Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Skip trial instead of aborting when CUDA graph capture fails #605

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
noahfarr wants to merge 1 commit into
base: 4.0
Choose a base branch
from
Open

Skip trial instead of aborting when CUDA graph capture fails #605

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
42 changes: 32 additions & 10 deletions src/pufferlib.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -707,12 +707,24 @@ extern "C" void net_callback_wrapper(void* ctx, int buf, int t) {
}

if (capturing) {
// Throw, don't assert: capture runs on the main thread during
// create_pufferl_impl warmup, so this reaches the existing OOM-skip
// handler in pufferl.py instead of abort()ing the whole sweep.
cudaGraph_t _graph;
assert(cudaStreamEndCapture(current_stream, &_graph) == cudaSuccess
&& "cudaStreamEndCapture failed");
assert(cudaGraphInstantiate(&pufferl->fused_rollout_cudagraphs[graph], _graph, 0) == cudaSuccess
&& "cudaGraphInstantiate failed");
assert(cudaGraphDestroy(_graph) == cudaSuccess && "cudaGraphDestroy failed");
cudaError_t end_err = cudaStreamEndCapture(current_stream, &_graph);
if (end_err != cudaSuccess)
throw std::runtime_error(std::string("cudaStreamEndCapture (rollout) failed: ")
+ cudaGetErrorString(end_err));
cudaError_t inst_err = cudaGraphInstantiate(&pufferl->fused_rollout_cudagraphs[graph], _graph, 0);
if (inst_err != cudaSuccess) {
cudaGraphDestroy(_graph);
throw std::runtime_error(std::string("cudaGraphInstantiate (rollout) failed: ")
+ cudaGetErrorString(inst_err));
}
cudaError_t destroy_err = cudaGraphDestroy(_graph);
if (destroy_err != cudaSuccess)
throw std::runtime_error(std::string("cudaGraphDestroy (rollout) failed: ")
+ cudaGetErrorString(destroy_err));
cudaDeviceSynchronize();
}
profile_end(hypers.profile);
Expand Down Expand Up @@ -1643,12 +1655,22 @@ void train_impl(PuffeRL& pufferl) {
pufferl.param_puf.data, pufferl.master_weights.data, n);
}
if (capturing) {
// Throw not assert (see net_callback_wrapper): skip the trial, don't abort.
cudaGraph_t _graph;
assert(cudaStreamEndCapture(train_stream, &_graph) == cudaSuccess
&& "cudaStreamEndCapture failed");
assert(cudaGraphInstantiate(&pufferl.train_cudagraph, _graph, 0) == cudaSuccess
&& "cudaGraphInstantiate failed");
assert(cudaGraphDestroy(_graph) == cudaSuccess && "cudaGraphDestroy failed");
cudaError_t end_err = cudaStreamEndCapture(train_stream, &_graph);
if (end_err != cudaSuccess)
throw std::runtime_error(std::string("cudaStreamEndCapture (train) failed: ")
+ cudaGetErrorString(end_err));
cudaError_t inst_err = cudaGraphInstantiate(&pufferl.train_cudagraph, _graph, 0);
if (inst_err != cudaSuccess) {
cudaGraphDestroy(_graph);
throw std::runtime_error(std::string("cudaGraphInstantiate (train) failed: ")
+ cudaGetErrorString(inst_err));
}
cudaError_t destroy_err = cudaGraphDestroy(_graph);
if (destroy_err != cudaSuccess)
throw std::runtime_error(std::string("cudaGraphDestroy (train) failed: ")
+ cudaGetErrorString(destroy_err));
cudaDeviceSynchronize();
pufferl.train_captured = true;
}
Expand Down
Loading