From: Aman Gupta <redacted>
Date: Tue, 4 Nov 2025 02:53:48 +0000 (+0800)
Subject: CUDA: avoid mul + bias fusion when doing fusion (llama/16935)
X-Git-Tag: upstream/0.9.4.185~34
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=da232870b280eaf79c4af9e595562852233fe375;p=pkg%2Fggml%2Fsources%2Fggml

CUDA: avoid mul + bias fusion when doing fusion (llama/16935)
---

diff --git a/src/ggml-cuda/ggml-cuda.cu b/src/ggml-cuda/ggml-cuda.cu
index 5667ec0c..415a7e96 100644
--- a/src/ggml-cuda/ggml-cuda.cu
+++ b/src/ggml-cuda/ggml-cuda.cu
@@ -2115,6 +2115,14 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
     const int cc      = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
 
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
     //we only support fusion for ncols_dst = 1
     if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
         return false;
@@ -2154,6 +2162,15 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
         return false;
     }
 
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
     return use_mul_mat_vec_q;
 }