whispercpp 1.3.5 → 1.3.7

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
@@ -32,6 +32,9 @@
 #include <aclnnop/aclnn_cat.h>
 #include <aclnnop/aclnn_clamp.h>
 #include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_cumsum.h>
+#include <aclnnop/aclnn_tril.h>
+#include <aclnnop/aclnn_triu.h>
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_gelu.h>
 #include <aclnnop/aclnn_gelu_v2.h>
@@ -47,6 +50,9 @@
 #include <aclnnop/aclnn_sign.h>
 #include <aclnnop/aclnn_silu.h>
 #include <aclnnop/aclnn_sin.h>
+#include <aclnnop/aclnn_softplus.h>
+#include <aclnnop/aclnn_swi_glu.h>
+#include <aclnnop/aclnn_geglu.h>
 #include <aclnnop/aclnn_slice.h>
 #include <aclnnop/aclnn_sqrt.h>
 #include <aclnnop/aclnn_tanh.h>
@@ -69,6 +75,9 @@
  */
 void ggml_cann_repeat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+void ggml_cann_swiglu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_geglu(ggml_backend_cann_context & ctx, ggml_tensor * dst, int64_t approximate);
+
 /**
  * @brief   Applies the Leaky ReLU activation function to a tensor using the CANN
  *          backend.
@@ -325,6 +334,48 @@ void ggml_cann_sum_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+/**
+ * @brief   Computes the cumulative sum of a ggml tensor along dim 0 using the
+ *          CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_CUMSUM`.
+ */
+void ggml_cann_cumsum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Computes a triangular mask (tril/triu) of a square ggml tensor
+ *          using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_TRI`.
+ */
+void ggml_cann_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Solves a triangular linear system AX=B using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_SOLVE_TRI`.
+ */
+void ggml_cann_solve_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Creates a diagonal matrix from a vector using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_DIAG`.
+ */
+void ggml_cann_diag(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Fills a tensor with a constant scalar value using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_FILL`.
+ */
+void ggml_cann_fill(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief   Upsamples a ggml tensor using nearest neighbor interpolation using
  *          the CANN backend.
@@ -461,6 +512,9 @@ void ggml_cann_timestep_embedding(ggml_backend_cann_context & ctx, ggml_tensor *
 // @see ggml_cann_dup.
 void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+// @see ggml_cann_acc, but copies src1 into dst instead of adding.
+void ggml_cann_set(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief   Computes the softmax activation with optional masking.
  *
@@ -543,6 +597,21 @@ void ggml_cann_mul_mat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  */
 void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+/**
+ * @brief Pre-load the RoPE cache before ACL graph capture.
+ *
+ * This function must be called outside of graph capture to perform
+ * host-to-device memory copies and device memory allocations that are
+ * not allowed on a captured stream.  After pre-loading, the rope cache
+ * metadata is updated so that the subsequent call to
+ * aclnn_rope_cache_init (inside graph capture) skips these operations
+ * and only records the on-device computations into the captured graph.
+ *
+ * @param ctx  CANN backend context.
+ * @param dst  A ROPE destination tensor from the computation graph.
+ */
+void ggml_cann_rope_cache_preload(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief   Computes the index of the maximum value along the specified dimension
  *          of a ggml tensor using the CANN backend.
@@ -798,6 +867,8 @@ void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *            dst->op is expected to be `GGML_OP_STEP`.
  */
 void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_softplus(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_geglu_quick(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Performs the Flash Attention extended operator using the CANN backend.
@@ -814,67 +885,20 @@ 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 +918,19 @@ 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 +971,9 @@ 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 +1130,13 @@ 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 +1159,13 @@ 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);
@@ -217,14 +216,16 @@ struct ggml_cann_pool_alloc {
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
     // dst tensor
-    void *  node_address;
-    int64_t ne[GGML_MAX_DIMS];
-    size_t  nb[GGML_MAX_DIMS];
+    void *    node_address;
+    ggml_type node_type;
+    int64_t   ne[GGML_MAX_DIMS];
+    size_t    nb[GGML_MAX_DIMS];
 
     // src tensor
-    void *  src_address[GGML_MAX_SRC];
-    int64_t src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
-    size_t  src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
+    void *    src_address[GGML_MAX_SRC];
+    ggml_type src_type[GGML_MAX_SRC];
+    int64_t   src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
+    size_t    src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
 
     // op
     ggml_op node_op;
@@ -248,6 +249,10 @@ struct ggml_graph_node_properties {
             return false;
         }
 
+        if (node->type != this->node_type) {
+            return false;
+        }
+
         for (int i = 0; i < GGML_MAX_DIMS; i++) {
             if (node->ne[i] != this->ne[i]) {
                 return false;
@@ -263,6 +268,10 @@ struct ggml_graph_node_properties {
                     return false;
                 }
 
+                if (node->src[i]->type != this->src_type[i]) {
+                    return false;
+                }
+
                 for (int d = 0; d < GGML_MAX_DIMS; d++) {
                     if (node->src[i]->ne[d] != this->src_ne[i][d]) {
                         return false;
@@ -278,10 +287,7 @@ struct ggml_graph_node_properties {
             }
         }
 
-        if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU) {
-            return memcmp(this->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
-        }
-        return true;
+        return memcmp(this->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
     }
 };
 
@@ -323,6 +329,7 @@ struct ggml_cann_graph {
 
             prop.node_address = node->data;
             prop.node_op      = node->op;
+            prop.node_type    = node->type;
 
             std::copy_n(node->ne, GGML_MAX_DIMS, prop.ne);
             std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
@@ -330,10 +337,12 @@ struct ggml_cann_graph {
             for (int src = 0; src < GGML_MAX_SRC; ++src) {
                 if (node->src[src]) {
                     prop.src_address[src] = node->src[src]->data;
+                    prop.src_type[src]    = node->src[src]->type;
                     std::copy_n(node->src[src]->ne, GGML_MAX_DIMS, prop.src_ne[src]);
                     std::copy_n(node->src[src]->nb, GGML_MAX_DIMS, prop.src_nb[src]);
                 } else {
                     prop.src_address[src] = nullptr;
+                    prop.src_type[src]    = GGML_TYPE_COUNT;
                     std::fill_n(prop.src_ne[src], GGML_MAX_DIMS, 0);
                     std::fill_n(prop.src_nb[src], GGML_MAX_DIMS, 0);
                 }
@@ -382,7 +391,7 @@ 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 +583,7 @@ 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