HIP: do not use MMQ on CDNA when GGML_HIP_MMQ_MFMA=Off is set

IMbackK · IMbackK · commit cf20caa3f942 · 2025-07-29T20:44:35.000+02:00
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
@@ -338,5 +338,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return false;
     }
 
-    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }

Original file line number	Diff line number	Diff line change
`@@ -338,5 +338,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {`
`338`	`338`	`return false;`
`339`	`339`	`}`
`340`	`340`
`341`		`- return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc)) \|\| ne11 < MMQ_DP4A_MAX_BATCH_SIZE;`
	`341`	`+ return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) \|\| ne11 < MMQ_DP4A_MAX_BATCH_SIZE;`
`342`	`342`	`}`