Use cublasGemmEx instead of cublasSgemm in ggml

Removes the latter symbol from our dependencies.
Mozilla-Ocho · Dec 11, 2023 · 252a255 · 252a255
1 parent 4fb0813
commit 252a255
Showing 1 changed file with 6 additions and 4 deletions.
diff --git a/llama.cpp/ggml-cuda.cu b/llama.cpp/ggml-cuda.cu
@@ -6762,11 +6762,13 @@ inline void ggml_cuda_op_mul_mat_cublas(
 
         CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
         CUBLAS_CHECK(
-            cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
+            cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
                     row_diff, src1_ncols, ne10,
-                    &alpha, src0_ddf_i, ne00,
-                            src1_ddf_i, ne10,
-                    &beta,  dst_dd_i,   ldc));
+                    &alpha, src0_ddf_i, CUDA_R_32F, ne00,
+                            src1_ddf_i, CUDA_R_32F, ne10,
+                    &beta,  dst_dd_i,   CUDA_R_32F, ldc,
+                    CUBLAS_COMPUTE_32F,
+                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
 
         if (src0_as != 0) {
             ggml_cuda_pool_free(src0_ddq_as_f32, src0_as);