Skip to content

[https://nvbugs/5930934][fix] Fix OOM hang with NCCL_SYMMETRIC fallback during long-context inference #11870

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
pcastonguay merged 1 commit into from
Mar 4, 2026
Merged

[https://nvbugs/5930934][fix] Fix OOM hang with NCCL_SYMMETRIC fallback during long-context inference #11870

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
6 changes: 5 additions & 1 deletion tensorrt_llm/_torch/custom_ops/torch_custom_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1769,7 +1769,11 @@ def forward(
do_preparation=True)
return input
if tactic == -1:
tactic = AllReduceStrategy.NCCL_SYMMETRIC.value
# tactic == -1 means the autotuner cache missed for this shape,
# so we fall back to plain NCCL instead of NCCL_SYMMETRIC.
# NCCL_SYMMETRIC requires ncclMemAlloc which can fail asymmetrically
# across ranks under OOM, causing a deadlock at ncclCommWindowRegister.
tactic = AllReduceStrategy.NCCL.value
Comment thread
Tabrizian marked this conversation as resolved.

return torch.ops.trtllm.allreduce(
input,
Expand Down
6 changes: 4 additions & 2 deletions tensorrt_llm/_torch/models/modeling_deepseekv3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,7 +139,8 @@ def weight_dequant(x: torch.Tensor,
@torch.compile(dynamic=True)
def moe_reduce_add_shared_output(routed_output, shared_output):
routed_output = torch.sum(routed_output, dim=1, keepdim=False)
return shared_output + routed_output
# In-place add to avoid allocating a temporary tensor, reducing peak memory
return shared_output.add_(routed_output)


class DeepseekV3WeightLoader:
Expand Down Expand Up @@ -1157,7 +1158,8 @@ def _compute_routed_output():
else:
assert shared_output.size() == routed_output.size(
), 'unmatched tensor shape'
final_hidden_states = shared_output + routed_output
# In-place add to avoid allocating a temporary tensor, reducing peak memory
final_hidden_states = shared_output.add_(routed_output)

if not self.use_dp and self.mapping.tp_size > 1:
final_hidden_states = self.allreduce(
Expand Down
Loading