musa: workaround for Guilty Lockup in cleaning src0

Signed-off-by: Xiaodong Ye <[email protected]>

Author: yeahdongcn

ggml/src/ggml-cuda.cu

@@ -1484,7 +1484,12 @@ static void ggml_cuda_op_mul_mat(
             const size_t nbytes_data    = ggml_nbytes(src0);
             const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
+        // TODO: remove this for MUSA once the Guilty Lockup issue is resolved
+#ifndef GGML_USE_MUSA
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream));
+#else // GGML_USE_MUSA
+            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
+#endif // !GGML_USE_MUSA
         }
 
         // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared: