deepseek-ai/DeepEP
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepEP.git synced 2026-11-05 00:47:32 +00:00
Code Issues 308 Projects Packages Wiki Activity

[GH-ISSUE #506] Internode combine_token using TMA fails assertion for GPTOSS hidden size of 2880 #327

New issue
Open
opened 2026-10-26 09:58:39 +00:00 by bsakaguchi · 0 comments
bsakaguchi commented 2026-10-26 09:58:39 +00:00
Owner
Copy link

Originally created by @hemildesai on GitHub (Nov 29, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/506

https://github.com/deepseek-ai/DeepEP/pull/320 added TMA for combine forwarder in combine_token. This also added the assertion that hidden_int4 % 32 == 0. However, this doesn't work for GPTOSS hidden size of 2880.

This can be reproduced by running the following command in a multi node setup

python tests/test_internode.py --hidden 2880

The decision to use TMA is controlled by the template argument here, but it's always set to true here. Setting kUseTMA to false works for now, but would be good to have a long term solution for this.

One possible workaround is to only set kUseTMA to true when hidden_int4 % 32 == 0, false otherwise.

Originally created by @hemildesai on GitHub (Nov 29, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/506 https://github.com/deepseek-ai/DeepEP/pull/320 added TMA for combine forwarder in `combine_token`. This also added the [assertion](https://github.com/deepseek-ai/DeepEP/blob/main/csrc/kernels/internode.cu#L1583) that `hidden_int4 % 32 == 0`. However, this doesn't work for GPTOSS hidden size of 2880. This can be reproduced by running the following command in a multi node setup ```bash python tests/test_internode.py --hidden 2880 ``` The decision to use TMA is controlled by the template argument [here](https://github.com/deepseek-ai/DeepEP/blob/main/csrc/kernels/internode.cu#L1545), but it's always set to `true` [here](https://github.com/deepseek-ai/DeepEP/blob/main/csrc/kernels/internode.cu#L2086). Setting `kUseTMA` to `false` works for now, but would be good to have a long term solution for this. One possible workaround is to only set `kUseTMA` to `true` when hidden_int4 % 32 == 0, `false` otherwise.
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
hybrid-ep
chore/remove-clang-format-yapf
v2.5
nvDev
try-fix-cu128
mori-ep
antgroup-opt
compat-old-torch
add-linter
eager-rdma
v1.2.1
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepEP#327
No description provided.