Skip to content

add NVSHMEM_DISABLE_NCCL to avoid cuda graph hang #1374

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

Merged
maocheng23 merged 1 commit into from
Jun 22, 2026
Merged

add NVSHMEM_DISABLE_NCCL to avoid cuda graph hang #1374

Show file tree
Hide file tree
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
5 changes: 5 additions & 0 deletions docs/platforms/nvidia.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,6 +65,11 @@ NVTE_FUSED_ATTN=1 # default, but verify
TORCHINDUCTOR_CACHE_DIR=/data/.inductor
```

miles sets `NVSHMEM_DISABLE_NCCL=1` on rollout and training actors automatically. DeepEP
initializes NVSHMEM, whose built-in NCCL path opens a second NCCL communicator that collides
with miles' own NCCL and hangs during SGLang CUDA-graph replay. Export `NVSHMEM_DISABLE_NCCL=0`
to opt out.

## NVLink + IB topology

For 8× GPUs per node:
Expand Down
2 changes: 2 additions & 0 deletions miles/ray/actor_group.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,6 +56,8 @@ def _allocate_gpus_for_actor(self, pg, num_gpus_per_actor):
# we need also set it to 0 to prevent nccl error.
"NCCL_CUMEM_ENABLE": os.environ.get("NCCL_CUMEM_ENABLE", "0"),
"NVTE_FP8_BLOCK_SCALING_FP32_SCALES": "1",
# DeepEP/NVSHMEM's internal NCCL conflicts with our NCCL and hangs under CUDA graphs.
"NVSHMEM_DISABLE_NCCL": os.environ.get("NVSHMEM_DISABLE_NCCL", "1"),
**{name: "1" for name in NOSET_VISIBLE_DEVICES_ENV_VARS_LIST},
**self.args.train_env_vars,
}
Expand Down
2 changes: 2 additions & 0 deletions miles/ray/rollout/server_group.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -99,6 +99,8 @@ def start_engines(
env_vars = {name: "1" for name in NOSET_VISIBLE_DEVICES_ENV_VARS_LIST} | {
key: os.environ.get(key, default_val)
for key, default_val in {
# DeepEP/NVSHMEM's internal NCCL conflicts with our NCCL and hangs under CUDA graphs.
"NVSHMEM_DISABLE_NCCL": "1",
"SGLANG_JIT_DEEPGEMM_PRECOMPILE": "false",
# TODO: this is hacky. Use env var SGLANG_DG_CACHE_DIR_PER_PROCESS=1
# to enable this isolation.
Expand Down
Loading