Skip to content

[Bugfix] Fix MoE Model DP+TP with NaiveAll2AllManager Bug #32705

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
River12 wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[Bugfix] Fix MoE Model DP+TP with NaiveAll2AllManager Bug #32705

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1f73a3b
Fix MoE Model DP+TP with NaiveAll2AllManager Bug (#32705)
Feb 2, 2026
451cd61
Merge branch 'main' into export-D91016491
River12 Feb 2, 2026
884d2e3
Merge branch 'main' into export-D91016491
River12 Feb 5, 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
6 changes: 4 additions & 2 deletions vllm/distributed/device_communicators/all2all.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,14 +48,15 @@ def naive_multicast(

rank = self.rank if is_sequence_parallel else self.dp_rank
world_size = self.world_size if is_sequence_parallel else self.dp_world_size
dist_group = get_ep_group() if is_sequence_parallel else get_dp_group()

start = 0 if rank == 0 else cu_tokens_across_sp_cpu[rank - 1]
end = cu_tokens_across_sp_cpu[rank]
buffer[start:end, :].copy_(x)
for idx in range(world_size):
start = 0 if idx == 0 else cu_tokens_across_sp_cpu[idx - 1]
end = cu_tokens_across_sp_cpu[idx]
get_ep_group().broadcast(buffer[start:end, :], idx)
dist_group.broadcast(buffer[start:end, :], idx)

return buffer

Expand Down Expand Up @@ -125,7 +126,8 @@ def combine(
start = 0 if ep_rank == 0 else cu_tokens_across_sp_cpu[ep_rank - 1]
end = cu_tokens_across_sp_cpu[ep_rank]

all_hidden_states = get_ep_group().all_reduce(hidden_states)
dist_group = get_ep_group() if is_sequence_parallel else get_dp_group()
all_hidden_states = dist_group.all_reduce(hidden_states)
hidden_states = all_hidden_states[start:end, :]
return hidden_states

Expand Down