Method commit

mlx/backend/cuda/device.cpp:463–536

}

void CommandEncoder::commit() {
  nvtx3::scoped_range r("CommandEncoder::commit");
  if (!temporaries_.empty()) {
    add_completed_handler([temporaries = std::move(temporaries_)]() {});
  }
  if (use_cuda_graphs() && node_count_ > 0) {
    if (!from_nodes_.empty()) {
#if CUDART_VERSION >= 13000
      CHECK_CUDA_ERROR(cudaGraphAddDependencies(
          graph_,
          from_nodes_.data(),
          to_nodes_.data(),
          nullptr, // edgeData
          from_nodes_.size()));
#else
      CHECK_CUDA_ERROR(cudaGraphAddDependencies(
          graph_, from_nodes_.data(), to_nodes_.data(), from_nodes_.size()));
#endif
    }

    device_.make_current();

    if (!is_graph_updatable_) {
      CudaGraphExec graph_exec;
      graph_exec.instantiate(graph_);
      CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
    } else {
      auto graph_key = graph_nodes_key_ + ":" + graph_deps_key_;
      auto& graph_exec = graph_cache_[graph_key];

      if (graph_exec != nullptr) {
        cudaGraphExecUpdateResult update_result;
#if CUDART_VERSION >= 12000
        cudaGraphExecUpdateResultInfo info;
        cudaGraphExecUpdate(graph_exec, graph_, &info);
        update_result = info.result;
#else
        cudaGraphNode_t error_node;
        cudaGraphExecUpdate(graph_exec, graph_, &error_node, &update_result);
#endif // CUDART_VERSION >= 12000
        if (update_result != cudaGraphExecUpdateSuccess) {
          cudaGetLastError(); // reset error
          graph_exec.reset();
        }
      }
      if (graph_exec == nullptr) {
        graph_exec.instantiate(graph_);
      }

      CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
    }

    // Save cuda graph to dot file
    if (const char* filename = save_cuda_graphs_dot_file(); filename) {
      static int count = 0;
      auto path = fmt::format("{}_{}.dot", filename, ++count);
      CHECK_CUDA_ERROR(cudaGraphDebugDotPrint(graph_, path.c_str(), 0));
    }