CUDA: skip fusion for repeating adds in bias (ggml-org#17080)

Author: am17an

ggml/src/ggml-cuda/CMakeLists.txt

@@ -124,6 +124,7 @@ if (CUDAToolkit_FOUND)
 
     if (GGML_CUDA_DEBUG)
         list(APPEND CUDA_FLAGS -lineinfo)
+        add_compile_definitions(GGML_CUDA_DEBUG)
     endif()
 
     if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")

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

@@ -3152,8 +3152,6 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
 
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
-
-
 #ifdef GGML_CUDA_DEBUG
                 const int nodes_fused = i - prev_i - 1;
                 prev_i = i;
@@ -3302,6 +3300,13 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                                 continue;
                             }
 
+                            // we don't support repeating adds
+                            if (bias_op == GGML_OP_ADD &&
+                                (!ggml_are_same_shape(gate_bias_n->src[0], gate_bias_n->src[1]) ||
+                                 !ggml_are_same_shape(up_bias_n->src[0], up_bias_n->src[1]))) {
+                                continue;
+                            }
+
                             const ggml_tensor * src0 = up_n->src[0];
                             const ggml_tensor * src1 = up_n->src[1];
                             const ggml_tensor * ids  = up_n->src[2];
@@ -3411,6 +3416,10 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                             continue;
                         }
 
+                        if (bias_op == GGML_OP_ADD && !ggml_are_same_shape(bias_node->src[0], bias_node->src[1])) {
+                            continue;
+                        }
+
                         ggml_cuda_mm_fusion_args_host fusion_data{};
                         fusion_data.x_bias = bias_tensor;
 

tests/test-backend-ops.cpp

@@ -4984,23 +4984,25 @@ struct test_mul_mat_vec_fusion : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         if (!use_id) {
-            std::array<int64_t, 4> ne = {k, m, 1, 1};
-            std::array<int64_t, 4> ne0 = {k, n, 1, 1};
+            const int              channels = 4;
+            const int              samples  = 2;
+            std::array<int64_t, 4> ne       = { k, m, channels, samples };
+            std::array<int64_t, 4> ne0      = { k, n, channels, samples };
 
             ggml_tensor * cur  = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne.data());
             ggml_tensor * gate = with_gate ? ggml_new_tensor(ctx, type, 4, ne0.data()) : nullptr;
             ggml_tensor * up   = ggml_new_tensor(ctx, type, 4, ne0.data());
 
             ggml_tensor * ffn_up = ggml_mul_mat(ctx, up, cur);
             if (with_bias) {
-                std::array<int64_t, 4> bias_ne = {ffn_up->ne[0], 1, 1, 1};
+                std::array<int64_t, 4> bias_ne = { ffn_up->ne[0], 1, channels, samples };
                 ggml_tensor * up_bias = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, bias_ne.data());
                 ffn_up = ggml_add(ctx, ffn_up, up_bias);
             }
 
             ggml_tensor * ffn_gate = with_gate ? ggml_mul_mat(ctx, gate, cur) : nullptr;
             if (with_bias && with_gate) {
-                std::array<int64_t, 4> bias_ne = {ffn_gate->ne[0], 1, 1, 1};
+                std::array<int64_t, 4> bias_ne   = { ffn_gate->ne[0], 1, channels, samples };
                 ggml_tensor * gate_bias = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, bias_ne.data());
                 ffn_gate = ggml_add(ctx, ffn_gate, gate_bias);
             }