CUDA/HIP: use mmv instead of rocblas for batch 3-5 mmv

IMbackK · IMbackK · commit 66f4c93b2c4c · 2025-06-23T17:15:00.000+02:00
diff --git a/ggml/src/ggml-cuda/mmv.cu b/ggml/src/ggml-cuda/mmv.cu
@@ -471,6 +471,9 @@ bool ggml_cuda_should_use_mmv(enum ggml_type type, int cc, const int64_t * src0_
                 }
                 return ne11 <= 8;
             } else if (fp16_mma_hardware_available(cc)) {
+                if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+                    return ne11 <= 5;
+                }
                 return ne11 <= 2;
             }
             return ne11 <= 8;

Original file line number	Diff line number	Diff line change
`@@ -471,6 +471,9 @@ bool ggml_cuda_should_use_mmv(enum ggml_type type, int cc, const int64_t * src0_`
`471`	`471`	`}`
`472`	`472`	`return ne11 <= 8;`
`473`	`473`	`} else if (fp16_mma_hardware_available(cc)) {`
	`474`	`+ if (GGML_CUDA_CC_IS_RDNA3(cc) \|\| GGML_CUDA_CC_IS_RDNA4(cc)) {`
	`475`	`+ return ne11 <= 5;`
	`476`	`+ }`
`474`	`477`	`return ne11 <= 2;`
`475`	`478`	`}`
`476`	`479`	`return ne11 <= 8;`