feat(cann): add optional support for ACL Graph execution

This commit adds support for executing ggml computational graphs using
Huawei's ACL graph mode via the USE_CANN_GRAPH flag. The support can be
enabled at compile time using the CMake option:

    -DUSE_CANN_GRAPH=ON

By default, ACL graph execution is **disabled**, and the fallback path
uses node-by-node execution.

Key additions:
- CMake option  to toggle graph mode
- Graph capture and execution logic using
- Tensor property matching to determine whether graph update is required
- Safe fallback and logging if the environment variable LLAMA_SET_ROWS
  is unset or invalid

This prepares the backend for performance improvements in repetitive graph
execution scenarios on Ascend devices.

Signed-off-by: noemotiovon <[email protected]>

Author: noemotiovon

ggml/src/ggml-cann/CMakeLists.txt

@@ -31,6 +31,7 @@ string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
 set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
 string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
 message(STATUS "CANN: SOC_VERSION =  ${SOC_VERSION}")
+option(USE_CANN_GRAPH "Enable CANN graph execution (ACL graph mode)" OFF)
 
 if (CANN_INSTALL_DIR)
     # Only Support Linux.
@@ -68,6 +69,13 @@ if (CANN_INSTALL_DIR)
 
     target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
 
+    if (USE_CANN_GRAPH)
+        target_compile_definitions(ggml-cann PRIVATE USE_CANN_GRAPH)
+        message(STATUS "CANN: USE_CANN_GRAPH is enabled.")
+    else()
+        message(STATUS "CANN: USE_CANN_GRAPH is disabled.")
+    endif()
+
     message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
     message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
 else()

ggml/src/ggml-cann/common.h

@@ -337,6 +337,29 @@ class cann_task_queue {
     int32_t device_;
 };
 
+#ifdef USE_CANN_GRAPH
+struct ggml_graph_node_properties {
+    void * node_address;
+    ggml_op node_op;
+    int64_t ne[GGML_MAX_DIMS];
+    size_t nb[GGML_MAX_DIMS];
+    void * src_address[GGML_MAX_SRC];
+    int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+};
+
+struct ggml_cann_graph {
+    ~ggml_cann_graph() {
+        if (graph != nullptr) {
+            aclmdlRIDestroy(graph);
+        }
+    }
+
+    aclmdlRI graph = nullptr;
+
+    std::vector<ggml_graph_node_properties> ggml_graph_properties;
+};
+#endif  // USE_CANN_GRAPH
+
 /**
  * @brief Context for managing CANN backend operations.
  */
@@ -345,6 +368,10 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+#ifdef USE_CANN_GRAPH
+    std::unique_ptr<ggml_cann_graph> cann_graph; /**< Cached CANN ACL graph used for executing the current ggml computation graph. */
+    bool set_row_log = true;
+#endif
     cann_task_queue task_queue;
     bool async_mode;
 

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2075,6 +2075,160 @@ static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }
 
+#ifdef USE_CANN_GRAPH
+/**
+ * @brief Populate the internal CANN graph node properties from the ggml computation graph.
+ *
+ * This function copies all node attributes (operation type, dimensions, strides, input sources,
+ * and operation parameters) into the cached CANN graph structure for later reuse or comparison.
+ *
+ * @param cann_ctx  The CANN backend context.
+ * @param cgraph    The ggml computational graph.
+ */
+static void set_ggml_graph_node_properties(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
+    for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
+        ggml_tensor * node = cgraph->nodes[node_idx];
+        cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_address = node->data;
+        cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_op = node->op;
+        
+        for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
+            cann_ctx->cann_graph->ggml_graph_properties[node_idx].ne[dim] = node->ne[dim];
+            cann_ctx->cann_graph->ggml_graph_properties[node_idx].nb[dim] = node->nb[dim];
+        }
+        for (int src = 0; src < GGML_MAX_SRC; src++) {
+            cann_ctx->cann_graph->ggml_graph_properties[node_idx].src_address[src] = 
+                node->src[src] ? node->src[src]->data : nullptr;
+        }
+        memcpy(cann_ctx->cann_graph->ggml_graph_properties[node_idx].op_params, node->op_params, GGML_MAX_OP_PARAMS);
+    }
+}
+
+/**
+ * @brief Check if a ggml tensor node matches a previously captured CANN graph node.
+ *
+ * This function compares all relevant fields (address, op type, shape, source inputs, op params)
+ * to determine whether the current node matches a previously recorded version.
+ *
+ * @param node                  The current ggml tensor node.
+ * @param graph_node_properties The stored properties of a CANN graph node.
+ * @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
+ */
+static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
+    if (node->data != graph_node_properties->node_address &&
+           node->op != GGML_OP_VIEW) {
+        return false;
+    }
+    if (node->op != graph_node_properties->node_op) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (node->ne[i] != graph_node_properties->ne[i]) {
+            return false;
+        }
+        if (node->nb[i] != graph_node_properties->nb[i]) {
+            return false;
+        }
+    }
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (node->src[i] &&
+            node->src[i]->data != graph_node_properties->src_address[i] &&
+            node->op != GGML_OP_VIEW
+        ) {
+            return false;
+        }
+    }
+    if (node->op == GGML_OP_SCALE &&
+        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+        return false;
+    }
+    return true;
+}
+
+/**
+ * @brief Determine if the CANN graph needs to be rebuilt due to graph changes.
+ *
+ * This checks whether the number or properties of ggml graph nodes have changed
+ * compared to the last captured CANN graph. If so, the CANN graph must be re-captured.
+ *
+ * @param cann_ctx  The CANN backend context.
+ * @param cgraph    The current ggml computation graph.
+ * @return true if an update is required; false otherwise.
+ */
+static bool is_cann_graph_update_required(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
+    // 节点个数不同，需重新构图
+    if (cann_ctx->cann_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
+        cann_ctx->cann_graph->ggml_graph_properties.resize(cgraph->n_nodes);
+        return true;
+    }
+
+    // 节点个数相同，便利每个节点，比较其是否匹配
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        bool has_matching_properties = ggml_graph_node_has_matching_properties(
+            cgraph->nodes[i], &cann_ctx->cann_graph->ggml_graph_properties[i]);
+        if(!has_matching_properties) {
+            return true;
+        }
+    }
+    return false;
+}
+#endif  // USE_CANN_GRAPH
+
+/**
+ * @brief Evaluate the computation graph and optionally capture or execute it using CANN graph API.
+ *
+ * If CANN graph execution is enabled and graph capture is required, this function begins
+ * graph capture, runs the graph, ends capture, and stores the captured graph.
+ *
+ * Otherwise, it falls back to op-by-op execution using the CANN compute kernel dispatcher.
+ *
+ * @param cann_ctx                 The CANN backend context.
+ * @param cgraph                   The ggml computation graph.
+ * @param use_cann_graph           Whether to use CANN graph execution.
+ * @param cann_graph_update_required Whether graph capture is needed due to graph changes.
+ */
+static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph,
+    bool & use_cann_graph, bool & cann_graph_update_required) {
+#ifdef USE_CANN_GRAPH
+    if (use_cann_graph && cann_graph_update_required) {
+        if (cann_ctx->cann_graph->graph != nullptr) {
+            ACL_CHECK(aclmdlRIDestroy(cann_ctx->cann_graph->graph));
+            cann_ctx->cann_graph->graph = nullptr;
+        }
+        ACL_CHECK(aclmdlRICaptureBegin(cann_ctx->stream(), ACL_MODEL_RI_CAPTURE_MODE_GLOBAL));
+    }
+#endif // USE_CANN_GRAPH
+
+    // Only perform the graph execution if CANN graphs are not enabled, or we are capturing the graph.
+    // With the use of CANN graphs, the execution will be performed by the graph launch.
+    if (!use_cann_graph || cann_graph_update_required) {
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            ggml_tensor * node = cgraph->nodes[i];
+
+            if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+                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));
+            }
+            GGML_ASSERT(ok);
+        }
+    }
+
+#ifdef USE_CANN_GRAPH
+    if (use_cann_graph && cann_graph_update_required) { // End CANN graph capture
+        ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &cann_ctx->cann_graph->graph));
+    }
+    
+    if (use_cann_graph) {
+        // Execute graph
+        ACL_CHECK(aclmdlRIExecuteAsync(cann_ctx->cann_graph->graph, cann_ctx->stream()));
+    }
+#endif
+}
+
+
 /**
  * @brief Computes a computational graph using a CANN backend.
  *
@@ -2089,28 +2243,52 @@ static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
  */
 static enum ggml_status ggml_backend_cann_graph_compute(
     ggml_backend_t backend, ggml_cgraph* cgraph) {
+
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
-
     ggml_cann_set_device(cann_ctx->device);
-    //release temp buffer create by set tensor.
     release_nz_workspace();
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_tensor* node = cgraph->nodes[i];
-
-        if (ggml_is_empty(node) || node->op == GGML_OP_NONE) {
-            continue;
+#ifdef USE_CANN_GRAPH
+    bool use_cann_graph = true;
+    bool cann_graph_update_required = false;
+
+    // 检查环境变量 LLAMA_SET_ROWS
+    const char* LLAMA_SET_ROWS_ENV = std::getenv("LLAMA_SET_ROWS");
+    bool supports_set_rows = LLAMA_SET_ROWS_ENV ? (std::atoi(LLAMA_SET_ROWS_ENV) != 0) : false;
+
+    if (!supports_set_rows) {
+        if(cann_ctx->set_row_log) {
+            GGML_LOG_ERROR(
+                "%s: CANN Graph disabled — environment variable LLAMA_SET_ROWS not set or invalid. "
+                "To enable CANN ACL Graph execution, export LLAMA_SET_ROWS=1. "
+                "Falling back to non-graph mode on device %d.\n",
+                __func__, cann_ctx->device
+            );
+            cann_ctx->set_row_log = false;
         }
-
-        bool ok = ggml_cann_compute_forward(*cann_ctx, node);
-
-        if (!ok) {
-            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__,
-                    node->name, ggml_op_name(node->op));
+        use_cann_graph = false;
+    }
+    
+    if (use_cann_graph) {
+        if (cann_ctx->cann_graph == nullptr) {
+            cann_ctx->cann_graph.reset(new ggml_cann_graph());
+            cann_graph_update_required = true;
         }
-        GGML_ASSERT(ok);
+
+        cann_graph_update_required = is_cann_graph_update_required(cann_ctx, cgraph);    
+        set_ggml_graph_node_properties(cann_ctx, cgraph);
     }
+#else
+    bool use_cann_graph = false;
+    bool cann_graph_update_required = false;
+#endif  // USE_CANN_GRAPH
+
+    evaluate_and_capture_cann_graph(
+        cann_ctx,
+        cgraph,
+        use_cann_graph,
+        cann_graph_update_required
+    );
 
     return GGML_STATUS_SUCCESS;
 }
@@ -2226,12 +2404,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // only support F32 and F16.
                 return false;
             }
-
-            if (!ggml_are_same_shape(op, src) && !ggml_is_contiguous(op)) {
-                // unsupport dst is not contiguous.
-                return false;
-            }
-
             return true;
         } break;
         case GGML_OP_CONT: {