Skip to content

[Bugfix] DeepEP-V2: expert_tokens_meta must be None on the decode/cudagraph path (empty recv_expert_num_tokens) #52632

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
ywang96 merged 5 commits into from
Aug 20, 2026
Merged

[Bugfix] DeepEP-V2: expert_tokens_meta must be None on the decode/cudagraph path (empty recv_expert_num_tokens) #52632

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
5b294d9
[BugFix] DeepEP-V2: return expert_tokens_meta=None on the decode/cuda…
dmvevents Aug 17, 2026
40ade20
Merge branch 'main' into fix/deepep-v2-empty-expert-tokens-meta
ywang96 Aug 19, 2026
6a05da7
[Bugfix] DeepEP-V2: assert expert_tokens_meta is not None on prefill …
ywang96 Aug 19, 2026
79bdbc7
Merge branch 'main' into fix/deepep-v2-empty-expert-tokens-meta
ywang96 Aug 19, 2026
b67e5b4
Merge branch 'main' into fix/deepep-v2-empty-expert-tokens-meta
ywang96 Aug 19, 2026
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
17 changes: 13 additions & 4 deletions vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -210,14 +210,23 @@ def _receiver(
else:
expert_x, expert_x_scale = recv_x, None

expert_tokens_meta = mk.ExpertTokensMetadata.make_from_list(
recv_expert_num_tokens,
device=expert_x.device,
)
if recv_expert_num_tokens:
expert_tokens_meta = mk.ExpertTokensMetadata.make_from_list(
recv_expert_num_tokens,
device=expert_x.device,
)
else:
# Decode/cudagraph path (do_cpu_sync=False) skips the CPU sync and
# leaves recv_expert_num_tokens empty. A present-but-empty
# ExpertTokensMetadata violates the decode-mode contract above
# (expert_tokens_meta must be None) and crashes DeepEP combine
# during profile_run when CUDA graphs are enabled.
expert_tokens_meta = None

if recv_topk_idx is None:
# do_expand=True (prefill mode): build topk_ids from
# per-expert token counts.
assert expert_tokens_meta is not None
total_tokens = sum(recv_expert_num_tokens)
if total_tokens > 0:
recv_topk_idx = torch.repeat_interleave(
Expand Down
Loading