CUDA: better error for FA kernel with 0 occupancy (#16643)

* ggml: add ggml_can_fuse_subgraph

* ggml-cuda: use ggml_can_fuse_subgraph for topk-moe

* format

* 1. remove inputs from signature as they are transient nodes
2. add check for views: view_src should be part of the subgraph

* - combine check into one loop
- check all view_src parents
- other minor review comments

* remove redudant if test

* - rename and other minor review comments

* add assert about count < 32

ggml/src/ggml-cuda/fattn-common.cuh

@@ -895,6 +895,7 @@ void launch_fattn(
     const dim3 block_dim(warp_size, nwarps, 1);
     int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
     CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
+    GGML_ASSERT(max_blocks_per_sm > 0);
     int parallel_blocks = max_blocks_per_sm;
 
     dim3 blocks_num;

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

@@ -2821,15 +2821,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     std::initializer_list<enum ggml_op> topk_moe_ops           = ggml_cuda_topk_moe_ops(false);
     std::initializer_list<enum ggml_op> topk_moe_ops_with_norm = ggml_cuda_topk_moe_ops(true);
 
-    if (ops.size() == topk_moe_ops_with_norm.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops_with_norm.begin())) {
-
-        if (node_idx + topk_moe_ops_with_norm.size() > (size_t)cgraph->n_nodes) {
-            return false;
-        }
-
-        for (size_t i = 0; i < topk_moe_ops_with_norm.size(); i++) {
-            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops_with_norm.begin()[i]) return false;
-        }
+    if (ops.size() == topk_moe_ops_with_norm.size() &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, topk_moe_ops_with_norm, { node_idx + 3, node_idx + 8 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx+8];
 
@@ -2838,16 +2831,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         }
     }
 
-    if (ops.size() == topk_moe_ops.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops.begin())) {
-
-        if (node_idx + topk_moe_ops.size() > (size_t)cgraph->n_nodes) {
-            return false;
-        }
-
-        for (size_t i = 0; i < topk_moe_ops.size(); i++) {
-            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops.begin()[i]) return false;
-        }
-
+    if (ops.size() == topk_moe_ops.size() &&
+        ggml_can_fuse_subgraph(cgraph, node_idx, topk_moe_ops, { node_idx + 3, node_idx + 4 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx+4];
         if (ggml_cuda_should_use_topk_moe(softmax, weights)) {

ggml/src/ggml-impl.h

@@ -647,6 +647,36 @@ static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx
     return ggml_can_fuse_ext(cgraph, idxs, ops, num_ops);
 }
 
+GGML_API bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
+                                         const int *                node_idxs,
+                                         int                        count,
+                                         const enum ggml_op *       ops,
+                                         const int *                outputs,
+                                         int                        num_outputs);
+
+// Returns true if the subgraph formed by {node_idxs} can be fused
+// checks whethers all nodes which are not part of outputs can be elided
+// by checking if their num_uses are confined to the subgraph
+static inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph * cgraph,
+                                          int                        node_idx,
+                                          int                        count,
+                                          const enum ggml_op *       ops,
+                                          const int *                outputs,
+                                          int                        num_outputs) {
+    GGML_ASSERT(count < 32);
+    if (node_idx + count > cgraph->n_nodes) {
+        return false;
+    }
+
+    int idxs[32];
+
+    for (int i = 0; i < count; ++i) {
+        idxs[i] = node_idx + i;
+    }
+
+    return ggml_can_fuse_subgraph_ext(cgraph, idxs, count, ops, outputs, num_outputs);
+}
+
 #ifdef __cplusplus
 }
 #endif
@@ -660,6 +690,13 @@ inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::
     return ggml_can_fuse(cgraph, node_idx, ops.begin(), (int)ops.size());
 }
 
+inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph *          cgraph,
+                                   int                                 start_idx,
+                                   std::initializer_list<enum ggml_op> ops,
+                                   std::initializer_list<int>          outputs = {}) {
+    return ggml_can_fuse_subgraph(cgraph, start_idx, ops.size(), ops.begin(), outputs.begin(), outputs.size());
+}
+
 // expose GGUF internals for test code
 GGML_API size_t gguf_type_size(enum gguf_type type);
 GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -15,13 +15,12 @@
 
 #include <CL/cl.h>
 
+#include <inttypes.h>
 #include <string.h>
 
 #include <cstddef>
 #include <cstdint>
-#include <atomic>
 #include <fstream>
-#include <limits>
 #include <vector>
 #include <string>
 #include <cmath>
@@ -533,25 +532,17 @@ struct ggml_backend_opencl_context {
         }
 
         // Dump a csv
-        float total_kernel_time = 0;
-        fprintf(fperf, "op name, kernel name, queued duration (ms), submit duration(ms), exec duration (ms), complete duration (ms), total duration (ms), global size, local size, output size\n");
+        fprintf(fperf, "op name, kernel name, exec duration (ms), global size, local size, output size\n");
         for (const ProfilingInfo & info : profiling_info) {
-            total_kernel_time += info.cmd_duration_ns/1.e6f;
-            fprintf(fperf, "%s,%s,%f,%f,%f,%f,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n",
+            fprintf(fperf, "%s,%s,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n",
                 info.op_name.c_str(), info.kernel_name.c_str(),
-                info.cmd_queued_duration_ns/1.e6f,
-                info.cmd_submit_duration_ns/1.e6f,
                 info.cmd_duration_ns/1.e6f,
-                info.cmd_complete_duration_ns/1.e6f,
-                info.cmd_total_duration_ns/1.e6f,
                 info.global_size[0], info.global_size[1], info.global_size[2],
                 info.local_size[0], info.local_size[1], info.local_size[2],
                 info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]);
         }
         fclose(fperf);
 
-        GGML_LOG_INFO("ggml_opencl: total kernel time: %f\n", total_kernel_time);
-
         // Dump a simple chrome trace
         FILE* ftrace = fopen("cl_trace.json", "w");
         if (!ftrace) {
@@ -561,14 +552,14 @@ struct ggml_backend_opencl_context {
 
         fprintf(ftrace, "[\n");
         for (const ProfilingInfo & info : profiling_info) {
-            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %llu, \"pid\": \"\", \"tid\": \"Host\"},\n",
+            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %" PRIu64 ", \"pid\": \"\", \"tid\": \"Host\"},\n",
                 info.kernel_name.c_str(), info.cmd_queued/1000);
-            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %llu, \"pid\": \"\", \"tid\": \"Host\"},\n",
+            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %" PRIu64 ", \"pid\": \"\", \"tid\": \"Host\"},\n",
                 info.kernel_name.c_str(), info.cmd_submit/1000);
 
-            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %llu, \"pid\": \"\", \"tid\": \"Device\"},\n",
+            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %" PRIu64 ", \"pid\": \"\", \"tid\": \"Device\"},\n",
                 info.kernel_name.c_str(), info.cmd_start/1000);
-            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %llu, \"pid\": \"\", \"tid\": \"Device\"},\n",
+            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %" PRIu64 ", \"pid\": \"\", \"tid\": \"Device\"},\n",
                 info.kernel_name.c_str(), info.cmd_end/1000);
         }
         fclose(ftrace);
@@ -7652,6 +7643,8 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
     const cl_ulong nb21 = src2->nb[1];
     const cl_ulong nb20 = src2->nb[0];
 
+    UNUSED(nb20);
+
     const int ne0 = dst->ne[0];
     const int ne1 = dst->ne[1];
 

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -345,7 +345,7 @@ void main() {
 
     float Lfrcp[Br];
     [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-        Lfrcp[r] = 1.0 / Lf[r];
+        Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]);
     }
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -380,7 +380,7 @@ void main() {
 
     float Lfrcp[rows_per_thread];
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        Lfrcp[r] = 1.0 / Lf[r];
+        Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]);
     }
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -121,7 +121,11 @@ void main() {
     const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
 
     L = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
+#if defined(ACC_TYPE_MAX)
+    M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(-ACC_TYPE_MAX / ACC_TYPE(2));
+#else
     M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(NEG_FLT_MAX_OVER_2);
+#endif
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> slopeMat = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(1.0);
 
@@ -294,7 +298,7 @@ void main() {
 
     [[unroll]]
     for (int k = 0; k < Ldiag.length(); ++k) {
-        Ldiag[k] = ACC_TYPE(1.0) / Ldiag[k];
+        Ldiag[k] = (Ldiag[k] == 0.0) ? ACC_TYPE(0.0) : (ACC_TYPE(1.0) / Ldiag[k]);
     }
 
     O = Ldiag*O;

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp

@@ -91,7 +91,7 @@ void main() {
         L = L*ms + vs;
     }
 
-    L = 1.0 / L;
+    L = (L == 0.0) ? 0.0 : 1.0 / L;
 
     // D dimension is split across workgroups in the y dimension
     uint d = tid + gl_WorkGroupID.y * BLOCK_SIZE;

ggml/src/ggml.c

@@ -6964,6 +6964,78 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     GGML_LOG_INFO("========================================\n");
 }
 
+static int ggml_node_list_find_tensor(const struct ggml_cgraph * cgraph,
+                                      const int *                idxs,
+                                      int                        count,
+                                      const struct ggml_tensor * tensor) {
+    GGML_ASSERT(cgraph && idxs);
+    for (int i = 0; i < count; ++i) {
+        const int node_idx = idxs[i];
+
+        if (node_idx >= cgraph->n_nodes) {
+            return -1;
+        }
+        if (cgraph->nodes[node_idx] == tensor) {
+            return i;
+        }
+    }
+    return -1;
+}
+
+bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
+                                const int *                node_idxs,
+                                int                        count,
+                                const enum ggml_op *       ops,
+                                const int *                outputs,
+                                int                        num_outputs) {
+    GGML_ASSERT(outputs && num_outputs > 0);
+
+    for (int i = 0; i < count; ++i) {
+        if (node_idxs[i] >= cgraph->n_nodes) {
+            return false;
+        }
+
+        const struct ggml_tensor * node = cgraph->nodes[node_idxs[i]];
+
+        if (node->op != ops[i]) {
+            return false;
+        }
+
+        if (ggml_node_list_find_tensor(cgraph, outputs, num_outputs, node) != -1) {
+            continue;
+        }
+
+        if (node->flags & GGML_TENSOR_FLAG_OUTPUT) {
+            return false;
+        }
+
+        int subgraph_uses = 0;
+        for (int j = i + 1; j < count; ++j) {
+            const struct ggml_tensor * other_node = cgraph->nodes[node_idxs[j]];
+            for (int src_idx = 0; src_idx < GGML_MAX_SRC; src_idx++) {
+                if (other_node->src[src_idx] == node) {
+                    subgraph_uses++;
+                }
+            }
+        }
+
+        if (subgraph_uses != ggml_node_get_use_count(cgraph, node_idxs[i])) {
+            return false;
+        }
+
+        // if node is a view, check if the view_src and all it's parent view_srcs are within the subgraph
+        struct ggml_tensor * view_src = node->view_src;
+        while (view_src) {
+            if (ggml_node_list_find_tensor(cgraph, node_idxs, count, view_src) == -1) {
+                return false;
+            }
+            view_src = view_src->view_src;
+        }
+    }
+
+    return true;
+}
+
 // check if node is part of the graph
 static bool ggml_graph_find(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
     if (cgraph == NULL) {