From: Johannes Gäßler <redacted>
Date: Sat, 6 Dec 2025 12:45:36 +0000 (+0100)
Subject: HIP: fix RDNA3 FP16/BF16 matrix multiplication (llama/17817)
X-Git-Tag: upstream/0.9.4.395~49
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=6f370af5bc69abe270f37d275c9590281c85f651;p=pkg%2Fggml%2Fsources%2Fggml

HIP: fix RDNA3 FP16/BF16 matrix multiplication (llama/17817)
---

diff --git a/src/ggml-cuda/mmf.cu b/src/ggml-cuda/mmf.cu
index be2ad1c6..7cf33f0d 100644
--- a/src/ggml-cuda/mmf.cu
+++ b/src/ggml-cuda/mmf.cu
@@ -160,9 +160,9 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc) || (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc));
         case GGML_TYPE_BF16:
-            return ampere_mma_available(cc) || amd_wmma_available(cc);
+            return ampere_mma_available(cc) || (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc));
         default:
             return false;
     }