Revert "with forced wmma for cu11"

Nexesenex · Nexesenex · commit 1b543916b266 · 2025-10-06T20:59:33.000+02:00
This reverts commit f545f4d.
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
@@ -396,9 +396,9 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         }
 
         //kcpp: use wmma to fix cu11 incoherence
-        if (fp16_mma_available(cc) && (ggml_cuda_highest_compiled_arch(cc) <= GGML_CUDA_CC_TURING || cc == GGML_CUDA_CC_TURING)) {
-            return BEST_FATTN_KERNEL_WMMA_F16;
-        }
+        // if (fp16_mma_available(cc) && (ggml_cuda_highest_compiled_arch(cc) <= GGML_CUDA_CC_TURING || cc == GGML_CUDA_CC_TURING)) {
+        //     return BEST_FATTN_KERNEL_WMMA_F16;
+        // }
         return BEST_FATTN_KERNEL_MMA_F16;
     }
 

Original file line number	Diff line number	Diff line change
`@@ -396,9 +396,9 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const`
`396`	`396`	`}`
`397`	`397`
`398`	`398`	`//kcpp: use wmma to fix cu11 incoherence`
`399`		`- if (fp16_mma_available(cc) && (ggml_cuda_highest_compiled_arch(cc) <= GGML_CUDA_CC_TURING \|\| cc == GGML_CUDA_CC_TURING)) {`
`400`		`- return BEST_FATTN_KERNEL_WMMA_F16;`
`401`		`- }`
	`399`	`+ // if (fp16_mma_available(cc) && (ggml_cuda_highest_compiled_arch(cc) <= GGML_CUDA_CC_TURING \|\| cc == GGML_CUDA_CC_TURING)) {`
	`400`	`+ // return BEST_FATTN_KERNEL_WMMA_F16;`
	`401`	`+ // }`
`402`	`402`	`return BEST_FATTN_KERNEL_MMA_F16;`
`403`	`403`	`}`
`404`	`404`