From: Aman Gupta <redacted>
Date: Sat, 8 Nov 2025 08:58:05 +0000 (+0800)
Subject: CUDA: skip fusion for repeating adds in bias (llama/17080)
X-Git-Tag: upstream/1.8.3~348
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=522b9bce33d6b20e239ef04cfe85a3bf7fbdb7d4;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp

CUDA: skip fusion for repeating adds in bias (llama/17080)
---

diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index 30247751..67af1d8c 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/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")
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 049aece1..2d4314fb 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/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;