint8_t> & data) {
+    if (spec == nullptr) {
+        return;
+    }
+
+    for (auto & impl : spec->impls) {
+        impl->set_state(seq_id, data);
+    }
+}
+
 void common_speculative_print_stats(const common_speculative * spec) {
     if (spec == nullptr) {
         return;

common/speculative.h

@@ -68,6 +68,10 @@ void common_speculative_draft(common_speculative * spec);
 // informs the speculative context that n_accepted tokens were accepted by the target model
 void common_speculative_accept(common_speculative * spec, llama_seq_id, uint16_t n_accepted);
 
+// (optional) get/set internal state
+bool common_speculative_get_state(common_speculative * spec, llama_seq_id seq_id, std::vector<uint8_t> & data);
+void common_speculative_set_state(common_speculative * spec, llama_seq_id seq_id, const std::vector<uint8_t> & data);
+
 // print statistics about the speculative decoding
 void common_speculative_print_stats(const common_speculative * spec);
 

conversion/bailingmoe.py

@@ -126,7 +126,7 @@ class BailingMoeV2Model(TextModel):
         if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
 
-        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.5)))
         self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
         self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])

conversion/base.py

@@ -1119,8 +1119,10 @@ class TextModel(ModelBase):
 
         rope_theta = self.find_hparam(["global_rope_theta", "rope_global_theta", "rope_theta_global", "rope_theta", "rotary_emb_base"], optional=True)
         local_rope_theta = self.find_hparam(["local_rope_theta", "rope_local_theta", "rope_theta_local", "swa_rope_theta", "rope_local_base_freq"], optional=True)
+        partial_rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct", "rope_percent"], optional=True)
+        original_max_position_embeddings = self.find_hparam(["original_max_position_embeddings"], optional=True)
 
-        # Ensure "rope_theta" and "rope_type" is mirrored in rope_parameters
+        # Ensure global params are mirrored in rope_parameters
         if "full_attention" not in self.rope_parameters and "sliding_attention" not in self.rope_parameters:
             if local_rope_theta is not None:
                 self.rope_parameters["sliding_attention"] = {"rope_theta": local_rope_theta}
@@ -1128,6 +1130,10 @@ class TextModel(ModelBase):
                 self.rope_parameters["rope_theta"] = rope_theta
             if "rope_type" not in self.rope_parameters and (rope_type := self.rope_parameters.get("type")) is not None:
                 self.rope_parameters["rope_type"] = rope_type
+            if "partial_rotary_factor" not in self.rope_parameters and partial_rotary_factor is not None:
+                self.rope_parameters["partial_rotary_factor"] = partial_rotary_factor
+            if "original_max_position_embeddings" not in self.rope_parameters and original_max_position_embeddings is not None:
+                self.rope_parameters["original_max_position_embeddings"] = original_max_position_embeddings
 
     @classmethod
     def __init_subclass__(cls):

conversion/chatglm.py

@@ -148,7 +148,7 @@ class ChatGLMModel(TextModel):
             rope_dim = self.hparams["attention_dim"]
         else:
             rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
-        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.5)))
         self.gguf_writer.add_add_bos_token(False)
         rope_freq = 10000
         if "rope_ratio" in self.hparams:

conversion/deci.py

@@ -161,7 +161,7 @@ class DeciModel(TextModel):
                 factor = rope_params.get("factor", 8.0)
                 low_freq_factor = rope_params.get("low_freq_factor", 1.0)
                 high_freq_factor = rope_params.get("high_freq_factor", 4.0)
-                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+                old_context_len = rope_params.get("original_max_position_embeddings", 8192)
 
                 low_freq_wavelen = old_context_len / low_freq_factor
                 high_freq_wavelen = old_context_len / high_freq_factor

conversion/exaone.py

@@ -24,7 +24,7 @@ class ExaoneModel(TextModel):
 
         assert (hparams["activation_function"] == "silu")
 
-        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True)
+        rotary_factor = self.rope_parameters.get("partial_rotary_factor")
         rotary_factor = rotary_factor if rotary_factor is not None else 1.0
         self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
 
@@ -39,7 +39,7 @@ class ExaoneModel(TextModel):
                 factor = rope_params.get("factor", 8.0)
                 low_freq_factor = rope_params.get("low_freq_factor", 1.0)
                 high_freq_factor = rope_params.get("high_freq_factor", 4.0)
-                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+                old_context_len = rope_params.get("original_max_position_embeddings", 8192)
 
                 low_freq_wavelen = old_context_len / low_freq_factor
                 high_freq_wavelen = old_context_len / high_freq_factor
@@ -104,7 +104,7 @@ class Exaone4Model(TextModel):
                 factor = rope_params.get("factor", 16.0)
                 low_freq_factor = rope_params.get("low_freq_factor", 1.0)
                 high_freq_factor = rope_params.get("high_freq_factor", 4.0)
-                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+                old_context_len = rope_params.get("original_max_position_embeddings", 8192)
 
                 low_freq_wavelen = old_context_len / low_freq_factor
                 high_freq_wavelen = old_context_len / high_freq_factor

conversion/gemma.py

@@ -693,7 +693,7 @@ class Gemma4Model(Gemma3Model):
             self.gguf_writer.add_head_count_kv(value_arr)
 
         # handle n_rot differently for global vs swa layers
-        partial_rotary_factor_swa = self.hparams.get("partial_rotary_factor", 1.0)
+        partial_rotary_factor_swa = self.rope_parameters.get("partial_rotary_factor", 1.0)
         n_rot_full = int(head_dim_full) # "proportional" is used, see generate_extra_tensors
         n_rot_swa = int(head_dim_swa * partial_rotary_factor_swa)
         self.gguf_writer.add_rope_dimension_count(n_rot_full)

conversion/glm.py

@@ -124,7 +124,7 @@ class Glm4MoeModel(TextModel):
                 self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
             )
         self.gguf_writer.add_rope_dimension_count(
-            int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5))
+            int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.5))
         )
 
         # MoE parameters - Use only routed expert count (shared experts handled separately)
@@ -226,7 +226,7 @@ class GlmMoeDsaModel(DeepseekV2Model):
         super().set_gguf_parameters()
 
         rope_dim = self.hparams["qk_rope_head_dim"]
-        partial_rotary_factor = self.hparams.get("partial_rotary_factor", 1.0)
+        partial_rotary_factor = self.rope_parameters.get("partial_rotary_factor", 1.0)
         self.gguf_writer.add_rope_dimension_count(int(rope_dim * partial_rotary_factor))
 
         # NextN/MTP prediction layers

conversion/llama.py

@@ -289,7 +289,7 @@ class LlamaModel(TextModel):
                 factor = rope_params.get("factor", 8.0)
                 low_freq_factor = rope_params.get("low_freq_factor", 1.0)
                 high_freq_factor = rope_params.get("high_freq_factor", 4.0)
-                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+                old_context_len = rope_params.get("original_max_position_embeddings", 8192)
 
                 low_freq_wavelen = old_context_len / low_freq_factor
                 high_freq_wavelen = old_context_len / high_freq_factor

conversion/mimo.py

@@ -154,7 +154,7 @@ class MimoV2Model(TextModel):
         self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
         self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
 
-        rope_dim = int(self.hparams["head_dim"] * self.hparams["partial_rotary_factor"])
+        rope_dim = int(self.hparams["head_dim"] * self.rope_parameters["partial_rotary_factor"])
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon", 1e-5))

conversion/minicpm.py

@@ -32,11 +32,9 @@ class MiniCPMModel(TextModel):
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         rope_dims = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
 
-        rope_scaling = self.find_hparam(['rope_scaling'], True)
-        if rope_scaling is not None:
-            long_factors = rope_scaling.get('long_factor', None)
-            short_factors = rope_scaling.get('short_factor', None)
-
+        long_factors = self.rope_parameters.get('long_factor')
+        short_factors = self.rope_parameters.get('short_factor')
+        if long_factors or short_factors:
             if long_factors is None or short_factors is None:
                 raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
 
@@ -85,13 +83,11 @@ class MiniCPM3Model(TextModel):
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
-        rope_scaling = self.find_hparam(['rope_scaling'], True)
-        if rope_scaling is not None:
+        long_factors = self.rope_parameters.get('long_factor')
+        short_factors = self.rope_parameters.get('short_factor')
+        if long_factors or short_factors:
             rope_dims = self.hparams["qk_rope_head_dim"]
 
-            long_factors = rope_scaling.get('long_factor', None)
-            short_factors = rope_scaling.get('short_factor', None)
-
             if long_factors is None or short_factors is None:
                 raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
 

conversion/nemotron.py

@@ -125,17 +125,18 @@ class NemotronModel(TextModel):
         self.gguf_writer.add_layer_norm_eps(f_norm_eps)
 
         # * Partial RoPE
-        rot_pct = self.find_hparam(["partial_rotary_factor", "rope_pct", "rope_percent"])
+        rot_pct = self.rope_parameters["partial_rotary_factor"]
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
         self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
 
         # * RopeScaling for Nemotron
-        if "rope_scaling" not in self.hparams or self.hparams["rope_scaling"] is None:
+        factor = self.hparams.get("factor") or self.rope_parameters.get("factor")
+        if factor is None:
             self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
         else:
             self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(self.hparams["factor"])
+            self.gguf_writer.add_rope_scaling_factor(factor)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # * Adding +1 to LayerNorm's weights here to implement layernorm1p w/o changing anything on the GGML engine side

conversion/phi.py

@@ -18,7 +18,7 @@ class Phi2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.PHI2
 
     def set_gguf_parameters(self):
-        rot_pct = self.find_hparam(["partial_rotary_factor"])
+        rot_pct = self.rope_parameters["partial_rotary_factor"]
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
 
@@ -149,8 +149,8 @@ class Phi3MiniModel(TextModel):
         n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
         rms_eps = self.find_hparam(["rms_norm_eps"])
         max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
-        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
-        rot_pct = self.hparams.get("partial_rotary_factor", 1.0)
+        orig_max_pos_embds = self.rope_parameters["original_max_position_embeddings"]
+        rot_pct = self.rope_parameters.get("partial_rotary_factor", 1.0)
         rope_dims = int(rot_pct * n_embd) // n_head
 
         self.gguf_writer.add_context_length(max_pos_embds)
@@ -174,18 +174,19 @@ class Phi3MiniModel(TextModel):
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
         max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
-        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
-        rot_pct = self.hparams.get("partial_rotary_factor", 1.0)
+        orig_max_pos_embds = self.rope_parameters["original_max_position_embeddings"]
+        rot_pct = self.rope_parameters.get("partial_rotary_factor", 1.0)
         rope_dims = int(rot_pct * n_embd) // n_head
 
         # write rope scaling for long context (128k) model
-        rope_scaling = self.find_hparam(['rope_scaling'], True)
-        if rope_scaling is None:
+        long_factors = self.rope_parameters.get('long_factor')
+        short_factors = self.rope_parameters.get('short_factor')
+        if not long_factors:
             return
 
         scale = max_pos_embds / orig_max_pos_embds
 
-        rope_scaling_type = rope_scaling.get('rope_type', rope_scaling.get('type', '')).lower()
+        rope_scaling_type = self.rope_parameters.get('rope_type', '').lower()
         if len(rope_scaling_type) == 0:
             raise KeyError('Missing the required key rope_scaling.type')
 
@@ -198,9 +199,6 @@ class Phi3MiniModel(TextModel):
 
         self.gguf_writer.add_rope_scaling_attn_factors(attn_factor)
 
-        long_factors = rope_scaling.get('long_factor', None)
-        short_factors = rope_scaling.get('short_factor', None)
-
         if long_factors is None or short_factors is None:
             raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
 

conversion/qwen.py

@@ -280,7 +280,7 @@ class Qwen3NextModel(Qwen2MoeModel):
         self.gguf_writer.add_full_attention_interval(self.hparams.get("full_attention_interval", 4))
         if (rope_dim := self.hparams.get("head_dim")) is None:
             rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
-        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25)))
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.25)))
 
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:

conversion/stablelm.py

@@ -28,7 +28,7 @@ class StableLMModel(TextModel):
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
+        rotary_factor = self.rope_parameters["partial_rotary_factor"]
         self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])

conversion/step3.py

@@ -314,7 +314,7 @@ class Step35Model(TextModel):
         factor = float(rope_params.get("factor", 8.0))
         low_freq_factor = float(rope_params.get("low_freq_factor", 1.0))
         high_freq_factor = float(rope_params.get("high_freq_factor", 4.0))
-        old_context_len = int(rope_params.get("original_max_position_embeddings", self.hparams.get("original_max_position_embeddings", 8192)))
+        old_context_len = int(rope_params.get("original_max_position_embeddings", 8192))
 
         low_freq_wavelen = old_context_len / low_freq_factor
         high_freq_wavelen = old_context_len / high_freq_factor

docs/multimodal.md

@@ -1,10 +1,11 @@
 # Multimodal
 
 llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools support this feature:
-- [llama-mtmd-cli](../tools/mtmd/README.md)
+- [llama-cli](../tools/cli/README.md)
 - [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API
+- [llama-mtmd-cli](../tools/mtmd/README.md), for testing and development
 
-Currently, we support **image** and **audio** input. Audio is highly experimental and may have reduced quality.
+Currently, we support **image**, **audio** and **video** input.
 
 To enable it, you can use one of the 2 methods below:
 

docs/preset.md

@@ -8,55 +8,53 @@ The INI preset feature, introduced in [PR#17859](https://github.com/ggml-org/lla
 
 When running multiple models on the server (router mode), INI preset files can be used to configure model-specific parameters. Please refer to the [server documentation](../tools/server/README.md) for more details.
 
-### Using a Remote Preset
+### Using a Hugging Face Preset
 
-> [!NOTE]
+> [!IMPORTANT]
 >
-> This feature is currently only supported via the `-hf` option.
+> Please only use presets that you can trust! Unknown presets may be unsafe
 
-For GGUF models hosted on Hugging Face, you can include a `preset.ini` file in the root directory of the repository to define specific configurations for that model.
+You can push your preset to Hugging Face Hub and share with other users by:
+1. Creating an empty model repository on Hugging Face
+2. Creating a `preset.ini` file in the root directory of the repository
 
-Example:
+Example of a `preset.ini`:
 
 ```ini
-hf-repo-draft = username/my-draft-model-GGUF
-temp = 0.5
-top-k = 20
-top-p = 0.95
+[*]
+ctx-size             = 0
+mmap                 = 1
+kv-unified           = 1
+parallel             = 4
+spec-default         = 1
+
+[Qwen3.5-4B]
+hf                   = unsloth/Qwen3.5-4B-GGUF:Q4_K_M
+ctx-size             = 262144
+batch-size           = 2048
+ubatch-size          = 2048
+top-p                = 1.0
+top-k                = 0
+min-p                = 0.01
+temp                 = 1.0
+
+[gpt-oss-120b-hf]
+hf                   = ggml-org/gpt-oss-120b-GGUF
+ctx-size             = 262144
+batch-size           = 2048
+ubatch-size          = 2048
+top-p                = 1.0
+top-k                = 0
+min-p                = 0.01
+temp                 = 1.0
+chat-template-kwargs = {"reasoning_effort": "high"}
 ```
 
-For security reasons, only certain options are allowed. Please refer to [preset.cpp](../common/preset.cpp) for the complete list of permitted options.
-
-Example usage:
-
-Assuming your repository `username/my-model-with-preset` contains a `preset.ini` with the configuration above:
-
-```sh
-llama-cli -hf username/my-model-with-preset
-
-# This is equivalent to:
-llama-cli -hf username/my-model-with-preset \
-  --hf-repo-draft username/my-draft-model-GGUF \
-  --temp 0.5 \
-  --top-k 20 \
-  --top-p 0.95
-```
-
-You can also override preset arguments by specifying them on the command line:
+The preset will be loaded similarly to the `--models-preset` option. Therefore, you can also override certain params via CLI arguments:
 
 ```sh
 # Force temp = 0.1, overriding the preset value
-llama-cli -hf username/my-model-with-preset --temp 0.1
-```
-
-If you want to define multiple preset configurations for one or more GGUF models, you can create a blank HF repo for each preset. Each HF repo should contain a `preset.ini` file that references the actual model(s):
-
-```ini
-hf-repo = user/my-model-main
-hf-repo-draft = user/my-model-draft
-temp = 0.8
-ctx-size = 1024
-; (and other configurations)
+llama-cli -hf username/my-preset --temp 0.1
 ```
 
 ### Named presets

ggml/CMakeLists.txt

@@ -5,7 +5,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 15)
-set(GGML_VERSION_PATCH 1)
+set(GGML_VERSION_PATCH 2)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -2417,15 +2417,14 @@ void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_te
             // Q4_K, Q5_K, Q6_K, IQ4_XS handles 8 TILE_K per blck_size
             GGML_ASSERT(TILE_K == blck_size || TILE_K * 8 == blck_size);
 
-            parallel_for_ggml(params, n_batch, [&](int begin, int end) {
-                for (int batch_idx = begin; batch_idx < end; ++batch_idx) {
+            parallel_for_ggml(params, n_batch * M, [&](int begin, int end) {
+                for (int idx = begin; idx < end; ++idx) {
+                    int batch_idx = idx / M;
+                    int m         = idx % M;
                     int64_t src1_offset = ggml_batch_offset(src1, batch_idx, ne2);
                     const float * A_data = (const float *)((const char *)src1->data + src1_offset);
                     char * wdata_batch = (char *)wdata + batch_idx * M * row_size_A;
-
-                    for (int m = 0; m < M; ++m) {
-                        from_float<vec_dot_type>(A_data + m * K, wdata_batch + m * row_size_A, K);
-                    }
+                    from_float<vec_dot_type>(A_data + m * K, wdata_batch + m * row_size_A, K);
                 }
             });
         });

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

@@ -2345,7 +2345,7 @@ class tinyBLAS_Q0_PPC {
             else if (n_aligned % 16 == 0) nc = 16;
             else                          nc = 8;
         }
-        bool can_use_tiled = n_aligned > 0 && (m % mc == 0) && (k % kc == 0);
+        bool can_use_tiled = n_aligned > 0 && (m % mc == 0);
         if (can_use_tiled) {
             matmul_tiled(m, n_aligned, mc, nc, kc);
             if (n > n_aligned) {
@@ -3063,13 +3063,14 @@ class tinyBLAS_Q0_PPC {
             int64_t ii = (job / xtiles) * mc;
             int64_t jj = (job % xtiles) * nc;
             for (int64_t kk = 0; kk < k; kk += kc) {
+                int64_t k_cur = MIN(kc, k - kk);
                 if constexpr(is_Ablock_q4) {
-                    packNormal_q4_fp16(A + ii * lda + kk, lda, mc, kc, (uint8_t *)A_pack);
+                    packNormal_q4_fp16(A + ii * lda + kk, lda, mc, k_cur, (uint8_t *)A_pack);
                 } else {
-                    packNormal_q8_fp16(A + ii * lda + kk, lda, mc, kc, (uint8_t *)A_pack);
+                    packNormal_q8_fp16(A + ii * lda + kk, lda, mc, k_cur, (uint8_t *)A_pack);
                 }
-                packNormal_q8_fp16(B + jj * ldb + kk, ldb, nc, kc, (uint8_t *)B_pack);
-                KERNEL_Q0(ii, jj, mc, nc, kc, kk, A_pack, B_pack);
+                packNormal_q8_fp16(B + jj * ldb + kk, ldb, nc, k_cur, (uint8_t *)B_pack);
+                KERNEL_Q0(ii, jj, mc, nc, k_cur, kk, A_pack, B_pack);
             }
         }
     }

ggml/src/ggml-cuda/col2im-1d.cu

@@ -0,0 +1,81 @@
+#include "col2im-1d.cuh"
+#include "convert.cuh"
+
+// col2im_1d: scatter-add GEMM columns to 1D signal (gather approach)
+// columns: [K*OC, T_in]  ->  output: [T_out, OC]
+// Supports F32, F16, BF16 data with F32 accumulator.
+
+template <typename T>
+static __global__ void col2im_1d_kernel(
+        const T * __restrict__ col,
+        T       * __restrict__ dst,
+        const int T_in, const uint3 T_out_fd,
+        const int OC, const int K, const int K_OC,
+        const int s0, const int p0, const int total) {
+
+    const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (idx >= total) return;
+
+    // dst layout: [T_out, OC], ne[0]=T_out fastest
+    const uint2 qr  = fast_div_modulo((uint32_t)idx, T_out_fd);  // qr.x = idx / T_out, qr.y = idx % T_out
+    const int oc    = (int)qr.x;
+    const int t_out = (int)qr.y;
+    const int t_abs = t_out + p0;  // absolute position in uncropped signal
+
+    // Gather: find all (t_in, k) where t_in*s + k == t_abs, 0 <= k < K
+    int t_in_min = (t_abs - K + s0) / s0;  // ceil((t_abs - K + 1) / s)
+    if (t_in_min < 0) t_in_min = 0;
+    int t_in_max = t_abs / s0;
+    if (t_in_max >= T_in) t_in_max = T_in - 1;
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * s0;
+        // col layout: [K*OC, T_in], column index = oc * K + k
+        sum += ggml_cuda_cast<float>(col[(oc * K + k) + t_in * K_OC]);
+    }
+
+    dst[idx] = ggml_cuda_cast<T>(sum);
+}
+
+void ggml_cuda_op_col2im_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+    const int32_t OC = ((const int32_t *)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
+
+    const int K_OC = (int) src0->ne[0];
+    const int T_in = (int) src0->ne[1];
+    const int K    = K_OC / OC;
+    const int T_out = (int) dst->ne[0];
+
+    const uint3 T_out_fd = init_fastdiv_values((uint32_t)T_out);
+
+    const int total = T_out * OC;
+    const int block_size = 256;
+    const int num_blocks = (total + block_size - 1) / block_size;
+
+    switch (src0->type) {
+        case GGML_TYPE_F32: {
+            col2im_1d_kernel<<<num_blocks, block_size, 0, stream>>>(
+                (const float *)src0->data, (float *)dst->data,
+                T_in, T_out_fd, OC, K, K_OC, s0, p0, total);
+        } break;
+        case GGML_TYPE_F16: {
+            col2im_1d_kernel<<<num_blocks, block_size, 0, stream>>>(
+                (const half *)src0->data, (half *)dst->data,
+                T_in, T_out_fd, OC, K, K_OC, s0, p0, total);
+        } break;
+        case GGML_TYPE_BF16: {
+            col2im_1d_kernel<<<num_blocks, block_size, 0, stream>>>(
+                (const nv_bfloat16 *)src0->data, (nv_bfloat16 *)dst->data,
+                T_in, T_out_fd, OC, K, K_OC, s0, p0, total);
+        } break;
+        default:
+            GGML_ABORT("col2im_1d: unsupported type");
+    }
+}

ggml/src/ggml-cuda/col2im-1d.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_col2im_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -11,6 +11,7 @@
 #include "ggml-cuda/argsort.cuh"
 #include "ggml-cuda/binbcast.cuh"
 #include "ggml-cuda/clamp.cuh"
+#include "ggml-cuda/col2im-1d.cuh"
 #include "ggml-cuda/concat.cuh"
 #include "ggml-cuda/conv-transpose-1d.cuh"
 #include "ggml-cuda/conv2d.cuh"
@@ -3051,6 +3052,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_CONV_TRANSPOSE_1D:
             ggml_cuda_op_conv_transpose_1d(ctx,dst);
             break;
+        case GGML_OP_COL2IM_1D:
+            ggml_cuda_op_col2im_1d(ctx, dst);
+            break;
         case GGML_OP_POOL_2D:
             ggml_cuda_op_pool2d(ctx, dst);
             break;
@@ -5316,6 +5320,14 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 }
                 return false;
             } break;
+        case GGML_OP_COL2IM_1D:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                return (src0_type == GGML_TYPE_F32 || src0_type == GGML_TYPE_F16 || src0_type == GGML_TYPE_BF16) &&
+                    op->type == src0_type &&
+                    ggml_is_contiguous(op->src[0]) &&
+                    ggml_is_contiguous(op);
+            } break;
         case GGML_OP_SILU_BACK:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
             break;

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -69,6 +69,7 @@ static int opt_opstage  = HTP_OPSTAGE_QUEUE | HTP_OPSTAGE_COMPUTE;
 static int opt_opbatch  = 1024; // max number of ops in a batch
 static int opt_opqueue  = 16;   // max number of pending batches
 static int opt_oppoll   = 0;    // polling for batch completions
+static int opt_optrace  = 0;    // trace buffer size per thread (0 means default)
 
 static std::regex* opt_opfilter = NULL; // regex of ops to not claim
 
@@ -118,20 +119,39 @@ static void ggml_hexagon_dump_op_supp(const std::string &sess_name, const struct
                 ggml_op_desc(op), fmt.names, fmt.dims, fmt.types, fmt.strides, fmt.buffs, supp ? "yes" : "no");
 }
 
+static const char * htp_event_name(uint16_t id) {
+    switch (id) {
+        case HTP_TRACE_EVT_DMA:            return "DMA";
+        case HTP_TRACE_EVT_HVX_COMP:       return "HVX_COMP";
+        case HTP_TRACE_EVT_HVX_A_QUANT:    return "HVX_A_QUANT";
+        case HTP_TRACE_EVT_HVX_A_PREP:     return "HVX_A_PREP";
+        case HTP_TRACE_EVT_HVX_W_DEQUANT:  return "HVX_W_DEQUANT";
+        case HTP_TRACE_EVT_HVX_W_PREP:     return "HVX_W_PREP";
+        case HTP_TRACE_EVT_HVX_O_PROC:     return "HVX_O_PROC";
+        case HTP_TRACE_EVT_HMX_COMP:       return "HMX_COMP";
+        default:                           return "UNKNOWN";
+    }
+}
+
 static void ggml_hexagon_dump_op_prof(const std::string &sess_name, const htp_opnode & node,
-                                      uint32_t op_usec, uint32_t op_cycles, const uint32_t pmu[]) {
+                                      const htp_prof_desc & pd) {
     if (!opt_profile) return;
 
+    uint32_t op_usec = pd.usecs;
+    uint32_t op_cycles = pd.cycles_stop - pd.cycles_start;
+    const uint32_t * pmu = pd.pmu;
+
     char pmu_str[256] = "";
-    if (opt_profile > 1) {
+    if (opt_profile == 2) {
         static_assert(HTP_PROF_PMU_NCNT == 8, "current implementation assumes 8 PMU counters");
         sprintf(pmu_str, " pmu [%u,%u,%u,%u,%u,%u,%u,%u]",
                 pmu[0], pmu[1], pmu[2], pmu[3], pmu[4], pmu[5], pmu[6], pmu[7]);
     }
 
     htp_opformat fmt(node);
-    GGML_LOG_DEBUG("ggml-hex: %s profile-op %s: %s : %s : %s : %s : usec %u cycles %u%s\n", sess_name.c_str(),
-            node.op_name().c_str(), fmt.names, fmt.dims, fmt.types, fmt.strides, op_usec, op_cycles, pmu_str);
+    float mhz = op_usec > 0 ? (float) op_cycles / op_usec : 0.0f;
+    GGML_LOG_DEBUG("ggml-hex: %s profile-op %s: %s : %s : %s : %s : usec %u cycles %u start %u mhz %.1f%s\n", sess_name.c_str(),
+            node.op_name().c_str(), fmt.names, fmt.dims, fmt.types, fmt.strides, op_usec, op_cycles, pd.cycles_start, mhz, pmu_str);
 }
 
 // ** backend sessions
@@ -1995,10 +2015,16 @@ struct ggml_hexagon_opqueue {
         size_t n_ops     = batch_size;
         size_t n_tensors = n_ops + n_ops * HTP_OP_MAX_INPUTS;
 
+        size_t tr_size = 0;
+        if (opt_profile == 3) {
+            tr_size = (HTP_MAX_NTHREADS + 1) * opt_optrace * sizeof(htp_trace_desc);
+        }
+
         shm_blk_size = sizeof(htp_buf_desc)  * n_bufs    +
                        sizeof(htp_tensor)    * n_tensors +
                        sizeof(htp_op_desc)   * n_ops     +
-                       sizeof(htp_prof_desc) * n_ops;
+                       sizeof(htp_prof_desc) * n_ops     +
+                       tr_size;
 
         shm_buf = new ggml_hexagon_shared_buffer(sess, shm_blk_size * depth, true /* pinned */);
 
@@ -2042,11 +2068,19 @@ struct ggml_hexagon_opqueue {
         const size_t o_size = sizeof(htp_op_desc)   * req.n_ops;
         const size_t p_size = sizeof(htp_prof_desc) * req.n_ops;
 
+        size_t tr_size = 0;
+        if (opt_profile == 3) {
+            req.n_traces = opt_optrace;
+            tr_size = (HTP_MAX_NTHREADS + 1) * req.n_traces * sizeof(htp_trace_desc);
+        } else {
+            req.n_traces = 0;
+        }
+
         dbuf.ptr      = shm_buf->base + (req.id * shm_blk_size);
         dbuf.fd       = shm_buf->fd;
         dbuf.flags    = DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT;
         dbuf.offset   = (uint8_t*) dbuf.ptr - (uint8_t*) shm_buf->base;
-        dbuf.size     = b_size + t_size + o_size + p_size;
+        dbuf.size     = b_size + t_size + o_size + p_size + tr_size;
 
         GGML_ASSERT(dbuf.size <= shm_blk_size);
 
@@ -2092,7 +2126,14 @@ struct ggml_hexagon_opqueue {
         const size_t o_size = sizeof(htp_op_desc)   * rsp.n_ops;
         const size_t p_size = sizeof(htp_prof_desc) * rsp.n_ops;
 
-        const size_t m_size = b_size + t_size + o_size + p_size;
+        size_t tr_size = 0;
+        uint32_t n_traces = 0;
+        if (opt_profile == 3) {
+            n_traces = opt_optrace;
+            tr_size = (HTP_MAX_NTHREADS + 1) * n_traces * sizeof(htp_trace_desc);
+        }
+
+        const size_t m_size = b_size + t_size + o_size + p_size + tr_size;
         GGML_ASSERT(m_size <= shm_blk_size);
 
         HEX_VERBOSE("ggml-hex: %s op-queue pop batch #%u : n-bufs %u n-tensors %u n-ops %u : m-size %zu b-size %zu t-size %zu o-size %zu\n",
@@ -2111,13 +2152,62 @@ struct ggml_hexagon_opqueue {
             GGML_ASSERT(rsp.n_ops <= ops.size());
 
             const htp_prof_desc * pd = (const htp_prof_desc *) p_ptr;
-            for (uint32_t i = 0; i < rsp.n_ops; i++) {
-                htp_usec += pd[i].usecs;
-                ggml_hexagon_dump_op_prof(shm_buf->sess->name, ops[i], pd[i].usecs, pd[i].cycles, pd[i].pmu);
+
+            const htp_trace_desc * trace_events = nullptr;
+
+            if (opt_profile == 3) {
+                trace_events = (const htp_trace_desc *) (p_ptr + p_size);
             }
 
-            GGML_LOG_DEBUG("ggml-hex: %s profile-batch n-ops %u batch-dur-usec %lld htp-ops-usec %u\n",
-                           shm_buf->sess->c_name(), rsp.n_ops, (long long) batch_usec, htp_usec);
+            uint32_t trace_idx[HTP_MAX_NTHREADS + 1] = {0};
+            uint32_t valid_cnt[HTP_MAX_NTHREADS + 1] = {0};
+
+            if (opt_profile == 3) {
+                for (uint32_t t = 0; t <= HTP_MAX_NTHREADS; t++) {
+                    uint32_t count = rsp.n_traces[t];
+                    valid_cnt[t] = count > n_traces ? n_traces : count;
+                }
+            }
+
+            for (uint32_t i = 0; i < rsp.n_ops; i++) {
+                htp_usec += pd[i].usecs;
+
+                ggml_hexagon_dump_op_prof(shm_buf->sess->name, ops[i], pd[i]);
+
+                if (opt_profile == 3) {
+                    uint32_t op_duration = pd[i].cycles_stop - pd[i].cycles_start;
+
+                    for (uint32_t t = 0; t <= HTP_MAX_NTHREADS; t++) {
+                        while (trace_idx[t] < valid_cnt[t]) {
+                            const auto & e = trace_events[t * n_traces + trace_idx[t]];
+                            uint32_t offset = e.cycles - pd[i].cycles_start;
+                            if (offset >= 0x80000000) {
+                                trace_idx[t]++;
+                                continue;
+                            }
+                            if (offset > op_duration) {
+                                break;
+                            }
+                            bool is_stop = (e.info & 0x8000) != 0;
+                            uint16_t info = e.info & 0x7FFF;
+                            GGML_LOG_DEBUG("ggml-hex: %s trace-op %s: thread %u event %s info %u %s %u\n",
+                                           shm_buf->sess->c_name(), ops[i].op_name().c_str(), t, htp_event_name(e.id), info, is_stop ? "stop" : "start", e.cycles);
+                            trace_idx[t]++;
+                        }
+                    }
+                }
+            }
+
+            char evt_str[256] = "";
+            if (opt_profile == 3) {
+                sprintf(evt_str, " evt [%u,%u,%u,%u,%u,%u,%u,%u,%u,%u,%u]",
+                        rsp.n_traces[0], rsp.n_traces[1], rsp.n_traces[2], rsp.n_traces[3],
+                        rsp.n_traces[4], rsp.n_traces[5], rsp.n_traces[6], rsp.n_traces[7],
+                        rsp.n_traces[8], rsp.n_traces[9], rsp.n_traces[10]);
+            }
+
+            GGML_LOG_DEBUG("ggml-hex: %s profile-batch n-ops %u batch-dur-usec %lld htp-ops-usec %u%s\n",
+                           shm_buf->sess->c_name(), rsp.n_ops, (long long) batch_usec, htp_usec, evt_str);
         }
     }
 };
@@ -3901,6 +3991,7 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
     const char * str_opbatch  = getenv("GGML_HEXAGON_OPBATCH");
     const char * str_opqueue  = getenv("GGML_HEXAGON_OPQUEUE");
     const char * str_oppoll   = getenv("GGML_HEXAGON_OPPOLL");
+    const char * str_optrace  = getenv("GGML_HEXAGON_OPTRACE");
     const char * str_opfilter = getenv("GGML_HEXAGON_OPFILTER");
     const char * str_profile  = getenv("GGML_HEXAGON_PROFILE");
     const char * str_etm      = getenv("GGML_HEXAGON_ETM");
@@ -3939,6 +4030,7 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
     opt_opbatch   = str_opbatch  ? strtoul(str_opbatch, NULL, 0)          : opt_opbatch;
     opt_opqueue   = str_opqueue  ? strtoul(str_opqueue, NULL, 0)          : opt_opqueue;
     opt_oppoll    = str_oppoll   ? strtoul(str_oppoll,  NULL, 0)          : opt_oppoll;
+    opt_optrace   = str_optrace  ? strtoul(str_optrace, NULL, 0)          : (opt_opbatch * 128);
     opt_profile   = str_profile  ? atoi(str_profile)                      : 0;
     opt_etm       = str_etm      ? atoi(str_etm)                          : 0;
     opt_nhvx      = str_nhvx     ? strtoul(str_nhvx, NULL, 0)             : opt_nhvx;

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

@@ -37,8 +37,8 @@ list(FIND HTP_HMX_VERSIONS ${DSP_VERSION} _hmx_idx)
 
 if (_hmx_idx GREATER_EQUAL 0)
     target_sources(${HTP_LIB} PRIVATE
-        hmx-matmul-ops.c
         hmx-flash-attn-ops.c
+        hmx-matmul-ops.c
         hmx-queue.c
     )
 

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

@@ -339,6 +339,9 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
 
     if (ir0 >= ir1) return;
 
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
+
     dma_queue * dma = octx->ctx->dma[ith];
 
     const uint32_t DK = nek0;
@@ -615,6 +618,7 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
             hvx_copy_f16_f32_ua(dst_ptr, (uint8_t *) VKQ32, DV);
         }
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
 }
 
 int op_flash_attn_ext(struct htp_ops_context * octx) {

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

@@ -6,6 +6,8 @@
 #include <stdbool.h>
 #include <stdint.h>
 
+#include "hex-profile.h"
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -88,6 +90,7 @@ typedef struct {
     uint32_t            pop_idx;
     uint32_t            capacity;
     uint32_t            idx_mask;
+    struct htp_thread_trace * trace;
 } dma_queue;
 
 dma_queue * dma_queue_create(size_t capacity);
@@ -152,6 +155,7 @@ static inline bool dma_queue_push_single_1d(dma_queue * q, dma_ptr dptr, size_t
     q->dptr[q->push_idx] = dptr;
 
     if (size) {
+        htp_trace_event_start(q->trace, HTP_TRACE_EVT_DMA, q->push_idx);
         dmlink(q->tail, desc);
         q->tail = (dma_descriptor_2d *) desc;
     } else {
@@ -202,6 +206,7 @@ static inline bool dma_queue_push_single_2d(dma_queue * q, dma_ptr dptr, size_t
     q->dptr[q->push_idx] = dptr;
 
     if (nrows) {
+        htp_trace_event_start(q->trace, HTP_TRACE_EVT_DMA, q->push_idx);
         dmlink(q->tail, desc);
         q->tail = desc;
     } else {
@@ -223,10 +228,12 @@ static inline dma_ptr dma_queue_pop(dma_queue * q) {
     dma_descriptor_2d * desc = &q->desc[q->pop_idx];
 
     // Wait for desc to complete
-    while (!desc->done) {
-        // FARF(ERROR, "dma-pop: waiting for DMA : %u\n", q->pop_idx);
-        dmpoll();
+    if (!desc->done) {
+        while (!desc->done) {
+            dmpoll();
+        }
     }
+    htp_trace_event_stop(q->trace, HTP_TRACE_EVT_DMA, q->pop_idx);
 
     dptr = q->dptr[q->pop_idx];
 

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

@@ -0,0 +1,64 @@
+#ifndef HEX_PROFILE_H
+#define HEX_PROFILE_H
+
+#include <stdbool.h>
+#include <stdint.h>
+#include <qurt.h>
+
+#include "hex-utils.h"
+#include "htp-ops.h"
+
+#define HTP_TRACE_EVT_START 0
+#define HTP_TRACE_EVT_STOP  1
+
+#ifndef HEX_NUM_PMU_COUNTERS
+#define HEX_NUM_PMU_COUNTERS 8
+#endif
+
+static inline void hex_get_pmu(uint32_t counters[]) {
+#if __HVX_ARCH__ >= 79
+    asm volatile("%0 = upmucnt0" : "=r"(counters[0]));
+    asm volatile("%0 = upmucnt1" : "=r"(counters[1]));
+    asm volatile("%0 = upmucnt2" : "=r"(counters[2]));
+    asm volatile("%0 = upmucnt3" : "=r"(counters[3]));
+    asm volatile("%0 = upmucnt4" : "=r"(counters[4]));
+    asm volatile("%0 = upmucnt5" : "=r"(counters[5]));
+    asm volatile("%0 = upmucnt6" : "=r"(counters[6]));
+    asm volatile("%0 = upmucnt7" : "=r"(counters[7]));
+#else
+    counters[0] = qurt_pmu_get(QURT_PMUCNT0);
+    counters[1] = qurt_pmu_get(QURT_PMUCNT1);
+    counters[2] = qurt_pmu_get(QURT_PMUCNT2);
+    counters[3] = qurt_pmu_get(QURT_PMUCNT3);
+    counters[4] = qurt_pmu_get(QURT_PMUCNT4);
+    counters[5] = qurt_pmu_get(QURT_PMUCNT5);
+    counters[6] = qurt_pmu_get(QURT_PMUCNT6);
+    counters[7] = qurt_pmu_get(QURT_PMUCNT7);
+#endif
+}
+
+struct htp_thread_trace {
+    uint32_t count;
+    uint32_t max_events;
+    struct htp_trace_desc * events;
+};
+
+static inline void htp_trace_event(struct htp_thread_trace * tr, uint16_t id, uint16_t info, uint32_t type) {
+    if (tr && tr->events && tr->count < tr->max_events) {
+        uint32_t idx = tr->count;
+        tr->events[idx].id = id;
+        tr->events[idx].info = info | (type == HTP_TRACE_EVT_STOP ? 0x8000 : 0);
+        tr->events[idx].cycles = (uint32_t) hex_get_cycles();
+        tr->count++;
+    }
+}
+
+static inline void htp_trace_event_start(struct htp_thread_trace * tr, uint16_t id, uint16_t info) {
+    htp_trace_event(tr, id, info, HTP_TRACE_EVT_START);
+}
+
+static inline void htp_trace_event_stop(struct htp_thread_trace * tr, uint16_t id, uint16_t info) {
+    htp_trace_event(tr, id, info, HTP_TRACE_EVT_STOP);
+}
+
+#endif /* HEX_PROFILE_H */

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

@@ -107,31 +107,4 @@ static inline void hex_pause() {
     asm volatile(" pause(#255)\n");
 }
 
-#ifndef HEX_NUM_PMU_COUNTERS
-#define HEX_NUM_PMU_COUNTERS 8
-#endif
-
-static inline void hex_get_pmu(uint32_t counters[]) {
-#if __HVX_ARCH__ >= 79
-    asm volatile("%0 = upmucnt0" : "=r"(counters[0]));
-    asm volatile("%0 = upmucnt1" : "=r"(counters[1]));
-    asm volatile("%0 = upmucnt2" : "=r"(counters[2]));
-    asm volatile("%0 = upmucnt3" : "=r"(counters[3]));
-    asm volatile("%0 = upmucnt4" : "=r"(counters[4]));
-    asm volatile("%0 = upmucnt5" : "=r"(counters[5]));
-    asm volatile("%0 = upmucnt6" : "=r"(counters[6]));
-    asm volatile("%0 = upmucnt7" : "=r"(counters[7]));
-#else
-    counters[0] = qurt_pmu_get(QURT_PMUCNT0);
-    counters[1] = qurt_pmu_get(QURT_PMUCNT1);
-    counters[2] = qurt_pmu_get(QURT_PMUCNT2);
-    counters[3] = qurt_pmu_get(QURT_PMUCNT3);
-    counters[4] = qurt_pmu_get(QURT_PMUCNT4);
-    counters[5] = qurt_pmu_get(QURT_PMUCNT5);
-    counters[6] = qurt_pmu_get(QURT_PMUCNT6);
-    counters[7] = qurt_pmu_get(QURT_PMUCNT7);
-    // qurt_pmu_get_pmucnt(counters);
-#endif
-}
-
 #endif /* HEX_UTILS_H */

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

@@ -18,7 +18,7 @@
 #include "ggml-common.h"
 #include "hex-dma.h"
 #include "hex-fastdiv.h"
-#include "hmx-profile.h"
+#include "hex-profile.h"
 #include "hmx-queue.h"
 #include "hmx-utils.h"
 #include "htp-ctx.h"
@@ -367,8 +367,11 @@ static void fa_k_interleave_thread(unsigned int n, unsigned int i, void * data)
         return;
     }
 
+    struct htp_thread_trace * tr = factx->octx->ctx ? &factx->octx->ctx->trace[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, start);
     hmx_interleave_rows_to_tiles(factx->vtcm_k_tiles, factx->vtcm_k_fp16[args->buf_idx], total_rows, (int) factx->DK,
                              (int) args->src_stride, start, end);
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, start);
 }
 
 static void fa_phase_k_interleave(struct hmx_fa_context * factx, int kv_rows, size_t src_stride, size_t buf_idx) {
@@ -408,8 +411,11 @@ static void fa_v_interleave_thread(unsigned int n, unsigned int i, void * data)
 
     __fp16 * v_tiles_dest = factx->use_pipeline ? factx->vtcm_v_tiles[args->buf_idx] : factx->vtcm_v_tiles[0];
 
+    struct htp_thread_trace * tr = factx->octx->ctx ? &factx->octx->ctx->trace[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, start);
     hmx_interleave_cols_to_tiles(v_tiles_dest, factx->vtcm_v_fp16[args->buf_idx], total_rows, (int) factx->DV,
                              (int) args->src_stride, (int) args->n_col_tiles, start, end);
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, start);
 }
 
 static void fa_phase_v_interleave(struct hmx_fa_context * factx,
@@ -462,6 +468,9 @@ static void fa_q_load_thread(unsigned int n, unsigned int i, void * data) {
         return;
     }
 
+    struct htp_thread_trace * tr = factx->octx->ctx ? &factx->octx->ctx->trace[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, start);
+
     const struct htp_tensor * q       = args->q;
     const uint32_t            q_start = args->q_start;
     const uint32_t            kv_head = args->kv_head;
@@ -515,6 +524,7 @@ static void fa_q_load_thread(unsigned int n, unsigned int i, void * data) {
             }
         }
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, start);
 }
 
 static void fa_phase_q_load(struct hmx_fa_context *   factx,
@@ -566,6 +576,9 @@ static void fa_o_store_thread(unsigned int n, unsigned int i, void * data) {
         return;
     }
 
+    struct htp_thread_trace * tr = factx->octx->ctx ? &factx->octx->ctx->trace[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, start);
+
     const struct htp_tensor * dst        = args->dst;
     const __fp16 *            o_tile_src = args->o_tile_src;
     const uint32_t            q_start    = args->q_start;
@@ -611,6 +624,7 @@ static void fa_o_store_thread(unsigned int n, unsigned int i, void * data) {
             }
         }
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, start);
 }
 
 static void fa_phase_o_store(struct hmx_fa_context *   factx,
@@ -680,6 +694,9 @@ static void fa_softmax_thread(unsigned int n, unsigned int i, void * data) {
         return;
     }
 
+    struct htp_thread_trace * tr = factx->octx->ctx ? &factx->octx->ctx->trace[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, vec_start);
+
     // Per-thread row scratch: thread i uses bufs at offset i * 2 * stride
     const size_t row_buf_stride = factx->row_buf_stride;
     HVX_Vector * my_row_buf0    = factx->vtcm_row_bufs + i * 2 * row_buf_stride;
@@ -950,6 +967,7 @@ static void fa_softmax_thread(unsigned int n, unsigned int i, void * data) {
         factx->vtcm_s_rowmax[r_vec_idx] = rowmax_acc_v;
         factx->vtcm_p_rowsum[r_vec_idx] = rowsum_acc_v;
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, vec_start);
 }
 
 // Serial m/l update + build_D.  Must run after softmax barrier (s_rowmax written by all threads).
@@ -1245,6 +1263,7 @@ static __attribute__((noinline)) void fa_compute_slopes(
 // ============================================================================
 
 int hmx_flash_attn_ext(struct htp_ops_context * octx) {
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[HTP_MAX_NTHREADS] : NULL;
     const struct htp_tensor * q    = octx->src[0];
     const struct htp_tensor * k    = octx->src[1];
     const struct htp_tensor * v    = octx->src[2];
@@ -1422,19 +1441,6 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
-    // Profiling timers
-    TIMER_DEFINE(total);
-    TIMER_DEFINE(q_load);
-    TIMER_DEFINE(kv_dma);
-    TIMER_DEFINE(k_interleave);
-    TIMER_DEFINE(v_interleave);
-    TIMER_DEFINE(qk_dot);
-    TIMER_DEFINE(softmax);
-    TIMER_DEFINE(o_update);
-    TIMER_DEFINE(o_norm);
-    TIMER_DEFINE(o_store);
-
-    TIMER_START(total);
 
     // ======== DMA setup ========
     dma_queue * const dma = ctx->dma[0];
@@ -1474,12 +1480,10 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                 const size_t   n_row_tiles = g_br_actual / HMX_FP16_TILE_N_ROWS;
 
                 // ---- Load Q block [g_br, D] -> tiles, interleaving G heads ----
-                TIMER_START(q_load);
                 if (n_rows_g < g_br) {
                     hvx_splat_u8_a(factx.vtcm_q_tiles, 0, q_tile_bytes);
                 }
                 fa_phase_q_load(&factx, q, q_start, kv_head, ib3, n_rows_g);
-                TIMER_STOP(q_load);
 
                 // ---- Initialize per-block state ----
                 hvx_splat_u8_a(factx.vtcm_l_vec,   0,      col_vec_bytes);
@@ -1558,10 +1562,8 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         const size_t   n_col_tiles = hmx_ceil_div(kv_rows, HMX_FP16_TILE_N_COLS);
 
                         // Wait for current KV DMA
-                        TIMER_START(kv_dma);
                         dma_queue_pop(dma);  // K
                         dma_queue_pop(dma);  // V
-                        TIMER_STOP(kv_dma);
 
                         // Push mask DMA for this block (single 2D DMA when broadcast)
                         bool has_mask_dma = false;
@@ -1583,10 +1585,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                             ou_job.DV               = DV;
                             hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job));
                         }
-
-                        TIMER_START(k_interleave);
                         fa_phase_k_interleave(&factx, kv_rows, k_src_stride, buf_idx);
-                        TIMER_STOP(k_interleave);
 
                         // ---- Phase 2: qk_dot(blk) on HMX ‖ V_int(blk) + DMA prefetch on HVX ----
                         qk_job.q_tiles        = factx.vtcm_q_tiles;
@@ -1597,15 +1596,11 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         qk_job.n_dot_tiles    = DK / 32;
                         qk_job.n_tiles_per_bc = n_tiles_per_bc;
                         qk_job.hmx_scales     = factx.vtcm_hmx_scales_qk;
-                        TIMER_START(qk_dot);
                         hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_qk_dot_worker, &qk_job));
 
                         // DMA push next block (non-blocking, before worker_pool)
                         DMA_PREFETCH_KV(kv_blk + 1);
-
-                        TIMER_START(v_interleave);
                         fa_phase_v_interleave(&factx, kv_rows, v_src_stride, buf_idx, n_tiles_per_bc);
-                        TIMER_STOP(v_interleave);
 
                         // Pop and swap previous block's output update (deferred HMX pop)
                         if (kv_blk > 0) {
@@ -1615,7 +1610,6 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
 
                         // Pop current block's dot product job
                         hmx_queue_pop(hmx_q);
-                        TIMER_STOP(qk_dot);
 
                         // ---- Phase 3: softmax(blk) + build_D(blk) | HMX idle ----
                         // Pop mask DMA before softmax (ensures VTCM buffer is ready)
@@ -1641,10 +1635,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         sargs.mask_vtcm            = has_mask_dma ? (const __fp16 *) factx.vtcm_mask_buf : NULL;
                         sargs.mask_vtcm_row_stride = factx.mask_buf_row_stride;
                         sargs.slopes               = factx.vtcm_slopes;
-
-                        TIMER_START(softmax);
                         fa_phase_softmax_and_build_d(&factx, &sargs, n_row_tiles, n_row_tiles_g_br);
-                        TIMER_STOP(softmax);
 
                         buf_idx = 1 - buf_idx;
                     }  // end KV block loop (pipeline)
@@ -1664,11 +1655,8 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         ou_job.n_row_tiles_g_br = n_row_tiles_g_br;
                         ou_job.n_tiles_per_bc   = n_tiles_per_bc;
                         ou_job.DV               = DV;
-
-                        TIMER_START(o_update);
                         hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job));
                         hmx_queue_pop(hmx_q);
-                        TIMER_STOP(o_update);
 
                         hex_swap_ptr((void **) &o_tile_curr, (void **) &o_tile_prev);
                     }
@@ -1683,23 +1671,14 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         const uint32_t kv_start    = kv_blk * Bc;
                         const uint32_t kv_rows     = hex_smin(Bc, nek1 - kv_start);
                         const size_t   n_col_tiles = hmx_ceil_div(kv_rows, HMX_FP16_TILE_N_COLS);
-
-                        TIMER_START(kv_dma);
                         dma_queue_pop(dma);  // K
                         dma_queue_pop(dma);  // V
-                        TIMER_STOP(kv_dma);
 
                         bool has_mask_dma = false;
                         MASK_DMA_PUSH(kv_start, kv_rows, has_mask_dma);
                         DMA_PREFETCH_KV(kv_blk + 1);
-
-                        // K interleave (multi-thread HVX)
-                        TIMER_START(k_interleave);
                         fa_phase_k_interleave(&factx, kv_rows, k_src_stride, buf_idx);
-                        TIMER_STOP(k_interleave);
 
-                        // QK dot (inline HMX on main thread)
-                        TIMER_START(qk_dot);
                         {
                             const size_t n_dot_tiles       = (size_t) (DK / 32);
                             const __fp16 * restrict q_base = factx.vtcm_q_tiles;
@@ -1709,6 +1688,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                             __builtin_assume(n_col_tiles > 0);
                             __builtin_assume(n_dot_tiles > 0);
 
+                            htp_trace_event_start(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
                             Q6_bias_mxmem2_A((void *) factx.vtcm_hmx_scales_qk);
                             for (size_t r = 0; r < n_row_tiles; ++r) {
                                 for (size_t c = 0; c < n_col_tiles; ++c) {
@@ -1724,8 +1704,8 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                                     Q6_mxmem_AR_after_hf(out_tile, 0);
                                 }
                             }
+                            htp_trace_event_stop(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
                         }
-                        TIMER_STOP(qk_dot);
 
                         // Pop mask DMA
                         MASK_DMA_POP(has_mask_dma);
@@ -1751,21 +1731,9 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         sargs.mask_vtcm            = has_mask_dma ? (const __fp16 *) factx.vtcm_mask_buf : NULL;
                         sargs.mask_vtcm_row_stride = factx.mask_buf_row_stride;
                         sargs.slopes               = factx.vtcm_slopes;
-
-                        TIMER_START(softmax);
                         fa_phase_softmax_and_build_d(&factx, &sargs, n_row_tiles, n_row_tiles_g_br);
-                        TIMER_STOP(softmax);
-
-                        // V interleave (multi-thread HVX)
-                        TIMER_START(v_interleave);
-                        // FIX(v-stride): use n_tiles_per_bc (block-invariant) as V tile layout
-                        // stride to match o_update's v_tile access.  Using per-block n_col_tiles
-                        // misplaces DV_tile 1..3 in the last partial KV block.
                         fa_phase_v_interleave(&factx, kv_rows, v_src_stride, buf_idx, n_tiles_per_bc);
-                        TIMER_STOP(v_interleave);
 
-                        // O update (inline HMX on main thread)
-                        TIMER_START(o_update);
                         {
                             const size_t DV_tiles           = (size_t) (DV / 32);
                             const __fp16 * restrict d_base  = factx.vtcm_d_tiles;
@@ -1777,6 +1745,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                             __builtin_assume(n_col_tiles > 0);
                             __builtin_assume(DV_tiles > 0);
 
+                            htp_trace_event_start(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
                             Q6_bias_mxmem2_A((void *) factx.vtcm_hmx_scales_id);
                             for (size_t r = 0; r < n_row_tiles; ++r) {
                                 for (size_t c = 0; c < DV_tiles; ++c) {
@@ -1798,16 +1767,15 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                                     Q6_mxmem_AR_after_hf(o_tile_out, 0);
                                 }
                             }
+                            htp_trace_event_stop(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
                             hex_swap_ptr((void **) &o_tile_curr, (void **) &o_tile_prev);
                         }
-                        TIMER_STOP(o_update);
 
                         buf_idx = 1 - buf_idx;
                     }  // end KV block loop (fallback)
                 }
 
                 // ---- Final normalization: O = diag(1/l) @ O ----
-                TIMER_START(o_norm);
                 {
                     fa_build_d_diag_inv_l(&factx, n_row_tiles, n_row_tiles_g_br);
 
@@ -1830,6 +1798,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         __builtin_assume(n_row_tiles > 0);
                         __builtin_assume(DV_tiles > 0);
 
+                        htp_trace_event_start(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
                         Q6_bias_mxmem2_A((void *) factx.vtcm_hmx_scales_id);
                         for (size_t r = 0; r < n_row_tiles; ++r) {
                             for (size_t c = 0; c < DV_tiles; ++c) {
@@ -1842,14 +1811,12 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                                 Q6_mxmem_AR_after_hf(o_out, 0);
                             }
                         }
+                        htp_trace_event_stop(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
                     }
                 }
-                TIMER_STOP(o_norm);
 
                 // ---- Store O block ----
-                TIMER_START(o_store);
                 fa_phase_o_store(&factx, dst, o_tile_curr, q_start, kv_head, ib3, n_rows_g);
-                TIMER_STOP(o_store);
 
 #undef MASK_DMA_PUSH
 #undef MASK_DMA_POP
@@ -1865,14 +1832,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
         HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
     }
 
-    TIMER_STOP(total);
 
-#if defined(ENABLE_PROFILE_TIMERS)
-    FARF(HIGH, "hmx-fa: %lld us, q_load=%lld kv_dma=%lld k_interleave=%lld v_interleave=%lld", TIMER_US(total),
-         TIMER_US(q_load), TIMER_US(kv_dma), TIMER_US(k_interleave), TIMER_US(v_interleave));
-    FARF(HIGH, "  qk_dot=%lld softmax=%lld o_update=%lld o_norm=%lld o_store=%lld", TIMER_US(qk_dot), TIMER_US(softmax),
-         TIMER_US(o_update), TIMER_US(o_norm), TIMER_US(o_store));
-#endif
 
     return HTP_STATUS_OK;
 }

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

@@ -27,7 +27,7 @@
 #include "hmx-ops.h"
 #include "hmx-utils.h"
 #include "hmx-queue.h"
-#include "hmx-profile.h"
+#include "hex-profile.h"
 
 #include "vtcm-utils.h"
 
@@ -430,6 +430,7 @@ typedef struct {
     int                      n_tasks;
     int                      n_k_tiles;
     struct fastdiv_values    n_k_tiles_div;
+    struct htp_thread_trace * traces;
 } x4x2_dequantize_state_t;
 
 // Dequantize a tile range from x4x2 weight data (already in VTCM) to tile-major FP16.
@@ -533,11 +534,14 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task_##suffix(
                                                                                                                \
 static void dequantize_x4x2_worker_loop_##suffix(unsigned int n, unsigned int i, void *data) {                 \
     x4x2_dequantize_state_t *state = (x4x2_dequantize_state_t *)data;                                          \
+    struct htp_thread_trace * tr = state->traces ? &state->traces[i] : NULL;                                   \
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);                                                 \
     for (unsigned int task_id = i; task_id < (unsigned int)state->n_tasks; task_id += n) {                     \
         int start = task_id * state->n_tiles_per_task;                                                         \
         int end   = hex_smin(start + state->n_tiles_per_task, state->n_tot_tiles);                             \
         dequantize_x4x2_weight_to_fp16_tiles_task_##suffix(state, start, end);                                 \
     }                                                                                                          \
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);                                                 \
 }
 
 DEFINE_DEQUANTIZE_Q4_TASK(q4_0,   q4_0_to_fp16_lut,   q4_0, HMX_X4X2_DBLK_SIZE, (int)sizeof(__fp16))
@@ -657,11 +661,14 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task_mxfp4(
 
 static void dequantize_x4x2_worker_loop_mxfp4(unsigned int n, unsigned int i, void *data) {
     x4x2_dequantize_state_t *state = (x4x2_dequantize_state_t *)data;
+    struct htp_thread_trace * tr = state->traces ? &state->traces[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);
     for (unsigned int task_id = i; task_id < (unsigned int)state->n_tasks; task_id += n) {
         int start = task_id * state->n_tiles_per_task;
         int end   = hex_smin(start + state->n_tiles_per_task, state->n_tot_tiles);
         dequantize_x4x2_weight_to_fp16_tiles_task_mxfp4(state, start, end);
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);
 }
 
 static void dequantize_x4x2_weight_to_fp16_tiles_task_q8_0(
@@ -717,11 +724,14 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task_q8_0(
 
 static void dequantize_x4x2_worker_loop_q8_0(unsigned int n, unsigned int i, void *data) {
     x4x2_dequantize_state_t *state = (x4x2_dequantize_state_t *)data;
+    struct htp_thread_trace * tr = state->traces ? &state->traces[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);
     for (unsigned int task_id = i; task_id < (unsigned int)state->n_tasks; task_id += n) {
         int start = task_id * state->n_tiles_per_task;
         int end   = hex_smin(start + state->n_tiles_per_task, state->n_tot_tiles);
         dequantize_x4x2_weight_to_fp16_tiles_task_q8_0(state, start, end);
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);
 }
 
 static void convert_f16_weight_to_fp16_tiles_task(
@@ -773,11 +783,14 @@ static void convert_f16_weight_to_fp16_tiles_task(
 
 static void convert_f16_worker_loop(unsigned int n, unsigned int i, void *data) {
     x4x2_dequantize_state_t *state = (x4x2_dequantize_state_t *)data;
+    struct htp_thread_trace * tr = state->traces ? &state->traces[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);
     for (unsigned int task_id = i; task_id < (unsigned int)state->n_tasks; task_id += n) {
         int start = task_id * state->n_tiles_per_task;
         int end   = hex_smin(start + state->n_tiles_per_task, state->n_tot_tiles);
         convert_f16_weight_to_fp16_tiles_task(state, start, end);
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);
 }
 
 static void quantize_f32_weight_to_fp16_tiles_task(
@@ -833,11 +846,14 @@ static void quantize_f32_weight_to_fp16_tiles_task(
 
 static void quantize_f32_worker_loop(unsigned int n, unsigned int i, void *data) {
     x4x2_dequantize_state_t *state = (x4x2_dequantize_state_t *)data;
+    struct htp_thread_trace * tr = state->traces ? &state->traces[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);
     for (unsigned int task_id = i; task_id < (unsigned int)state->n_tasks; task_id += n) {
         int start = task_id * state->n_tiles_per_task;
         int end   = hex_smin(start + state->n_tiles_per_task, state->n_tot_tiles);
         quantize_f32_weight_to_fp16_tiles_task(state, start, end);
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_W_DEQUANT, i);
 }
 
 
@@ -868,6 +884,7 @@ static void dequantize_x4x2_weight_chunk_to_fp16_tiles(
     state.weight_type      = weight_type;
     state.n_k_tiles        = n_k_tiles;
     state.n_k_tiles_div    = n_k_tiles_div;
+    state.traces           = ctx ? ctx->trace : NULL;
 
     if (state.n_tasks == 1 || n_threads == 1) {
         dequant_worker_fn(1, 0, &state);
@@ -985,10 +1002,13 @@ typedef struct {
     int            n_chunks_per_task;
     int            n_cols;
     int            n;  // DDR row stride (total output columns)
+    struct htp_thread_trace * traces;
 } output_transfer_task_state_t;
 
 static void transfer_output_chunk_worker_fn(unsigned int n, unsigned int i, void *data) {
     output_transfer_task_state_t *st = (output_transfer_task_state_t *) data;
+    struct htp_thread_trace * tr = st->traces ? &st->traces[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_O_PROC, i);
 
     for (unsigned int task_id = i; task_id < (unsigned int)st->n_tasks; task_id += n) {
         int    chunk_idx  = task_id * st->n_chunks_per_task;
@@ -998,6 +1018,7 @@ static void transfer_output_chunk_worker_fn(unsigned int n, unsigned int i, void
         const __fp16 *vtcm_src = st->vtcm_src + chunk_idx * st->n_cols;
         transfer_output_chunk_fp16_to_fp32(dst, vtcm_src, chunk_size, st->n_cols, st->n);
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_O_PROC, i);
 }
 
 static void transfer_output_chunk_threaded(struct htp_context *ctx, float *dst, const __fp16 *vtcm_src,
@@ -1015,6 +1036,7 @@ static void transfer_output_chunk_threaded(struct htp_context *ctx, float *dst,
     state.vtcm_src          = vtcm_src;
     state.n_cols            = n_cols;
     state.n                 = n;
+    state.traces            = ctx ? ctx->trace : NULL;
 
     if (state.n_tasks == 1 || n_threads == 1) {
         transfer_output_chunk_worker_fn(1, 0, &state);
@@ -1086,10 +1108,13 @@ typedef struct {
     int          n_chunks_per_task;
     int          k_block;
     int          k_stride;
+    struct htp_thread_trace * traces;
 } activation_transfer_task_state_t;
 
 static void transfer_activation_chunk_worker_fn(unsigned int n, unsigned int i, void *data) {
     activation_transfer_task_state_t *st = (activation_transfer_task_state_t *) data;
+    struct htp_thread_trace * tr = st->traces ? &st->traces[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_A_PREP, i);
 
     for (unsigned int task_id = i; task_id < (unsigned int)st->n_tasks; task_id += n) {
         // one chunk: one row
@@ -1100,6 +1125,7 @@ static void transfer_activation_chunk_worker_fn(unsigned int n, unsigned int i,
         const float *src = st->src + chunk_idx * st->k_stride;
         transfer_activation_chunk_fp32_to_fp16(dst, src, chunk_size, st->k_block, st->k_stride);
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_A_PREP, i);
 }
 
 static void transfer_activation_chunk_threaded(struct htp_context *ctx, __fp16 *dst, const float *src, int n_rows, int k_block, int k_stride, int n_threads) {
@@ -1117,6 +1143,7 @@ static void transfer_activation_chunk_threaded(struct htp_context *ctx, __fp16 *
     state.src               = src;
     state.k_block           = k_block;
     state.k_stride          = k_stride;
+    state.traces            = ctx ? ctx->trace : NULL;
 
     if (state.n_tasks == 1 || n_threads == 1) {
         transfer_activation_chunk_worker_fn(1, 0, &state);
@@ -1245,13 +1272,7 @@ int hmx_matmul_2d_f32(struct htp_context *ctx, float *restrict dst, const float
     FARF(HIGH, "hmx-mm-2d: standard : m %d k %d n %d wtype %d mc %zu nc %zu vtcm %zu/%zu",
          m, k, n, weight_type, m_chunk_n_rows, n_chunk_n_cols, vtcm_used, vtcm_budget);
 
-    TIMER_DEFINE(activation_load);
-    TIMER_DEFINE(weight_load);
-    TIMER_DEFINE(hmx_core);
-    TIMER_DEFINE(output_store);
 
-    TIMER_DEFINE(total);
-    TIMER_START(total);
 
     int n_chunk_cnt = hmx_ceil_div(n, n_chunk_n_cols);
 
@@ -1370,7 +1391,12 @@ int hmx_matmul_2d_f32(struct htp_context *ctx, float *restrict dst, const float
                 dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_scratch0, vtcm_weight, n_cols, k, row_stride, weight_type, n_k_tiles, n_k_tiles_div, dequant_worker_fn, num_threads);
 
                 // C: HMX Compute (Synchronous)
-                core_dot_chunk_fp16(vtcm_output, vtcm_activation, vtcm_scratch0, vtcm_scales, n_row_tiles, n_col_tiles, k / HMX_FP16_TILE_N_ROWS);
+                {
+                    struct htp_thread_trace * tr = ctx ? &ctx->trace[HTP_MAX_NTHREADS] : NULL;
+                    htp_trace_event_start(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
+                    core_dot_chunk_fp16(vtcm_output, vtcm_activation, vtcm_scratch0, vtcm_scales, n_row_tiles, n_col_tiles, k / HMX_FP16_TILE_N_ROWS);
+                    htp_trace_event_stop(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
+                }
 
                 // D: Output Store
                 float *output_chunk = dst + (mr * n + nc);
@@ -1380,18 +1406,7 @@ int hmx_matmul_2d_f32(struct htp_context *ctx, float *restrict dst, const float
         HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
     }
 
-    TIMER_STOP(total);
 
-#if defined(ENABLE_PROFILE_TIMERS)
-    FARF(HIGH, "hex-mm-2d: %lld us : m %d k %d n %d", TIMER_US(total), m, k, n);
-    if (!use_pipeline) {
-        FARF(HIGH, "  activation_load: %lld us, weight_load: %lld us, hmx_core: %lld us, output_store: %lld us",
-             TIMER_US(activation_load), TIMER_US(weight_load), TIMER_US(hmx_core), TIMER_US(output_store));
-        size_t weight_size = (size_t)n * row_stride;
-        float  bandwidth   = 1e-3f * weight_size / (float)TIMER_US(weight_load);
-        FARF(HIGH, "  weight load bandwidth: %.2f GB/s", bandwidth);
-    }
-#endif
 
     return 0;
 }
@@ -1523,13 +1538,7 @@ int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32
             m_chunk_n_rows, n_chunk_n_cols,
             (size_t) (vtcm_ptr - (uint8_t *) ctx->vtcm_base), vtcm_budget);
 
-    TIMER_DEFINE(activation_load);
-    TIMER_DEFINE(weight_load);
-    TIMER_DEFINE(hmx_core);
-    TIMER_DEFINE(output_store);
-    TIMER_DEFINE(total);
 
-    TIMER_START(total);
 
     const size_t fp16_row_bytes   = (size_t) params->k * sizeof(__fp16);
     const size_t weight_row_bytes = (size_t) params->weight_stride * sizeof(__fp16);
@@ -1549,7 +1558,6 @@ int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32
                 // contiguous rows into a VTCM scratch buffer first, then HVX
                 // converts from the contiguous VTCM buffer.  This avoids L2 cache
                 // thrashing from HVX loads at large strides.
-                TIMER_START(activation_load);
                 for (int g = 0; g < group_size; ++g) {
                     const float *activation_chunk = hmx_matmul_activation_batch_ptr(params, b2_base + g, b3) + mr * params->act_stride;
                     __fp16 *vtcm_act_g = vtcm_activation + (size_t) g * act_head_stride;
@@ -1569,7 +1577,6 @@ int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32
                                                               params->k, params->act_stride, ctx->n_threads);
                     }
                 }
-                TIMER_STOP(activation_load);
 
                 void *buf_curr = vtcm_scratch0;
                 void *buf_next = vtcm_scratch1;
@@ -1584,7 +1591,6 @@ int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32
                     const size_t n_cols = hex_smin((size_t) params->n - nc, n_chunk_n_cols);
                     const size_t n_col_tiles = hmx_ceil_div((int) n_cols, HMX_FP16_TILE_N_COLS);
 
-                    TIMER_START(weight_load);
                     {
                         dma_queue_pop(ctx->dma[0]);
 
@@ -1601,24 +1607,22 @@ int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32
                                                  0, n_cols);
                         hex_swap_ptr(&buf_curr, &buf_next);
                     }
-                    TIMER_STOP(weight_load);
 
                     // Reuse the interleaved weight for every q_head in this GQA group
                     for (int g = 0; g < group_size; ++g) {
-                        TIMER_START(hmx_core);
                         {
                             const __fp16 * vtcm_act_g = vtcm_activation + (size_t) g * act_head_stride;
+                            struct htp_thread_trace * tr = ctx ? &ctx->trace[HTP_MAX_NTHREADS] : NULL;
+                            htp_trace_event_start(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
                             core_dot_chunk_fp16(vtcm_output, vtcm_act_g, vtcm_weight, vtcm_scales, n_row_tiles, n_col_tiles,
                                                 params->k / 32);
+                            htp_trace_event_stop(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
                         }
-                        TIMER_STOP(hmx_core);
 
-                        TIMER_START(output_store);
                         {
                             float *output = hmx_matmul_dst_batch_ptr(params, b2_base + g, b3) + mr * params->dst_stride + nc;
                             transfer_output_chunk_threaded(ctx, output, vtcm_output, (int) n_rows, (int) n_cols, params->dst_stride, ctx->n_threads);
                         }
-                        TIMER_STOP(output_store);
                     }
                 }
             }
@@ -1627,14 +1631,7 @@ int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32
 
     HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
 
-    TIMER_STOP(total);
 
-#if defined(ENABLE_PROFILE_TIMERS)
-    FARF(HIGH, "%s: %lld us, m=%d k=%d n=%d group=%d", __func__, TIMER_US(total),
-         params->m, params->k, params->n, group_size);
-    FARF(HIGH, "  activation_load: %lld us, weight_load: %lld us, hmx_core: %lld us, output_store: %lld us",
-         TIMER_US(activation_load), TIMER_US(weight_load), TIMER_US(hmx_core), TIMER_US(output_store));
-#endif
 
     return 0;
 }
@@ -1668,6 +1665,7 @@ typedef struct {
     size_t                          nb12;
     int                             start_row;
     int                             cne1;
+    struct htp_thread_trace        *traces;
 } activation_transfer_gathered_task_state_t;
 
 typedef struct {
@@ -1684,6 +1682,7 @@ typedef struct {
     size_t                          dst_nb2;
     int                             start_row;
     int                             cne1;
+    struct htp_thread_trace        *traces;
 } output_transfer_scattered_task_state_t;
 
 static void transfer_activation_chunk_fp32_to_fp16_gathered(
@@ -1780,6 +1779,9 @@ static void transfer_activation_chunk_fp32_to_fp16_gathered(
 
 static void transfer_activation_chunk_gathered_worker_fn(unsigned int n, unsigned int i, void *data) {
     activation_transfer_gathered_task_state_t *st = data;
+    struct htp_thread_trace * tr = st->traces ? &st->traces[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_A_PREP, i);
+
     int chunk_idx = i;
     int chunk_size = st->n_chunks_per_task;
     int start_row = st->start_row + chunk_idx * chunk_size;
@@ -1791,6 +1793,7 @@ static void transfer_activation_chunk_gathered_worker_fn(unsigned int n, unsigne
             st->matrix_rows, st->cur_a, st->mapping_stride,
             st->ne11, &st->ne11_div, st->nb11, st->nb12, st->cne1);
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_A_PREP, i);
 }
 
 static void transfer_activation_chunk_gathered_threaded(
@@ -1830,6 +1833,7 @@ static void transfer_activation_chunk_gathered_threaded(
         .nb12              = nb12,
         .start_row         = start_row,
         .cne1              = cne1,
+        .traces            = ctx ? ctx->trace : NULL,
     };
 
     if (actual_threads <= 1) {
@@ -1895,6 +1899,9 @@ static void transfer_output_chunk_fp16_to_fp32_scattered(
 
 static void transfer_output_chunk_scattered_worker_fn(unsigned int n, unsigned int i, void *data) {
     output_transfer_scattered_task_state_t *st = data;
+    struct htp_thread_trace * tr = st->traces ? &st->traces[i] : NULL;
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_O_PROC, i);
+
     int chunk_idx = i;
     int chunk_size = st->n_chunks_per_task;
     int start_row = st->start_row + chunk_idx * chunk_size;
@@ -1906,6 +1913,7 @@ static void transfer_output_chunk_scattered_worker_fn(unsigned int n, unsigned i
             st->matrix_rows, st->cur_a, st->mapping_stride,
             st->dst_nb1, st->dst_nb2, st->cne1);
     }
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_O_PROC, i);
 }
 
 static void transfer_output_chunk_scattered_threaded(
@@ -1942,6 +1950,7 @@ static void transfer_output_chunk_scattered_threaded(
         .dst_nb2           = dst_nb2,
         .start_row         = start_row,
         .cne1              = cne1,
+        .traces            = ctx ? ctx->trace : NULL,
     };
 
     if (actual_threads <= 1) {
@@ -2053,7 +2062,12 @@ int hmx_matmul_id_2d_f32(struct htp_context *ctx,
 
             dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_scratch0, vtcm_weight, n_cols, k, row_stride, weight_type, n_k_tiles, n_k_tiles_div, dequant_worker_fn, num_threads);
 
-            core_dot_chunk_fp16(vtcm_output, vtcm_activation, vtcm_scratch0, vtcm_scales, n_row_tiles, n_col_tiles, k / HMX_FP16_TILE_N_ROWS);
+            {
+                struct htp_thread_trace * tr = ctx ? &ctx->trace[HTP_MAX_NTHREADS] : NULL;
+                htp_trace_event_start(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
+                core_dot_chunk_fp16(vtcm_output, vtcm_activation, vtcm_scratch0, vtcm_scales, n_row_tiles, n_col_tiles, k / HMX_FP16_TILE_N_ROWS);
+                htp_trace_event_stop(tr, HTP_TRACE_EVT_HMX_COMP, HTP_MAX_NTHREADS);
+            }
 
             transfer_output_chunk_scattered_threaded(
                 ctx, dst, vtcm_output, (int) mr, (int) n_rows, (int) n_cols,

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

@@ -1,34 +0,0 @@
-// Conditional fine-grained profiling macros for HMX operations.
-//
-// Define ENABLE_PROFILE_TIMERS (via compiler flag or before including this
-// header) to instrument sub-operation latencies with HAP qtimer.  When the
-// macro is not defined the TIMER_* helpers expand to nothing so there is zero
-// overhead.
-//
-// Usage:
-//   TIMER_DEFINE(my_phase);          // declare accumulator variable
-//   TIMER_START(my_phase);           // snapshot start time
-//   ... work ...
-//   TIMER_STOP(my_phase);            // accumulate elapsed ticks
-//   FARF(ALWAYS, "my_phase: %lld us", TIMER_US(my_phase));
-
-#ifndef HMX_PROFILE_H
-#define HMX_PROFILE_H
-
-#include <HAP_perf.h>
-
-// #define ENABLE_PROFILE_TIMERS
-
-#if defined(ENABLE_PROFILE_TIMERS)
-#  define TIMER_DEFINE(name) int64_t name##_ticks = 0
-#  define TIMER_START(name)  int64_t name##_t0 = HAP_perf_get_qtimer_count()
-#  define TIMER_STOP(name)   name##_ticks += HAP_perf_get_qtimer_count() - name##_t0
-#  define TIMER_US(name)     HAP_perf_qtimer_count_to_us(name##_ticks)
-#else
-#  define TIMER_DEFINE(name)
-#  define TIMER_START(name)
-#  define TIMER_STOP(name)
-#  define TIMER_US(name)     0LL
-#endif
-
-#endif // HMX_PROFILE_H

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

@@ -44,7 +44,9 @@ static inline void hmx_queue_process(struct hmx_queue *q, bool* killed) {
                 case HMX_QUEUE_SUSPEND: hmx_unlock(q);  break;
                 default:
                     hmx_lock(q);
+                    htp_trace_event_start(q->trace, HTP_TRACE_EVT_HMX_COMP, ir);
                     d->func(d->data);
+                    htp_trace_event_stop(q->trace, HTP_TRACE_EVT_HMX_COMP, ir);
                     break;
             }
 

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

@@ -11,6 +11,7 @@
 #include <HAP_farf.h>
 
 #include "hex-utils.h"
+#include "hex-profile.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -47,6 +48,7 @@ struct hmx_queue {
     void *           stack;
     uint32_t         hap_rctx;
     bool             hmx_locked;
+    struct htp_thread_trace * trace;
 };
 
 struct hmx_queue * hmx_queue_create(size_t capacity, uint32_t hap_rctx);

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

@@ -4,6 +4,7 @@
 #include "hex-dma.h"
 #include "hmx-queue.h"
 #include "htp-ops.h"
+#include "hex-profile.h"
 #include "worker-pool.h"
 
 #include <assert.h>
@@ -70,6 +71,7 @@ struct htp_context {
     bool                   hmx_enabled;
     bool                   etm;
     uint32_t               profiler;
+    struct htp_thread_trace trace[HTP_MAX_NTHREADS + 1];
 
     uint8_t *              vtcm_base;
     size_t                 vtcm_size;

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

@@ -146,10 +146,36 @@ struct htp_op_desc {
     uint16_t dst;                       // Output tensor index
 };
 
+#ifndef HTP_MAX_NTHREADS
+#define HTP_MAX_NTHREADS 10
+#endif
+
+#define HTP_TRACE_MAX_EVENTS 256
+
 enum htp_profiler_mode {
     HTP_PROF_DISABLED = 0,
     HTP_PROF_BASIC    = 1,
     HTP_PROF_PMU      = 2,
+    HTP_PROF_TRACE    = 3,
+};
+
+enum htp_trace_event_id {
+    HTP_TRACE_EVT_DMA                 = 0,
+
+    HTP_TRACE_EVT_HVX_COMP            = 20,
+    HTP_TRACE_EVT_HVX_A_QUANT         = 21,
+    HTP_TRACE_EVT_HVX_A_PREP          = 22,
+    HTP_TRACE_EVT_HVX_W_DEQUANT       = 23,
+    HTP_TRACE_EVT_HVX_W_PREP          = 24,
+    HTP_TRACE_EVT_HVX_O_PROC          = 25,
+
+    HTP_TRACE_EVT_HMX_COMP            = 40,
+};
+
+struct htp_trace_desc {
+    uint32_t cycles;  // lower 32-bits of cycle counter
+    uint16_t id;      // Event ID
+    uint16_t info;    // bit 15: is_stop. bits 14-0: tile/chunk index or other metadata.
 };
 
 #define HTP_PROF_PMU_NCNT 8
@@ -158,8 +184,8 @@ enum htp_profiler_mode {
 struct htp_prof_desc {
     uint32_t opcode;                 // GGML/HTP Op
     uint32_t usecs;                  // Number of usec
-    uint32_t cycles;                 // Number of cycles
-    uint32_t pad;                    // Unused
+    uint32_t cycles_start;           // Start cycle counter
+    uint32_t cycles_stop;            // Stop cycle counter
     uint32_t pmu[HTP_PROF_PMU_NCNT]; // PMU counters
 };
 
@@ -168,7 +194,7 @@ struct htp_opbatch_req {
     uint32_t n_bufs;      // Number of buffers
     uint32_t n_tensors;   // Number of tensors
     uint32_t n_ops;       // Number of ops
-    uint32_t flags;       // unused
+    uint32_t n_traces;    // Number of trace descriptors per thread
     uint32_t pad;         // unused
     // struct htp_buf_desc  bufs[];    -- dspqueue buf 0
     // struct htp_tensor    tensors[]; -- dspqueue buf 0
@@ -181,7 +207,8 @@ struct htp_opbatch_rsp {
     uint32_t n_bufs;     // Number of buffers
     uint32_t n_tensors;  // Number of tensors
     uint32_t n_ops;      // Number of op profile descriptors
-    uint32_t pad;        // unused
+    uint32_t n_traces[HTP_MAX_NTHREADS + 1];
+    uint8_t  pad[8];     // align to 8 bytes
     // struct htp_prof_desc profs[];  -- dspqueue buf 0
 };
 

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

@@ -400,7 +400,9 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
     ctx->hmx_queue   = NULL;
     if (use_hmx) {
         ctx->hmx_queue = hmx_queue_create(16, ctx->vtcm_rctx);
-        if (!ctx->hmx_queue) {
+        if (ctx->hmx_queue) {
+            ctx->hmx_queue->trace = &ctx->trace[HTP_MAX_NTHREADS];
+        } else {
             FARF(ERROR, "hmx-queue-create failed");
             ctx->hmx_enabled = false;
         }
@@ -425,6 +427,9 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
     ctx->n_threads = n_hvx;
     for (int i = 0; i < ctx->n_threads; i++) {
         ctx->dma[i] = dma_queue_create(256); // queue depth
+        if (ctx->dma[i]) {
+            ctx->dma[i]->trace = &ctx->trace[i];
+        }
     }
 
     ctx->ddr_spad_size = 512 * 1024; // 512 KB
@@ -502,7 +507,8 @@ static void htp_error_callback(dspqueue_t queue, int error, void * context) {
 
 struct profile_data {
     uint64_t usecs;
-    uint64_t cycles;
+    uint64_t cycles_start;
+    uint64_t cycles_stop;
     uint32_t pmu_counters[HEX_NUM_PMU_COUNTERS];
 };
 
@@ -512,8 +518,9 @@ static inline void profile_start(uint32_t mode, struct profile_data * d) {
             hex_get_pmu(d->pmu_counters);
             // fallthrough
         case HTP_PROF_BASIC:
+        case HTP_PROF_TRACE:
             d->usecs  = HAP_perf_get_qtimer_count();
-            d->cycles = hex_get_cycles();
+            d->cycles_start = hex_get_cycles();
             break;
         default:
             break;
@@ -530,8 +537,9 @@ static inline void profile_stop(uint32_t mode, struct profile_data * d) {
             }
             // fallthrough
         case HTP_PROF_BASIC:
+        case HTP_PROF_TRACE:
             d->usecs  = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - d->usecs);
-            d->cycles = hex_get_cycles() - d->cycles;
+            d->cycles_stop = hex_get_cycles();
             break;
         default:
             break;
@@ -845,14 +853,15 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
         const uint32_t t_size = sizeof(struct htp_tensor)    * n_tens;
         const uint32_t o_size = sizeof(struct htp_op_desc)   * n_ops;
         const uint32_t p_size = sizeof(struct htp_prof_desc) * n_ops;
+        const uint32_t tr_size = (HTP_MAX_NTHREADS + 1) * req.n_traces * sizeof(struct htp_trace_desc);
 
-        if (dbuf.size < b_size + t_size + o_size + p_size) {
-            FARF(ERROR, "invalid opbatch memory block size %u", dbuf.size);
+        if (dbuf.size < b_size + t_size + o_size + p_size + tr_size) {
+            FARF(ERROR, "invalid opbatch memory block size %u (req %u)", dbuf.size, b_size + t_size + o_size + p_size + tr_size);
             break;
         }
 
-        FARF(HIGH, "processing opbatch #%u: n-bufs %u n-tensors %u n-ops %u : m-size %u b-size %u t-size %u o-size %u", req.id,
-                n_bufs, n_tens, n_ops, dbuf.size, b_size, t_size, o_size);
+        FARF(HIGH, "processing opbatch #%u: n-bufs %u n-tensors %u n-ops %u n-traces %u : m-size %u b-size %u t-size %u o-size %u", req.id,
+                n_bufs, n_tens, n_ops, req.n_traces, dbuf.size, b_size, t_size, o_size);
 
         // Setup descriptor pointers
         uint8_t * m_ptr = dbuf.ptr;
@@ -869,6 +878,20 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
         octx->n_threads = ctx->n_threads;
         octx->ctx       = ctx;
 
+        if (ctx->profiler == HTP_PROF_TRACE) {
+            memset(ctx->trace, 0, sizeof(ctx->trace));
+            struct htp_trace_desc * trace_events = (struct htp_trace_desc *) (m_ptr + p_size);
+            for (int t = 0; t <= HTP_MAX_NTHREADS; t++) {
+                ctx->trace[t].events = &trace_events[t * req.n_traces];
+                ctx->trace[t].max_events = req.n_traces;
+            }
+        } else {
+            for (int t = 0; t <= HTP_MAX_NTHREADS; t++) {
+                ctx->trace[t].events = NULL;
+                ctx->trace[t].max_events = 0;
+            }
+        }
+
         for (uint32_t i=0; i < n_ops; i++) {
             struct profile_data prof;
 
@@ -886,7 +909,8 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
             if (ctx->profiler) {
                 pds[i].opcode = ops[i].opcode;
                 pds[i].usecs  = prof.usecs;
-                pds[i].cycles = prof.cycles;
+                pds[i].cycles_start = prof.cycles_start;
+                pds[i].cycles_stop = prof.cycles_stop;
                 for (int j = 0; j < HEX_NUM_PMU_COUNTERS; j++) {
                     pds[i].pmu[j] = prof.pmu_counters[j];
                 }
@@ -899,6 +923,14 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
         rsp.n_bufs    = n_bufs;
         rsp.n_tensors = n_tens;
         rsp.n_ops     = n_ops;
+        memset(rsp.pad, 0, sizeof(rsp.pad));
+        if (ctx->profiler == HTP_PROF_TRACE) {
+            for (int t = 0; t <= HTP_MAX_NTHREADS; t++) {
+                rsp.n_traces[t] = ctx->trace[t].count;
+            }
+        } else {
+            memset(rsp.n_traces, 0, sizeof(rsp.n_traces));
+        }
 
         dbuf.flags = DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT;
 

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

@@ -3350,6 +3350,7 @@ static void vec_dot_f16_f32_uu_1x1(const int n, float * restrict s, const void *
 
 static void matmul_4d(unsigned int nth, unsigned int ith, void * data) {
     htp_matmul_preamble;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     uint64_t t1, t2;
     t1 = HAP_perf_get_qtimer_count();
@@ -3411,10 +3412,12 @@ static void matmul_4d(unsigned int nth, unsigned int ith, void * data) {
                 float * dst_col = (float *) ((uint8_t * restrict) dst->data + (i1 * nb1 + i2 * nb2 + i3 * nb3));
 
                 const uint32_t ir0_block_end = MIN(iir0 + blck_0, ir0_end);
+                htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, iir0);
                 for (uint32_t ir0 = iir0; ir0 < ir0_block_end; ir0++) {
                     const uint8_t * restrict src0_row = src0_base + ir0 * nb01;
                     mmctx->vec_dot_1x1(ne00, &dst_col[ir0], src0_row, src1_col);
                 }
+                htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, iir0);
             }
         }
     }
@@ -3430,6 +3433,7 @@ static void matmul_4d(unsigned int nth, unsigned int ith, void * data) {
 // src1 tensor is already in VTCM spad
 static void matmul_2d(unsigned int nth, unsigned int ith, void * data) {
     htp_matmul_preamble;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
     const uint32_t src1_nrows = ne11 * ne12 * ne13;  // src1 rows
@@ -3477,6 +3481,8 @@ static void matmul_2d(unsigned int nth, unsigned int ith, void * data) {
     for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x2; ir0 += 2) {
         const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
+
         // Process src1 columns in pairs (2×2 tiling)
         uint32_t ir1 = 0;
         for (; ir1 + 1 < src1_nrows; ir1 += 2) {
@@ -3494,6 +3500,8 @@ static void matmul_2d(unsigned int nth, unsigned int ith, void * data) {
             mmctx->vec_dot_2x1(ne00, &dst_row[ir0], ss0, ss0 + src0_stride, src1_col);
         }
 
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
+
         // Prefetch next (n + spad_nrows) row
         const int pr0 = (ir0 + MM_SPAD_SRC0_NROWS);
         const int is0 = (pr0 - src0_start_row) % MM_SPAD_SRC0_NROWS;
@@ -3511,12 +3519,14 @@ static void matmul_2d(unsigned int nth, unsigned int ith, void * data) {
                        src0_stride, src0_row_size, 1);
         const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
         #pragma unroll(2)
         for (uint32_t ir1 = 0; ir1 < src1_nrows; ++ir1) {
             const uint8_t * restrict src1_col = (const uint8_t *) (src1_data + ir1 * src1_stride);
             float * restrict dst_row          = (float *) (dst->data + (ir1 * dst_row_size));
             mmctx->vec_dot_1x1(ne00, &dst_row[ir0], ss0, src1_col);
         }
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
     }
 
     t2 = HAP_perf_get_qtimer_count();
@@ -3530,6 +3540,7 @@ static void matmul_2d(unsigned int nth, unsigned int ith, void * data) {
 // q8x4x2 src1 tensor is already in VTCM spad
 static void matvec_2d(unsigned int nth, unsigned int ith, void * data) {
     htp_matmul_preamble;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const uint32_t src0_nrows = ne01;
 
@@ -3581,7 +3592,9 @@ static void matvec_2d(unsigned int nth, unsigned int ith, void * data) {
         // Process src0 rows
         for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x4; ir0 += 4) {
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             mmctx->vec_dot_4x1(ne00, &tmp[ir0 - src0_start_row], ss0, ss0 + src0_stride, ss0 + 2 * src0_stride, ss0 + 3 * src0_stride, src1_col);
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
 
             // Prefetch next (n + spad_nrows) row
             const uint32_t pr0 = (ir0 + MM_SPAD_SRC0_NROWS);
@@ -3599,7 +3612,9 @@ static void matvec_2d(unsigned int nth, unsigned int ith, void * data) {
             dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_stride, src0_row + ir0 * src0_row_size),
                            src0_stride, src0_row_size, 2);
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             mmctx->vec_dot_2x1(ne00, &tmp[ir0 - src0_start_row], ss0, ss0 + src0_stride, src1_col);
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             ir0 += 2;
         }
         if (ir0 < src0_end_row) {
@@ -3607,7 +3622,9 @@ static void matvec_2d(unsigned int nth, unsigned int ith, void * data) {
             dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_stride, src0_row + ir0 * src0_row_size),
                            src0_stride, src0_row_size, 1);
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             mmctx->vec_dot_1x1(ne00, &tmp[ir0 - src0_start_row], ss0, src1_col);
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             ir0 += 1;
         }
     } else {
@@ -3627,7 +3644,9 @@ static void matvec_2d(unsigned int nth, unsigned int ith, void * data) {
         // Process src0 rows
         for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x2; ir0 += 2) {
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             mmctx->vec_dot_2x1(ne00, &tmp[ir0 - src0_start_row], ss0, ss0 + src0_stride, src1_col);
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
 
             // Prefetch next (n + spad_nrows) row
             const uint32_t pr0 = (ir0 + MM_SPAD_SRC0_NROWS);
@@ -3645,7 +3664,9 @@ static void matvec_2d(unsigned int nth, unsigned int ith, void * data) {
             dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_stride, src0_row + ir0 * src0_row_size),
                            src0_stride, src0_row_size, 1);
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             mmctx->vec_dot_1x1(ne00, &tmp[ir0 - src0_start_row], ss0, src1_col);
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
         }
     }
 
@@ -3669,6 +3690,7 @@ struct mmid_row_mapping {
 // src1 tensor is already in VTCM spad
 static void matmul_id(unsigned int nth, unsigned int ith, void * data) {
     htp_matmul_preamble;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const struct htp_tensor * restrict ids = octx->src[2];
     struct htp_spad * restrict   src2_spad = &octx->src2_spad;
@@ -3735,6 +3757,7 @@ static void matmul_id(unsigned int nth, unsigned int ith, void * data) {
         for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x2; ir0 += 2) {
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             for (uint32_t cid = 0; cid < cne1; ++cid) {
                 struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, cid);
                 const int               rm1         = row_mapping.i1;  // expert idx
@@ -3746,6 +3769,7 @@ static void matmul_id(unsigned int nth, unsigned int ith, void * data) {
 
                 mmctx->vec_dot_2x1(ne00, &dst_row[ir0], ss0, ss0 + src0_row_size_padded, src1_col);
             }
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
 
             // Prefetch next (n + spad_nrows) row
             const int pr0 = (ir0 + MM_SPAD_SRC0_NROWS);
@@ -3764,6 +3788,7 @@ static void matmul_id(unsigned int nth, unsigned int ith, void * data) {
                            src0_row_size_padded, src0_row_size, 1);
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
 
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             for (uint32_t cid = 0; cid < cne1; ++cid) {
                 struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, cid);
                 const int               rm1         = row_mapping.i1;  // expert idx
@@ -3775,6 +3800,7 @@ static void matmul_id(unsigned int nth, unsigned int ith, void * data) {
 
                 mmctx->vec_dot_1x1(ne00, &dst_row[ir0], ss0, src1_col);
             }
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
         }
     }
 
@@ -3789,6 +3815,7 @@ static void matmul_id(unsigned int nth, unsigned int ith, void * data) {
 // src1 tensor is already in VTCM spad
 static void matvec_id(unsigned int nth, unsigned int ith, void * data) {
     htp_matmul_preamble;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const struct htp_tensor * restrict ids = octx->src[2];
     struct htp_spad * restrict   src2_spad = &octx->src2_spad;
@@ -3847,7 +3874,9 @@ static void matvec_id(unsigned int nth, unsigned int ith, void * data) {
         // Process src0 rows
         for (uint32_t ir0 = src0_start_row; ir0 < src0_end_row_x2; ir0 += 2) {
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             mmctx->vec_dot_2x1(ne00, &dst_row[ir0], ss0, ss0 + src0_row_size_padded, src1_col);
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
 
             // Prefetch next (n + spad_nrows) row
             const int pr0 = (ir0 + MM_SPAD_SRC0_NROWS);
@@ -3865,7 +3894,9 @@ static void matvec_id(unsigned int nth, unsigned int ith, void * data) {
             dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(spad_src0 + is0 * src0_row_size_padded, src0_row + ir0 * src0_row_size),
                            src0_row_size_padded, src0_row_size, 1);
             const uint8_t * ss0 = dma_queue_pop(dma_queue).dst;
+            htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
             mmctx->vec_dot_1x1(ne00, &dst_row[ir0], ss0, src1_col);
+            htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, ir0);
         }
     }
 
@@ -4147,6 +4178,7 @@ static void quantize_row_f32_q8x4x2(float * restrict x, uint8_t * restrict y, ui
 static void quantize_f32_q8x4x2(unsigned int nth, unsigned int ith, void * data) {
     struct htp_matmul_context * mmctx = data;
     struct htp_ops_context * octx = mmctx->octx;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const struct htp_tensor * src = octx->src[1];
     uint8_t * restrict dst = octx->src1_spad.data;
@@ -4163,6 +4195,7 @@ static void quantize_f32_q8x4x2(unsigned int nth, unsigned int ith, void * data)
     const uint32_t nrows = ne1 * ne2 * ne3;                             // total n_rows
 
     const uint32_t ir_first = nrows_per_thread * ith;                   // first row
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
     const uint32_t ir_last  = MIN(ir_first + nrows_per_thread, nrows);  // last row
 
     const size_t src_row_size = src->nb[1];
@@ -4189,6 +4222,7 @@ static void quantize_f32_q8x4x2(unsigned int nth, unsigned int ith, void * data)
 
     FARF(HIGH, "quantize-f32-q8x4: %u/%u : n-rows %u (%u:%u) row-size %u -> %u usec %u\n", ith, nth, nrows, ir_first,
          ir_last, src_row_size, dst_row_size, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
 }
 
 static void quantize_row_f32_q8_1x4x2(float * restrict x, uint8_t * restrict y, uint32_t k) {
@@ -4219,6 +4253,7 @@ static void quantize_row_f32_q8_1x4x2(float * restrict x, uint8_t * restrict y,
 static void quantize_f32_q8_1x4x2(unsigned int nth, unsigned int ith, void * data) {
     struct htp_matmul_context * mmctx = data;
     struct htp_ops_context * octx = mmctx->octx;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const struct htp_tensor * src = octx->src[1];
     uint8_t * restrict dst = octx->src1_spad.data;
@@ -4235,6 +4270,7 @@ static void quantize_f32_q8_1x4x2(unsigned int nth, unsigned int ith, void * dat
     const uint32_t nrows = ne1 * ne2 * ne3;                             // total n_rows
 
     const uint32_t ir_first = nrows_per_thread * ith;                   // first row
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
     const uint32_t ir_last  = MIN(ir_first + nrows_per_thread, nrows);  // last row
 
     const size_t src_row_size = src->nb[1];
@@ -4260,11 +4296,13 @@ static void quantize_f32_q8_1x4x2(unsigned int nth, unsigned int ith, void * dat
 
     FARF(HIGH, "quantize-f32-q8_1x4: %u/%u : n-rows %u (%u:%u) row-size %u -> %u usec %u\n", ith, nth, nrows, ir_first,
          ir_last, src_row_size, dst_row_size, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
 }
 
 static void quantize_f32_f32(unsigned int nth, unsigned int ith, void * data) {
     struct htp_matmul_context * mmctx = data;
     struct htp_ops_context * octx = mmctx->octx;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const struct htp_tensor * src = octx->src[1];
     uint8_t * restrict dst = octx->src1_spad.data;
@@ -4281,6 +4319,7 @@ static void quantize_f32_f32(unsigned int nth, unsigned int ith, void * data) {
     const uint32_t nrows = ne1 * ne2 * ne3;                             // total n_rows
 
     const uint32_t ir_first = nrows_per_thread * ith;                   // first row
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
     const uint32_t ir_last  = MIN(ir_first + nrows_per_thread, nrows);  // last row
 
     const size_t src_row_size = ne0 * sizeof(float);
@@ -4301,11 +4340,13 @@ static void quantize_f32_f32(unsigned int nth, unsigned int ith, void * data) {
 
     FARF(HIGH, "quantize-f32-f32: %u/%u : n-rows %u (%u:%u) row-size %u (%u) -> %u usec %u\n", ith, nth, nrows, ir_first,
         ir_last, src_row_size, src_stride, dst_stride, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
 }
 
 static void quantize_f32_f16(unsigned int nth, unsigned int ith, void * data) {
     struct htp_matmul_context * mmctx = data;
     struct htp_ops_context * octx = mmctx->octx;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const struct htp_tensor * src = octx->src[1];
     uint8_t * restrict dst = octx->src1_spad.data;
@@ -4322,6 +4363,7 @@ static void quantize_f32_f16(unsigned int nth, unsigned int ith, void * data) {
     const uint32_t nrows = ne1 * ne2 * ne3;                             // total n_rows
 
     const uint32_t ir_first = nrows_per_thread * ith;                   // first row
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
     const uint32_t ir_last  = MIN(ir_first + nrows_per_thread, nrows);  // last row
 
     const size_t src_row_size = ne0 * sizeof(float);
@@ -4342,12 +4384,14 @@ static void quantize_f32_f16(unsigned int nth, unsigned int ith, void * data) {
 
     FARF(HIGH, "quantize-f32-f16: %u/%u : n-rows %u (%u:%u) row-size %u (%u) -> %u usec %u\n", ith, nth, nrows, ir_first,
         ir_last, src_row_size, src_stride, dst_stride, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
 }
 
 // TODO just a plain copy that should be done via the DMA during the Op setup
 static void quantize_f16_f16(unsigned int nth, unsigned int ith, void * data) {
     struct htp_matmul_context * mmctx = data;
     struct htp_ops_context * octx = mmctx->octx;
+    struct htp_thread_trace * tr = octx->ctx ? &octx->ctx->trace[ith] : NULL;
 
     const struct htp_tensor * src = octx->src[1];
     uint8_t * restrict dst = octx->src1_spad.data;
@@ -4364,6 +4408,7 @@ static void quantize_f16_f16(unsigned int nth, unsigned int ith, void * data) {
     const uint32_t nrows = ne1 * ne2 * ne3;                             // total n_rows
 
     const uint32_t ir_first = nrows_per_thread * ith;                   // first row
+    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
     const uint32_t ir_last  = MIN(ir_first + nrows_per_thread, nrows);  // last row
 
     const size_t src_row_size = ne0 * sizeof(float);
@@ -4384,6 +4429,7 @@ static void quantize_f16_f16(unsigned int nth, unsigned int ith, void * data) {
 
     FARF(HIGH, "quantize-f16-f16: %u/%u : n-rows %u (%u:%u) row-size %u (%u) -> %u usec %u\n", ith, nth, nrows, ir_first,
         ir_last, src_row_size, src_stride, dst_stride, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+    htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_A_QUANT, ir_first);
 }
 
 

ggml/src/ggml-hexagon/htp/ssm-conv.c

@@ -183,24 +183,25 @@ static inline void hvx_transpose_32x32_f32(HVX_Vector m[32]) {
 // transposed into VTCM.
 //
 // VTCM layouts (per thread):
-//   src1_T : {d_inner_per_thread, d_conv}   — staged once per launch (small).
-//   src0_T : {d_inner_tile,     ncs}        — staged per d_inner-tile.
+//   src1_T : {d_inner_stride, d_conv}       - staged once per launch (small).
+//   src0_T : {d_inner_tile,     ncs}        - staged per d_inner-tile.
 //
 // d_inner_tile is chosen so that per-thread VTCM stays under the budget.
 // Each thread iterates ceil(d_inner_per_thread d_inner_tile) tiles serially.
 #define HTP_SSM_CONV_VTCM_BUDGET (1u << 20) // 1 MiB per thread
 
-// Scalar transpose: src1 {d_conv, d_inner} (DDR) -> {d_inner_per_thread, d_conv} (VTCM)
+// Scalar transpose: src1 {d_conv, d_inner} (DDR) -> {d_inner_stride, d_conv} (VTCM)
 static inline void transpose_src1(const float * src1_data,
                                   uint32_t      src1_stride_inner,
                                   uint32_t      i1_off,
                                   uint32_t      d_inner_per_thread,
+                                  uint32_t      d_inner_stride,
                                   uint32_t      d_conv,
                                   float *       src1_T) {
     for (uint32_t i = 0; i < d_inner_per_thread; ++i) {
         const float * src_row = src1_data + (i1_off + i) * src1_stride_inner;
         for (uint32_t j = 0; j < d_conv; ++j) {
-            src1_T[j * d_inner_per_thread + i] = src_row[j];
+            src1_T[j * d_inner_stride + i] = src_row[j];
         }
     }
 }
@@ -280,6 +281,7 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
     }
 
     const uint32_t d_inner_per_thread = ir1 - ir0;
+    const uint32_t d_inner_stride     = scctx->nrows_per_thread;
     const uint32_t d_inner_tile       = scctx->d_inner_tile;
 
     const float * src0_data = (const float *) src0->data;
@@ -290,8 +292,8 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
     float * src0_T = (float *)(octx->src0_spad.data + ith * octx->src0_spad.size_per_thread);
     float * src1_T = (float *)(octx->src1_spad.data + ith * octx->src1_spad.size_per_thread);
 
-    // Stage src1 weights once into VTCM in {d_inner_per_thread, d_conv} layout.
-    transpose_src1(src1_data, src1_stride_inner, ir0, d_inner_per_thread, d_conv, src1_T);
+    // Stage src1 weights once into VTCM in {d_inner_stride, d_conv} layout.
+    transpose_src1(src1_data, src1_stride_inner, ir0, d_inner_per_thread, d_inner_stride, d_conv, src1_T);
 
     const uint32_t C_TILE = VLEN_FP32;
 
@@ -314,7 +316,7 @@ static void ssm_conv_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void
                     HVX_Vector acc = hvx_vec_splat_f32(0.0f);
                     for (uint32_t j = 0; j < d_conv; ++j) {
                         HVX_Vector x = *(const HVX_Vector *) (src0_T + (t + j) * d_inner_tile + cb);
-                        HVX_Vector w = *(const HVX_Vector *) (src1_T + j * d_inner_per_thread + tile_off + cb);
+                        HVX_Vector w = *(const HVX_Vector *) (src1_T + j * d_inner_stride + tile_off + cb);
                         acc          = Q6_Vqf32_vadd_Vqf32Vqf32(acc, Q6_Vqf32_vmpy_VsfVsf(x, w));
                     }
                     HVX_Vector res = Q6_Vsf_equals_Vqf32(acc);
@@ -362,8 +364,7 @@ int op_ssm_conv_f32(struct htp_ops_context * octx) {
             use_hvx = 1;
         }
 
-        scctx.nrows_per_thread  = (d_inner + n_threads - 1) / n_threads;
-        scctx.nrows_per_thread += (scctx.nrows_per_thread & 1);
+        scctx.nrows_per_thread = hex_round_up((d_inner + n_threads - 1) / n_threads, VLEN_FP32);
 
         const uint32_t d_inner_per_thread = scctx.nrows_per_thread;
         const uint32_t ncs                = src0->ne[0];

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

@@ -3788,7 +3788,7 @@ static void ggml_webgpu_init_memset_pipeline(webgpu_global_context & ctx) {
     ctx->memset_pipeline = ggml_webgpu_create_pipeline(ctx->device, wgsl_memset, "memset", constants);
 }
 
-static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
+static void ggml_backend_webgpu_request_adapter(wgpu::Instance & instance, wgpu::Adapter & adapter) {
     wgpu::RequestAdapterOptions options = {};
 
 #ifndef __EMSCRIPTEN__
@@ -3800,17 +3800,20 @@ static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
     options.nextInChain                   = &adapterTogglesDesc;
 #endif
 
-    ctx->webgpu_global_ctx->instance.WaitAny(
-        ctx->webgpu_global_ctx->instance.RequestAdapter(
-            &options, wgpu::CallbackMode::AllowSpontaneous,
-            [&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) {
-                if (status != wgpu::RequestAdapterStatus::Success) {
-                    GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
-                    return;
-                }
-                ctx->webgpu_global_ctx->adapter = std::move(adapter);
-            }),
-        UINT64_MAX);
+    instance.WaitAny(instance.RequestAdapter(
+                         &options, wgpu::CallbackMode::AllowSpontaneous,
+                         [&adapter](wgpu::RequestAdapterStatus status, wgpu::Adapter _adapter, const char * message) {
+                             if (status != wgpu::RequestAdapterStatus::Success) {
+                                 GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
+                                 return;
+                             }
+                             adapter = std::move(_adapter);
+                         }),
+                     UINT64_MAX);
+}
+
+static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
+    ggml_backend_webgpu_request_adapter(ctx->webgpu_global_ctx->instance, ctx->webgpu_global_ctx->adapter);
     GGML_ASSERT(ctx->webgpu_global_ctx->adapter != nullptr);
 
     ctx->webgpu_global_ctx->adapter.GetLimits(&ctx->webgpu_global_ctx->capabilities.limits);
@@ -4543,20 +4546,7 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {
     // Probe for adapter support
     wgpu::Adapter adapter;
     if (ctx->webgpu_global_ctx->instance != nullptr) {
-        wgpu::RequestAdapterOptions options = {};
-
-        // probe for adapter support
-        ctx->webgpu_global_ctx->instance.WaitAny(
-            ctx->webgpu_global_ctx->instance.RequestAdapter(
-                &options, wgpu::CallbackMode::AllowSpontaneous,
-                [&adapter](wgpu::RequestAdapterStatus status, wgpu::Adapter _adapter, const char * message) {
-                    if (status != wgpu::RequestAdapterStatus::Success) {
-                        GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
-                        return;
-                    }
-                    adapter = std::move(_adapter);
-                }),
-            UINT64_MAX);
+        ggml_backend_webgpu_request_adapter(ctx->webgpu_global_ctx->instance, adapter);
     }
 
     // WebGPU backend requires f16 support and, on native, implicit device synchronization.

ggml/src/ggml.c

@@ -600,18 +600,15 @@ FILE * ggml_fopen(const char * fname, const char * mode) {
     // convert fname (UTF-8)
     wchar_t * wfname = ggml_mbstowcs(fname);
     if (wfname) {
-        // convert mode (ANSI)
-        wchar_t * wmode = GGML_MALLOC((strlen(mode) + 1) * sizeof(wchar_t));
-        wchar_t * wmode_p = wmode;
-        do {
-            *wmode_p++ = (wchar_t)*mode;
-        } while (*mode++);
-
-        // open file
-        file = _wfopen(wfname, wmode);
+        // convert mode (UTF-8)
+        wchar_t * wmode = ggml_mbstowcs(mode);
+        if (wmode) {
+            // open file
+            file = _wfopen(wfname, wmode);
+            GGML_FREE(wmode);
+        }
 
         GGML_FREE(wfname);
-        GGML_FREE(wmode);
     }
 
     return file;

scripts/snapdragon/ggml-hexagon-profile.py

@@ -6,6 +6,7 @@ import re
 import argparse
 import statistics
 import logging
+from typing import Any, Dict, List, Optional
 
 from collections import defaultdict
 
@@ -25,12 +26,47 @@ COL_MAP = {
 }
 
 op_pattern = re.compile(
-    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+.*?\s+(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?"
+    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+.*?\s+(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?"
+)
+
+trace_pattern = re.compile(
+    r"trace-op\s+(?P<op_name>[A-Z_0-9+]+):\s+thread\s+(?P<thread>\d+)\s+event\s+(?P<event>[A-Z_0-9\-]+)\s+info\s+(?P<info>\d+)\s+(?P<state>start|stop)\s+(?P<cycles>\d+)"
 )
 
 logger = logging.getLogger("ggml-hexagon-profile")
 
 
+def normalize_event_name(evt_type):
+    if evt_type == "HVX_COMP":
+        return "V-COMP"
+    if evt_type == "HMX_COMP":
+        return "M-COMP"
+
+    # Strip HVX_ or HMX_ prefixes
+    name = evt_type
+    if name.startswith("HVX_") or name.startswith("HMX_"):
+        name = name[4:]
+    return name.replace("_", "-")
+
+
+class CycleUnwrapper:
+    def __init__(self):
+        self.last_raw = None
+        self.high_part = 0
+
+    def unwrap(self, raw):
+        if self.last_raw is None:
+            self.last_raw = raw
+            return raw
+        diff = raw - self.last_raw
+        if diff < -0x80000000:
+            self.high_part += 0x100000000
+        elif diff > 0x80000000:
+            self.high_part -= 0x100000000
+        self.last_raw = raw
+        return raw + self.high_part
+
+
 def parse_log(file_path, pmu_index=None):
     try:
         if file_path != "-":
@@ -41,35 +77,211 @@ def parse_log(file_path, pmu_index=None):
         logger.error(f"file '{file_path}' not found.")
         sys.exit(1)
 
-    all_ops = []
+    all_ops: List[Dict[str, Any]] = []
+    current_op: Optional[Dict[str, Any]] = None
+
+    timestamp_pattern = re.compile(r"^(?P<min>\d+)\.(?P<sec>\d+)\.(?P<ms>\d+)\.(?P<us>\d+)\s+[A-Z]\s+")
+    unwrapper = CycleUnwrapper()
+
     for line in f:
-        match = op_pattern.search(line)
-        if not match: continue
+        ts_match = timestamp_pattern.match(line)
+        abs_usec = 0
+        if ts_match:
+            abs_usec = (
+                (int(ts_match.group('min')) * 60 + int(ts_match.group('sec'))) * 1000000
+                + int(ts_match.group('ms')) * 1000
+                + int(ts_match.group('us'))
+            )
 
-        pmu_raw = match.group('pmu')
-        pmu_val = None
-        if pmu_raw and pmu_index is not None:
-            try:
-                pmu_list = [int(x.strip()) for x in pmu_raw.split(',')]
-                if len(pmu_list) > pmu_index:
-                    pmu_val = pmu_list[pmu_index]
-            except (ValueError, IndexError):
-                pmu_val = None
+        op_match = op_pattern.search(line)
+        if op_match:
+            pmu_raw = op_match.group('pmu')
+            pmu_val = None
+            if pmu_raw and pmu_index is not None:
+                try:
+                    pmu_list = [int(x.strip()) for x in pmu_raw.split(',')]
+                    if len(pmu_list) > pmu_index:
+                        pmu_val = pmu_list[pmu_index]
+                except (ValueError, IndexError):
+                    pmu_val = None
 
-        all_ops.append({
-            'name':    match.group('op_name'),
-            'dims':    match.group('dims').strip(),
-            'types':   match.group('types').strip(),
-            'usec':    int(match.group('usec')),
-            'cycles':  int(match.group('cycles')),
-            'pmu_val': pmu_val
-        })
+            evt_raw = op_match.group('evt')
+            evt_val = None
+            if evt_raw:
+                try:
+                    evt_val = [int(x.strip()) for x in evt_raw.split(',')]
+                except ValueError:
+                    evt_val = None
+
+            cycles_start_raw = op_match.group('start')
+            unwrapped_cycles_start = None
+            if cycles_start_raw:
+                unwrapped_cycles_start = unwrapper.unwrap(int(cycles_start_raw))
+
+            idx = line.find("profile-op ")
+            op_text = line[idx + 11:].strip() if idx != -1 else line.strip()
+
+            current_op = {
+                'name':         op_match.group('op_name'),
+                'dims':         op_match.group('dims').strip(),
+                'types':        op_match.group('types').strip(),
+                'op_text':      op_text,
+                'usec':         int(op_match.group('usec')),
+                'cycles':       int(op_match.group('cycles')),
+                'cycles_start': int(cycles_start_raw) if cycles_start_raw else None,
+                'unwrapped_cycles_start': unwrapped_cycles_start,
+                'pmu_val':      pmu_val,
+                'evt_val':      evt_val,
+                'abs_usec':     abs_usec,
+                'trace_events': []
+            }
+            all_ops.append(current_op)
+            continue
+
+        trace_match = trace_pattern.search(line)
+        if trace_match and current_op:
+            if trace_match.group('op_name') == current_op['name']:
+                raw_cyc = int(trace_match.group('cycles'))
+                current_op['trace_events'].append({
+                    'thread': int(trace_match.group('thread')),
+                    'event':  trace_match.group('event'),
+                    'info':   int(trace_match.group('info')),
+                    'cycles': raw_cyc,
+                    'unwrapped_cycles': unwrapper.unwrap(raw_cyc),
+                    'state':  trace_match.group('state')
+                })
 
     f.close()
-
     return all_ops
 
 
+def print_ascii_timeline(op_name, dims, types, usec, cycles, events, evt_val=None):
+    evt_str = ""
+    if evt_val:
+        evt_str = " - evt [" + ",".join(str(x) for x in evt_val) + "]"
+    logger.info("=" * 100)
+    logger.info(f"{op_name} ({dims} : {types}) - {usec} usec {cycles} cycles{evt_str}")
+    logger.info("=" * 100)
+
+    events = sorted(events, key=lambda e: e['cycles'])
+    if not events:
+        logger.info("  No trace events recorded.")
+        return
+
+    min_cycles = events[0]['cycles']
+
+    logger.info("Cycles      %-30s" % "EventDetails" + " ".join(f"T{i:<2}" for i in range(10)) + " HMX")
+    logger.info("-" * 100)
+
+    thread_stacks = [[] for _ in range(11)]
+
+    for e in events:
+        t = e['thread']
+        if t < 0 or t > 10:
+            continue
+
+        if e['cycles'] >= min_cycles:
+            rel_cycles = e['cycles'] - min_cycles
+        else:
+            rel_cycles = (e['cycles'] + 0x100000000) - min_cycles
+
+        state = e['state']
+        evt_type = e['event']
+
+        # Determine char representing the event
+        norm_evt = normalize_event_name(evt_type)
+        char = '?'
+        if norm_evt == 'V-COMP':
+            char = 'V'
+        elif norm_evt == 'M-COMP':
+            char = 'H'
+        elif norm_evt == 'A-QUANT':
+            char = 'Q'
+        elif norm_evt == 'A-PREP':
+            char = 'A'
+        elif norm_evt == 'W-DEQUANT':
+            char = 'D'
+        elif norm_evt == 'O-PROC':
+            char = 'O'
+        elif norm_evt == 'W-PREP':
+            char = 'P'
+        elif norm_evt == 'DMA':
+            char = 'M'
+
+        if state == 'start':
+            thread_stacks[t].append(char)
+        elif state == 'stop':
+            if thread_stacks[t]:
+                if thread_stacks[t][-1] == char:
+                    thread_stacks[t].pop()
+                elif char in thread_stacks[t]:
+                    thread_stacks[t].remove(char)
+                else:
+                    thread_stacks[t].pop()
+
+        cols = []
+        for i in range(11):
+            if thread_stacks[i]:
+                cols.append(f"[{thread_stacks[i][-1]}]")
+            else:
+                cols.append(" | ")
+
+        evt_desc = f"T{t}: {evt_type} {state} ({e['info']})"
+        logger.info(f"{rel_cycles:10d}  %-30s" % evt_desc + " ".join(cols[:10]) + "  " + cols[10])
+    logger.info("-" * 100)
+
+
+def print_ascii_summary(op_name, dims, types, usec, cycles, events, evt_val=None):
+    evt_str = ""
+    if evt_val:
+        evt_str = " - evt [" + ",".join(str(x) for x in evt_val) + "]"
+    logger.info("=" * 100)
+    logger.info(f"{op_name} ({dims} : {types}) - {usec} usec {cycles} cycles{evt_str}")
+    logger.info("=" * 100)
+
+    events = sorted(events, key=lambda e: e['cycles'])
+    if not events:
+        logger.info("  No trace events recorded.")
+        return
+
+    active_starts = {}
+    thread_totals = defaultdict(lambda: defaultdict(int))
+
+    for e in events:
+        t = e['thread']
+        evt = e['event']
+        info = e['info']
+        cyc = e['cycles']
+        state = e['state']
+
+        key = (t, evt, info)
+        if state == 'start':
+            active_starts[key] = cyc
+        elif state == 'stop':
+            if key in active_starts:
+                start_cyc = active_starts[key]
+                del active_starts[key]
+
+                if cyc >= start_cyc:
+                    dur = cyc - start_cyc
+                else:
+                    dur = (cyc + 0x100000000) - start_cyc
+
+                norm_evt = normalize_event_name(evt)
+                thread_totals[t][norm_evt] += dur
+
+    for t in sorted(thread_totals.keys()):
+        thread_name = f"Thread {t} (HVX)" if t != 10 else "Thread 10 (HMX)"
+        sorted_evts = sorted(thread_totals[t].items(), key=lambda item: item[0])
+
+        evt_strs = []
+        for evt, dur in sorted_evts:
+            pct = (dur / cycles * 100) if cycles > 0 else 0
+            evt_strs.append(f"{evt} {dur} ({pct:.1f}%)")
+
+        logger.info(f"  {thread_name:<16}: " + " | ".join(evt_strs))
+
+
 def generate_report(ops, top_n, width_overrides, sort_col, pmu_name=None):
     if not ops:
         logger.info("No valid records found.")
@@ -115,7 +327,6 @@ def generate_report(ops, top_n, width_overrides, sort_col, pmu_name=None):
 
     # Sorting logic
     actual_sort_key = COL_MAP[sort_col][2]
-    # We sort numeric fields descending, strings (op/dims) ascending
     is_numeric    = actual_sort_key.startswith("_") or actual_sort_key == "count"
     sorted_groups = sorted(group_stats, key=lambda x: x[actual_sort_key], reverse=is_numeric)[:top_n]
 
@@ -132,7 +343,7 @@ def generate_report(ops, top_n, width_overrides, sort_col, pmu_name=None):
         if "pmu" in col_name and pmu_name:
             header_text = header_text.replace("PMU", pmu_name)
 
-        natural_width = max([len(row[data_key]) for row in sorted_groups] + [len(header_text)])
+        natural_width = max([len(str(row[data_key])) for row in sorted_groups] + [len(header_text)])
         target_width  = width_overrides.get(col_name, natural_width)
 
         if target_width == 0:
@@ -152,7 +363,7 @@ def generate_report(ops, top_n, width_overrides, sort_col, pmu_name=None):
     for group in sorted_groups:
         row_vals = []
         for i, key in enumerate(final_keys):
-            val = group[key]
+            val = str(group[key])
             if len(val) > final_widths[i]:
                 val = val[:final_widths[i] - 3] + "..."
             row_vals.append(f"{val:<{final_widths[i]}}")
@@ -167,12 +378,18 @@ def main():
     parser.add_argument("--pmu-index", type=int)
     parser.add_argument("--pmu-name", type=str)
     parser.add_argument("--width", action='append', default=['dims:40'], help="Override column width, e.g. --width dims:50")
+    parser.add_argument("--timeline", type=str, nargs='?', const='summary', choices=["summary", "diagram"],
+                        help="Output ASCII art event summary or timing diagram (default: summary)")
+    parser.add_argument("--filter", type=str, help="Regex filter matching against the original profile-op line")
+
+    group = parser.add_mutually_exclusive_group()
+    group.add_argument("--head", type=int, help="Limit to first N ops")
+    group.add_argument("--tail", type=int, help="Limit to last N ops")
 
     args = parser.parse_args()
 
     logging.basicConfig(level=logging.INFO, format='%(message)s')
 
-    # Sort validation: can't sort by PMU if index isn't provided
     if "pmu" in args.sort and args.pmu_index is None:
         logger.error(f"Cannot sort by '{args.sort}' without --pmu-index.")
         sys.exit(1)
@@ -188,7 +405,33 @@ def main():
 
     final_pmu_name = (args.pmu_name or f"#{args.pmu_index}") if args.pmu_index is not None else None
     ops = parse_log(args.logfile, pmu_index=args.pmu_index)
-    generate_report(ops, args.top, overrides, args.sort, pmu_name=final_pmu_name)
+
+    if args.filter:
+        try:
+            filter_re = re.compile(args.filter)
+        except re.error as e:
+            logger.error(f"Invalid regex filter: {e}")
+            sys.exit(1)
+        ops = [op for op in ops if filter_re.search(op['op_text'])]
+
+    if args.head is not None:
+        ops = ops[:args.head]
+    elif args.tail is not None:
+        ops = ops[-args.tail:]
+
+    if args.timeline:
+        logger.info(f"\n# ASCII Timing {args.timeline.capitalize()}\n")
+        printed_cnt = 0
+        for op in ops:
+            if args.timeline == "summary":
+                print_ascii_summary(op['name'], op['dims'], op['types'], op['usec'], op['cycles'], op['trace_events'], op.get('evt_val'))
+            elif args.timeline == "diagram":
+                print_ascii_timeline(op['name'], op['dims'], op['types'], op['usec'], op['cycles'], op['trace_events'], op.get('evt_val'))
+            printed_cnt += 1
+            if printed_cnt >= args.top:
+                break
+    else:
+        generate_report(ops, args.top, overrides, args.sort, pmu_name=final_pmu_name)
 
 
 if __name__ == "__main__":

scripts/snapdragon/ggml-hexagon-trace.py

@@ -0,0 +1,463 @@
+#!/usr/bin/env python3
+
+import sys
+import os
+import re
+import argparse
+import statistics
+import logging
+from typing import Any, Dict, List, Optional
+from collections import defaultdict
+
+logger = logging.getLogger("ggml-hexagon-trace")
+
+op_pattern = re.compile(
+    r"profile-op\s+(?P<op_name>[A-Z_0-9+]+):\s+.*?\s+:\s+(?P<dims>[\d:x\s\->!]+)\s+:\s+(?P<types>[a-z\d_\s\->x]+)\s+:\s+(?P<strides>[\d:x\s\->!]+)\s+:\s+(?:op-)?usec\s+(?P<usec>\d+)\s+(?:op-)?cycles\s+(?P<cycles>\d+)(?:\s+start\s+(?P<start>\d+))?(?:\s+mhz\s+(?P<mhz>[\d.]+))?(?:\s+pmu\s+\[(?P<pmu>[\d,\s]+)\])?(?:\s+evt\s+\[(?P<evt>[\d,\s]+)\])?"
+)
+
+trace_pattern = re.compile(
+    r"trace-op\s+(?P<op_name>[A-Z_0-9+]+):\s+thread\s+(?P<thread>\d+)\s+event\s+(?P<event>[A-Z_0-9\-]+)\s+info\s+(?P<info>\d+)\s+(?P<state>start|stop)\s+(?P<cycles>\d+)"
+)
+
+
+def normalize_event_name(evt_type):
+    if evt_type == "HVX_COMP":
+        return "V-COMP"
+    if evt_type == "HMX_COMP":
+        return "M-COMP"
+    name = evt_type
+    if name.startswith("HVX_") or name.startswith("HMX_"):
+        name = name[4:]
+    return name.replace("_", "-")
+
+
+class CycleUnwrapper:
+    def __init__(self):
+        self.last_raw = None
+        self.high_part = 0
+
+    def unwrap(self, raw):
+        if self.last_raw is None:
+            self.last_raw = raw
+            return raw
+        diff = raw - self.last_raw
+        if diff < -0x80000000:
+            self.high_part += 0x100000000
+        elif diff > 0x80000000:
+            self.high_part -= 0x100000000
+        self.last_raw = raw
+        return raw + self.high_part
+
+
+def parse_log(file_path):
+    try:
+        if file_path != "-":
+            f = open(file_path, 'r', encoding='utf-8', errors='ignore')
+        else:
+            f = os.fdopen(0, 'r', encoding='utf-8', errors='ignore')
+    except FileNotFoundError:
+        logger.error(f"file '{file_path}' not found.")
+        sys.exit(1)
+
+    all_ops: List[Dict[str, Any]] = []
+    current_op: Optional[Dict[str, Any]] = None
+    unwrapper = CycleUnwrapper()
+    line_idx = 0
+
+    for line in f:
+        line_idx += 1
+        op_match = op_pattern.search(line)
+        if op_match:
+            cycles_start_raw = op_match.group('start')
+            unwrapped_cycles_start = None
+            if cycles_start_raw:
+                unwrapped_cycles_start = unwrapper.unwrap(int(cycles_start_raw))
+
+            idx = line.find("profile-op ")
+            op_text = line[idx + 11:].strip() if idx != -1 else line.strip()
+
+            current_op = {
+                'name':         op_match.group('op_name'),
+                'dims':         op_match.group('dims').strip() if op_match.group('dims') else '',
+                'types':        op_match.group('types').strip() if op_match.group('types') else '',
+                'strides':      op_match.group('strides').strip() if op_match.group('strides') else '',
+                'op_text':      op_text,
+                'usec':         int(op_match.group('usec')),
+                'cycles':       int(op_match.group('cycles')),
+                'cycles_start': int(cycles_start_raw) if cycles_start_raw else None,
+                'unwrapped_cycles_start': unwrapped_cycles_start,
+                'trace_events': [],
+                'line_num':     line_idx
+            }
+            all_ops.append(current_op)
+            continue
+
+        trace_match = trace_pattern.search(line)
+        if trace_match and current_op:
+            if trace_match.group('op_name') == current_op['name']:
+                raw_cyc = int(trace_match.group('cycles'))
+                current_op['trace_events'].append({
+                    'thread': int(trace_match.group('thread')),
+                    'event':  trace_match.group('event'),
+                    'info':   int(trace_match.group('info')),
+                    'cycles': raw_cyc,
+                    'unwrapped_cycles': unwrapper.unwrap(raw_cyc),
+                    'state':  trace_match.group('state')
+                })
+
+    f.close()
+    return all_ops
+
+# --- Simple protobuf encoder ---
+
+
+def write_varint(val):
+    if val < 0:
+        val = (1 << 64) + val
+    res = bytearray()
+    while True:
+        towrite = val & 0x7f
+        val >>= 7
+        if val > 0:
+            res.append(towrite | 0x80)
+        else:
+            res.append(towrite)
+            break
+    return bytes(res)
+
+
+def pb_field(num, wire, data):
+    return write_varint((num << 3) | wire) + data
+
+
+def pb_varint(num, val):
+    return pb_field(num, 0, write_varint(val))
+
+
+def pb_length_delimited(num, data):
+    return pb_field(num, 2, write_varint(len(data)) + data)
+
+
+def pb_string(num, text):
+    return pb_length_delimited(num, text.encode('utf-8'))
+
+
+# Message Encoders
+def make_process_descriptor(pid, name):
+    return pb_varint(1, pid) + pb_string(6, name)
+
+
+def make_thread_descriptor(pid, tid, name, sort_index=None):
+    payload = pb_varint(1, pid) + pb_varint(2, tid) + pb_string(5, name)
+    if sort_index is not None:
+        payload += pb_varint(3, sort_index)
+    return payload
+
+
+def make_track_descriptor(uuid, name=None, parent_uuid=None, thread=None, process=None, sibling_merge_behavior=None, child_ordering=None, sibling_order_rank=None):
+    payload = pb_varint(1, uuid)
+    if name is not None:
+        payload += pb_string(2, name)
+    if parent_uuid is not None:
+        payload += pb_varint(5, parent_uuid)
+    if process is not None:
+        payload += pb_length_delimited(3, process)
+    if thread is not None:
+        payload += pb_length_delimited(4, thread)
+    if sibling_merge_behavior is not None:
+        payload += pb_varint(15, sibling_merge_behavior)
+    if child_ordering is not None:
+        payload += pb_varint(11, child_ordering)
+    if sibling_order_rank is not None:
+        payload += pb_varint(12, sibling_order_rank)
+    return payload
+
+
+def make_debug_annotation(name, string_val=None, int_val=None):
+    payload = pb_string(10, name)
+    if string_val is not None:
+        payload += pb_string(6, string_val)
+    elif int_val is not None:
+        payload += pb_varint(4, int_val)
+    return payload
+
+
+def make_track_event(event_type, track_uuid, name=None, category=None, debug_annotations=None):
+    payload = pb_varint(9, event_type)
+    payload += pb_varint(11, track_uuid)
+    if name is not None:
+        payload += pb_string(23, name)
+    if category is not None:
+        payload += pb_string(22, category)
+    if debug_annotations is not None:
+        for da in debug_annotations:
+            payload += pb_length_delimited(4, da)
+    return payload
+
+
+def make_trace_packet(timestamp, track_event=None, track_descriptor=None, seq_id=1):
+    payload = pb_varint(8, timestamp)
+    payload += pb_varint(10, seq_id)
+    if track_event is not None:
+        payload += pb_length_delimited(11, track_event)
+    if track_descriptor is not None:
+        payload += pb_length_delimited(60, track_descriptor)
+    return payload
+
+
+def write_trace_packet_to_file(f, packet_bytes):
+    # Write as field 1 of top-level Trace message
+    f.write(pb_length_delimited(1, packet_bytes))
+
+# --- End Protobuf Encoder ---
+
+
+def generate_perfetto_trace(filtered_ops, output_path):
+    if not filtered_ops:
+        logger.warning("No operators found after filtering.")
+        return
+
+    # Compute average frequency
+    frequencies = []
+    for op in filtered_ops:
+        if op['usec'] > 0 and op['cycles'] > 0:
+            frequencies.append(op['cycles'] / op['usec'])
+    avg_freq_mhz = statistics.mean(frequencies) if frequencies else 1000.0
+    if avg_freq_mhz <= 0:
+        avg_freq_mhz = 1000.0
+
+    # Assign start and end cycles to each operator
+    for op in filtered_ops:
+        op['start_cycles'] = op['unwrapped_cycles_start']
+        op['end_cycles'] = op['start_cycles'] + op['cycles']
+
+    global_min_cyc = min(op['start_cycles'] for op in filtered_ops if op['start_cycles'] is not None)
+
+    # Process events
+    completed_events = []
+    for op in filtered_ops:
+        events = op['trace_events']
+        if not events:
+            continue
+        events = sorted(events, key=lambda e: e['unwrapped_cycles'])
+
+        active_starts = {}
+        for e in events:
+            t = e['thread']
+            evt = e['event']
+            info = e['info']
+            state = e['state']
+            cyc = e['unwrapped_cycles']
+
+            key = (t, evt, info)
+            if state == 'start':
+                active_starts[key] = cyc
+            elif state == 'stop':
+                if key in active_starts:
+                    start_cyc = active_starts[key]
+                    del active_starts[key]
+                    completed_events.append({
+                        'thread': t,
+                        'event': evt,
+                        'info': info,
+                        'start_cyc': start_cyc,
+                        'end_cyc': cyc,
+                        'op_name': op['name']
+                    })
+
+    completed_events.sort(key=lambda e: e['start_cyc'])
+
+    # Convert event times to microseconds and apply clamp rounded to 1ns resolution (3 decimals)
+    for e in completed_events:
+        start_us = (e['start_cyc'] - global_min_cyc) / avg_freq_mhz
+        dur_us = (e['end_cyc'] - e['start_cyc']) / avg_freq_mhz
+        e['ts_ns'] = int(round(start_us * 1000))
+        e['dur_ns'] = int(round(max(dur_us, 0.1) * 1000))
+
+    # Allocate slots (sub-tracks) to prevent overlaps on same virtual track
+    active_slots = defaultdict(list)
+    for e in completed_events:
+        t = e['thread']
+        evt = e['event']
+        ts = e['ts_ns']
+        dur = e['dur_ns']
+
+        norm_evt = normalize_event_name(evt)
+        if norm_evt == "DMA":
+            track_key = (t, "DMA")
+        elif t == 10:
+            track_key = (t, "HMX")
+        else:
+            track_key = (t, "HVX")
+
+        slots = active_slots[track_key]
+        allocated_slot = -1
+        for idx, slot_end_ns in enumerate(slots):
+            if ts >= slot_end_ns:
+                slots[idx] = ts + dur
+                allocated_slot = idx
+                break
+        if allocated_slot == -1:
+            slots.append(ts + dur)
+            allocated_slot = len(slots) - 1
+        e['slot'] = allocated_slot
+
+    # Generate Track IDs and track definitions
+    used_tracks = {}
+    for e in completed_events:
+        t = e['thread']
+        evt = e['event']
+        slot = e['slot']
+
+        norm_evt = normalize_event_name(evt)
+        if norm_evt == "DMA":
+            track_evt = "DMA"
+            evt_id = 1
+        elif t == 10:
+            track_evt = "HMX"
+            evt_id = 3
+        else:
+            track_evt = "HVX"
+            evt_id = 2
+
+        t_sort = 1 if t == 10 else t + 2
+        # Unique UUID for each sub-track
+        if t == 10:
+            uuid = 20  # HMX thread track UUID
+        else:
+            uuid = int(t_sort * 1000000 + evt_id * 1000 + slot)
+        e['uuid'] = uuid
+        used_tracks[uuid] = (t, track_evt, slot)
+
+    with open(output_path, "wb") as f:
+        # Define Process with EXPLICIT child sorting
+        proc_desc = make_process_descriptor(1, "HTP NPU")
+        proc_packet = make_trace_packet(0, track_descriptor=make_track_descriptor(1, process=proc_desc, child_ordering=3))
+        write_trace_packet_to_file(f, proc_packet)
+
+        # Define Operators Track (UUID = 2) as a thread track at rank 1, tid 8
+        op_thread_desc = make_thread_descriptor(1, 8, "Ops", sort_index=1)
+        op_packet = make_trace_packet(0, track_descriptor=make_track_descriptor(2, parent_uuid=1, thread=op_thread_desc))
+        write_trace_packet_to_file(f, op_packet)
+
+        # Define HMX Thread Track (UUID = 20) at rank 2, tid 9
+        hmx_thread_desc = make_thread_descriptor(1, 9, "HMX", sort_index=2)
+        hmx_packet = make_trace_packet(0, track_descriptor=make_track_descriptor(20, parent_uuid=1, thread=hmx_thread_desc))
+        write_trace_packet_to_file(f, hmx_packet)
+
+        # Define Thread Tracks (T0, T1, ..., T9)
+        unique_threads = sorted(list(set(t for (t, _, _) in used_tracks.values() if t != 10)))
+        for t in unique_threads:
+            thread_uuid = 10 + t
+            thread_name = f"T{t}"
+            # Sort order starts from index 3 (T0 -> 3, T1 -> 4, etc.)
+            sort_index = 3 + t
+            tid = 10 + t
+            thread_desc = make_thread_descriptor(1, tid, thread_name, sort_index=sort_index)
+            thread_packet = make_trace_packet(0, track_descriptor=make_track_descriptor(
+                thread_uuid,
+                parent_uuid=1,
+                thread=thread_desc,
+                sibling_order_rank=sort_index,
+                child_ordering=3  # Explicit child sorting for sub-tracks
+            ))
+            write_trace_packet_to_file(f, thread_packet)
+
+        # Define Track descriptors for sub-tracks parented to thread tracks
+        for uuid in sorted(used_tracks.keys()):
+            if uuid == 20:
+                continue
+            t, evt, slot = used_tracks[uuid]
+            name = f"T{t} {evt}"
+            rank = 0 if evt == "HVX" else 1
+            parent_thread_uuid = 10 + t
+            # Sibling merge behavior: 1 (SIBLING_MERGE_BEHAVIOR_BY_TRACK_NAME)
+            track_desc = make_track_descriptor(
+                uuid=uuid,
+                name=name,
+                parent_uuid=parent_thread_uuid,
+                sibling_merge_behavior=1,
+                sibling_order_rank=rank
+            )
+            track_packet = make_trace_packet(0, track_descriptor=track_desc)
+            write_trace_packet_to_file(f, track_packet)
+
+        # Emit Operators
+        last_op_end_ns = 0
+        for op in filtered_ops:
+            op_start_ns = int(round(((op['start_cycles'] - global_min_cyc) / avg_freq_mhz) * 1000))
+            op_dur_ns = int(round((op['cycles'] / avg_freq_mhz) * 1000))
+            if op_start_ns < last_op_end_ns:
+                op_start_ns = last_op_end_ns
+            clamped_dur = max(op_dur_ns, 100) # Clamp to 100ns (0.1us)
+
+            # Debug annotations for Ops
+            debug_annots = []
+            if 'line_num' in op:
+                debug_annots.append(make_debug_annotation("line", int_val=op['line_num']))
+            if 'strides' in op and op['strides']:
+                debug_annots.append(make_debug_annotation("strides", string_val=op['strides']))
+
+            # Slice Begin
+            evt_begin = make_track_event(1, 2, name=f"{op['name']} ({op['dims']})", category="operator", debug_annotations=debug_annots)
+            packet_begin = make_trace_packet(op_start_ns, track_event=evt_begin)
+            write_trace_packet_to_file(f, packet_begin)
+
+            # Slice End
+            evt_end = make_track_event(2, 2)
+            packet_end = make_trace_packet(op_start_ns + clamped_dur, track_event=evt_end)
+            write_trace_packet_to_file(f, packet_end)
+
+            last_op_end_ns = op_start_ns + clamped_dur
+
+        # Emit Thread Trace Events
+        for e in completed_events:
+            norm_name = normalize_event_name(e['event'])
+            name = f"DMA {e['info']}" if norm_name == "DMA" else norm_name
+
+            # Slice Begin
+            evt_begin = make_track_event(1, e['uuid'], name=name, category="trace")
+            packet_begin = make_trace_packet(e['ts_ns'], track_event=evt_begin)
+            write_trace_packet_to_file(f, packet_begin)
+
+            # Slice End
+            evt_end = make_track_event(2, e['uuid'])
+            packet_end = make_trace_packet(e['ts_ns'] + e['dur_ns'], track_event=evt_end)
+            write_trace_packet_to_file(f, packet_end)
+
+    logger.info(f"Successfully generated Perfetto trace at {output_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Convert Hexagon Op profile logs to native Perfetto Protobuf traces.")
+    parser.add_argument("logfile", help="Path to hex-log profile file")
+    parser.add_argument("-o", "--output", default="optrace.perfetto-trace", help="Output trace file path (default: optrace.perfetto-trace)")
+    parser.add_argument("--filter", type=str, help="Regex filter matching against the original profile-op line")
+
+    group = parser.add_mutually_exclusive_group()
+    group.add_argument("--head", type=int, help="Limit to first N ops")
+    group.add_argument("--tail", type=int, help="Limit to last N ops")
+
+    args = parser.parse_args()
+    logging.basicConfig(level=logging.INFO, format='%(message)s')
+
+    ops = parse_log(args.logfile)
+
+    if args.filter:
+        try:
+            filter_re = re.compile(args.filter)
+        except re.error as e:
+            logger.error(f"Invalid regex filter: {e}")
+            sys.exit(1)
+        ops = [op for op in ops if filter_re.search(op['op_text'])]
+
+    if args.head is not None:
+        ops = ops[:args.head]
+    elif args.tail is not None:
+        ops = ops[-args.tail:]
+
+    generate_perfetto_trace(ops, args.output)
+
+
+if __name__ == "__main__":
+    main()

scripts/sync-ggml.last

@@ -1 +1 @@
-3af5f5760e19a96427f5f7a93b79cbdf3d4b265b
+707321c4cf6d21cb4bc831aa8b687dbf01a521ce

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "refs/tags/v0.47.0"
+HTTPLIB_VERSION = "refs/tags/v0.48.0"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",

scripts/ui-assets.cmake

@@ -20,6 +20,7 @@ set(LLAMA_UI_GZIP     "" CACHE STRING "Apply gzip compress to assets to save ban
 
 set(DIST_DIR     "${UI_BINARY_DIR}/dist")
 set(SRC_DIST_DIR "${UI_SOURCE_DIR}/dist")
+set(WORK_DIR     "${UI_BINARY_DIR}/ui-src")
 set(STAMP_FILE   "${UI_BINARY_DIR}/.ui-stamp")
 set(UI_CPP       "${UI_BINARY_DIR}/ui.cpp")
 set(UI_H         "${UI_BINARY_DIR}/ui.h")
@@ -64,6 +65,22 @@ function(npm_build_should_skip out_var)
     set(${out_var} TRUE PARENT_SCOPE)
 endfunction()
 
+function(stage_sources)
+    if(EXISTS "${WORK_DIR}")
+        file(GLOB staged RELATIVE "${WORK_DIR}" "${WORK_DIR}/*")
+        list(REMOVE_ITEM staged "node_modules")
+        foreach(entry ${staged})
+            file(REMOVE_RECURSE "${WORK_DIR}/${entry}")
+        endforeach()
+    endif()
+
+    file(COPY "${UI_SOURCE_DIR}/"
+        DESTINATION "${WORK_DIR}"
+        NO_SOURCE_PERMISSIONS
+        PATTERN "node_modules" EXCLUDE
+    )
+endfunction()
+
 function(npm_build out_var)
     set(${out_var} FALSE PARENT_SCOPE)
 
@@ -89,14 +106,16 @@ function(npm_build out_var)
         return()
     endif()
 
+    stage_sources()
+
     # npm writes node_modules/.package-lock.json on every successful install,
     # so a package-lock.json newer than this marker means node_modules is stale
-    set(NPM_MARKER "${UI_SOURCE_DIR}/node_modules/.package-lock.json")
+    set(NPM_MARKER "${WORK_DIR}/node_modules/.package-lock.json")
     set(need_install FALSE)
     if(NOT EXISTS "${NPM_MARKER}")
         set(need_install TRUE)
     else()
-        file(TIMESTAMP "${UI_SOURCE_DIR}/package-lock.json" lock_ts)
+        file(TIMESTAMP "${WORK_DIR}/package-lock.json" lock_ts)
         file(TIMESTAMP "${NPM_MARKER}" marker_ts)
         if(lock_ts STRGREATER marker_ts)
             set(need_install TRUE)
@@ -107,7 +126,7 @@ function(npm_build out_var)
         message(STATUS "UI: running npm install")
         execute_process(
             COMMAND ${NPM_EXECUTABLE} install
-            WORKING_DIRECTORY "${UI_SOURCE_DIR}"
+            WORKING_DIRECTORY "${WORK_DIR}"
             RESULT_VARIABLE rc
             ERROR_VARIABLE  err
         )
@@ -124,7 +143,7 @@ function(npm_build out_var)
     execute_process(
         COMMAND ${CMAKE_COMMAND} -E env "LLAMA_UI_OUT_DIR=${DIST_DIR}" "LLAMA_UI_VERSION=${HF_VERSION}" "LLAMA_BUILD_NUMBER=${LLAMA_BUILD_NUMBER}"
                 ${NPM_EXECUTABLE} run build
-        WORKING_DIRECTORY "${UI_SOURCE_DIR}"
+        WORKING_DIRECTORY "${WORK_DIR}"
         RESULT_VARIABLE rc
         ERROR_VARIABLE  err
     )

src/llama-quant.cpp

@@ -932,8 +932,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
     // copy the KV pairs from the input file
     gguf_set_kv     (ctx_out.get(), ml.metadata);
-    gguf_set_val_u32(ctx_out.get(), "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV
-    gguf_set_val_u32(ctx_out.get(), "general.file_type", ftype); // TODO: use LLM_KV
+    gguf_set_val_u32(ctx_out.get(), ml.llm_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION).c_str(), GGML_QNT_VERSION);
+    gguf_set_val_u32(ctx_out.get(), ml.llm_kv(LLM_KV_GENERAL_FILE_TYPE).c_str(), ftype);
 
     // Remove split metadata
     gguf_remove_key(ctx_out.get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str());

src/models/glm-dsa.cpp

@@ -101,11 +101,11 @@ void llama_model_glm_dsa::load_arch_tensors(llama_model_loader &) {
         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
 
         // DSA indexer
-        layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags);
-        layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags);
-        layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags);
-        layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags);
-        layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
+        layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
+        layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
+        layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags | TENSOR_NOT_REQUIRED);
+        layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
+        layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
         if (i < (int) hparams.n_layer_dense_lead) {
             layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
             layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);

src/models/qwen35.cpp

@@ -156,6 +156,8 @@ llama_model_qwen35::graph::graph(const llama_model & model, const llm_graph_para
 
     // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.
     for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = inpL;
+
         ggml_tensor * inpSA = inpL;
 
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);

src/models/qwen35moe.cpp

@@ -179,6 +179,8 @@ llama_model_qwen35moe::graph::graph(const llama_model & model, const llm_graph_p
 
     // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.
     for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = inpL;
+
         ggml_tensor * inpSA = inpL;
 
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);

tests/test-arg-parser.cpp

@@ -10,7 +10,7 @@
 #undef NDEBUG
 #include <cassert>
 
-int main(void) {
+static void test(void) {
     common_params params;
 
     printf("test-arg-parser: make sure there is no duplicated arguments in any examples\n\n");
@@ -210,3 +210,13 @@ int main(void) {
 
     printf("test-arg-parser: all tests OK\n\n");
 }
+
+int main(void) {
+    try {
+        test();
+    } catch (std::exception & e) {
+        fprintf(stderr, "test-arg-parser: exception: %s\n", e.what());
+        return 1;
+    }
+    return 0;
+}

tools/cli/README.md

@@ -161,7 +161,7 @@
 | `-mmu, --mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) |
 | `--mmproj-auto, --no-mmproj, --no-mmproj-auto` | whether to use multimodal projector file (if available), useful when using -hf (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_AUTO) |
 | `--mmproj-offload, --no-mmproj-offload` | whether to enable GPU offloading for multimodal projector (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_OFFLOAD) |
-| `--image, --audio FILE` | path to an image or audio file. use with multimodal models, use comma-separated values for multiple files |
+| `--image, --audio, --video FILE` | path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files |
 | `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MIN_TOKENS) |
 | `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MAX_TOKENS) |
 | `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_KWARGS) |
@@ -174,6 +174,7 @@
 | `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek-ocr, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, granite-4.0, granite-4.1, grok-2, hunyuan-dense, hunyuan-moe, hunyuan-vl, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--skip-chat-parsing, --no-skip-chat-parsing` | force a pure content parser, even if a Jinja template is specified; model will output everything in the content section, including any reasoning and/or tool calls (default: disabled)<br/>(env: LLAMA_ARG_SKIP_CHAT_PARSING) |
 | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
+| `--log-prompts-dir PATH` | Log prompts to directory (only used for debugging, default: disabled) |
 | `--spec-draft-hf, -hfd, -hfrd, --hf-repo-draft <user>/<model>[:quant]` | Same as --hf-repo, but for the draft model (default: unused)<br/>(env: LLAMA_ARG_SPEC_DRAFT_HF_REPO) |
 | `--spec-draft-threads, -td, --threads-draft N` | number of threads to use during generation (default: same as --threads) |
 | `--spec-draft-threads-batch, -tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) |

tools/cli/cli.cpp

@@ -202,7 +202,7 @@ struct cli_context {
 
     // TODO: support remote files in the future (http, https, etc)
     std::string load_input_file(const std::string & fname, bool is_media) {
-        std::ifstream file(fname, std::ios::binary);
+        std::ifstream file = fs_open_ifstream(fname, std::ios::binary);
         if (!file) {
             return "";
         }

tools/export-lora/README.md

@@ -6,11 +6,10 @@ Apply LORA adapters to base model and export the resulting model.
 usage: llama-export-lora [options]
 
 options:
-  -m,    --model                  model path from which to load base model (default '')
-         --lora FNAME             path to LoRA adapter  (can be repeated to use multiple adapters)
-         --lora-scaled FNAME S    path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)
-  -t,    --threads N              number of threads to use during computation (default: 4)
-  -o,    --output FNAME           output file (default: 'ggml-lora-merged-f16.gguf')
+  -m,    --model FNAME                  model path from which to load base model
+         --lora FNAME                   path to LoRA adapter (use comma-separated values to load multiple adapters)
+         --lora-scaled FNAME:SCALE,...  path to LoRA adapter with user defined scaling (format: FNAME:SCALE,...)
+  -o,    --output, --output-file FNAME  output file (default: 'ggml-lora-merged-f16.gguf')
 ```
 
 For example:
@@ -22,12 +21,11 @@ For example:
     --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
 ```
 
-Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:
+Multiple LORA adapters can be applied by passing comma-separated values to `--lora FNAME` or `--lora-scaled FNAME:SCALE,...`:
 
 ```bash
 ./bin/llama-export-lora \
     -m your_base_model.gguf \
     -o your_merged_model.gguf \
-    --lora-scaled lora_task_A.gguf 0.5 \
-    --lora-scaled lora_task_B.gguf 0.5
+    --lora-scaled lora_task_A.gguf:0.5,lora_task_B.gguf:0.5
 ```

tools/mtmd/README-dev.md

@@ -0,0 +1,35 @@
+# libmtmd dev guide
+
+## History
+
+Please refer to [multimodal.md](../../docs/multimodal.md) for a broader context.
+
+In short:
+- `libmtmd` started as a wrapper around `libllava` / `clip.cpp`
+- Various components that used to be in `clip.cpp` are moved progressively to mtmd. For example, preprocessor is now part of mtmd
+
+## Terminologies
+
+- mtmd: **M**ul**T**i**M**o**D**al
+- bitmap: representing a raw input data, for example: RGB image, PCM audio
+- tiles / slices: for llava-uhd-style models, the preprocessor breaks a large input into smaller square images called tiles or slices
+- chunk: a mtmd_input_chunk represents a preprocessed input that can then be passed through `mtmd_encode()`
+
+## Pipeline
+
+A typical pipeline of the core libmtmd is as follows:
+- A bitmap (RGB image or PCM audio) is created
+- Bitmap and the text prompt is provided to `mtmd_tokenize()` that breaks the input into chunks
+    - The tokenizer function first expands a "lazy" bitmap if it finds one. Typically, this is used by video, so that one media token corresponds to one input bitmap
+    - For models that support "fused" temporal frames like Qwen-VL, the tokenizer tries to merge pair of consecutive frames into one batch
+    - The preprocessor will then be called, which produces a list of chunks
+    - Depending on the model itself, special tokens will be injected to separate image chunks (i.e. llava-uhd-style models)
+- Multiple bitmaps may be batched together to form a larger `mtmd_batch()`
+- Single image or batch is encoded, via `mtmd_encode()` or `mtmd_batch_encode()`
+- Get the output embeddings
+
+## Helper
+
+We provide a set of helper functions via `mtmd_helper` to make using libmtmd easier. The helper provides:
+- Image, audio and video file decoding (for example, decode raw JPEG into RGB bitmap)
+- Manage `llama_batch` and calls to `llama_decode`

tools/mtmd/clip-impl.h

@@ -13,6 +13,14 @@
 #include <sstream>
 #include <vector>
 #include <memory>
+#include <fstream>
+
+#ifdef _WIN32
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#endif
 
 // Internal header for clip.cpp
 
@@ -367,56 +375,56 @@ enum projector_type {
 };
 
 static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
-    { PROJECTOR_TYPE_MLP,       "mlp" },
-    { PROJECTOR_TYPE_LDP,       "ldp" },
-    { PROJECTOR_TYPE_LDPV2,     "ldpv2"},
-    { PROJECTOR_TYPE_MINICPMV,  "resampler"},
-    { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
-    { PROJECTOR_TYPE_QWEN2VL,   "qwen2vl_merger"},
-    { PROJECTOR_TYPE_QWEN25VL,  "qwen2.5vl_merger"},
-    { PROJECTOR_TYPE_QWEN3VL,   "qwen3vl_merger"},
-    { PROJECTOR_TYPE_STEP3VL,   "step3vl"},
-    { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
-    { PROJECTOR_TYPE_GEMMA3NV,  "gemma3nv"},
-    { PROJECTOR_TYPE_GEMMA3NA,  "gemma3na"},
-    { PROJECTOR_TYPE_GEMMA4V,   "gemma4v"},
-    { PROJECTOR_TYPE_GEMMA4A,   "gemma4a"},
-    { PROJECTOR_TYPE_GEMMA4UV,  "gemma4uv"},
-    { PROJECTOR_TYPE_GEMMA4UA,  "gemma4ua"},
-    { PROJECTOR_TYPE_PHI4,      "phi4"},
-    { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
-    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
-    { PROJECTOR_TYPE_ULTRAVOX,  "ultravox"},
-    { PROJECTOR_TYPE_INTERNVL,  "internvl"},
-    { PROJECTOR_TYPE_LLAMA4,    "llama4"},
-    { PROJECTOR_TYPE_QWEN2A,    "qwen2a"},
-    { PROJECTOR_TYPE_QWEN3A,    "qwen3a"},
-    { PROJECTOR_TYPE_GLMA,      "glma"},
-    { PROJECTOR_TYPE_QWEN25O,   "qwen2.5o"},
-    { PROJECTOR_TYPE_VOXTRAL,   "voxtral"},
-    { PROJECTOR_TYPE_MERALION,  "meralion"},
-    { PROJECTOR_TYPE_MUSIC_FLAMINGO, "musicflamingo"},
-    { PROJECTOR_TYPE_LFM2,      "lfm2"},
-    { PROJECTOR_TYPE_KIMIVL,    "kimivl"},
-    { PROJECTOR_TYPE_PADDLEOCR, "paddleocr"},
-    { PROJECTOR_TYPE_LIGHTONOCR,"lightonocr"},
-    { PROJECTOR_TYPE_COGVLM,    "cogvlm"},
-    { PROJECTOR_TYPE_JANUS_PRO, "janus_pro"},
-    { PROJECTOR_TYPE_DOTS_OCR,  "dots_ocr"},
-    { PROJECTOR_TYPE_DEEPSEEKOCR,"deepseekocr"},
-    { PROJECTOR_TYPE_DEEPSEEKOCR2,"deepseekocr2"},
-    { PROJECTOR_TYPE_LFM2A,     "lfm2a"},
-    { PROJECTOR_TYPE_GLM4V,     "glm4v"},
-    { PROJECTOR_TYPE_YOUTUVL,   "youtuvl"},
-    { PROJECTOR_TYPE_YASA2,     "yasa2"},
-    { PROJECTOR_TYPE_KIMIK25,   "kimik25"},
-    { PROJECTOR_TYPE_NEMOTRON_V2_VL, "nemotron_v2_vl"},
-    { PROJECTOR_TYPE_EXAONE4_5, "exaone4_5"},
-    { PROJECTOR_TYPE_HUNYUANVL,  "hunyuanvl"},
-    { PROJECTOR_TYPE_MINICPMV4_6, "minicpmv4_6"},
-    { PROJECTOR_TYPE_GRANITE_SPEECH, "granite_speech"},
-    { PROJECTOR_TYPE_MIMOVL,     "mimovl"},
-    { PROJECTOR_TYPE_GRANITE4_VISION, "granite4_vision"},
+    { PROJECTOR_TYPE_MLP,               "mlp" },
+    { PROJECTOR_TYPE_LDP,               "ldp" },
+    { PROJECTOR_TYPE_LDPV2,             "ldpv2"},
+    { PROJECTOR_TYPE_MINICPMV,          "resampler"},
+    { PROJECTOR_TYPE_GLM_EDGE,          "adapter"},
+    { PROJECTOR_TYPE_QWEN2VL,           "qwen2vl_merger"},
+    { PROJECTOR_TYPE_QWEN25VL,          "qwen2.5vl_merger"},
+    { PROJECTOR_TYPE_QWEN3VL,           "qwen3vl_merger"},
+    { PROJECTOR_TYPE_STEP3VL,           "step3vl"},
+    { PROJECTOR_TYPE_GEMMA3,            "gemma3"},
+    { PROJECTOR_TYPE_GEMMA3NV,          "gemma3nv"},
+    { PROJECTOR_TYPE_GEMMA3NA,          "gemma3na"},
+    { PROJECTOR_TYPE_GEMMA4V,           "gemma4v"},
+    { PROJECTOR_TYPE_GEMMA4A,           "gemma4a"},
+    { PROJECTOR_TYPE_GEMMA4UV,          "gemma4uv"},
+    { PROJECTOR_TYPE_GEMMA4UA,          "gemma4ua"},
+    { PROJECTOR_TYPE_PHI4,              "phi4"},
+    { PROJECTOR_TYPE_IDEFICS3,          "idefics3"},
+    { PROJECTOR_TYPE_PIXTRAL,           "pixtral"},
+    { PROJECTOR_TYPE_ULTRAVOX,          "ultravox"},
+    { PROJECTOR_TYPE_INTERNVL,          "internvl"},
+    { PROJECTOR_TYPE_LLAMA4,            "llama4"},
+    { PROJECTOR_TYPE_QWEN2A,            "qwen2a"},
+    { PROJECTOR_TYPE_QWEN3A,            "qwen3a"},
+    { PROJECTOR_TYPE_GLMA,              "glma"},
+    { PROJECTOR_TYPE_QWEN25O,           "qwen2.5o"},
+    { PROJECTOR_TYPE_VOXTRAL,           "voxtral"},
+    { PROJECTOR_TYPE_MERALION,          "meralion"},
+    { PROJECTOR_TYPE_MUSIC_FLAMINGO,    "musicflamingo"},
+    { PROJECTOR_TYPE_LFM2,              "lfm2"},
+    { PROJECTOR_TYPE_KIMIVL,            "kimivl"},
+    { PROJECTOR_TYPE_PADDLEOCR,         "paddleocr"},
+    { PROJECTOR_TYPE_LIGHTONOCR,        "lightonocr"},
+    { PROJECTOR_TYPE_COGVLM,            "cogvlm"},
+    { PROJECTOR_TYPE_JANUS_PRO,         "janus_pro"},
+    { PROJECTOR_TYPE_DOTS_OCR,          "dots_ocr"},
+    { PROJECTOR_TYPE_DEEPSEEKOCR,       "deepseekocr"},
+    { PROJECTOR_TYPE_DEEPSEEKOCR2,      "deepseekocr2"},
+    { PROJECTOR_TYPE_LFM2A,             "lfm2a"},
+    { PROJECTOR_TYPE_GLM4V,             "glm4v"},
+    { PROJECTOR_TYPE_YOUTUVL,           "youtuvl"},
+    { PROJECTOR_TYPE_YASA2,             "yasa2"},
+    { PROJECTOR_TYPE_KIMIK25,           "kimik25"},
+    { PROJECTOR_TYPE_NEMOTRON_V2_VL,    "nemotron_v2_vl"},
+    { PROJECTOR_TYPE_EXAONE4_5,         "exaone4_5"},
+    { PROJECTOR_TYPE_HUNYUANVL,         "hunyuanvl"},
+    { PROJECTOR_TYPE_MINICPMV4_6,       "minicpmv4_6"},
+    { PROJECTOR_TYPE_GRANITE_SPEECH,    "granite_speech"},
+    { PROJECTOR_TYPE_MIMOVL,            "mimovl"},
+    { PROJECTOR_TYPE_GRANITE4_VISION,   "granite4_vision"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {
@@ -640,47 +648,18 @@ static void clip_log_internal(enum ggml_log_level level, const char * format, ..
 // cpp wrappers
 //
 
-// wrapper for clip_image_size
-struct clip_image_size_deleter {
-    void operator()(clip_image_size * val) { clip_image_size_free(val); }
-};
-typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
-
-// wrapper for clip_image_u8
-struct clip_image_u8_deleter {
-    void operator()(clip_image_u8 * val) { clip_image_u8_free(val); }
-};
-typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;
-
-// wrapper for clip_image_f32
-struct clip_image_f32_deleter {
-    void operator()(clip_image_f32 * val) { clip_image_f32_free(val); }
-};
-typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;
-
-struct clip_image_u8_batch {
-    std::vector<clip_image_u8_ptr> entries;
-};
-
 struct clip_image_f32_batch {
-    std::vector<clip_image_f32_ptr> entries;
+    std::vector<clip_image_f32> entries;
     bool is_audio = false;
 
-    // for llava-uhd style models, we need to know the grid size
-    // note: entries.size() == grid_x * grid_y + 1 (one overview image)
-    int grid_x = 0;
-    int grid_y = 0;
-
     clip_image_f32_batch clone() const {
         clip_image_f32_batch new_batch{
             /* entries  */ {},
             /* is_audio */ is_audio,
-            /* grid_x   */ grid_x,
-            /* grid_y   */ grid_y,
         };
         new_batch.entries.reserve(entries.size());
         for (const auto & entry : entries) {
-            new_batch.entries.emplace_back(new clip_image_f32(*entry));
+            new_batch.entries.emplace_back(entry); // copy
         }
         return new_batch;
     }
@@ -690,6 +669,22 @@ struct clip_image_f32_batch {
 // common utils
 //
 
+#ifdef _WIN32
+static std::ifstream open_ifstream_binary(const std::string & fname) {
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, fname.c_str(), -1, NULL, 0);
+    if (!wlen) {
+        throw std::runtime_error("failed to convert filename to UTF-16: " + fname);
+    }
+    std::vector<wchar_t> wfname(wlen);
+    (void)MultiByteToWideChar(CP_UTF8, 0, fname.c_str(), -1, wfname.data(), wlen);
+    return std::ifstream(wfname.data(), std::ios::binary);
+}
+#else
+static std::ifstream open_ifstream_binary(const std::string & fname) {
+    return std::ifstream(fname, std::ios::binary);
+}
+#endif
+
 static std::string string_format(const char * fmt, ...) {
     va_list ap;
     va_list ap2;

tools/mtmd/clip.cpp

@@ -534,7 +534,7 @@ ggml_tensor * clip_graph::build_vit(
 ggml_tensor * clip_graph::build_inp() {
     ggml_tensor * inp_raw = build_inp_raw();
     ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
-    inp = ggml_reshape_2d(ctx0, inp, n_patches, n_embd);
+    inp = ggml_reshape_3d(ctx0, inp, n_patches, n_embd, n_batch);
     inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
     if (model.patch_bias) {
         inp = ggml_add(ctx0, inp, model.patch_bias);
@@ -865,7 +865,7 @@ ggml_tensor * clip_graph::build_patch_merge_permute(ggml_tensor * cur, int scale
 }
 
 static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
-    const clip_image_f32 & img = *imgs.entries[0];
+    const clip_image_f32 & img = imgs.entries[0];
     std::unique_ptr<clip_graph> builder;
 
     switch (ctx->proj_type()) {
@@ -1675,6 +1675,9 @@ struct clip_model_loader {
                     // note: some models having hparams.image_size == 0, which means the image size is dynamic
                     throw std::runtime_error(string_format("%s: image_size (%d) cannot be negative\n", __func__, hparams.image_size));
                 }
+                if (hparams.image_size > 65536) {
+                    throw std::runtime_error(string_format("%s: image_size (%d) is too large (max 65536)\n", __func__, hparams.image_size));
+                }
                 if (hparams.patch_size <= 0) {
                     throw std::runtime_error(string_format("%s: patch_size (%d) must be greater than 0\n", __func__, hparams.patch_size));
                 }
@@ -1723,6 +1726,19 @@ struct clip_model_loader {
                 LOG_INF("%s: audio_n_fft:        %d\n", __func__, hparams.audio_n_fft);
                 LOG_INF("%s: audio_window_len:   %d\n", __func__, hparams.audio_window_len);
                 LOG_INF("%s: audio_hop_len:      %d\n", __func__, hparams.audio_hop_len);
+
+                // GEMMA4UA is encoder-free: it uses n_mel_bins as a raw-waveform frame size (640) and has no FFT/filterbank, so the mel-range and FFT
+                // checks below do not apply to it.
+                const bool fft_based = model.proj_type != PROJECTOR_TYPE_GEMMA4UA;
+
+                // Validate audio hparams loaded from GGUF metadata
+                if (hparams.n_mel_bins <= 0 || (fft_based && hparams.n_mel_bins > 256)) {
+                    throw std::runtime_error(string_format("%s: n_mel_bins (%d) must be in range [1, 256]\n", __func__, hparams.n_mel_bins));
+                }
+                if (fft_based && (hparams.audio_sample_rate <= 0 || hparams.audio_n_fft <= 0 || hparams.audio_hop_len <= 0 || hparams.audio_window_len <= 0)) {
+                    throw std::runtime_error(string_format("%s: audio hparams invalid: sample_rate=%d n_fft=%d window_len=%d hop_len=%d\n",
+                        __func__, hparams.audio_sample_rate, hparams.audio_n_fft, hparams.audio_window_len, hparams.audio_hop_len));
+                }
             }
             LOG_INF("\n");
             LOG_INF("%s: model size:         %.2f MiB\n", __func__, model_size / 1024.0 / 1024.0);
@@ -1736,7 +1752,7 @@ struct clip_model_loader {
         std::map<std::string, size_t> tensor_offset;
         std::vector<ggml_tensor *> tensors_to_load;
 
-        auto fin = std::ifstream(fname, std::ios::binary);
+        auto fin = open_ifstream_binary(fname);
         if (!fin) {
             throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
         }
@@ -2825,16 +2841,22 @@ struct clip_model_loader {
         // create a fake batch
         const auto & hparams = ctx_clip.model.hparams;
         clip_image_f32_batch batch;
-        clip_image_f32_ptr img(clip_image_f32_init());
+        clip_image_f32 img;
         if (ctx_clip.model.modality == CLIP_MODALITY_VISION) {
             const int sz = hparams.warmup_image_size;
-            img->set_size({sz, sz}, false, false);
+            img.set_size({sz, sz}, false, false);
             LOG_INF("%s: warmup with image size = %d x %d\n", __func__, sz, sz);
         } else {
-            img->set_size({hparams.warmup_audio_size, hparams.n_mel_bins}, false, false);
+            // GEMMA4UA uses n_mel_bins as a raw-waveform frame size (640), not a mel-bin count,
+            // so the [1, 256] bound only applies to FFT-based models.
+            const bool fft_based = ctx_clip.model.proj_type != PROJECTOR_TYPE_GEMMA4UA;
+            if (hparams.n_mel_bins <= 0 || (fft_based && hparams.n_mel_bins > 256)) {
+                throw std::runtime_error(string_format("%s: invalid n_mel_bins (%d), must be in [1, 256]\n", __func__, hparams.n_mel_bins));
+            }
+            img.set_size({hparams.warmup_audio_size, hparams.n_mel_bins}, false, false);
             LOG_INF("%s: warmup with audio size = %d\n", __func__, hparams.warmup_audio_size);
         }
-        batch.entries.push_back(std::move(img));
+        batch.entries.push_back(img);
         return batch;
     }
 
@@ -2994,7 +3016,13 @@ struct clip_model_loader {
             }
             return;
         }
-        output = gguf_get_val_u32(ctx_gguf.get(), i);
+        const uint32_t val = gguf_get_val_u32(ctx_gguf.get(), i);
+        // sanity check
+        if (val > (uint32_t) INT32_MAX) {
+            throw std::runtime_error(string_format("%s: value %u for key '%s' exceeds INT32_MAX\n",
+                __func__, val, key.c_str()));
+        }
+        output = (int) val;
     }
 
     void get_f32(const std::string & key, float & output, bool required = true) const {
@@ -3124,64 +3152,6 @@ struct clip_cap clip_get_cap(const char * fname) {
     return res;
 }
 
-struct clip_image_size * clip_image_size_init() {
-    struct clip_image_size * load_image_size = new struct clip_image_size();
-    load_image_size->width = 448;
-    load_image_size->height = 448;
-    return load_image_size;
-}
-
-struct clip_image_u8 * clip_image_u8_init() {
-    return new clip_image_u8();
-}
-
-struct clip_image_f32 * clip_image_f32_init() {
-    return new clip_image_f32();
-}
-
-struct clip_image_f32_batch * clip_image_f32_batch_init() {
-    return new clip_image_f32_batch();
-}
-
-void clip_image_size_free(struct clip_image_size * load_image_size) {
-    if (load_image_size == nullptr) {
-        return;
-    }
-    delete load_image_size;
-}
-void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
-void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
-void clip_image_u8_batch_free(struct clip_image_u8_batch * batch) { delete batch; }
-void clip_image_f32_batch_free(struct clip_image_f32_batch * batch) { delete batch; }
-
-size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch) {
-    return batch->entries.size();
-}
-
-size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx) {
-    if (idx < 0 || idx >= (int)batch->entries.size()) {
-        LOG_ERR("%s: invalid index %d\n", __func__, idx);
-        return 0;
-    }
-    return batch->entries[idx]->nx();
-}
-
-size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx) {
-    if (idx < 0 || idx >= (int)batch->entries.size()) {
-        LOG_ERR("%s: invalid index %d\n", __func__, idx);
-        return 0;
-    }
-    return batch->entries[idx]->ny();
-}
-
-clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx) {
-    if (idx < 0 || idx >= (int)batch->entries.size()) {
-        LOG_ERR("%s: invalid index %d\n", __func__, idx);
-        return nullptr;
-    }
-    return batch->entries[idx].get();
-}
-
 void clip_free(clip_ctx * ctx) {
     if (ctx == nullptr) {
         return;
@@ -3189,23 +3159,11 @@ void clip_free(clip_ctx * ctx) {
     delete ctx;
 }
 
-int32_t clip_get_image_size(const struct clip_ctx * ctx) {
-    return ctx->model.hparams.image_size;
-}
-
-int32_t clip_get_patch_size(const struct clip_ctx * ctx) {
-    return ctx->model.hparams.patch_size;
-}
-
-int32_t clip_get_hidden_size(const struct clip_ctx * ctx) {
-    return ctx->model.hparams.n_embd;
-}
-
 const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
     return ctx->model.hparams.mm_patch_merge_type == PATCH_MERGE_SPATIAL_UNPAD ? "spatial_unpad" : "flat";
 }
 
-int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
+int clip_n_output_tokens_x(const clip_ctx * ctx, const clip_image_f32 * img) {
     const auto & params = ctx->model.hparams;
     const int n_total = clip_n_output_tokens(ctx, img);
     const auto & proj = ctx->proj_type();
@@ -3228,7 +3186,7 @@ int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 *
     return n_total;
 }
 
-int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
+int clip_n_output_tokens_y(const clip_ctx * ctx, const clip_image_f32 * img) {
     const auto & params = ctx->model.hparams;
     const auto & proj = ctx->proj_type();
     switch (proj) {
@@ -3250,7 +3208,7 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 *
     return 1;
 }
 
-int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
+int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
     const auto & params = ctx->model.hparams;
 
     // for models with fixed size image, the input image is already pre-processed and resized to square
@@ -3500,16 +3458,15 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
     return n_patches;
 }
 
-bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, std::vector<float> & out_vec) {
+bool clip_image_encode(struct clip_ctx * ctx, int n_threads, const clip_image_f32 * img, std::vector<float> & out_vec) {
     clip_image_f32_batch imgs;
-    clip_image_f32_ptr img_copy(clip_image_f32_init());
-    *img_copy = *img;
+    clip_image_f32 img_copy = *img;
     imgs.entries.push_back(std::move(img_copy));
 
     return clip_image_batch_encode(ctx, n_threads, &imgs, out_vec);
 }
 
-bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs_c_ptr, std::vector<float> & out_batch_embd) {
+bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32_batch * imgs_c_ptr, std::vector<float> & out_batch_embd) {
     const clip_image_f32_batch & imgs = *imgs_c_ptr;
     int n_batch_cur = imgs.entries.size();
 
@@ -3533,8 +3490,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     const auto & model   = ctx->model;
     const auto & hparams = model.hparams;
 
-    const int image_size_width  = imgs.entries[0]->nx();
-    const int image_size_height = imgs.entries[0]->ny();
+    const int image_size_width  = imgs.entries[0].nx();
+    const int image_size_height = imgs.entries[0].ny();
 
     const int patch_size    = hparams.patch_size;
     const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
@@ -3572,7 +3529,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     if (!imgs.is_audio) {
         size_t nelem = 0;
         for (const auto & img : imgs.entries) {
-            nelem += img->nx() * img->ny() * 3;
+            nelem += img.nx() * img.ny() * 3;
         }
         std::vector<float> inp_raw(nelem);
 
@@ -3590,13 +3547,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         // IMPORTANT: [QWEN_VIDEO] the batch dim is currently used for temporal dim in Qwen-VL models
         // All entries must have the same spatial size (enforced by can_batch_with() during merging)
         {
-            const int nx = imgs.entries[0]->nx();
-            const int ny = imgs.entries[0]->ny();
+            const int nx = imgs.entries[0].nx();
+            const int ny = imgs.entries[0].ny();
             const int n  = nx * ny;
 
             for (int b = 0; b < n_batch_cur; b++) {
                 LOG_DBG("%s: copying image %d/%d to input buffer (nx=%d, ny=%d)\n", __func__, b+1, n_batch_cur, nx, ny);
-                const auto & buf = imgs.entries[b]->get_ro_buf();
+                const auto & buf = imgs.entries[b].get_ro_buf();
                 float * batch_entry = inp_raw.data() + b * (3*n);
                 for (int y = 0; y < ny; y++) {
                     for (int x = 0; x < nx; x++) {
@@ -3616,9 +3573,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         GGML_ASSERT(imgs.entries.size() == 1);
 
         const auto & mel_inp = imgs.entries[0];
-        const auto & buf = mel_inp->get_ro_buf();
-        const int n_step = mel_inp->nx();
-        const int n_mel  = mel_inp->ny();
+        const auto & buf = mel_inp.get_ro_buf();
+        const int n_step = mel_inp.nx();
+        const int n_mel  = mel_inp.ny();
         GGML_ASSERT((size_t)n_step * n_mel == buf.size());
 
         set_input_f32("inp_raw", buf);
@@ -4232,7 +4189,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 GGML_ASSERT(imgs.entries.size() == 1);
                 const auto & img0 = imgs.entries.front();
                 // Compute n_pos matching SSCP output: two stride-2 convs
-                int n_pos = img0->nx();
+                int n_pos = img0.nx();
                 for (int i = 0; i < 2; i++) { n_pos = (n_pos - 1) / 2 + 1; }
 
                 // Chunked local attention: blocked causal mask and RPE
@@ -4280,7 +4237,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_LFM2A:
             {
                 GGML_ASSERT(imgs.entries.size() == 1);
-                const auto n_frames = clip_n_output_tokens(ctx, imgs.entries.front().get());
+                const auto n_frames = clip_n_output_tokens(ctx, &imgs.entries.front());
 
                 auto d_model = 512;
                 auto seq_len = n_frames * 2 - 1;
@@ -4338,7 +4295,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 // reshapes as ggml_get_rows gathers. The names are set
                 // by g4v_gather() in models/granite4-vision.cpp.
                 const int patch_size  = model.hparams.patch_size;
-                const int image_side  = imgs.entries.front()->nx() / patch_size;
+                const int image_side  = imgs.entries.front().nx() / patch_size;
                 const int window_side = hparams.downsample_window_side;
                 const int query_side  = hparams.downsample_query_side;
                 const int n           = image_side / window_side;
@@ -4432,7 +4389,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
 
     // sanity check (assuming that all images in batch have the same number of tokens, so we only check the first one)
     const int n_tokens_out = embeddings->ne[1];
-    const int expected_n_tokens_out = clip_n_output_tokens(ctx, imgs.entries[0].get());
+    const int expected_n_tokens_out = clip_n_output_tokens(ctx, &imgs.entries[0]);
     if (n_tokens_out != expected_n_tokens_out) {
         LOG_ERR("%s: expected output %d tokens, got %d\n", __func__, expected_n_tokens_out, n_tokens_out);
         GGML_ABORT("Invalid number of output tokens");

tools/mtmd/clip.h

@@ -24,12 +24,14 @@ struct clip_image_size {
         return !(*this == other);
     }
     int area() const {
+        // avoid overflow when computing area
+        GGML_ASSERT(width  >= 0 && width  <= 46000);
+        GGML_ASSERT(height >= 0 && height <= 46000);
         return width * height;
     }
 };
 
 struct clip_image_f32;
-struct clip_image_u8_batch;
 struct clip_image_f32_batch;
 
 enum clip_modality {
@@ -63,41 +65,21 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
 
 void clip_free(struct clip_ctx * ctx);
 
-int32_t clip_get_image_size (const struct clip_ctx * ctx);
-int32_t clip_get_patch_size (const struct clip_ctx * ctx);
-int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
-
 // TODO: should be enum, not string
 const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 
-int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+int clip_n_output_tokens(const clip_ctx * ctx, const 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
-int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
-int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+int clip_n_output_tokens_x(const clip_ctx * ctx, const clip_image_f32 * img);
+int clip_n_output_tokens_y(const clip_ctx * ctx, const clip_image_f32 * img);
 
 // this should be equal to the embedding dimension of the text model
 int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
-struct clip_image_size      * clip_image_size_init(void);
-struct clip_image_u8        * clip_image_u8_init (void);
-struct clip_image_f32       * clip_image_f32_init(void);
-struct clip_image_f32_batch * clip_image_f32_batch_init(void); // only used by libllava
-
-void clip_image_size_free (struct clip_image_size * img_size);
-void clip_image_u8_free (struct clip_image_u8  * img);
-void clip_image_f32_free(struct clip_image_f32 * img);
-void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
-void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
-
-// use for accessing underlay data of clip_image_f32_batch
-size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
-size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
-size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
-struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
-
-bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, std::vector<float> & out_vec);
+// TODO: remove clip_image_encode() and always use batched version
+bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, const clip_image_f32 * img, std::vector<float> & out_vec);
 bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, std::vector<float> & out_batch_embd);
 
 bool clip_is_llava(const struct clip_ctx * ctx);

tools/mtmd/models/internvl.cpp

@@ -8,7 +8,9 @@ ggml_cgraph * clip_graph_internvl::build() {
     ggml_tensor * inp = build_inp();
 
     // add CLS token
-    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
+    ggml_tensor * cls_repeated = ggml_repeat_4d(ctx0, model.class_embedding,
+            model.class_embedding->ne[0], 1, n_batch, 1);
+    inp = ggml_concat(ctx0, inp, cls_repeated, 1);
 
     // The larger models use a different ViT, which uses RMS norm instead of layer norm
     // ref: https://github.com/ggml-org/llama.cpp/pull/13443#issuecomment-2869786188
@@ -24,14 +26,15 @@ ggml_cgraph * clip_graph_internvl::build() {
                             nullptr);
 
     // remove CLS token
-    cur = ggml_view_2d(ctx0, cur,
-        n_embd, n_patches,
-        ggml_row_size(cur->type, n_embd), 0);
+    cur = ggml_view_3d(ctx0, cur,
+        n_embd, n_patches, n_batch,
+        cur->nb[1], cur->nb[2], 0);
+    cur = ggml_cont(ctx0, cur);
 
     // pixel shuffle
     {
         const int scale_factor = model.hparams.n_merge;
-        const int bsz    = 1; // batch size, always 1 for now since we don't support batching
+        const int bsz    = n_batch;
         const int height = n_patches_y;
         const int width  = n_patches_x;
         GGML_ASSERT(scale_factor > 0);
@@ -44,9 +47,10 @@ ggml_cgraph * clip_graph_internvl::build() {
             bsz);
         cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
         // flatten to 2D
-        cur = ggml_cont_2d(ctx0, cur,
+        cur = ggml_cont_3d(ctx0, cur,
             n_embd * scale_factor * scale_factor,
-            cur->ne[1] * cur->ne[2]);
+            cur->ne[1] * cur->ne[2],
+            cur->ne[3]);
     }
 
     // projector (always using GELU activation)

tools/mtmd/models/models.h

@@ -80,6 +80,7 @@ struct clip_graph_minicpmv4_6 : clip_graph {
 struct clip_graph_internvl : clip_graph {
     clip_graph_internvl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
+    bool support_batch() const override { return true; }
 };
 
 struct clip_graph_nemotron_v2_vl : clip_graph {

tools/mtmd/mtmd-audio.cpp

@@ -32,8 +32,8 @@ void mtmd_audio_cache::fill_hann_window(uint32_t length, bool periodic) {
     }
 }
 
-void mtmd_audio_cache::fill_mel_filterbank_matrix(int   n_mel,
-                                                  int   n_fft,
+void mtmd_audio_cache::fill_mel_filterbank_matrix(int64_t n_mel,
+                                                  int64_t n_fft,
                                                   int   sample_rate,
                                                   float fmin,
                                                   float fmax,
@@ -86,11 +86,16 @@ void mtmd_audio_cache::fill_mel_filterbank_matrix(int   n_mel,
         hz_pts[i] = mel_to_hz(mel_pts[i]);
     }
 
-    const int n_fft_bins = n_fft / 2 + 1;
+    const int64_t n_fft_bins = n_fft / 2 + 1;
+
+    // Validate allocation size
+    if ((size_t)n_mel * (size_t)n_fft_bins > SIZE_MAX) {
+        GGML_ASSERT(false && "mel filterbank allocation too large");
+    }
 
     // filterbank
-    std::vector<float> out(n_mel * n_fft_bins, 0);
-    for (int m = 0; m < n_mel; ++m) {
+    std::vector<float> out((size_t)n_mel * (size_t)n_fft_bins, 0);
+    for (int64_t m = 0; m < n_mel; ++m) {
         const double f_left   = hz_pts[m];
         const double f_center = hz_pts[m + 1];
         const double f_right  = hz_pts[m + 2];
@@ -266,8 +271,8 @@ static void ifft(const mtmd_audio_cache & cache, float * in, int N, float * out)
 }
 
 struct filter_params {
-    int32_t n_mel;
-    int32_t n_fft_bins;
+    int64_t n_mel;
+    int64_t n_fft_bins;
     int32_t hann_window_size;
     int32_t hop_length;
     int32_t sample_rate;
@@ -293,8 +298,8 @@ static void log_mel_spectrogram_worker_thread(int                        ith,
     std::vector<float> fft_in(frame_size * 2, 0.0);
     std::vector<float> fft_out(frame_size * 2 * 2 * 2);
 
-    int n_fft_bins = params.n_fft_bins;
-    int i = ith;
+    int64_t n_fft_bins = params.n_fft_bins;
+    int64_t i = ith;
 
     const auto & filters = cache.filters;
 
@@ -302,17 +307,18 @@ static void log_mel_spectrogram_worker_thread(int                        ith,
     GGML_ASSERT(n_fft_bins == 1 + (frame_size / 2));
     GGML_ASSERT(cache.sin_vals.size() == cache.cos_vals.size());
     // calculate FFT only when fft_in are not all zero
-    for (; i < std::min(n_samples / frame_step + 1, out.n_len); i += n_threads) {
-        const int offset = i * frame_step;
+    for (; i < std::min((int64_t)(n_samples / frame_step + 1), out.n_len); i += n_threads) {
+        const int64_t offset = i * frame_step;
 
         // apply Hann window (~10% faster)
-        for (int j = 0; j < std::min(frame_size, n_samples - offset); j++) {
+        const int valid_len = std::min(frame_size, std::max(0, n_samples - (int)offset));
+        for (int j = 0; j < valid_len; j++) {
             fft_in[j] = hann[j] * samples[offset + j];
         }
 
         // fill the rest with zeros
-        if (n_samples - offset < frame_size) {
-            std::fill(fft_in.begin() + (n_samples - offset), fft_in.end(), 0.0);
+        if (valid_len < frame_size) {
+            std::fill(fft_in.begin() + valid_len, fft_in.end(), 0.0);
         }
 
         // FFT
@@ -325,7 +331,7 @@ static void log_mel_spectrogram_worker_thread(int                        ith,
         }
 
         // mel spectrogram
-        for (int j = 0; j < out.n_mel; j++) {
+        for (int64_t j = 0; j < out.n_mel; j++) {
             double sum = 0.0;
             // unroll loop (suggested by GH user @lunixbochs)
             int k = 0;
@@ -339,21 +345,21 @@ static void log_mel_spectrogram_worker_thread(int                        ith,
             }
             // handle n_fft remainder
             for (; k < n_fft_bins; k++) {
-                sum += fft_out[k] * filters.data[j * n_fft_bins + k];
+                sum += fft_out[k] * filters.data[(size_t)j * n_fft_bins + k];
             }
             sum = std::max(sum, (double)params.mel_floor);
             sum = params.use_natural_log
                 ? log(sum)
                 : log10(sum);
-            out.data[j * out.n_len + i] = sum;
+            out.data[(size_t)j * out.n_len + i] = sum;
         }
     }
 
     // Otherwise fft_out are all zero
     double sum = params.use_natural_log ? log(1e-10) : log10(1e-10);
     for (; i < out.n_len; i += n_threads) {
-        for (int j = 0; j < out.n_mel; j++) {
-            out.data[j * out.n_len + i] = sum;
+        for (int64_t j = 0; j < out.n_mel; j++) {
+            out.data[(size_t)j * out.n_len + i] = sum;
         }
     }
 }
@@ -437,16 +443,21 @@ static bool log_mel_spectrogram(
     GGML_ASSERT(params.hop_length > 0);
     out.n_mel = params.n_mel;
     out.n_len = (n_samples - frame_size) / frame_step + 1;
-    // TODO: handle these checks better
-    if (out.n_mel > 0 && (unsigned long)out.n_len > SIZE_MAX / out.n_mel) {
-        LOG_ERR("%s: size overflow\n", __func__);
+    // Validate dimensions before allocation to prevent integer overflow
+    if (out.n_mel <= 0 || out.n_len <= 0) {
+        LOG_ERR("%s: invalid mel dimensions n_mel=%lld n_len=%lld\n", __func__, (long long)out.n_mel, (long long)out.n_len);
+        return false;
+    }
+    const size_t total_size = (size_t)out.n_mel * (size_t)out.n_len;
+    if (total_size > SIZE_MAX / sizeof(float)) {
+        LOG_ERR("%s: size overflow: n_mel=%lld n_len=%lld\n", __func__, (long long)out.n_mel, (long long)out.n_len);
         return false;
     }
     if (n_samples < frame_size) {
         LOG_ERR("%s: not enough samples after padding\n", __func__);
         return false;
     }
-    out.data.resize(out.n_mel * out.n_len);
+    out.data.resize(total_size);
 
     {
         std::vector<std::thread> workers(n_threads - 1);
@@ -464,38 +475,39 @@ static bool log_mel_spectrogram(
         }
     }
 
-    const int effective_n_len = n_samples_in / frame_step;
+    const int64_t effective_n_len = n_samples_in / frame_step;
     if (params.norm_per_feature) {
         GGML_ASSERT(effective_n_len > 1);
-        for (int i = 0; i < out.n_mel; i++) {
+        for (int64_t i = 0; i < out.n_mel; i++) {
             double mean = 0;
-            for (int j = 0; j < effective_n_len; ++j) {
-                mean += out.data[i * out.n_len + j];
+            for (int64_t j = 0; j < effective_n_len; ++j) {
+                mean += out.data[(size_t)i * out.n_len + j];
             }
             mean /= effective_n_len;
 
             double var = 0.0;
-            for (int j = 0; j < effective_n_len; ++j) {
-                const double value = out.data[i * out.n_len + j] - mean;
+            for (int64_t j = 0; j < effective_n_len; ++j) {
+                const double value = out.data[(size_t)i * out.n_len + j] - mean;
                 var += value * value;
             }
             var /= effective_n_len - 1;  // unbiased
             const double mstd = std::sqrt(var + 1e-5);
 
-            for (int j = 0; j < effective_n_len; ++j) {
-                auto &value = out.data[i * out.n_len + j];
+            for (int64_t j = 0; j < effective_n_len; ++j) {
+                auto &value = out.data[(size_t)i * out.n_len + j];
                 value        = (value - mean) / mstd;
             }
 
             // pad the rest with zeros
-            for (int j = effective_n_len; j < out.n_len; ++j) {
-                out.data[i * out.n_len + j] = 0.0;
+            for (int64_t j = effective_n_len; j < out.n_len; ++j) {
+                out.data[(size_t)i * out.n_len + j] = 0.0;
             }
         }
     } else if (!params.no_padding) {
         // Whisper-style clamping and normalization (NOT used by Gemma4)
         double mmax = -1e20;
-        for (int i = 0; i < out.n_mel*out.n_len; i++) {
+        const size_t mel_size = (size_t)out.n_mel * (size_t)out.n_len;
+        for (size_t i = 0; i < mel_size; i++) {
             if (out.data[i] > mmax) {
                 mmax = out.data[i];
             }
@@ -503,7 +515,7 @@ static bool log_mel_spectrogram(
 
         mmax -= 8.0;
 
-        for (int i = 0; i < out.n_mel*out.n_len; i++) {
+        for (size_t i = 0; i < mel_size; i++) {
             if (out.data[i] < mmax) {
                 out.data[i] = mmax;
             }
@@ -582,13 +594,13 @@ bool mtmd_audio_preprocessor_whisper::preprocess(const float *                 s
     // because the cgraph in clip.cpp only accepts 3000 frames each, we need to split the mel
     // we always expect the mel to have 3000 silent frames at the end
     if (DEBUG) {
-        printf("output: n_mel = %d, n_len = %d\n", out_full.n_mel, out_full.n_len);
+        printf("output: n_mel = %d, n_len = %d\n", (int) out_full.n_mel, (int) out_full.n_len);
     }
     const size_t frames_per_chunk = 3000;
     GGML_ASSERT((size_t) out_full.n_len > frames_per_chunk);
     for (size_t off = 0; off < (size_t) out_full.n_len; off += frames_per_chunk) {
-        int n_len = std::min(frames_per_chunk, (size_t) out_full.n_len - off);
-        if ((size_t) n_len < frames_per_chunk) {
+        int64_t n_len = std::min((int64_t)frames_per_chunk, out_full.n_len - (int64_t)off);
+        if (n_len < (int64_t)frames_per_chunk) {
             break;  // last incomplete chunk will always be a padded chunk, safe to ignore
         }
 
@@ -596,10 +608,10 @@ bool mtmd_audio_preprocessor_whisper::preprocess(const float *                 s
         out_chunk.n_len     = n_len;
         out_chunk.n_mel     = out_full.n_mel;
         out_chunk.n_len_org = out_full.n_mel;  // unused
-        out_chunk.data.reserve(out_chunk.n_mel * out_chunk.n_len);
+        out_chunk.data.reserve((size_t)out_chunk.n_mel * (size_t)out_chunk.n_len);
 
-        for (int i = 0; i < out_full.n_mel; i++) {
-            auto src = out_full.data.begin() + i * out_full.n_len + off;
+        for (int64_t i = 0; i < out_full.n_mel; i++) {
+            auto src = out_full.data.begin() + (size_t)i * out_full.n_len + off;
             out_chunk.data.insert(out_chunk.data.end(), src, src + frames_per_chunk);
         }
 
@@ -681,8 +693,8 @@ bool mtmd_audio_preprocessor_qwen3a::preprocess(const float *                 sa
 
     // The effective frame count: center-padded STFT gives ~n_samples/hop_length frames.
     // We take min(mel_full.n_len, n_samples/hop + 1) to avoid including excess frames.
-    const int n_eff = std::min(mel_full.n_len,
-                               (int)(n_samples / hparams.audio_hop_len) + 1);
+    const int64_t n_eff = std::min(mel_full.n_len,
+                               (int64_t)(n_samples / hparams.audio_hop_len) + 1);
 
     // Split into inference windows matching n_window_infer=800 from model config.
     // Each window is padded to the next multiple of chunk_size for the cgraph.
@@ -690,18 +702,18 @@ bool mtmd_audio_preprocessor_qwen3a::preprocess(const float *                 sa
     const int chunk_size  = 100; // conv sub-chunk size (n_window * 2, n_window=50)
     const int window_size = 800; // mel frames per forward pass (n_window_infer=800)
 
-    for (int off = 0; off < n_eff; off += window_size) {
-        const int win_eff    = std::min(window_size, n_eff - off);
-        const int n_chunks   = (win_eff + chunk_size - 1) / chunk_size;
-        const int n_padded   = n_chunks * chunk_size;
+    for (int64_t off = 0; off < n_eff; off += window_size) {
+        const int64_t win_eff  = std::min((int64_t)window_size, n_eff - off);
+        const int64_t n_chunks  = (win_eff + chunk_size - 1) / chunk_size;
+        const int64_t n_padded  = n_chunks * chunk_size;
 
         mtmd_audio_mel out;
         out.n_mel     = mel_full.n_mel;
         out.n_len     = n_padded;
         out.n_len_org = win_eff;
-        out.data.assign(out.n_mel * out.n_len, 0.0f);
-        for (int m = 0; m < out.n_mel; m++) {
-            const int copy_len = std::min(win_eff, mel_full.n_len - off);
+        out.data.assign((size_t)out.n_mel * (size_t)out.n_len, 0.0f);
+        for (int64_t m = 0; m < out.n_mel; m++) {
+            const int64_t copy_len = std::min((int64_t)win_eff, mel_full.n_len - off);
             if (copy_len > 0) {
                 std::copy(mel_full.data.begin() + (size_t)m * mel_full.n_len + off,
                           mel_full.data.begin() + (size_t)m * mel_full.n_len + off + copy_len,
@@ -823,37 +835,38 @@ bool mtmd_audio_preprocessor_granite_speech::preprocess(const float *
     }
 
     double mmax = -1e20;
-    for (int i = 0; i < mel.n_mel * mel.n_len; i++) {
+    const size_t mel_size = (size_t)mel.n_mel * (size_t)mel.n_len;
+    for (size_t i = 0; i < mel_size; i++) {
         if (mel.data[i] > mmax) {
             mmax = mel.data[i];
         }
     }
     mmax -= 8.0;
 
-    for (int i = 0; i < mel.n_mel * mel.n_len; i++) {
+    for (size_t i = 0; i < mel_size; i++) {
         if (mel.data[i] < mmax) {
             mel.data[i] = mmax;
         }
         mel.data[i] = (mel.data[i] + 4.0) / 4.0;
     }
 
-    int n_frames = mel.n_len;
+    int64_t n_frames = mel.n_len;
     if (n_frames % 2 == 1) {
         n_frames--;
     }
-    const int n_mel     = mel.n_mel;
-    const int n_stacked = n_frames / 2;
+    const int64_t n_mel     = mel.n_mel;
+    const int64_t n_stacked = n_frames / 2;
 
     mtmd_audio_mel stacked;
     stacked.n_mel     = 2 * n_mel;
     stacked.n_len     = n_stacked;
-    stacked.n_len_org = (int)n_samples;
-    stacked.data.resize(2 * n_mel * n_stacked);
+    stacked.n_len_org = (int64_t)n_samples;
+    stacked.data.resize((size_t)2 * (size_t)n_mel * (size_t)n_stacked);
 
-    for (int t = 0; t < n_stacked; t++) {
-        for (int m = 0; m < n_mel; m++) {
-            stacked.data[m * n_stacked + t] = mel.data[m * mel.n_len + 2 * t];
-            stacked.data[(m + n_mel) * n_stacked + t] = mel.data[m * mel.n_len + 2 * t + 1];
+    for (int64_t t = 0; t < n_stacked; t++) {
+        for (int64_t m = 0; m < n_mel; m++) {
+            stacked.data[(size_t)m * n_stacked + t] = mel.data[(size_t)m * mel.n_len + 2 * t];
+            stacked.data[(size_t)(m + n_mel) * n_stacked + t] = mel.data[(size_t)m * mel.n_len + 2 * t + 1];
         }
     }
 
@@ -921,8 +934,8 @@ bool mtmd_audio_preprocessor_gemma4a::preprocess(const float *                 s
         const int hop = hparams.audio_hop_len;
         const int n_with_left = (int)chunk_len + pad_left;
         // PyTorch: unfold(size=frame_length+1, step=hop) on semicausal-padded waveform
-        const int pt_frames = (n_with_left - (hparams.audio_window_len + 1)) / hop + 1;
-        const int n_padded_needed = (pt_frames - 1) * hop + fft_size;
+        const int64_t pt_frames = (n_with_left - (hparams.audio_window_len + 1)) / hop + 1;
+        const int64_t n_padded_needed = (pt_frames - 1) * hop + fft_size;
         const int total_pad = std::max((int)(n_padded_needed - (int)chunk_len), pad_left);
         std::vector<float> padded_samples(total_pad + chunk_len, 0.0f);
         std::copy(chunk_ptr, chunk_ptr + chunk_len, padded_samples.data() + pad_left);

tools/mtmd/mtmd-audio.h

@@ -10,16 +10,16 @@
 #define MTMD_INTERNAL_HEADER
 
 struct mtmd_audio_mel {
-    int n_len;
-    int n_len_org;
-    int n_mel;
+    int64_t n_len;
+    int64_t n_len_org;
+    int64_t n_mel;
 
     std::vector<float> data;
 };
 
 struct mtmd_audio_mel_filters {
-    int32_t n_mel;
-    int32_t n_fft;
+    int64_t n_mel;
+    int64_t n_fft;
 
     std::vector<float> data;
 };
@@ -39,8 +39,8 @@ struct mtmd_audio_cache {
 
     // Build mel filterbank matrix [n_mel × n_fft_bins] at runtime.
     // n_fft_bins must be (N_fft / 2 + 1). Example: if N_fft=512 -> n_fft_bins=257.
-    void fill_mel_filterbank_matrix(int   n_mel,
-                                    int   n_fft,
+    void fill_mel_filterbank_matrix(int64_t n_mel,
+                                    int64_t n_fft,
                                     int   sample_rate,               // e.g. 16000
                                     float fmin             = 0.0f,   // e.g. 0.0
                                     float fmax             = -1.0f,  // e.g. sr/2; pass -1 for auto

tools/mtmd/mtmd-cli.cpp

@@ -32,9 +32,9 @@ static volatile bool g_is_generating = false;
 static volatile bool g_is_interrupted = false;
 
 /**
- * Please note that this is NOT a production-ready stuff.
+ * Please note that this is NOT a production-ready binary.
  * It is a playground for trying multimodal support in llama.cpp.
- * For contributors: please keep this code simple and easy to understand.
+ * For contributors: please keep this code simple and easy to understand. Do not add unnecessary complexity. The goal is to have a simple CLI for testing multimodal support.
  */
 
 static void show_additional_info(int /*argc*/, char ** argv) {
@@ -65,6 +65,14 @@ static void sigint_handler(int signo) {
 }
 #endif
 
+// this is only used by tests.sh to capture the response ; it's not meant to be used in production
+static void inject_test_response_marker() {
+    const char * env = std::getenv("MTMD_TEST_RESPONSE_MARKER");
+    if (env) {
+        LOG("%s\n", env);
+    }
+}
+
 struct mtmd_cli_context {
     mtmd::context_ptr ctx_vision;
     common_init_result_ptr llama_init;
@@ -79,6 +87,8 @@ struct mtmd_cli_context {
     mtmd::bitmaps bitmaps;
     std::vector<mtmd_helper::video_ptr> videos;
 
+    mtmd::batch_ptr mbatch;
+
     // chat template
     common_chat_templates_ptr tmpls;
     std::vector<common_chat_msg> chat_history;
@@ -233,6 +243,8 @@ static std::string chat_add_and_format(mtmd_cli_context & ctx, common_chat_msg &
 }
 
 static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg) {
+    inject_test_response_marker();
+
     bool add_bos = ctx.chat_history.empty();
     auto formatted_chat = chat_add_and_format(ctx, msg);
     LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.c_str());
@@ -259,20 +271,95 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg) {
     ctx.bitmaps.entries.clear();
     ctx.videos.clear();
 
-    llama_pos new_n_past;
-    if (mtmd_helper_eval_chunks(ctx.ctx_vision.get(),
-                ctx.lctx, // lctx
-                chunks.ptr.get(), // chunks
-                ctx.n_past, // n_past
-                0, // seq_id
-                ctx.n_batch, // n_batch
-                true, // logits_last
-                &new_n_past)) {
-        LOG_ERR("Unable to eval prompt\n");
-        return 1;
-    }
+    // batch encode all media chunks, then decode each
+    size_t n_chunks = mtmd_input_chunks_size(chunks.ptr.get());
+    for (size_t i = 0; i < n_chunks; i++) {
+        auto chunk = mtmd_input_chunks_get(chunks.ptr.get(), i);
+        auto chunk_type = mtmd_input_chunk_get_type(chunk);
 
-    ctx.n_past = new_n_past;
+        if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            // decode text chunk
+            llama_pos new_n_past = ctx.n_past;
+            res = mtmd_helper_eval_chunk_single(ctx.ctx_vision.get(),
+                        ctx.lctx,
+                        chunk,
+                        ctx.n_past,
+                        0, // seq_id
+                        ctx.n_batch,
+                        i == n_chunks - 1, // logits_last
+                        &new_n_past);
+            if (res != 0) {
+                LOG_ERR("Unable to eval text chunk %zu\n", i);
+                return 1;
+            }
+            ctx.n_past = new_n_past;
+        } else {
+            // media chunk: try to get embd from existing batch, or create a new batch
+            float * embd = nullptr;
+            if (ctx.mbatch) {
+                embd = mtmd_batch_get_output_embd(ctx.mbatch.get(), chunk);
+
+                if (embd) {
+                    LOG_DBG("found embd for media chunk %zu in existing batch\n", i);
+                } else {
+                    LOG_DBG("media chunk %zu not found in existing batch, creating new batch\n", i);
+                }
+            }
+
+            if (!embd) {
+                // create and encode a new batch with as many media chunks as possible
+                ctx.mbatch.reset(mtmd_batch_init(ctx.ctx_vision.get()));
+                res = mtmd_batch_add_chunk(ctx.mbatch.get(), chunk);
+                GGML_ASSERT(res == 0); // first chunk must always succeed
+
+                int n_added = 1;
+                // add as many subsequent media chunks as possible
+                for (size_t j = i + 1; j < n_chunks; j++) {
+                    auto next_chunk = mtmd_input_chunks_get(chunks.ptr.get(), j);
+                    auto next_type = mtmd_input_chunk_get_type(next_chunk);
+                    if (next_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+                        break; // text chunk splits the batch
+                    }
+                    res = mtmd_batch_add_chunk(ctx.mbatch.get(), next_chunk);
+                    if (res != 0) {
+                        break; // batch full or incompatible
+                    }
+                    n_added++;
+                }
+
+                int64_t time_start = ggml_time_ms();
+                LOG_INF("encoding mtmd batch, n_chunks = %d (done = %zu, total = %zu)\n", n_added, i, n_chunks);
+                res = mtmd_batch_encode(ctx.mbatch.get());
+                if (res != 0) {
+                    LOG_ERR("Failed to encode mtmd batch, res = %d\n", res);
+                    return 1;
+                }
+                LOG_INF("mtmd batch encoding done in %d ms\n", (int)(ggml_time_ms() - time_start));
+
+                embd = mtmd_batch_get_output_embd(ctx.mbatch.get(), chunk);
+            }
+
+            GGML_ASSERT(embd != nullptr);
+
+            llama_pos new_n_past = ctx.n_past;
+            res = mtmd_helper_decode_image_chunk(ctx.ctx_vision.get(),
+                        ctx.lctx,
+                        chunk,
+                        embd,
+                        ctx.n_past,
+                        0, // seq_id
+                        ctx.n_batch,
+                        &new_n_past,
+                        nullptr, // callback
+                        nullptr  // user_data
+                    );
+            if (res != 0) {
+                LOG_ERR("Unable to decode media chunk %zu\n", i);
+                return 1;
+            }
+            ctx.n_past = new_n_past;
+        }
+    }
 
     LOG("\n");
 
@@ -309,6 +396,9 @@ int main(int argc, char ** argv) {
 
     int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
 
+    console::init(params.simple_io, params.use_color);
+    atexit([]() { console::cleanup(); });
+
     // Ctrl+C handling
     {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))

tools/mtmd/mtmd-helper.cpp

@@ -582,13 +582,29 @@ mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx,
 }
 
 mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname, bool placeholder) {
-    std::vector<unsigned char> buf;
+#ifdef _WIN32
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
+    if (!wlen) {
+        LOG_ERR("Unable to convert filename to UTF-16: %s\n", fname);
+        return {nullptr, nullptr};
+    }
+    std::vector<wchar_t> wfname(wlen);
+    wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wfname.data(), wlen);
+    if (!wlen) {
+        LOG_ERR("Unable to convert filename to UTF-16: %s\n", fname);
+        return {nullptr, nullptr};
+    }
+    FILE * f = _wfopen(wfname.data(), L"rb");
+#else
     FILE * f = fopen(fname, "rb");
+#endif
     if (!f) {
         LOG_ERR("Unable to open file %s: %s\n", fname, strerror(errno));
         return {nullptr, nullptr};
     }
 
+    std::vector<unsigned char> buf;
+
     fseek(f, 0, SEEK_END);
     long file_size = ftell(f);
     fseek(f, 0, SEEK_SET);

tools/mtmd/mtmd-image.cpp

@@ -4,17 +4,33 @@
 #include <cmath>
 #include <vector>
 
-//
-// base implementation
-//
-
-void mtmd_image_preprocessor::img_u8_to_f32(const clip_image_u8 & src, clip_image_f32 & dst, const float mean[3], const float std[3]) {
-    dst.from_u8(src);
-    dst.normalize(mean, std);
+void mtmd_image_preproc_out::append(const clip_hparams & hparams, const clip_image_u8 & img, bool normalized) {
+    clip_image_f32 dst;
+    dst.from_u8(img);
+    if (normalized) {
+        dst.normalize(hparams.image_mean, hparams.image_std);
+    }
+    entries.push_back(std::move(dst));
 }
 
-void mtmd_image_preprocessor::img_u8_to_f32(const clip_image_u8 & src, clip_image_f32 & dst) {
-    dst.from_u8(src);
+void mtmd_image_preproc_out::append(const clip_hparams & hparams, const std::vector<clip_image_u8> & imgs, bool normalized) {
+    for (const auto & img : imgs) {
+        append(hparams, img, normalized);
+    }
+}
+
+void mtmd_image_preproc_out::append(const clip_hparams & hparams, clip_image_f32 & img, bool normalized) {
+    if (normalized) {
+        img.normalize(hparams.image_mean, hparams.image_std);
+    }
+    entries.push_back(std::move(img));
+}
+
+void mtmd_image_preproc_out::append_overview(const clip_hparams & hparams, const clip_image_u8 & img, bool normalized) {
+    overview.from_u8(img);
+    if (normalized) {
+        overview.normalize(hparams.image_mean, hparams.image_std);
+    }
 }
 
 // set of tools to manipulate images
@@ -595,21 +611,18 @@ private:
 // mtmd_image_preprocessor_llava_uhd
 //
 
-bool mtmd_image_preprocessor_llava_uhd::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_llava_uhd::preprocess(const clip_image_u8 & img) {
     const clip_image_size original_size = img.get_size();
     auto const inst = get_slice_instructions(original_size);
-    std::vector<clip_image_u8_ptr> imgs = slice_image(img, inst);
-
-    for (size_t i = 0; i < imgs.size(); ++i) {
-        // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
-        clip_image_f32_ptr res(clip_image_f32_init());
-        img_u8_to_f32(*imgs[i], *res, hparams.image_mean, hparams.image_std);
-        output.entries.push_back(std::move(res));
-    }
+    auto sliced = slice_image(img, inst);
 
+    mtmd_image_preproc_out output;
+    output.append_overview(hparams, sliced.overview, true);
+    output.append(hparams, sliced.slices, true);
     output.grid_x = inst.grid_size.width;
     output.grid_y = inst.grid_size.height;
-    return true;
+
+    return output;
 }
 
 mtmd_image_preprocessor_llava_uhd::slice_instructions mtmd_image_preprocessor_llava_uhd::get_slice_instructions(const clip_image_size & original_size) {
@@ -717,28 +730,21 @@ mtmd_image_preprocessor_llava_uhd::slice_instructions mtmd_image_preprocessor_ll
     return res;
 }
 
-std::vector<clip_image_u8_ptr> mtmd_image_preprocessor_llava_uhd::slice_image(const clip_image_u8 & img, const mtmd_image_preprocessor_llava_uhd::slice_instructions & inst, bool overview_first) {
-    std::vector<clip_image_u8_ptr> output;
+mtmd_image_preprocessor_llava_uhd::slice_output mtmd_image_preprocessor_llava_uhd::slice_image(const clip_image_u8 & img, const mtmd_image_preprocessor_llava_uhd::slice_instructions & inst) {
+    slice_output output;
 
     // resize to overview size
-    clip_image_u8_ptr resized_img(clip_image_u8_init());
-    img_tool::resize(img, *resized_img, inst.overview_size, hparams.image_resize_algo_ov,
+    img_tool::resize(img, output.overview, inst.overview_size, hparams.image_resize_algo_ov,
                         hparams.image_pad_ov, hparams.image_pad_color_ov);
-    if (overview_first) {
-        output.push_back(std::move(resized_img));
-    }
 
     if (inst.slices.empty()) {
-        // no slices, just return the resized image
-        if (!overview_first) {
-            output.push_back(std::move(resized_img));
-        }
+        // no slices, just return the overview image
         return output;
     }
 
     // resize to refined size
-    clip_image_u8_ptr refined_img(clip_image_u8_init());
-    img_tool::resize(img, *refined_img, inst.refined_size, hparams.image_resize_algo_rf,
+    clip_image_u8 refined_img;
+    img_tool::resize(img, refined_img, inst.refined_size, hparams.image_resize_algo_rf,
                         hparams.image_pad_rf, hparams.image_pad_color_rf);
 
     // create slices
@@ -748,13 +754,9 @@ std::vector<clip_image_u8_ptr> mtmd_image_preprocessor_llava_uhd::slice_image(co
         int w = slice.size.width;
         int h = slice.size.height;
 
-        clip_image_u8_ptr img_slice(clip_image_u8_init());
-        img_tool::crop(*refined_img, *img_slice, x, y, w, h);
-        output.push_back(std::move(img_slice));
-    }
-
-    if (!overview_first) {
-        output.push_back(std::move(resized_img));
+        clip_image_u8 img_slice;
+        img_tool::crop(refined_img, img_slice, x, y, w, h);
+        output.slices.push_back(std::move(img_slice));
     }
 
     return output;
@@ -871,24 +873,23 @@ clip_image_size mtmd_image_preprocessor_llava_uhd::get_best_grid(const int max_s
 // mtmd_image_preprocessor_fixed_size
 //
 
-bool mtmd_image_preprocessor_fixed_size::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_fixed_size::preprocess(const clip_image_u8 & img) {
     clip_image_u8 resized_image;
     int sz = hparams.image_size;
     img_tool::resize(img, resized_image, {sz, sz},
                         hparams.image_resize_algo,
                         hparams.image_resize_pad,
                         hparams.image_pad_color);
-    clip_image_f32_ptr img_f32(clip_image_f32_init());
-    img_u8_to_f32(resized_image, *img_f32, hparams.image_mean, hparams.image_std);
-    output.entries.push_back(std::move(img_f32));
-    return true;
+    mtmd_image_preproc_out output;
+    output.append(hparams, resized_image, true);
+    return output;
 }
 
 //
 // mtmd_image_preprocessor_dyn_size
 //
 
-bool mtmd_image_preprocessor_dyn_size::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_dyn_size::preprocess(const clip_image_u8 & img) {
     GGML_ASSERT(hparams.image_min_pixels > 0 && hparams.image_max_pixels > 0);
     clip_image_u8 resized_image;
     const clip_image_size original_size = img.get_size();
@@ -903,17 +904,16 @@ bool mtmd_image_preprocessor_dyn_size::preprocess(const clip_image_u8 & img, cli
                         hparams.image_resize_algo,
                         hparams.image_resize_pad,
                         hparams.image_pad_color);
-    clip_image_f32_ptr img_f32(clip_image_f32_init());
-    img_u8_to_f32(resized_image, *img_f32, hparams.image_mean, hparams.image_std);
-    output.entries.push_back(std::move(img_f32));
-    return true;
+    mtmd_image_preproc_out output;
+    output.append(hparams, resized_image, true);
+    return output;
 }
 
 //
 // mtmd_image_preprocessor_longest_edge
 //
 
-bool mtmd_image_preprocessor_longest_edge::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_longest_edge::preprocess(const clip_image_u8 & img) {
     GGML_ASSERT(hparams.image_longest_edge > 0);
     clip_image_u8 resized_image;
     const clip_image_size original_size = img.get_size();
@@ -927,10 +927,9 @@ bool mtmd_image_preprocessor_longest_edge::preprocess(const clip_image_u8 & img,
                         hparams.image_resize_algo,
                         hparams.image_resize_pad,
                         hparams.image_pad_color);
-    clip_image_f32_ptr img_f32(clip_image_f32_init());
-    img_u8_to_f32(resized_image, *img_f32, hparams.image_mean, hparams.image_std);
-    output.entries.push_back(std::move(img_f32));
-    return true;
+    mtmd_image_preproc_out output;
+    output.append(hparams, resized_image, true);
+    return output;
 }
 
 //
@@ -1040,7 +1039,7 @@ clip_image_size mtmd_image_preprocessor_lfm2::get_grid_layout(int height, int wi
 // mtmd_image_preprocessor_idefics3
 //
 
-bool mtmd_image_preprocessor_idefics3::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_idefics3::preprocess(const clip_image_u8 & img) {
     // The refined size has two steps:
     // 1. Resize w/ aspect-ratio preserving such that the longer side is
     //      the preprocessor longest size
@@ -1075,46 +1074,40 @@ bool mtmd_image_preprocessor_idefics3::preprocess(const clip_image_u8 & img, cli
             });
         }
     }
-    auto imgs = slice_image(img, instructions);
-
-    // cast and normalize to f32
-    for (size_t i = 0; i < imgs.size(); ++i) {
-        // clip_image_save_to_bmp(*imgs[i], "slice_" + std::to_string(i) + ".bmp");
-        clip_image_f32_ptr res(clip_image_f32_init());
-        img_u8_to_f32(*imgs[i], *res, hparams.image_mean, hparams.image_std);
-        output.entries.push_back(std::move(res));
-    }
+    auto sliced = slice_image(img, instructions);
 
+    mtmd_image_preproc_out output;
+    output.append_overview(hparams, sliced.overview, true);
+    output.append(hparams, sliced.slices, true);
     output.grid_x = instructions.grid_size.width;
     output.grid_y = instructions.grid_size.height;
-    return true;
+    return output;
 }
 
 //
 // mtmd_image_preprocessor_internvl
 //
 
-bool mtmd_image_preprocessor_internvl::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_internvl::preprocess(const clip_image_u8 & img) {
     GGML_ASSERT(!hparams.image_res_candidates.empty());
     const clip_image_size original_size = img.get_size();
     auto const inst = get_slice_instructions(original_size);
-    std::vector<clip_image_u8_ptr> imgs = slice_image(img, inst, false);
+    auto sliced = slice_image(img, inst);
 
-    for (size_t i = 0; i < imgs.size(); ++i) {
-        clip_image_f32_ptr res(clip_image_f32_init());
-        img_u8_to_f32(*imgs[i], *res, hparams.image_mean, hparams.image_std);
-        output.entries.push_back(std::move(res));
-    }
+    mtmd_image_preproc_out output;
+    // InternVL: slices first, then overview
+    output.append(hparams, sliced.slices, true);
+    output.append_overview(hparams, sliced.overview, true);
     output.grid_x = inst.grid_size.width;
     output.grid_y = inst.grid_size.height;
-    return true;
+    return output;
 }
 
 //
 // mtmd_image_preprocessor_deepseekocr
 //
 
-bool mtmd_image_preprocessor_deepseekocr::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_deepseekocr::preprocess(const clip_image_u8 & img) {
     static constexpr int native_resolutions[] = { 1024 /* base */, 1280 /* large */ };
     // TODO: support 512 (tiny) and 640 (small) once we have eval data for them
 
@@ -1137,14 +1130,12 @@ bool mtmd_image_preprocessor_deepseekocr::preprocess(const clip_image_u8 & img,
     clip_image_u8 padded;
     img_tool::resize(img, padded, {image_size, image_size}, RESIZE_ALGO_BICUBIC_PILLOW,
                      PAD_NEAREST, hparams.image_pad_color);
-
-    clip_image_f32_ptr res(clip_image_f32_init());
-    img_u8_to_f32(padded, *res, hparams.image_mean, hparams.image_std);
-    output.entries.push_back(std::move(res));
-
-    output.grid_x = 1;
-    output.grid_y = 1;
-    return true;
+    mtmd_image_preproc_out output;
+    output.append_overview(hparams, padded, true);
+    output.grid_x = 0;
+    output.grid_y = 0;
+    // TODO @ngxson : support slicing for DeepSeek-OCR, to do in another PR
+    return output;
 }
 
 //
@@ -1207,10 +1198,11 @@ clip_image_size mtmd_image_preprocessor_deepseekocr2::find_closest_aspect_ratio(
     return best_ratio;
 }
 
-bool mtmd_image_preprocessor_deepseekocr2::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_deepseekocr2::preprocess(const clip_image_u8 & img) {
     // emit 768x768 local tiles when the image is larger than a tile in either
     // dimension, then always a 1024x1024 global view. order: [tiles..., global].
 
+    mtmd_image_preproc_out output;
     const auto img_size = img.get_size();
     if (img_size.width > tile_size || img_size.height > tile_size) {
         const float           aspect_ratio  = static_cast<float>(img_size.width) / img_size.height;
@@ -1226,9 +1218,7 @@ bool mtmd_image_preprocessor_deepseekocr2::preprocess(const clip_image_u8 & img,
             for (int col = 0; col < grid.width; col++) {
                 clip_image_u8 tile;
                 img_tool::crop(refined, tile, col * tile_size, row * tile_size, tile_size, tile_size);
-                clip_image_f32_ptr res(clip_image_f32_init());
-                img_u8_to_f32(tile, *res, hparams.image_mean, hparams.image_std);
-                output.entries.push_back(std::move(res));
+                output.append(hparams, tile, true);
             }
         }
     }
@@ -1237,14 +1227,9 @@ bool mtmd_image_preprocessor_deepseekocr2::preprocess(const clip_image_u8 & img,
     clip_image_u8 padded;
     img_tool::resize(img, padded, { base_size, base_size }, RESIZE_ALGO_BICUBIC_PILLOW,
                      PAD_NEAREST, hparams.image_pad_color);
-    clip_image_f32_ptr global(clip_image_f32_init());
-    img_u8_to_f32(padded, *global, hparams.image_mean, hparams.image_std);
-    global->add_viewsep = true;
-    output.entries.push_back(std::move(global));
-
-    output.grid_x = 1;
-    output.grid_y = 1;
-    return true;
+    output.append_overview(hparams, padded, true);
+    output.overview.add_viewsep = true;
+    return output;
 }
 
 //
@@ -1260,7 +1245,8 @@ void mtmd_image_preprocessor_step3vl::img_u8_resize_bilinear_to_f32(
         const float std[3]) {
     const auto src_size = src.get_size();
     if (src_size.width == target_width && src_size.height == target_height) {
-        img_u8_to_f32(src, dst, mean, std);
+        dst.from_u8(src);
+        dst.normalize(mean, std);
         return;
     }
 
@@ -1455,24 +1441,24 @@ mtmd_image_preprocessor_step3vl::slice_instructions mtmd_image_preprocessor_step
     return instructions;
 }
 
-bool mtmd_image_preprocessor_step3vl::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_step3vl::preprocess(const clip_image_u8 & img) {
     clip_image_u8 prepared = prepare_image(img, hparams);
     const auto instructions = build_slice_instructions(hparams, prepared.get_size());
 
-    clip_image_f32_ptr overview_f32(clip_image_f32_init());
+    mtmd_image_preproc_out output;
+    // overview (normalized f32, already includes mean/std)
     img_u8_resize_bilinear_to_f32(
         prepared,
-        *overview_f32,
+        output.overview,
         hparams.image_size,
         hparams.image_size,
         hparams.image_mean,
         hparams.image_std);
-    output.entries.push_back(std::move(overview_f32));
 
     if (instructions.slices.empty()) {
         output.grid_x = 0;
         output.grid_y = 0;
-        return true;
+        return output;
     }
 
     clip_image_u8 img_for_crop = prepared;
@@ -1488,28 +1474,28 @@ bool mtmd_image_preprocessor_step3vl::preprocess(const clip_image_u8 & img, clip
         // If the requested patch extends past the source image, pad the out-of-bounds area with black.
         clip_image_u8 patch = crop_with_black_padding(img_for_crop, slice.x, slice.y, slice.size.width, slice.size.height);
 
-        clip_image_f32_ptr patch_f32(clip_image_f32_init());
+        clip_image_f32 patch_f32;
         img_u8_resize_bilinear_to_f32(
             patch,
-            *patch_f32,
+            patch_f32,
             crop_size,
             crop_size,
             hparams.image_mean,
             hparams.image_std);
-        output.entries.push_back(std::move(patch_f32));
+        output.append(hparams, patch_f32, false);
     }
 
     output.grid_x = instructions.grid_size.width;
     output.grid_y = instructions.grid_size.height;
 
-    return true;
+    return output;
 }
 
 //
 // mtmd_image_preprocessor_youtuvl
 //
 
-bool mtmd_image_preprocessor_youtuvl::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
+mtmd_image_preproc_out mtmd_image_preprocessor_youtuvl::preprocess(const clip_image_u8 & img) {
     const int patch_size = hparams.patch_size;   // typically 16
     const int merge_size = hparams.n_merge;      // typically 2
     const int align_size = patch_size * merge_size;  // 32
@@ -1553,29 +1539,22 @@ bool mtmd_image_preprocessor_youtuvl::preprocess(const clip_image_u8 & img, clip
     clip_image_u8 resized;
     img_tool::resize(img, resized, new_size, hparams.image_resize_algo, hparams.image_resize_pad);
 
-    // Normalize to float32
-    clip_image_f32_ptr img_f32(clip_image_f32_init());
-    img_u8_to_f32(resized, *img_f32, hparams.image_mean, hparams.image_std);
-    // Add to results
-    output.entries.push_back(std::move(img_f32));
-    return true;
+    mtmd_image_preproc_out output;
+    output.append(hparams, resized, true);
+    return output;
 }
 
-bool mtmd_image_preprocessor_granite::preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) {
-    // call super class preprocessor
-    bool ok = mtmd_image_preprocessor_llava_uhd::preprocess(img, output);
-    if (!ok) {
-        return false;
-    }
-    if (output.entries.size() == 1) {
+mtmd_image_preproc_out mtmd_image_preprocessor_granite::preprocess(const clip_image_u8 & img) {
+    auto output = mtmd_image_preprocessor_llava_uhd::preprocess(img);
+    if (output.entries.size() == 0) {
         // Single-tile (overview only): append one newline row.
-        output.entries[0]->add_newline = true;
+        output.overview.add_newline = true;
     } else {
         // Multi-tile: overview gets no newline, grid tiles get one.
-        output.entries[0]->add_newline = false;
-        for (size_t i = 1; i < output.entries.size(); ++i) {
-            output.entries[i]->add_newline = true;
+        output.overview.add_newline = false;
+        for (size_t i = 0; i < output.entries.size(); ++i) {
+            output.entries[i].add_newline = true;
         }
     }
-    return true;
+    return output;
 }

tools/mtmd/mtmd-image.h

@@ -8,6 +8,24 @@
 
 #define MTMD_INTERNAL_HEADER
 
+struct mtmd_image_preproc_out {
+    std::vector<clip_image_f32> entries;
+    // grid size is required for llava-uhd style models
+
+    clip_image_f32 overview; // overview image (downscaled image)
+    int grid_x = 0;
+    int grid_y = 0;
+
+    void append(const clip_hparams & hparams, const clip_image_u8 & img, bool normalized = true);
+    void append(const clip_hparams & hparams, const std::vector<clip_image_u8> & imgs, bool normalized = true);
+    void append(const clip_hparams & hparams, clip_image_f32 & img, bool normalized = true);
+
+    void append_overview(const clip_hparams & hparams, const clip_image_u8 & img, bool normalized = true);
+    bool has_overview() const {
+        return overview.nx() > 0 || overview.ny() > 0;
+    }
+};
+
 // base class, models must inherit from this class
 struct mtmd_image_preprocessor {
     const clip_hparams & hparams;
@@ -15,10 +33,7 @@ struct mtmd_image_preprocessor {
     mtmd_image_preprocessor(const clip_ctx * ctx): hparams(*clip_get_hparams(ctx)) {}
 
     virtual ~mtmd_image_preprocessor() = default;
-    virtual bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) = 0;
-
-    void img_u8_to_f32(const clip_image_u8 & src, clip_image_f32 & dst, const float mean[3], const float std[3]);
-    void img_u8_to_f32(const clip_image_u8 & src, clip_image_f32 & dst);
+    virtual mtmd_image_preproc_out preprocess(const clip_image_u8 & img) = 0;
 };
 
 /**
@@ -39,10 +54,12 @@ struct mtmd_image_preprocessor {
  * [overview] --> [slice 1] --> [slice 2]
  *           |                |
  *           +--> [slice 3] --> [slice 4]
+ *
+ * NOTE: for the ordering of overview, set "ov_img_first" on the mtmd_context
  */
 struct mtmd_image_preprocessor_llava_uhd : mtmd_image_preprocessor {
     mtmd_image_preprocessor_llava_uhd(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 
     struct slice_coordinates {
         int x;
@@ -60,7 +77,11 @@ struct mtmd_image_preprocessor_llava_uhd : mtmd_image_preprocessor {
     // LFM2 override this function to implement its custom slicing logic
     virtual slice_instructions get_slice_instructions(const clip_image_size & original_size);
 
-    std::vector<clip_image_u8_ptr> slice_image(const clip_image_u8 & img, const slice_instructions & inst, bool overview_first = true);
+    struct slice_output {
+        clip_image_u8 overview;
+        std::vector<clip_image_u8> slices;
+    };
+    slice_output slice_image(const clip_image_u8 & img, const slice_instructions & inst);
 
 private:
     clip_image_size get_best_resize(const clip_image_size & original_size, int scale_resolution, int patch_size, bool allow_upscale = false);
@@ -91,7 +112,7 @@ private:
 // downscale or upscale the input image to fixed size
 struct mtmd_image_preprocessor_fixed_size : mtmd_image_preprocessor {
     mtmd_image_preprocessor_fixed_size(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 };
 
 // resize image to multiple of patch_size*n_merge, while preserving aspect ratio
@@ -99,13 +120,13 @@ struct mtmd_image_preprocessor_fixed_size : mtmd_image_preprocessor {
 // this is used by models with native support for dynamic image size, for example: Qwen-VL, Pixtral, Kimi-VL, etc
 struct mtmd_image_preprocessor_dyn_size : mtmd_image_preprocessor {
     mtmd_image_preprocessor_dyn_size(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 };
 
 // similar to mtmd_image_preprocessor_dyn_size, but resize the image to have longest edge equal to hparams.image_longest_edge, while preserving aspect ratio
 struct mtmd_image_preprocessor_longest_edge : mtmd_image_preprocessor {
     mtmd_image_preprocessor_longest_edge(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 };
 
 // custom llava-uhd slicing logic for LFM2
@@ -131,17 +152,17 @@ private:
 
 struct mtmd_image_preprocessor_idefics3 : mtmd_image_preprocessor_llava_uhd {
     mtmd_image_preprocessor_idefics3(const clip_ctx * ctx) : mtmd_image_preprocessor_llava_uhd(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 };
 
 struct mtmd_image_preprocessor_internvl : mtmd_image_preprocessor_llava_uhd {
     mtmd_image_preprocessor_internvl(const clip_ctx * ctx) : mtmd_image_preprocessor_llava_uhd(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 };
 
 struct mtmd_image_preprocessor_deepseekocr : mtmd_image_preprocessor {
     mtmd_image_preprocessor_deepseekocr(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 };
 
 // DeepSeek-OCR-2: a 1024x1024 global view, plus InternVL-style 768x768 local
@@ -153,7 +174,7 @@ struct mtmd_image_preprocessor_deepseekocr2 : mtmd_image_preprocessor {
     static constexpr int max_tiles = 6;
 
     mtmd_image_preprocessor_deepseekocr2(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 
 private:
     static std::vector<clip_image_size> get_target_ratios();
@@ -168,7 +189,7 @@ private:
 // ref: https://huggingface.co/stepfun-ai/Step3-VL-10B/blob/main/processing_step3.py
 struct mtmd_image_preprocessor_step3vl : mtmd_image_preprocessor_llava_uhd {
     mtmd_image_preprocessor_step3vl(const clip_ctx * ctx) : mtmd_image_preprocessor_llava_uhd(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
     static slice_instructions build_slice_instructions(const clip_hparams & params, const clip_image_size & prepared_size);
 
 private:
@@ -195,11 +216,11 @@ private:
 
 struct mtmd_image_preprocessor_youtuvl : mtmd_image_preprocessor {
     mtmd_image_preprocessor_youtuvl(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 };
 
 // similar to llava_uhd, but has add_newline
 struct mtmd_image_preprocessor_granite : mtmd_image_preprocessor_llava_uhd {
     mtmd_image_preprocessor_granite(const clip_ctx * ctx) : mtmd_image_preprocessor_llava_uhd(ctx) {}
-    bool preprocess(const clip_image_u8 & img, clip_image_f32_batch & output) override;
+    mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
 };

tools/mtmd/mtmd.cpp

@@ -114,7 +114,7 @@ struct mtmd_image_tokens {
     // true if one of entries in batch_f32 is a placeholder
     bool is_placeholder() const {
         for (const auto & entry : batch_f32.entries) {
-            if (entry->is_placeholder()) {
+            if (entry.is_placeholder()) {
                 return true;
             }
         }
@@ -147,7 +147,7 @@ struct mtmd_audio_tokens {
     // true if one of entries in batch_f32 is a placeholder
     bool is_placeholder() const {
         for (const auto & entry : batch_f32.entries) {
-            if (entry->is_placeholder()) {
+            if (entry.is_placeholder()) {
                 return true;
             }
         }
@@ -516,6 +516,7 @@ struct mtmd_context {
                     LOG_WRN("%s: llama 4 vision is known to have degraded quality:\n"
                             "    https://github.com/ggml-org/llama.cpp/pull/13282\n", __func__);
                     image_preproc = std::make_unique<mtmd_image_preprocessor_llava_uhd>(ctx_v);
+                    ov_img_first = false;
                 } break;
             case PROJECTOR_TYPE_STEP3VL:
                 {
@@ -539,6 +540,7 @@ struct mtmd_context {
                     img_beg = "<img>";
                     img_end = "</img>";
                     image_preproc = std::make_unique<mtmd_image_preprocessor_internvl>(ctx_v);
+                    ov_img_first = false;
                 } break;
             case PROJECTOR_TYPE_KIMIVL:
                 {
@@ -615,11 +617,13 @@ struct mtmd_context {
                 {
                     img_end = "\n"; // prevent empty batch on llama-server
                     image_preproc = std::make_unique<mtmd_image_preprocessor_deepseekocr>(ctx_v);
+                    ov_img_first = false;
                 } break;
             case PROJECTOR_TYPE_DEEPSEEKOCR2:
                 {
                     img_end = "\n"; // prevent empty batch on llama-server
                     image_preproc = std::make_unique<mtmd_image_preprocessor_deepseekocr2>(ctx_v);
+                    ov_img_first = false;
                 } break;
             case PROJECTOR_TYPE_HUNYUANVL:
                 {
@@ -640,6 +644,7 @@ struct mtmd_context {
                     img_beg = "<image>";
                     img_end = "";
                     image_preproc = std::make_unique<mtmd_image_preprocessor_granite>(ctx_v);
+                    ov_img_first = true;
                 } break;
             default:
                 throw std::runtime_error(string_format("%s: unexpected vision projector type %d\n", __func__, proj));
@@ -1050,7 +1055,7 @@ struct mtmd_tokenizer {
 
             // TODO @ngxson : this is quite hacky because preprocessor only support batch with one single element, that need to be fixed in the future (e.g. by changing the preprocessor interface always take single input)
 
-            clip_image_f32_batch batch_f32;
+            mtmd_image_preproc_out preproc_out;
 
             for (const auto * bmp : bitmaps) {
                 // sanity check
@@ -1063,44 +1068,54 @@ struct mtmd_tokenizer {
                 }
 
                 // convert mtmd_bitmap to clip_image_u8
-                clip_image_u8_ptr img_u8(clip_image_u8_init());
-                img_u8->set_size(
+                clip_image_u8 img_u8;
+                img_u8.set_size(
                     {(int)bmp->nx, (int)bmp->ny},
                     bmp->is_placeholder());
-                img_u8->cpy_buf(bmp->get_ro_buf());
+                img_u8.cpy_buf(bmp->get_ro_buf());
 
                 // preprocess image
-                clip_image_f32_batch tmp_batch;
-                bool ok = ctx->image_preproc->preprocess(*img_u8, tmp_batch);
-                if (!ok) {
-                    LOG_ERR("Unable to preprocess image\n");
-                    return 2;
-                }
+                mtmd_image_preproc_out tmp_preproc_out = ctx->image_preproc->preprocess(img_u8);
 
-                // move entries and grid dimensions to the "global" batch_f32
-                for (auto & entry : tmp_batch.entries) {
-                    batch_f32.entries.emplace_back(std::move(entry));
+                // move entries and grid dimensions to the "global" preproc_out
+                for (auto & entry : tmp_preproc_out.entries) {
+                    preproc_out.entries.emplace_back(std::move(entry));
                 }
 
                 // for llava-uhd style, we need to handle grid too
-                // we don't care about overwriting these values for now because llama-uhd doesn't support batching anyway
-                batch_f32.grid_x = tmp_batch.grid_x;
-                batch_f32.grid_y = tmp_batch.grid_y;
+                // we don't care about overwriting these values for now because the case where bitmaps.size() > 1 is only for frame merging (qwen-vl), not supported by llava-uhd
+                if ((tmp_preproc_out.grid_x > 0 && tmp_preproc_out.grid_y > 0)
+                        || tmp_preproc_out.has_overview()) {
+                    GGML_ASSERT(bitmaps.size() == 1);
+                    preproc_out.grid_x = tmp_preproc_out.grid_x;
+                    preproc_out.grid_y = tmp_preproc_out.grid_y;
+                    preproc_out.overview = std::move(tmp_preproc_out.overview);
+                }
             }
 
+            LOG_DBG("%s: preproc_out has %zu entries, grid_x = %d, grid_y = %d, has_overview = %d\n",
+                    __func__, preproc_out.entries.size(), preproc_out.grid_x, preproc_out.grid_y,
+                    preproc_out.has_overview() ? 1 : 0);
+
             // handle llava-uhd style preprocessing
-            const bool has_tiling_grid = batch_f32.grid_x > 0 && batch_f32.grid_y > 0;
+            // (output either a grid, or overview-only)
+            const bool has_tiling_grid = (preproc_out.grid_x > 0 && preproc_out.grid_y > 0)
+                || preproc_out.has_overview();
+
             if (has_tiling_grid) {
                 // [QWEN_VIDEO] we do not support "frame merging" for llama-uhd style, so no batching for now
                 GGML_ASSERT(bitmaps.size() == 1);
 
-                const int n_col = batch_f32.grid_x;
-                const int n_row = batch_f32.grid_y;
+                const int n_col = preproc_out.grid_x;
+                const int n_row = preproc_out.grid_y;
+
                 // split batch into chunks of single images
-                // NOTE: batch_f32 will be invalidated after this call
-                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[0]->id);
+                auto chunks = split_batch_to_chunk(std::move(preproc_out), bitmaps[0]->id);
                 GGML_ASSERT(chunks.size() > 0);
 
+                // NOTE: preproc_out is invalidated after this point, do not use it anymore
+
+                // split_batch_to_chunk must always put the overview image first
                 auto ov_chunk = std::move(chunks.front());
                 chunks.erase(chunks.begin());
 
@@ -1127,7 +1142,16 @@ struct mtmd_tokenizer {
                                 std::snprintf(buf.get(), sz, ctx->sli_img_start_tmpl.c_str(), y+1, x+1);
                                 add_text(std::string(buf.get(), buf.get() + sz - 1), true);
                             }
-                            cur.entries.emplace_back(std::move(chunks[y * n_col + x]));
+
+                            auto & curr_chunk = chunks[y * n_col + x];
+                            auto & curr_batch = curr_chunk.tokens_image->batch_f32;
+                            if (curr_batch.entries.size() != 1) {
+                                throw std::runtime_error(string_format("%s: expect 1 image in batch_f32", __func__));
+                            }
+
+                            LOG_DBG("%s: adding slice image at row %d col %d\n", __func__, y, x);
+                            cur.entries.emplace_back(std::move(curr_chunk));
+
                             add_text(ctx->tok_sli_img_end);
                             if (!is_last_in_row) {
                                 add_text(ctx->tok_sli_img_mid);
@@ -1149,9 +1173,14 @@ struct mtmd_tokenizer {
 
             } else {
 
+                if (preproc_out.entries.size() == 0) {
+                    LOG_ERR("%s: no image tokens produced by preprocessor (ref: https://github.com/ggml-org/llama.cpp/pull/24769)\n", __func__);
+                    return 2;
+                }
+
                 size_t n_tokens = 0;
-                for (const auto & e : batch_f32.entries) {
-                    n_tokens += clip_n_output_tokens(ctx->ctx_v, e.get());
+                for (auto & e : preproc_out.entries) {
+                    n_tokens += clip_n_output_tokens(ctx->ctx_v, &e);
                     if (clip_model_n_temporal_merge(ctx->ctx_v) == 2) {
                         // [QWEN_VIDEO] pair input is merged to the same embd, so only count as one image
                         break;
@@ -1165,8 +1194,8 @@ struct mtmd_tokenizer {
 
                 if (mtmd_decode_use_mrope(ctx)) {
                     // for Qwen2VL, we need this information for M-RoPE decoding positions
-                    image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_v, batch_f32.entries[0].get());
-                    image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_v, batch_f32.entries[0].get());
+                    image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_v, &preproc_out.entries[0]);
+                    image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_v, &preproc_out.entries[0]);
                 } else {
                     // other models, we only need the total number of tokens
                     image_tokens->nx = n_tokens;
@@ -1181,6 +1210,12 @@ struct mtmd_tokenizer {
                     image_tokens->image_idx = n_images_added;
                     GGML_ASSERT(n_tokens == (size_t)image_tokens->n_tokens());
                 }
+
+                clip_image_f32_batch batch_f32;
+                batch_f32.is_audio = false;
+                batch_f32.entries = std::move(preproc_out.entries);
+                // do NOT use preproc_out from this point on, it's moved
+
                 image_tokens->batch_f32 = std::move(batch_f32);
                 image_tokens->id = bitmaps[0]->id; // optional
 
@@ -1260,13 +1295,16 @@ struct mtmd_tokenizer {
             for (auto & mel_spec : mel_spec_chunks) {
                 const bool is_placeholder = mel_spec.data.empty();
 
-                clip_image_f32_ptr mel_f32(clip_image_f32_init());
-                mel_f32->set_size(
-                    {mel_spec.n_len, mel_spec.n_mel},
+                // Validate dimensions fit in clip_image_size (int)
+                GGML_ASSERT(mel_spec.n_len <= INT32_MAX && mel_spec.n_len >= 0);
+                GGML_ASSERT(mel_spec.n_mel <= INT32_MAX && mel_spec.n_mel >= 0);
+                clip_image_f32 mel_f32;
+                mel_f32.set_size(
+                    {(int)mel_spec.n_len, (int)mel_spec.n_mel},
                     is_placeholder, /* is_audio */ true);
-                mel_f32->cpy_buf(mel_spec.data);
+                mel_f32.cpy_buf(mel_spec.data);
 
-                size_t n_tokens = clip_n_output_tokens(ctx->ctx_a, mel_f32.get());
+                size_t n_tokens = clip_n_output_tokens(ctx->ctx_a, &mel_f32);
 
                 clip_image_f32_batch batch_f32;
                 batch_f32.is_audio = true;
@@ -1296,16 +1334,18 @@ struct mtmd_tokenizer {
         return 0;
     }
 
-    std::vector<mtmd_input_chunk> split_batch_to_chunk(clip_image_f32_batch && batch_f32, const std::string & id) {
+    std::vector<mtmd_input_chunk> split_batch_to_chunk(mtmd_image_preproc_out && preproc_out, const std::string & id) {
         std::vector<mtmd_input_chunk> chunks;
 
-        for (auto & entry : batch_f32.entries) {
+        auto process_chunk = [&](clip_image_f32 && img) {
             mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-            image_tokens->nx = clip_n_output_tokens(ctx->ctx_v, entry.get());
+            image_tokens->nx = clip_n_output_tokens(ctx->ctx_v, &img);
             image_tokens->ny = 1;
-            image_tokens->batch_f32.entries.push_back(std::move(entry));
+            image_tokens->batch_f32.entries.push_back(std::move(img));
             image_tokens->id = id;
 
+            GGML_ASSERT(image_tokens->nx > 0);
+
             mtmd_input_chunk chunk{
                 MTMD_INPUT_CHUNK_TYPE_IMAGE,
                 {}, // text tokens
@@ -1313,6 +1353,21 @@ struct mtmd_tokenizer {
                 nullptr, // audio tokens
             };
             chunks.emplace_back(std::move(chunk));
+        };
+
+        // overview image first
+        auto & overview = preproc_out.overview;
+        if (overview.nx() == 0 || overview.ny() == 0) {
+            throw std::runtime_error(string_format("%s: invalid overview image for llava-uhd style preprocessing\n", __func__));
+        }
+        process_chunk(std::move(preproc_out.overview));
+
+        // then, process slices
+        for (auto & entry : preproc_out.entries) {
+            if (entry.nx() == 0 || entry.ny() == 0) {
+                throw std::runtime_error(string_format("%s: invalid image slice for llava-uhd style preprocessing\n", __func__));
+            }
+            process_chunk(std::move(entry));
         }
 
         return chunks;
@@ -1386,57 +1441,22 @@ static int32_t mtmd_encode_impl(mtmd_context * ctx, const mtmd_image_tokens * im
         LOG_ERR("%s: this API does not support non-vision input, please use mtmd_encode_chunk instead\n", __func__);
         return 1;
     }
-    auto proj_type = clip_get_projector_type(ctx_clip);
 
     int n_embd_out = ctx->n_embd_out();
     auto n_tokens_out = image_tokens->n_tokens();
     out_embd.resize((size_t)n_embd_out * n_tokens_out);
 
-    bool ok = false;
-
-    if (clip_is_llava(ctx_clip)
-        || proj_type == PROJECTOR_TYPE_MINICPMV
-        || proj_type == PROJECTOR_TYPE_GLM_EDGE
-        || proj_type == PROJECTOR_TYPE_INTERNVL
-        || proj_type == PROJECTOR_TYPE_DEEPSEEKOCR2
-        || proj_type == PROJECTOR_TYPE_GRANITE4_VISION) {
-        // 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;
-        // entries may have different token counts
-        // e.g., DeepSeek-OCR-2: 144 per tile views, 257 for the global view
-        size_t offset = 0;
-        for (size_t i = 0; i < entries.size(); i++) {
-            if (entries[i]->is_placeholder()) {
-                LOG_ERR("%s: image tokens batch entry %zu is placeholder\n", __func__, i);
-                return 1;
-            }
-            int n_tokens_per_image = clip_n_output_tokens(ctx_clip, entries[i].get());
-            std::vector<float> tmp_embd((size_t)n_tokens_per_image * n_embd_out);
-            bool ok_i = clip_image_encode(
-                ctx_clip,
-                ctx->n_threads,
-                entries[i].get(),
-                tmp_embd);
-            if (!ok_i) {
-                LOG_ERR("%s: failed to encode image %zu\n", __func__, i);
-                return 1;
-            }
-            ok = true;
-            std::copy(tmp_embd.begin(), tmp_embd.end(), out_embd.begin() + offset);
-            offset += static_cast<size_t>(n_embd_out) * n_tokens_per_image;
-        }
-    } else {
-        if (image_tokens->is_placeholder()) {
-            LOG_ERR("%s: image tokens batch is placeholder\n", __func__);
-            return 1;
-        }
-        ok = clip_image_batch_encode(
-            ctx_clip,
-            ctx->n_threads,
-            &image_tokens->batch_f32,
-            out_embd);
+    if (image_tokens->is_placeholder()) {
+        LOG_ERR("%s: image tokens batch is placeholder\n", __func__);
+        return 1;
     }
 
+    bool ok = clip_image_batch_encode(
+        ctx_clip,
+        ctx->n_threads,
+        &image_tokens->batch_f32,
+        out_embd);
+
     return ok ? 0 : 1;
 }
 
@@ -2063,16 +2083,18 @@ void mtmd_debug_preprocess_image(mtmd_context * ctx, const std::vector<uint8_t>
     clip_image_u8 img_u8;
     img_u8.set_size({nx, ny}, false);
     img_u8.cpy_buf(rgb_values);
-    clip_image_f32_batch batch_f32;
     GGML_ASSERT(ctx->image_preproc != nullptr);
-    bool ok = ctx->image_preproc->preprocess(img_u8, batch_f32);
-    if (!ok) {
-        LOG_ERR("%s: failed to preprocess image\n", __func__);
-        return;
+    mtmd_image_preproc_out preproc_out = ctx->image_preproc->preprocess(img_u8);
+
+    clip_image_f32_batch batch_f32;
+    batch_f32.is_audio = false;
+    for (auto & entry : preproc_out.entries) {
+        batch_f32.entries.push_back(std::move(entry));
     }
+
     LOG_INF("%s: preprocessed image to batch_f32 with %d entries\n", __func__, (int)batch_f32.entries.size());
     for (size_t i = 0; i < batch_f32.entries.size(); i++) {
-        LOG_INF("%s: entry %zu has nx=%d, ny=%d\n", __func__, i, batch_f32.entries[i]->nx(), batch_f32.entries[i]->ny());
+        LOG_INF("%s: entry %zu has nx=%d, ny=%d\n", __func__, i, batch_f32.entries[i].nx(), batch_f32.entries[i].ny());
         // TODO: better way to dump entry content?
     }
 }

tools/mtmd/tests.sh

@@ -13,6 +13,8 @@ mkdir -p $SCRIPT_DIR/output
 PROJ_ROOT="$SCRIPT_DIR/../.."
 cd $PROJ_ROOT
 
+export MTMD_TEST_RESPONSE_MARKER="<MTMD_TEST_RESPONSE_MARKER>"
+
 # Check if the first argument is "big", then run test with big models
 # This is useful if we're running the script on a larger machine, so we can test the big models
 RUN_BIG_TESTS=false
@@ -28,6 +30,15 @@ if [ "${1:-}" = "huge" ]; then
     echo "Include BIG and HUGE models..."
 fi
 
+USE_VIDEO=false
+if [ "${1:-}" = "video" ]; then
+    USE_VIDEO=true
+    echo "Using video as input..."
+    # behavior of USE_VIDEO:
+    # do NOT check if the output contains "new york", only verify if the exit code is 0
+    # when printing the result, print the OK/FAIL line then print the generated text
+fi
+
 # Check if the second argument is "flash", then enable flash attention
 # This is useful to test if flash attention off works correctly
 FLASH_ATTN="on"
@@ -50,13 +61,20 @@ add_test_vision() {
     if [ $# -gt 0 ]; then
         extra_args=$(printf " %q" "$@")
     fi
+    if [ "$USE_VIDEO" = true ]; then
+        arr_file+=("test-3.mp4")
+    else
+        arr_file+=("test-1.jpeg")
+    fi
     arr_prefix+=("[vision]")
     arr_hf+=("$hf")
     arr_extra_args+=("$extra_args")
-    arr_file+=("test-1.jpeg")
 }
 
 add_test_audio() {
+    if [ "$USE_VIDEO" = true ]; then
+        return 0
+    fi
     local hf=$1
     shift
     local extra_args=""
@@ -166,19 +184,35 @@ for i in "${!arr_hf[@]}"; do
         cmd+=" -p \"what is the publisher name of the newspaper?\""
     fi
 
-    output=$(eval "$cmd" 2>&1 | tee /dev/tty)
+    exit_code=0
+    output=$(eval "$cmd" 2>&1 | tee /dev/tty) || exit_code=$?
 
     echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log
 
-    # either contains "new york" or both "men" and "walk"
-    if echo "$output" | grep -iq "new york" \
-            || (echo "$output" | grep -iq "men" && echo "$output" | grep -iq "walk")
-    then
-        result="$prefix \033[32mOK\033[0m:   $hf"
+    if [ "$USE_VIDEO" = true ]; then
+        # for video, only check exit code; do not grep for "new york"
+        if [ $exit_code -eq 0 ]; then
+            result="$prefix \033[32mOK\033[0m:   $hf"
+        else
+            result="$prefix \033[31mFAIL\033[0m: $hf"
+        fi
+        # append generated text (after the response marker)
+        generated_text=$(echo "$output" | sed "1,/${MTMD_TEST_RESPONSE_MARKER}/d" | tail -10)
+        if [ -n "$generated_text" ]; then
+            result+="\n$generated_text"
+        fi
+        echo -e "$result"
     else
-        result="$prefix \033[31mFAIL\033[0m: $hf"
+        # either contains "new york" or both "men" and "walk"
+        if echo "$output" | grep -iq "new york" \
+                || (echo "$output" | grep -iq "men" && echo "$output" | grep -iq "walk")
+        then
+            result="$prefix \033[32mOK\033[0m:   $hf"
+        else
+            result="$prefix \033[31mFAIL\033[0m: $hf"
+        fi
+        echo -e "$result"
     fi
-    echo -e "$result"
     arr_res+=("$result")
 
     echo ""

tools/server/CMakeLists.txt

@@ -17,6 +17,8 @@ add_library(${TARGET} STATIC
     server-context.h
     server-tools.cpp
     server-tools.h
+    server-schema.cpp
+    server-schema.h
 )
 
 if (BUILD_SHARED_LIBS)

tools/server/README-dev.md

@@ -180,6 +180,17 @@ That requires `JSON.stringify` when formatted to message content:
 }
 ```
 
+### Router mode: how child <--> router communicates
+
+Upon spawning a new child process using `subprocess`, both child and router listen to the stdout/stderr (combined)
+
+For the direction from child to router:
+- Generic messages are logs, it will be forwarded to router's stdout
+- Special state update messages are prefixed by `cmd_child_to_router:state:`, followed by a JSON. See `server_models::handle_child_state` for more
+
+For the direction from router to child:
+- When server sends `cmd_router_to_child:exit`, the child should exit gracefully --> if after `DEFAULT_STOP_TIMEOUT` and the child is still running, force-kill it
+
 ### Model management API (router mode)
 
 Model management API was added via PR [#23976](https://github.com/ggml-org/llama.cpp/pull/23976)

tools/server/README.md

@@ -175,13 +175,12 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `-np, --parallel N` | number of server slots (default: -1, -1 = auto)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `-cb, --cont-batching, -nocb, --no-cont-batching` | whether to enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
 | `-mm, --mmproj FILE` | path to a multimodal projector file. see tools/mtmd/README.md<br/>note: if -hf is used, this argument can be omitted<br/>(env: LLAMA_ARG_MMPROJ) |
-| `-tk, --talker-model FILE` | path to the qwen3-omni talker gguf, enables the /v1/audio/speech endpoint<br/>(env: LLAMA_ARG_TALKER_MODEL) |
-| `-c2w, --code2wav-model FILE` | path to the qwen3-omni code2wav gguf, the talker code detokenizer<br/>(env: LLAMA_ARG_CODE2WAV_MODEL) |
 | `-mmu, --mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) |
 | `--mmproj-auto, --no-mmproj, --no-mmproj-auto` | whether to use multimodal projector file (if available), useful when using -hf (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_AUTO) |
 | `--mmproj-offload, --no-mmproj-offload` | whether to enable GPU offloading for multimodal projector (default: enabled)<br/>(env: LLAMA_ARG_MMPROJ_OFFLOAD) |
 | `--image-min-tokens N` | minimum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MIN_TOKENS) |
 | `--image-max-tokens N` | maximum number of tokens each image can take, only used by vision models with dynamic resolution (default: read from model)<br/>(env: LLAMA_ARG_IMAGE_MAX_TOKENS) |
+| `--mtmd-batch-max-tokens N` | maximum number of image tokens per batch when encoding images (default: 1024)<br/>(env: LLAMA_ARG_MTMD_BATCH_MAX_TOKENS) |
 | `-a, --alias STRING` | set model name aliases, comma-separated (to be used by API)<br/>(env: LLAMA_ARG_ALIAS) |
 | `--tags STRING` | set model tags, comma-separated (informational, not used for routing)<br/>(env: LLAMA_ARG_TAGS) |
 | `--embd-normalize N` | normalisation for embeddings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm) |
@@ -190,23 +189,21 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--reuse-port` | allow multiple sockets to bind to the same port (default: disabled)<br/>(env: LLAMA_ARG_REUSE_PORT) |
 | `--path PATH` | path to serve static files from (default: )<br/>(env: LLAMA_ARG_STATIC_PATH) |
 | `--api-prefix PREFIX` | prefix path the server serves from, without the trailing slash (default: )<br/>(env: LLAMA_ARG_API_PREFIX) |
-| `--webui-config JSON` | [DEPRECATED: use --ui-config] JSON that provides default WebUI settings (overrides WebUI defaults)<br/>(env: LLAMA_ARG_WEBUI_CONFIG) |
-| `--ui-config JSON` | JSON that provides default UI settings (overrides UI defaults)<br/>(env: LLAMA_ARG_UI_CONFIG) |
-| `--webui-config-file PATH` | [DEPRECATED: use --ui-config-file] JSON file that provides default WebUI settings (overrides WebUI defaults)<br/>(env: LLAMA_ARG_WEBUI_CONFIG_FILE) |
-| `--ui-config-file PATH` | JSON file that provides default UI settings (overrides UI defaults)<br/>(env: LLAMA_ARG_UI_CONFIG_FILE) |
-| `--webui-mcp-proxy, --no-webui-mcp-proxy` | [DEPRECATED: use --ui-mcp-proxy/--no-ui-mcp-proxy] experimental: whether to enable MCP CORS proxy<br/>(env: LLAMA_ARG_WEBUI_MCP_PROXY) |
-| `--ui-mcp-proxy, --no-ui-mcp-proxy` | experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: disabled)<br/>(env: LLAMA_ARG_UI_MCP_PROXY) |
+| `--ui-config, --webui-config JSON` | JSON that provides default UI settings (overrides UI defaults)<br/>(env: LLAMA_ARG_UI_CONFIG) |
+| `--ui-config-file, --webui-config-file PATH` | JSON file that provides default UI settings (overrides UI defaults)<br/>(env: LLAMA_ARG_UI_CONFIG_FILE) |
+| `--ui-mcp-proxy, --webui-mcp-proxy, --no-ui-mcp-proxy, --no-webui-mcp-proxy` | experimental: whether to enable MCP CORS proxy - do not enable in untrusted environments (default: disabled)<br/>(env: LLAMA_ARG_UI_MCP_PROXY) |
 | `--tools TOOL1,TOOL2,...` | experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)<br/>specify "all" to enable all tools<br/>available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, apply_diff, get_datetime<br/>(env: LLAMA_ARG_TOOLS) |
-| `--webui, --no-webui` | [DEPRECATED: use --ui/--no-ui] whether to enable the Web UI<br/>(env: LLAMA_ARG_WEBUI) |
-| `--ui, --no-ui` | whether to enable the Web UI (default: enabled)<br/>(env: LLAMA_ARG_UI) |
+| `-ag, --agent, -no-ag, --no-agent` | whether to enable CORS proxy and all built-in tools - do not enable in untrusted environments (default: disabled)<br/>(env: LLAMA_ARG_AGENT) |
+| `--ui, --webui, --no-ui, --no-webui` | whether to enable the Web UI (default: enabled)<br/>(env: LLAMA_ARG_UI) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
 | `--rerank, --reranking` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
 | `--api-key KEY` | API key to use for authentication, multiple keys can be provided as a comma-separated list (default: none)<br/>(env: LLAMA_API_KEY) |
-| `--api-key-file FNAME` | path to file containing API keys (default: none)<br/>(env: LLAMA_ARG_API_KEY_FILE) |
+| `--api-key-file FNAME` | path to file containing API keys, one per line; lines starting with a hash are treated as comments (default: none)<br/>(env: LLAMA_ARG_API_KEY_FILE) |
 | `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key<br/>(env: LLAMA_ARG_SSL_KEY_FILE) |
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
 | `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_KWARGS) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 3600)<br/>(env: LLAMA_ARG_TIMEOUT) |
+| `--sse-ping-interval N` | server SSE ping interval in seconds (-1 = disabled, default: 30)<br/>(env: LLAMA_ARG_SSE_PING_INTERVAL) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
 | `--cache-prompt, --no-cache-prompt` | whether to enable prompt caching (default: enabled)<br/>(env: LLAMA_ARG_CACHE_PROMPT) |
 | `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting, requires prompt caching to be enabled (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
@@ -231,6 +228,7 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled) |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
 | `--sleep-idle-seconds SECONDS` | number of seconds of idleness after which the server will sleep (default: -1; -1 = disabled) |
+| `--log-prompts-dir PATH` | Log prompts to directory (only used for debugging, default: disabled) |
 | `--spec-draft-hf, -hfd, -hfrd, --hf-repo-draft <user>/<model>[:quant]` | Same as --hf-repo, but for the draft model (default: unused)<br/>(env: LLAMA_ARG_SPEC_DRAFT_HF_REPO) |
 | `--spec-draft-threads, -td, --threads-draft N` | number of threads to use during generation (default: same as --threads) |
 | `--spec-draft-threads-batch, -tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) |

tools/server/server-common.cpp

@@ -12,6 +12,7 @@
 #include <random>
 #include <sstream>
 #include <fstream>
+#include <limits>
 
 json format_error_response(const std::string & message, const enum error_type type) {
     std::string type_str;
@@ -1238,7 +1239,7 @@ json format_response_rerank(
 // other utils
 //
 
-std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx) {
+std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx, size_t n_top) {
     std::vector<llama_token_data> cur;
 
     const auto * logits = llama_get_logits_ith(ctx, idx);
@@ -1257,21 +1258,34 @@ std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int i
         }
     }
 
-    // sort tokens by logits
-    std::sort(cur.begin(), cur.end(), [](const llama_token_data & a, const llama_token_data & b) {
-        return a.logit > b.logit;
-    });
+    // sort tokens by logits (partial: only the leading `n_top` need ordering)
+    if (n_top > cur.size()) {
+        n_top = cur.size();
+    }
+    if (n_top > 0) {
+        std::partial_sort(cur.begin(), cur.begin() + n_top, cur.end(),
+            [](const llama_token_data & a, const llama_token_data & b) {
+                return a.logit > b.logit;
+            });
+    }
 
     // apply softmax
-    float max_l = cur[0].logit;
+    float max_l = -std::numeric_limits<float>::infinity();
+    if (n_top > 0) {
+        max_l = cur[0].logit; // partial_sort guarantees the absolute maximum is at index 0
+    } else {
+        for (const auto & t : cur) {
+            max_l = std::max(max_l, t.logit);
+        }
+    }
     float cum_sum = 0.0f;
-    for (size_t i = 0; i < cur.size(); ++i) {
-        float p = expf(cur[i].logit - max_l);
-        cur[i].p = p;
+    for (auto & t : cur) {
+        float p = expf(t.logit - max_l);
+        t.p = p;
         cum_sum += p;
     }
-    for (size_t i = 0; i < cur.size(); ++i) {
-        cur[i].p /= cum_sum;
+    for (auto & t : cur) {
+        t.p /= cum_sum;
     }
 
     return cur;

tools/server/server-common.h

@@ -326,7 +326,7 @@ json format_response_rerank(
 // other utils
 //
 
-std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx);
+std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx, size_t n_top);
 
 std::string safe_json_to_str(const json & data);
 

tools/server/server-context.cpp

@@ -4,6 +4,7 @@
 #include "server-http.h"
 #include "server-task.h"
 #include "server-queue.h"
+#include "server-schema.h"
 
 #include "build-info.h"
 #include "common.h"
@@ -62,11 +63,6 @@ enum slot_state {
     SLOT_STATE_GENERATING,
 };
 
-enum server_state {
-    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
-    SERVER_STATE_READY,          // Server is ready and model is loaded
-};
-
 struct server_slot {
     int id;
 
@@ -189,9 +185,10 @@ struct server_slot {
     // stats
     size_t n_sent_text = 0; // number of sent text character
 
-    int64_t t_print_last = 0;
     int64_t t_start_process_prompt;
     int64_t t_start_generation;
+    int64_t t_print_last = 0;
+    int32_t n_decoded_last = 0;
 
     double t_prompt_processing = 0.0; // ms
     double t_token_generation = 0.0;  // ms
@@ -470,11 +467,13 @@ struct server_slot {
             return;
         }
 
+        const double n_gen_second     = 1e3 / (t_token_generation)   * (n_decoded);
+        const double n_gen_second_win = 1e6 / (t_now - t_print_last) * (n_decoded - n_decoded_last);
+
         t_print_last = t_now;
+        n_decoded_last = n_decoded;
 
-        const double n_gen_second = 1e3 / t_token_generation * n_decoded;
-
-        SLT_INF(*this, "n_decoded = %6d, tg = %6.2f t/s\n", n_decoded, n_gen_second);
+        SLT_INF(*this, "n_decoded = %6d, tg = %6.2f t/s, tg_3s = %6.2f t/s\n", n_decoded, n_gen_second, n_gen_second_win);
     }
 
     void print_timings_pp() const {
@@ -769,6 +768,8 @@ public:
     // note: chat_params must not be refreshed upon existing sleeping state
     server_chat_params chat_params;
 
+    server_state_callback_t callback_state = [](server_state, json) -> void {};
+
     server_context_impl() {
         mtmd_helper_log_set(common_log_default_callback, nullptr);
     }
@@ -821,8 +822,7 @@ private:
 
     server_metrics metrics;
 
-    json json_ui_settings = json::object();    // Primary: new name
-    json json_webui_settings = json::object();    // Deprecated: use json_ui_settings instead (kept for compat)
+    json json_ui_settings = json::object();
 
     // Necessary similarity of prompt for slot selection
     float slot_prompt_similarity = 0.0f;
@@ -1241,8 +1241,8 @@ private:
         if (!params_base.model_alias.empty()) {
             // backward compat: use first alias as model name
             model_name = *params_base.model_alias.begin();
-        } else if (!params_base.model.name.empty()) {
-            model_name = params_base.model.name;
+        } else if (!params_base.model.get_name().empty()) {
+            model_name = params_base.model.get_name();
         } else {
             // fallback: derive model name from file name
             auto model_path = std::filesystem::path(params_base.model.path);
@@ -1298,16 +1298,12 @@ private:
             }
         }
 
-        // populate UI settings (from either new ui_config_json or deprecated webui_config_json)
         {
-            const std::string & cfg = !params_base.ui_config_json.empty()
-                ? params_base.ui_config_json
-                : params_base.webui_config_json;
+            const std::string & cfg = params_base.ui_config_json;
             if (!cfg.empty()) {
                 try {
                     json json_settings = json::parse(cfg);
                     json_ui_settings = json_settings;
-                    json_webui_settings = json_settings; // deprecated: keep in sync
                 } catch (const std::exception & e) {
                     SRV_ERR("%s: failed to parse UI config: %s\n", __func__, e.what());
                     return false;
@@ -1391,11 +1387,23 @@ private:
 
         bool update_cache = false;
 
+        // if a specific slot is requested, use it (still goes through cache update logic below)
+        if (task.id_slot != -1) {
+            ret = get_slot_by_id(task.id_slot);
+            if (ret) {
+                SLT_INF(*ret, "selected slot by id (%d)\n", task.id_slot);
+            }
+        }
+
         // find the slot that has at least n% prompt similarity
-        if (ret == nullptr && slot_prompt_similarity != 0.0f) {
+        if (slot_prompt_similarity != 0.0f) {
             float sim_best = 0;
 
             for (server_slot & slot : slots) {
+                if (task.id_slot != -1 && slot.id != task.id_slot) {
+                    continue;
+                }
+
                 // skip the slot if it is not available
                 if (slot.is_processing()) {
                     continue;
@@ -1422,8 +1430,10 @@ private:
             if (ret != nullptr) {
                 const float f_keep = (sim_best*task.tokens.size()) / ret->prompt.tokens.size();
 
-                SLT_INF(*ret, "selected slot by LCP similarity, sim_best = %.3f (> %.3f thold), f_keep = %.3f\n",
-                        sim_best, slot_prompt_similarity, f_keep);
+                if (task.id_slot == -1) {
+                    SLT_INF(*ret, "selected slot by LCP similarity, sim_best = %.3f (> %.3f thold), f_keep = %.3f\n",
+                            sim_best, slot_prompt_similarity, f_keep);
+                }
 
                 // if we are about to lose a large portion of the existing context - save it in the prompt cache
                 if (f_keep < 0.5f) {
@@ -1811,8 +1821,7 @@ private:
                 });
             }
         } else {
-            // TODO: optimize this with min-p optimization
-            std::vector<llama_token_data> cur = get_token_probabilities(ctx_tgt, idx);
+            std::vector<llama_token_data> cur = get_token_probabilities(ctx_tgt, idx, n_probs_request);
             const size_t max_probs = cur.size();
             const size_t n_probs = std::min(max_probs, n_probs_request);
 
@@ -2154,6 +2163,8 @@ private:
 
         cur.update_tgt(ctx_tgt,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
         cur.update_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        // stash the draft's speculative state with the checkpoint
+        common_speculative_get_state(spec.get(), slot.id, cur.data_spec);
 
         SLT_INF(slot,
                 "created context checkpoint %d of %d (pos_min = %d, pos_max = %d, n_tokens = %" PRId64 ", size = %.3f MiB)\n",
@@ -2176,10 +2187,9 @@ private:
                         }
                     }
 
-                    const int id_slot = task.id_slot;
                     const int id_task = task.id;
 
-                    server_slot * slot = id_slot != -1 ? get_slot_by_id(id_slot) : get_available_slot(task);
+                    server_slot * slot = get_available_slot(task);
 
                     //
                     // slot scheduling logic
@@ -2548,7 +2558,10 @@ private:
                 n_keep = std::min(slot.n_ctx - 4, n_keep);
 
                 const int n_left    = slot.prompt.n_tokens() - n_keep;
-                const int n_discard = slot.task->params.n_discard ? slot.task->params.n_discard : (n_left / 2);
+                int       n_discard = slot.task->params.n_discard ? slot.task->params.n_discard : (n_left / 2);
+
+                // ref: https://github.com/ggml-org/llama.cpp/pull/24786
+                n_discard = std::clamp(n_discard, 0, std::max(0, n_left - 1));
 
                 SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard);
 
@@ -2978,6 +2991,8 @@ private:
                                         // restore the context checkpoint
                                         it->load_tgt(ctx_tgt,       slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
                                         it->load_dft(ctx_dft.get(), slot.id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                                        // restore the draft's speculative state
+                                        common_speculative_set_state(spec.get(), slot.id, it->data_spec);
 
                                         pos_next = std::min(pos_next, std::max(it->pos_min + 1, it->pos_max));
                                         n_past   = std::min(slot.prompt.tokens.size_up_to_pos(pos_next), (size_t) it->n_tokens);
@@ -3038,8 +3053,8 @@ private:
                         }
                     }
 
-                    const int64_t t_current = ggml_time_us();
-                    slot.t_prompt_processing = (t_current - slot.t_start_process_prompt) / 1e3;
+                    const int64_t t_now = ggml_time_us();
+                    slot.t_prompt_processing = (t_now - slot.t_start_process_prompt) / 1e3;
                     slot.print_timings_pp();
 
                     // truncate any tokens that are beyond n_past for this slot
@@ -3447,17 +3462,19 @@ private:
                 common_sampler_accept(slot.smpl.get(), id, true);
 
                 // here we have synchronized the llama_context (due to the sampling above), so we can do time measurement
-                const int64_t t_current = ggml_time_us();
+                const int64_t t_now = ggml_time_us();
 
                 slot.n_decoded += 1;
 
                 if (slot.n_decoded == 1) {
-                    slot.t_start_generation = t_current;
+                    slot.t_start_generation = t_now;
+                    slot.t_print_last = t_now;
+                    slot.n_decoded_last = 0;
                     slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
                     metrics.on_prompt_eval(slot);
                 }
 
-                slot.t_token_generation = std::max<int64_t>(1, t_current - slot.t_start_generation) / 1e3;
+                slot.t_token_generation = std::max<int64_t>(1, t_now - slot.t_start_generation) / 1e3;
 
                 completion_token_output result;
                 result.tok          = id;
@@ -3551,11 +3568,11 @@ private:
                     slot.spec_draft = std::move(accepted);
                 }
 
-                const int64_t t_current = ggml_time_us();
+                const int64_t t_now = ggml_time_us();
 
                 const auto ids = std::move(slot.spec_draft);
 
-                slot.t_token_generation = std::max<int64_t>(1, t_current - slot.t_start_generation) / 1e3;
+                slot.t_token_generation = std::max<int64_t>(1, t_now - slot.t_start_generation) / 1e3;
 
                 // update how many tokens out of those tested were accepted
                 slot.n_draft_accepted += ids.size() - 1;
@@ -3664,7 +3681,6 @@ server_context_meta server_context::get_meta() const {
         /* has_inp_audio          */ impl->chat_params.allow_audio,
         /* has_inp_video          */ impl->chat_params.allow_video,
         /* json_ui_settings       */ impl->json_ui_settings,
-        /* json_webui_settings    */ impl->json_webui_settings,  // Deprecated
         /* slot_n_ctx             */ impl->get_slot_n_ctx(),
         /* pooling_type           */ llama_pooling_type(impl->ctx_tgt),
 
@@ -3715,8 +3731,11 @@ struct server_res_generator : server_http_res {
     }
 };
 
-void server_context::on_sleeping_changed(std::function<void(bool)> callback) {
-    impl->queue_tasks.on_sleeping_state(std::move(callback));
+void server_context::set_state_callback(server_state_callback_t callback) {
+    impl->callback_state = std::move(callback);
+    impl->queue_tasks.on_sleeping_state([this](bool sleeping) {
+        impl->callback_state(sleeping ? SERVER_STATE_SLEEPING : SERVER_STATE_READY, {});
+    });
 }
 
 // compute the number of tokens before the last user message in the prompt
@@ -3820,7 +3839,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
             task.id = rd.get_new_id();
 
             task.tokens = std::move(inputs[i]);
-            task.params = server_task::params_from_json_cmpl(
+            task.params = server_schema::eval_llama_cmpl_schema(
                     ctx_server.vocab,
                     params,
                     meta->slot_n_ctx,
@@ -4277,19 +4296,15 @@ void server_routes::init_routes() {
             { "endpoint_slots",              params.endpoint_slots },
             { "endpoint_props",              params.endpoint_props },
             { "endpoint_metrics",            params.endpoint_metrics },
-            // New keys
-            { "ui",                           params.ui },
-            { "ui_settings",                  meta->json_ui_settings },
-            // Deprecated: use ui/ui_settings instead (kept for backward compat)
-            { "webui",                        params.webui },
-            { "webui_settings",               meta->json_webui_settings },
+            { "ui",                          params.ui },
+            { "ui_settings",                 meta->json_ui_settings },
             { "chat_template",               tmpl_default },
             { "chat_template_caps",          meta->chat_template_caps },
             { "bos_token",                   meta->bos_token_str },
             { "eos_token",                   meta->eos_token_str },
             { "build_info",                  meta->build_info },
             { "is_sleeping",                 queue_tasks.is_sleeping() },
-            { "cors_proxy_enabled",          params.ui_mcp_proxy || params.webui_mcp_proxy },
+            { "cors_proxy_enabled",          params.ui_mcp_proxy },
         };
         if (params.use_jinja) {
             if (!tmpl_tools.empty()) {

tools/server/server-context.h

@@ -22,8 +22,7 @@ struct server_context_meta {
     bool has_inp_image;
     bool has_inp_audio;
     bool has_inp_video;
-    json json_ui_settings;            // Primary: new name
-    json json_webui_settings;            // Deprecated: use json_ui_settings instead (kept for backward compat)
+    json json_ui_settings;
     int slot_n_ctx;
     enum llama_pooling_type pooling_type;
 
@@ -53,6 +52,31 @@ struct server_context_meta {
     uint64_t model_size;
 };
 
+enum server_state {
+    // SERVER_STATE_DOWNLOADING,
+    SERVER_STATE_LOADING,
+    SERVER_STATE_READY,
+    SERVER_STATE_SLEEPING,
+};
+
+static std::string server_state_to_str(server_state state) {
+    switch (state) {
+        case SERVER_STATE_LOADING:     return "loading";
+        case SERVER_STATE_READY:       return "ready";
+        case SERVER_STATE_SLEEPING:    return "sleeping";
+        default: GGML_ASSERT(false && "invalid server_state");
+    }
+}
+
+static server_state server_state_from_str(const std::string & str) {
+    if (str == "loading")     return SERVER_STATE_LOADING;
+    if (str == "ready")       return SERVER_STATE_READY;
+    if (str == "sleeping")    return SERVER_STATE_SLEEPING;
+    GGML_ASSERT(false && "invalid server_state string");
+}
+
+using server_state_callback_t = std::function<void(server_state, json /* payload */)>;
+
 struct server_context {
     std::unique_ptr<server_context_impl> impl;
 
@@ -80,9 +104,8 @@ struct server_context {
     // not thread-safe, should only be used from the main thread
     server_context_meta get_meta() const;
 
-    // register a callback to be called when sleeping state changes
-    // must be set before load_model() is called
-    void on_sleeping_changed(std::function<void(bool)> callback);
+    // note: must be set before load_model() is called
+    void set_state_callback(server_state_callback_t callback);
 };
 
 

tools/server/server-cors-proxy.h

@@ -7,9 +7,18 @@
 #include <unordered_set>
 #include <list>
 #include <map>
+#include <algorithm>
+#include <cctype>
 
 #include "server-http.h"
 
+static std::string proxy_header_to_lower(std::string header) {
+    std::transform(header.begin(), header.end(), header.begin(), [](unsigned char c) {
+        return std::tolower(c);
+    });
+    return header;
+}
+
 static server_http_res_ptr proxy_request(const server_http_req & req, std::string method) {
     std::string target_url = req.get_param("url");
     common_http_url parsed_url = common_http_parse_url(target_url);
@@ -33,11 +42,18 @@ static server_http_res_ptr proxy_request(const server_http_req & req, std::strin
     SRV_INF("proxying %s request to %s://%s:%i%s\n", method.c_str(), parsed_url.scheme.c_str(), parsed_url.host.c_str(), parsed_url.port, parsed_url.path.c_str());
 
     std::map<std::string, std::string> headers;
+    const std::string proxy_header_prefix = "x-llama-server-proxy-header-";
     for (auto [key, value] : req.headers) {
-        auto new_key = key;
-        if (string_starts_with(new_key, "x-proxy-header-")) {
-            string_replace_all(new_key, "x-proxy-header-", "");
+        const std::string lowered_key = proxy_header_to_lower(key);
+        if (!string_starts_with(lowered_key, proxy_header_prefix)) {
+            continue;
         }
+
+        auto new_key = key.substr(proxy_header_prefix.size());
+        if (new_key.empty()) {
+            continue;
+        }
+
         headers[new_key] = value;
     }
 

tools/server/server-http.cpp

@@ -492,6 +492,8 @@ using server_http_req_ptr = std::unique_ptr<server_http_req>;
 static void process_handler_response(server_http_req_ptr && request, server_http_res_ptr & response, httplib::Response & res) {
     if (response->is_stream()) {
         res.status = response->status;
+        // Tell Nginx to not buffer any streamed response
+        response->headers["X-Accel-Buffering"] = "no";
         set_headers(res, response->headers);
         const std::string content_type = response->content_type;
         // convert to shared_ptr as both chunked_content_provider() and on_complete() need to use it