feat: sync multnpu backend/runtime changes into feat/npu

.gitignore

@@ -33,7 +33,7 @@
 /.vs/
 /.vscode/
 /nppBackup
-
+CMakeUserPresets.json
 
 # Coverage
 
@@ -140,3 +140,5 @@ poetry.toml
 /.windsurf/
 # emscripten
 a.out.*
+
+pkg-adb/

CMakeLists.txt

@@ -109,6 +109,7 @@ option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_DEFAULT})
+option(IGNITE_USE_SYSTEM_DVFS "Use system DVFS library" ON)
 
 # 3rd party libs
 option(LLAMA_HTTPLIB    "llama: httplib for downloading functionality" ON)
@@ -248,6 +249,9 @@ set_target_properties(llama
         PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
 
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
+# llama-ignite-npu links against dvfs at runtime, so package installs need to
+# ship the library alongside the rest of the shared objects.
+install(TARGETS dvfs LIBRARY RUNTIME ARCHIVE)
 
 configure_package_config_file(
         ${CMAKE_CURRENT_SOURCE_DIR}/cmake/llama-config.cmake.in

common/arg.cpp

@@ -1039,7 +1039,91 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.max_query_number = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+
+    // ----------------------------------------------------------------------------------------
+    // 20260406 IGNITE DVFS
+    // ----------------------------------------------------------------------------------------
+    add_opt(common_arg(
+        {"--device-name", "--dvfs-device"}, "DN",
+        "DVFS target device name (e.g. S25, S24, Pixel9)",
+        [](common_params & params, const std::string & value) {
+            params.device_name = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--cpu-p"}, "IDX",
+        "prefill CPU DVFS index",
+        [](common_params & params, int value) {
+            params.cpu_clk_idx_p = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--ram-p"}, "IDX",
+        "prefill RAM DVFS index",
+        [](common_params & params, int value) {
+            params.ram_clk_idx_p = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--cpu-d"}, "IDX",
+        "decode CPU DVFS index",
+        [](common_params & params, int value) {
+            params.cpu_clk_idx_d = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--ram-d"}, "IDX",
+        "decode RAM DVFS index",
+        [](common_params & params, int value) {
+            params.ram_clk_idx_d = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--phase-pause"}, "MS",
+        "pause time between prefill and decode phases in milliseconds",
+        [](common_params & params, int value) {
+            params.phase_pause = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--token-pause"}, "MS",
+        "pause time between generated decode tokens in milliseconds",
+        [](common_params & params, int value) {
+            params.token_pause = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--layer-pause"}, "MS",
+        "pause time for layer-wise experiments in milliseconds",
+        [](common_params & params, int value) {
+            params.layer_pause = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--ignite-verbose"},
+        "enable verbose ignite markers for layer pause debugging",
+        [](common_params & params) {
+            params.ignite_verbose = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--backend-compute-profile"},
+        "enable backend scheduler compute profiling in ignite CSV output",
+        [](common_params & params) {
+            params.backend_compute_profile = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+    add_opt(common_arg(
+        {"--backend-op-breakdown"},
+        "append per-op backend scheduler counters to ignite CSV output",
+        [](common_params & params) {
+            params.backend_compute_profile = true;
+            params.backend_op_breakdown = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
+
+    // ----------------------------------------------------------------------------------------
     add_opt(common_arg(
         {"--strict"}, "ST",
         "enable strict mode",

common/common.cpp

@@ -374,6 +374,46 @@ void common_init() {
     LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
 }
 
+void common_ignite_init(llama_context * ctx, common_params & params) {
+    if (!ctx) {
+        return;
+    }
+
+    llama_igparams ig{};
+
+    llama_ignite_set_active(ctx, params.is_ignite_active);
+    llama_ignite_set_layer_pause(ctx, params.layer_pause);
+
+    ig.max_query_number    = params.max_query_number;
+    ig.strict_limit        = params.strict_limit;
+    ig.strict_limit_length = params.strict_limit_length;
+    ig.enable_thinking     = params.enable_thinking;
+    ig.layer_pause         = params.layer_pause;
+    ig.phase_pause         = params.phase_pause;
+    ig.token_pause         = params.token_pause;
+    ig.query_interval      = params.query_interval;
+    ig.prefill_phase       = params.prefill_phase;
+    ig.prefill_speed       = params.prefill_speed;
+    ig.decode_speed        = params.decode_speed;
+    ig.backend_compute_profile = params.backend_compute_profile;
+    ig.backend_op_breakdown    = params.backend_op_breakdown;
+
+    std::strcpy(ig.input_path, params.input_path.c_str());
+    std::strcpy(ig.output_dir, params.output_dir.c_str());
+    std::strcpy(ig.output_path_hard, params.output_path_hard.c_str());
+    std::strcpy(ig.output_path_infer, params.output_path_infer.c_str());
+
+    std::strcpy(ig.device_name, params.device_name.c_str());
+    ig.is_ignite_active = params.is_ignite_active;
+    ig.ignite_verbose   = params.ignite_verbose;
+    ig.cpu_clk_idx_p    = params.cpu_clk_idx_p;
+    ig.ram_clk_idx_p    = params.ram_clk_idx_p;
+    ig.cpu_clk_idx_d    = params.cpu_clk_idx_d;
+    ig.ram_clk_idx_d    = params.ram_clk_idx_d;
+
+    init_ignite_params(ctx, &ig);
+}
+
 std::string common_params_get_system_info(const common_params & params) {
     std::ostringstream os;
 
@@ -1225,6 +1265,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
         return res;
     }
+    common_ignite_init(lctx, params);
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 

common/common.h

@@ -621,15 +621,28 @@ struct common_params {
     bool enable_thinking = false;
 
     // llm plane
-    // int phase_pause = 0; // ms
-    // int token_pause = 0; // ms
-    // int layer_pause = 0; // ms
-    // int query_interval = 0; // ms
-    // bool prefill_phase = true; // prefill phase or not
-    // double prefill_speed = 0.0; // tokens/s
-    // double decode_speed = 0.0; // tokens/s
-    // bool is_ignite_active = false;
-    // bool ignite_verbose = false;
+    std::string device_name = "S25";
+    int cpu_clk_idx_p = -1;
+    int ram_clk_idx_p = -1;
+    int cpu_clk_idx_d = -1;
+    int ram_clk_idx_d = -1;
+
+    int phase_pause = 0; // ms
+    int token_pause = 0; // ms
+    int layer_pause = 0; // ms
+    bool backend_compute_profile = false;
+    bool backend_op_breakdown = false;
+    int query_interval = 0; // ms
+    bool prefill_phase = true; // prefill phase or not
+    double prefill_speed = 0.0; // tokens/s
+    double decode_speed = 0.0; // tokens/s
+#if defined (IGNITE_USE_SYSTEM_DVFS)
+    bool is_ignite_active = true;
+    bool ignite_verbose = false;
+#else
+    bool is_ignite_active = false;
+    bool ignite_verbose = false;
+#endif
 
     // basic measure configs
     int max_query_number = -1;       // limit of CSV questions (0=no limit) // deprecated in future
@@ -643,6 +656,7 @@ struct common_params {
 // call once at the start of a program if it uses libcommon
 // initializes the logging system and prints info about the build
 void common_init();
+void common_ignite_init(llama_context * ctx, common_params & params);
 
 std::string common_params_get_system_info(const common_params & params);
 

docs/backend/hexagon/CMakeUserPresets.json

@@ -0,0 +1,51 @@
+{
+  "version": 4,
+  "configurePresets": [
+    {
+        "name": "arm64-android-snapdragon",
+        "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "ANDROID_ABI":      "arm64-v8a",
+            "ANDROID_PLATFORM": "android-31",
+            "CMAKE_TOOLCHAIN_FILE": "$env{ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake",
+            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
+            "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
+            "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
+            "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
+            "CMAKE_CXX_FLAGS_RELWITHDEBINFO": "-O3 -DNDEBUG -g",
+            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
+            "PREBUILT_LIB_DIR": "android_aarch64",
+            "GGML_OPENMP":      "OFF",
+            "GGML_LLAMAFILE":   "OFF",
+            "GGML_OPENCL":      "ON",
+            "GGML_HEXAGON":     "ON",
+            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
+            "LLAMA_OPENSSL":    "OFF"
+        }
+    },
+
+    {
+        "name": "arm64-windows-snapdragon",
+        "inherits": [ "base", "arm64-windows-llvm" ],
+        "cacheVariables": {
+            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
+            "PREBUILT_LIB_DIR": "windows_aarch64",
+            "GGML_OPENMP":      "OFF",
+            "GGML_LLAMAFILE":   "OFF",
+            "GGML_OPENCL":      "ON",
+            "GGML_HEXAGON":     "ON",
+            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
+            "LLAMA_OPENSSL":    "OFF"
+        }
+    },
+
+    { "name": "arm64-android-snapdragon-debug"  , "inherits": [ "base", "arm64-android-snapdragon", "debug" ] },
+    { "name": "arm64-android-snapdragon-release", "inherits": [ "base", "arm64-android-snapdragon", "release" ] },
+
+    { "name": "arm64-windows-snapdragon-debug"  , "inherits": [ "base", "arm64-windows-snapdragon", "debug" ] },
+    { "name": "arm64-windows-snapdragon-release", "inherits": [ "base", "arm64-windows-snapdragon", "release" ] }
+  ]
+}