HIP: workaround runtime bug in hipGraph support

ggml-org · Feb 19, 2025 · dbf12a6 · dbf12a6
1 parent 300907b
commit dbf12a6
Showing 1 changed file with 7 additions and 0 deletions.
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -2558,7 +2558,14 @@ static void maintain_cuda_graph(ggml_backend_cuda_context * cuda_ctx, std::vecto
         for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
             if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
                 char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
+#if defined(__HIP_PLATFORM_AMD__)
+                // Workaroudnd for https://github.com/ROCm/clr/issues/138
+                // The hip runtime fails to copy this and calls delete on it later, so we must
+                // perform an alloc here.
+                cuda_ctx->cuda_graph->params[i].kernelParams[1] = new void*(*updated_kernel_arg_ptr);
+#else
                 cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
+#endif // defined(__HIP_PLATFORM_AMD__)
                 CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
             }
         }