Skip to content

fix(inkling): support HF 5.14 attention fields and embed norm FSDP - #3281

Merged
HuiyingLi merged 2 commits into
mainfrom
huiyingl/fix/inkling-hf514-fsdp
Jul 29, 2026
Merged

HuiyingLi merged 2 commits into
mainfrom
huiyingl/fix/inkling-hf514-fsdp

Conversation

@HuiyingLi

Copy link
Copy Markdown
Contributor

Summary

  • accept both the legacy attention_type field and the Hugging Face 5.14 layer_type field when selecting Inkling causal masks
  • map hybrid and hybrid_sliding layer types to their corresponding causal-mask entries
  • give an optional text embed_norm its own FSDP unit and extend the parallelizer unit test

Root cause

Hugging Face 5.14 renamed the per-layer attention selector and stores Inkling-specific layer values instead of causal-mask keys. Separately, the multimodal forward invokes language_model.embed_norm(...) before entering the text model's FSDP root forward, so a root-owned sharded weight can be used with a plain activation before the root pre-forward unshard hook runs.

Validation

  • targeted parallelizer FSDP unit test
  • Inkling logit parity test on Transformers 5.14.1
  • one-node, eight-GPU EP8 distributed smoke: two training steps plus validation
  • git diff --check

@copy-pr-bot

copy-pr-bot Bot commented Jul 29, 2026

Copy link
Copy Markdown

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.

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>
@HuiyingLi

Copy link
Copy Markdown
Contributor Author

/claude review

@HuiyingLi
HuiyingLi force-pushed the huiyingl/fix/inkling-hf514-fsdp branch from 7b84f5d to c01b6c2 Compare July 29, 2026 03:38
@HuiyingLi
HuiyingLi marked this pull request as ready for review July 29, 2026 03:40
@HuiyingLi
HuiyingLi requested a review from a team as a code owner July 29, 2026 03:40
@HuiyingLi

Copy link
Copy Markdown
Contributor Author

/ok to test 7b84f5d

@copy-pr-bot

copy-pr-bot Bot commented Jul 29, 2026

Copy link
Copy Markdown

/ok to test 7b84f5d

@HuiyingLi, there was an error processing your request: E2

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/2/

@HuiyingLi

Copy link
Copy Markdown
Contributor Author

/ok to test c01b6c2

@claude

claude Bot commented Jul 29, 2026

Copy link
Copy Markdown
Contributor

LGTM

@HuiyingLi

Copy link
Copy Markdown
Contributor Author

/ok to test 7d698b4

@HuiyingLi
HuiyingLi merged commit e111add into main Jul 29, 2026
83 checks passed
@HuiyingLi
HuiyingLi deleted the huiyingl/fix/inkling-hf514-fsdp branch July 29, 2026 07:23

This branch was previously deployed

3 inactive deployments
public — 7d698b40 Deployed Jul 29, 2026 by copy-pr-bot[bot] via release / finalize / notify #3035
test — 7d698b40 Deployed Jul 29, 2026 by copy-pr-bot[bot] via cicd-wait-in-queue #9218
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants