Fix issue with setAttribute and int8_t vs int32_t variables (pytorch#…

…143693) Test Plan: Sandcastle Pull Request resolved: pytorch#143693 Approved by: https://github.com/huydhn
abhishek-iitmadras · Dec 21, 2024 · 9f3c291 · 9f3c291
1 parent 518b505
commit 9f3c291
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/aten/src/ATen/cuda/CUDABlas.cpp b/aten/src/ATen/cuda/CUDABlas.cpp
@@ -1453,7 +1453,8 @@ void scaled_gemm(
     computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_D_SCALE_POINTER, result_scale_ptr);
   }
 #ifndef USE_ROCM
-  computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_FAST_ACCUM, use_fast_accum ? 1 : 0);
+  const int8_t fastAccuMode = use_fast_accum ? 1 : 0;
+  computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_FAST_ACCUM, fastAccuMode);
 #endif
   CuBlasLtMatrixLayout Adesc(ScalarTypeToCudaDataType(mat1_dtype), m, k, mat1_ld, transa == 't');
   CuBlasLtMatrixLayout Bdesc(ScalarTypeToCudaDataType(mat2_dtype), k, n, mat2_ld, transb == 't');