whispercpp 1.3.5 → 1.3.6

data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h

@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
@@ -814,67 +814,20 @@ void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  */
 void ggml_cann_flash_attn_ext(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
-/*
- * @brief A generic wrapper for ACL resources with custom deleter support.
- */
-using any_acl_resource = std::unique_ptr<void, std::function<void(void *)>>;
-
 /**
- * @brief Trait structure used to define how to destroy a given ACL resource type.
+ * @brief Forward Gated Linear Attention on the CANN backend.
  *
- * @tparam T ACL resource type.
- */
-template <typename T> struct acl_resource_traits;
-
-/**
- * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
- */
-template <> struct acl_resource_traits<aclTensor> {
-    static void destroy(void * p) { ACL_CHECK(aclDestroyTensor(static_cast<aclTensor *>(p))); }
-};
-
-/**
- * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
- */
-template <> struct acl_resource_traits<aclIntArray> {
-    static void destroy(void * p) { ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray *>(p))); }
-};
-
-/**
- * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
- */
-template <> struct acl_resource_traits<aclScalar> {
-    static void destroy(void * p) { ACL_CHECK(aclDestroyScalar(static_cast<aclScalar *>(p))); }
-};
-
-/**
- * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
- */
-template <> struct acl_resource_traits<aclTensorList> {
-    static void destroy(void * p) { ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList *>(p))); }
-};
-
-/**
- * @brief Creates a generic ACL resource wrapper with proper destruction logic.
+ * Expects dst->src[0..4] = {k, v, q, g, s} with shape conventions:
+ *   k, v, q, g: [D] with outer dims T x H batched as ne[2]=T, ne[1]=H
+ *   s: initial state [B, H, D, D], where B is batch and D=C/H
+ * dst holds both outputs (o) and updated state; a scale factor is read from op params.
  *
- * @tparam T ACL resource type.
- * @param ptr Raw pointer to ACL resource.
- * @return any_acl_resource Smart pointer that handles destruction.
- */
-template <typename T> any_acl_resource make_acl_resource(T * ptr) {
-    return any_acl_resource(static_cast<void *>(ptr), [](void * p) { acl_resource_traits<T>::destroy(p); });
-}
-
-/**
- * @brief Registers multiple ACL resources into a vector for lifetime management.
+ * The kernel updates per time step l: S_new = g ⊗ S_old + k ⊗ v, then computes o = (S_new^T q) * scale.
  *
- * @tparam Args Variadic list of ACL resource types.
- * @param vec Target vector to hold ACL resources.
- * @param args Raw pointers to ACL resources.
+ * @param ctx Backend context providing stream/allocator utilities.
+ * @param dst Output tensor; src deps are k, v, q, g, s as above.
  */
-template <typename... Args> void register_acl_resources(std::vector<any_acl_resource> & vec, Args *... args) {
-    (vec.emplace_back(make_acl_resource(args)), ...);
-}
+void ggml_cann_gated_linear_attn(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief Launches an asynchronous task using the memory allocator.
@@ -894,19 +847,19 @@ template <typename... Args> void register_acl_resources(std::vector<any_acl_reso
  * same stream are executed in queue order.
  */
 
-#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                           \
-    do {                                                                                     \
-        uint64_t        workspaceSize = 0;                                                   \
-        aclOpExecutor * executor;                                                            \
-        void *          workspaceAddr = nullptr;                                             \
-        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor)); \
-        /* workspace should alloced in main thread to keep malloc order when using vmm. */   \
-        if (workspaceSize > 0) {                                                             \
-            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);             \
-            workspaceAddr = workspace_allocator.get();                                       \
-        }                                                                                    \
-        ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));     \
-    } while (0)
+#    define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                           \
+        do {                                                                                     \
+            uint64_t        workspaceSize = 0;                                                   \
+            aclOpExecutor * executor;                                                            \
+            void *          workspaceAddr = nullptr;                                             \
+            ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor)); \
+            /* workspace should alloced in main thread to keep malloc order when using vmm. */   \
+            if (workspaceSize > 0) {                                                             \
+                ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);             \
+                workspaceAddr = workspace_allocator.get();                                       \
+            }                                                                                    \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));     \
+        } while (0)
 
 /**
  * @brief   Performs sparse expert-based matrix multiplication using the CANN backend.
@@ -947,7 +900,9 @@ void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  * @param rms_norm_tensor The RMS_NORM operation node, contains the gamma weights
  *                        and epsilon parameter.
  */
-void ggml_cann_op_add_rms_norm_fused(ggml_backend_cann_context & ctx, ggml_tensor * add_node, ggml_tensor * rms_norm_node);
+void ggml_cann_op_add_rms_norm_fused(ggml_backend_cann_context & ctx,
+                                     ggml_tensor *               add_node,
+                                     ggml_tensor *               rms_norm_node);
 
 /**
  * @brief   Check whether a tensor is a weight tensor for matrix multiplication.
@@ -1104,13 +1059,13 @@ void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, ac
  * @see ggml_cann_op_unary
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_OP_UNARY(OP_NAME)                                                              \
-    do {                                                                                              \
-        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
-            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
-        };                                                                                            \
-        ggml_cann_op_unary(lambda, ctx, dst);                                                         \
-    } while (0)
+#    define GGML_CANN_CALL_OP_UNARY(OP_NAME)                                                              \
+        do {                                                                                              \
+            auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
+                GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
+            };                                                                                            \
+            ggml_cann_op_unary(lambda, ctx, dst);                                                         \
+        } while (0)
 
 /**
  * @brief Helper macro to call a gated unary ACL operator via ggml_cann_op_unary_gated.
@@ -1133,13 +1088,13 @@ void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, ac
  * @see ggml_cann_op_unary_gated
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                                                        \
-    do {                                                                                              \
-        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
-            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
-        };                                                                                            \
-        ggml_cann_op_unary_gated(lambda, ctx, dst);                                                   \
-    } while (0)
+#    define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME)                                                        \
+        do {                                                                                              \
+            auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) { \
+                GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);                                  \
+            };                                                                                            \
+            ggml_cann_op_unary_gated(lambda, ctx, dst);                                                   \
+        } while (0)
 
 #endif  // CANN_ACLNN_OPS
 

data/ext/sources/ggml/src/ggml-cann/common.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
@@ -101,7 +101,6 @@ struct ggml_cann_device_info {
 const ggml_cann_device_info & ggml_cann_info();
 
 void    ggml_cann_set_device(int32_t device);
-int32_t ggml_cann_get_device();
 
 std::optional<std::string> get_env_as_lowercase(const std::string & name);
 bool                       parse_bool(const std::string & value);
@@ -382,7 +381,7 @@ struct ggml_cann_graph_lru_cache {
 
     std::list<ggml_cann_graph *> cache_list; /**< List storing cached graphs as raw pointers. */
 
-    ggml_cann_graph_lru_cache() { capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12")); }
+    ggml_cann_graph_lru_cache() { capacity = parse_integer(get_env_as_lowercase("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12")); }
 
     /**
      * @brief Push a new graph to the front of the cache.
@@ -574,7 +573,7 @@ struct ggml_backend_cann_context {
         description = aclrtGetSocName();
 
 #ifdef USE_ACL_GRAPH
-        acl_graph_mode = parse_bool(get_env("GGML_CANN_ACL_GRAPH").value_or("on"));
+        acl_graph_mode = parse_bool(get_env_as_lowercase("GGML_CANN_ACL_GRAPH").value_or("on"));
         GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n", __func__, device, acl_graph_mode ? "GRAPH" : "EAGER",
                       acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
 #endif

data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to
@@ -93,17 +93,6 @@ void ggml_cann_set_device(const int32_t device) {
     g_current_cann_device = device;
 }
 
-/**
- * @brief Retrieves the current device ID.
- *
- * @return The current device ID.
- */
-int32_t ggml_cann_get_device() {
-    int32_t id;
-    ACL_CHECK(aclrtGetDevice(&id));
-    return id;
-}
-
 /**
  * @brief Get the value of the specified environment variable (name) as lowercase.
  *        if not empty, return a std::string object
@@ -805,19 +794,44 @@ struct ggml_backend_cann_buffer_context {
     ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
 };
 
+// cann buffer type
 /**
- * @brief Check if a buffer is a CANN buffer.
+ * @brief Structure representing context information for a specific backend
+ * buffer type.
+ */
+struct ggml_backend_cann_buffer_type_context {
+    int32_t     device; /**< Device identifier associated with the buffer context. */
+    std::string name;   /**< Name associated with the buffer context. */
+};
+
+/**
+ * @brief Retrieves the name associated with a CANN buffer type.
  *
- * This function checks if a given buffer is a CANN buffer by comparing its
- * `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
+ * This function returns the descriptive name associated with the specified
+ * CANN buffer type context.
  *
- * @param buffer The buffer to check.
- * @return true if the buffer is a CANN buffer, false otherwise.
+ * @param buft Pointer to the buffer type context.
+ * @return Const pointer to the C-style string containing the name.
  */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
+static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
 
-static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_is_cann(buffer->buft);
+    return buft_ctx->name.c_str();
+}
+
+/**
+ * @brief Checks if the backend buffer type is associated with the CANN backend.
+ *
+ * This function checks whether the provided backend buffer type is associated
+ * with the CANN backend based on the comparison of its name retrieval function
+ * pointer.
+ *
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the buffer type is associated with the CANN
+ * backend, otherwise false.
+ */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
 }
 
 /**
@@ -1282,7 +1296,7 @@ static void ggml_backend_cann_buffer_get_tensor(ggml_backend_buffer_t buffer,
 static bool ggml_backend_cann_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                                 const ggml_tensor *   src,
                                                 ggml_tensor *         dst) {
-    if (ggml_backend_buffer_is_cann(src->buffer)) {
+    if (ggml_backend_buft_is_cann(src->buffer->buft)) {
         ggml_backend_cann_buffer_context * src_ctx = (ggml_backend_cann_buffer_context *) src->buffer->context;
         ggml_backend_cann_buffer_context * dst_ctx = (ggml_backend_cann_buffer_context *) buffer->context;
 
@@ -1346,31 +1360,6 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .reset           = */ NULL,
 };
 
-// cann buffer type
-/**
- * @brief Structure representing context information for a specific backend
- * buffer type.
- */
-struct ggml_backend_cann_buffer_type_context {
-    int32_t     device; /**< Device identifier associated with the buffer context. */
-    std::string name;   /**< Name associated with the buffer context. */
-};
-
-/**
- * @brief Retrieves the name associated with a CANN buffer type.
- *
- * This function returns the descriptive name associated with the specified
- * CANN buffer type context.
- *
- * @param buft Pointer to the buffer type context.
- * @return Const pointer to the C-style string containing the name.
- */
-static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
-    ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
-
-    return buft_ctx->name.c_str();
-}
-
 /**
  * @brief Allocates a new CANN buffer of the specified type and size.
  *
@@ -1889,6 +1878,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
         case GGML_OP_OUT_PROD:
             ggml_cann_out_prod(ctx, dst);
             break;
+        case GGML_OP_GATED_LINEAR_ATTN:
+            ggml_cann_gated_linear_attn(ctx, dst);
+            break;
         case GGML_OP_SSM_CONV:
             ggml_cann_ssm_conv(ctx, dst);
             break;
@@ -2005,7 +1997,7 @@ static bool ggml_backend_cann_cpy_tensor_async(ggml_backend_t      backend_src,
 
     GGML_ASSERT(!is_matmul_weight((const ggml_tensor *) src));
 
-    if (!ggml_backend_buffer_is_cann(src->buffer) || !ggml_backend_buffer_is_cann(dst->buffer)) {
+    if (!ggml_backend_buft_is_cann(src->buffer->buft) || !ggml_backend_buft_is_cann(dst->buffer->buft)) {
         return false;
     }
 
@@ -2154,6 +2146,10 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
                 continue;
             }
 
+            if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                continue;
+            }
+
             bool ok = ggml_cann_compute_forward(*cann_ctx, node);
             if (!ok) {
                 GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
@@ -2454,6 +2450,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
         case GGML_OP_MEAN:
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
+        case GGML_OP_GATED_LINEAR_ATTN:
             return true;
         case GGML_OP_OUT_PROD:
             {
@@ -2526,21 +2523,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
     GGML_UNUSED(dev);
 }
 
-/**
- * @brief Checks if the backend buffer type is associated with the CANN backend.
- *
- * This function checks whether the provided backend buffer type is associated
- * with the CANN backend based on the comparison of its name retrieval function
- * pointer.
- *
- * @param buft Pointer to the backend buffer type to check.
- * @return bool Returns true if the buffer type is associated with the CANN
- * backend, otherwise false.
- */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
-}
-
 /**
  * @brief Records an event on the CANN backend stream.
  *

data/ext/sources/ggml/src/ggml-common.h

@@ -102,6 +102,9 @@ typedef sycl::half2 ggml_half2;
 #define QI_MXFP4 (QK_MXFP4 / (4 * QR_MXFP4))
 #define QR_MXFP4 2
 
+#define QI_NVFP4 (QK_NVFP4 / (4 * QR_NVFP4))
+#define QR_NVFP4 2
+
 #define QI5_0 (QK5_0 / (4 * QR5_0))
 #define QR5_0 2
 
@@ -194,6 +197,14 @@ typedef struct {
 } block_mxfp4;
 static_assert(sizeof(block_mxfp4) == sizeof(uint8_t) + QK_MXFP4/2, "wrong mxfp4 block size/padding");
 
+#define QK_NVFP4 64
+#define QK_NVFP4_SUB 16  // sub-block size for per-group scales
+typedef struct {
+    uint8_t d[QK_NVFP4/QK_NVFP4_SUB]; // UE4M3 scales (4 bytes, one per 16-element sub-block)
+    uint8_t qs[QK_NVFP4/2];           // packed 4-bit E2M1 values (32 bytes)
+} block_nvfp4;
+static_assert(sizeof(block_nvfp4) == sizeof(uint8_t)*(QK_NVFP4/QK_NVFP4_SUB) + QK_NVFP4/2, "wrong nvfp4 block size/padding");
+
 #define QK5_0 32
 typedef struct {
     ggml_half d;           // delta

data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt

@@ -9,6 +9,11 @@ function(ggml_add_cpu_backend_features cpu_name arch)
     target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
     target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
     set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    # Disable LTO for the feature detection code to prevent cross-module optimization
+    # from inlining architecture-specific instructions into the score function.
+    # Without this, LTO can cause SIGILL when loading backends on older CPUs
+    # (e.g., loading power10 backend on power9 crashes before feature check runs).
+    target_compile_options(${GGML_CPU_FEATS_NAME} PRIVATE -fno-lto)
     target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME})
 endfunction()
 
@@ -561,35 +566,32 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.16.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.22.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "0a9e9008adb6031f9e8cf70dff4a3321")
+        set(KLEIDIAI_ARCHIVE_MD5  "54049037570ab0ee0a0d126b2ba5ece1")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
         endif()
 
+        # TODO: Use FetchContent_MakeAvailable with EXCLUDE_FROM_ALL after bumping minimum CMake version to 3.28+
+        # Using FetchContent_Populate instead to avoid EXCLUDE_FROM_ALL which requires CMake 3.28
         FetchContent_Declare(KleidiAI_Download
             URL ${KLEIDIAI_DOWNLOAD_URL}
             DOWNLOAD_EXTRACT_TIMESTAMP NEW
             URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5})
 
-        FetchContent_MakeAvailable(KleidiAI_Download)
         FetchContent_GetProperties(KleidiAI_Download
             SOURCE_DIR  KLEIDIAI_SRC
             POPULATED   KLEIDIAI_POPULATED)
 
         if (NOT KLEIDIAI_POPULATED)
-            message(FATAL_ERROR "KleidiAI source downloaded failed.")
+            FetchContent_Populate(KleidiAI_Download)
+            FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
         endif()
 
         add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
 
-        # Remove kleidiai target after fetching it
-        if (TARGET kleidiai)
-            set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE)
-        endif()
-
         list(APPEND GGML_CPU_SOURCES
             ggml-cpu/kleidiai/kleidiai.cpp
             ggml-cpu/kleidiai/kernels.cpp
@@ -606,6 +608,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
 
         set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
@@ -646,7 +649,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         if (NOT SME_ENABLED MATCHES -1)
             list(APPEND GGML_KLEIDIAI_SOURCES
-                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
@@ -654,10 +656,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
                 ${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S)
-            set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
+            set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2+sme2+fp16")
         endif()
 
         if (NOT SVE_ENABLED MATCHES -1)

data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp

@@ -141,27 +141,50 @@ static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_typ
 namespace ggml::cpu::amx {
 class extra_buffer_type : ggml::cpu::extra_buffer_type {
     bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
-        // handle only 2d gemm for now
-        auto is_contiguous_2d = [](const struct ggml_tensor * t) {
-            return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
-        };
-
-        if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) &&  // src0 must be contiguous
-            is_contiguous_2d(op->src[1]) &&                               // src1 must be contiguous
-            op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
-            op->src[0]->ne[0] % (TILE_K * 2 * 32) == 0 && // TODO: not sure if correct (https://github.com/ggml-org/llama.cpp/pull/16315)
-            op->ne[0] % (TILE_N * 2) == 0 &&                              // out_features is 32x
-            (qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
-            // src1 must be host buffer
-            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+        if (op->op != GGML_OP_MUL_MAT) {
+            return false;
+        }
+        auto * src0 = op->src[0];
+        auto * src1 = op->src[1];
+
+        if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+            return false;
+        }
+        if (!src0->buffer || src0->buffer->buft != ggml_backend_amx_buffer_type()) {
+            return false;
+        }
+        if (src1->buffer && !ggml_backend_buft_is_host(src1->buffer->buft)) {
+            return false;
+        }
+        if (op->ne[0] % (TILE_N * 2)) {
+            return false;
+        }
+        int alignment;
+        switch (src0->type) {
+            case GGML_TYPE_Q4_0:
+            case GGML_TYPE_Q4_1:
+            case GGML_TYPE_Q8_0:
+                alignment = TILE_K;
+                break;
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+            case GGML_TYPE_Q6_K:
+            case GGML_TYPE_IQ4_XS:
+                alignment = 256; // QK_K
+                break;
+            case GGML_TYPE_F16:
+                alignment = 16;
+                break;
+            default:
                 return false;
-            }
-            // src1 must be float32
-            if (op->src[1]->type == GGML_TYPE_F32) {
-                return true;
-            }
         }
-        return false;
+        if (src0->ne[0] % alignment) {
+            return false;
+        }
+        if (src1->type != GGML_TYPE_F32) {
+            return false;
+        }
+        return true;
     }
 
     ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {

data/ext/sources/ggml/src/ggml-cpu/amx/common.h

@@ -9,6 +9,8 @@
 
 #if defined(GGML_USE_OPENMP)
 #include <omp.h>
+#else
+#include <thread>
 #endif
 
 #define TILE_M 16
@@ -56,18 +58,40 @@ inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
 }
 
 template <typename func_t>
-inline void parallel_for(int n, const func_t& f) {
+inline void parallel_for(int n, const func_t & f) {
+    if (n <= 0) {
+        return;
+    }
 #if defined(GGML_USE_OPENMP)
-#pragma omp parallel
-{
-    int nth = omp_get_num_threads();
-    int ith = omp_get_thread_num();
-    int tbegin, tend;
-    balance211(n, nth, ith, tbegin, tend);
-    f(tbegin, tend);
-}
+    #pragma omp parallel
+    {
+        int nth = omp_get_num_threads();
+        int ith = omp_get_thread_num();
+        int tbegin, tend;
+        balance211(n, nth, ith, tbegin, tend);
+        f(tbegin, tend);
+    }
 #else
-    f(0, n);
+    int nth = std::thread::hardware_concurrency();
+    if (nth <= 1) {
+        f(0, n);
+        return;
+    }
+    if (nth > n) {
+        nth = n;
+    }
+    std::vector<std::thread> threads;
+    threads.reserve(nth);
+    for (int ith = 0; ith < nth; ++ith) {
+        threads.emplace_back([&f, n, ith, nth] {
+            int tbegin, tend;
+            balance211(n, nth, ith, tbegin, tend);
+            f(tbegin, tend);
+        });
+    }
+    for (auto & t : threads) {
+        t.join();
+    }
 #endif
 }