Skip to content

fix: fix gemma3#2185

Merged
terrykong merged 8 commits intomainfrom
yukih/fix-fsdp2-move-device
Apr 5, 2026
Merged

fix: fix gemma3#2185
terrykong merged 8 commits intomainfrom
yukih/fix-fsdp2-move-device

Conversation

@yuki-97
Copy link
Copy Markdown
Contributor

@yuki-97 yuki-97 commented Apr 1, 2026 •
edited
Loading

as title, fix gemma3 after transformers v5 bump

  1. fix model layer name and add missing token_type_ids for gemma3
  2. fix broadcast buffers in a deterministic order in fsdp2 (dtensor v1)

fix the following release/perf tests:

  • grpo-gemma3-27b-it-8n8g-fsdp2tp8-actckpt-long
  • grpo-gemma3-27b-it-8n4g-fsdp2tp4-actckpt-long

@yuki-97 yuki-97 requested review from a team as code owners April 1, 2026 15:54
@copy-pr-bot
Copy link
Copy Markdown

copy-pr-bot Bot commented Apr 1, 2026

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@yuki-97 yuki-97 added the CI:Lfast Runs a fast test suite and re-use nightly `main` container (but sync dependencies to PRs version) label Apr 1, 2026
@yuki-97
Copy link
Copy Markdown
Contributor Author

yuki-97 commented Apr 1, 2026

/ok to test 55b9d86

@yuki-97 yuki-97 requested a review from terrykong April 1, 2026 15:55
terrykong
terrykong previously approved these changes Apr 1, 2026
View reviewed changes
@terrykong terrykong enabled auto-merge (squash) April 1, 2026 16:00
@yuki-97 yuki-97 marked this pull request as draft April 1, 2026 17:02
auto-merge was automatically disabled April 1, 2026 17:02

Pull request was converted to draft

@yuki-97 yuki-97 force-pushed the yukih/fix-fsdp2-move-device branch from 55b9d86 to c4f3002 Compare April 3, 2026 08:48
@yuki-97 yuki-97 changed the title fix: fix fsdp2 (dtensor v1) move to device fix: fix fsdp2 (dtensor v1) gemma3 Apr 3, 2026
@yuki-97 yuki-97 force-pushed the yukih/fix-fsdp2-move-device branch 3 times, most recently from 9d81d75 to b974855 Compare April 3, 2026 09:21
@yuki-97
Copy link
Copy Markdown
Contributor Author

yuki-97 commented Apr 3, 2026

/ok to test b974855

@yuki-97
Copy link
Copy Markdown
Contributor Author

yuki-97 commented Apr 3, 2026

/ok to test 007a51d

yuki-97 added 4 commits April 3, 2026 06:04
@yuki-97
fix gemma3 model layer name
d38c301 
Signed-off-by: Yuki Huang <yukih@nvidia.com>
@yuki-97
fix missing token_type_ids
50051ea 
Signed-off-by: Yuki Huang <yukih@nvidia.com>
@yuki-97
clear tie weights
6d96cf6 
Signed-off-by: Yuki Huang <yukih@nvidia.com>
@yuki-97
add missing token_type_ids in automodel
e88218a 
Signed-off-by: Yuki Huang <yukih@nvidia.com>
@yuki-97 yuki-97 force-pushed the yukih/fix-fsdp2-move-device branch from 007a51d to e88218a Compare April 3, 2026 13:49
@yuki-97
clear tie weights in automodel
b901c47 
Signed-off-by: Yuki Huang <yukih@nvidia.com>
@yuki-97 yuki-97 changed the title fix: fix fsdp2 (dtensor v1) gemma3 fix: fix gemma3 Apr 3, 2026
yuki-97 added 2 commits April 4, 2026 10:36
@yuki-97
add functional test
491f095 
Signed-off-by: Yuki Huang <yukih@nvidia.com>
@yuki-97
fix broadcast buffers in a deterministic order
c9a8066 
Signed-off-by: Yuki Huang <yukih@nvidia.com>
@yuki-97
Copy link
Copy Markdown
Contributor Author

yuki-97 commented Apr 4, 2026

/ok to test c9a8066

@yuki-97 yuki-97 marked this pull request as ready for review April 5, 2026 04:10
@yuki-97
update unit test
880a0b8 
Signed-off-by: Yuki Huang <yukih@nvidia.com>
@yuki-97
Copy link
Copy Markdown
Contributor Author

yuki-97 commented Apr 5, 2026

/ok to test 880a0b8

@yuki-97 yuki-97 mentioned this pull request Apr 5, 2026
Merged
@yuki-97 yuki-97 mentioned this pull request Apr 5, 2026
Merged
@yuki-97 yuki-97 requested a review from yfw April 5, 2026 05:53
@terrykong terrykong enabled auto-merge (squash) April 5, 2026 06:12
@terrykong terrykong merged commit 986d48a into main Apr 5, 2026
45 of 48 checks passed
@terrykong terrykong deleted the yukih/fix-fsdp2-move-device branch April 5, 2026 06:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@terrykong terrykong terrykong approved these changes

@yfw yfw Awaiting requested review from yfw

Assignees

No one assigned

Labels

CI:Lfast Runs a fast test suite and re-use nightly `main` container (but sync dependencies to PRs version)

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@yuki-97 @terrykong