[bridge] Fix off-by-one in sliding window size for Gemma2, Gemma3, Mistral, and GPT-OSS

[cuichenx]

What does this PR do?

Fix off-by-one error in sliding window attention size for Gemma2, Gemma3, Mistral, and GPT-OSS bridges.

FlashAttention's window_size tuple (left, right) uses inclusive bounds — (W, 0) attends to W + 1 tokens (W preceding + current). Since HuggingFace sliding_window defines the total window size (including the current token), we must subtract 1 when converting to the tuple form. This convention is also observed by Transformer Engine.

Changelog

• Fix Gemma2 bridge to subtract 1 from sliding_window (was passing raw HF value)
• Fix Gemma3 provider to subtract 1 when converting int window_size to tuple at the layer level (was passing raw int)
• Fix Mistral bridge to subtract 1 from sliding_window (was passing raw HF value)
• Fix GPT-OSS bridge to read sliding_window from HF config instead of hardcoding 128, and subtract 1
• Fix GPT-OSS predefined provider default from (128, 0) to (127, 0)

GitHub Actions CI

See the CI section in the Contributing doc for how to trigger the CI.
A Nvidia developer will need to approve and trigger the CI for external contributors.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

Additional Information

Thanks to @returnL for catching this in NVIDIA/Megatron-LM#2771

[coderabbitai]

📝 Walkthrough

Walkthrough

The changes adjust sliding window size calculations in three model bridge implementations. In Gemma2Bridge, Gemma3TEDotProductAttention, and GPT-OSS bridge, window_size is modified to subtract 1 from the computed or configured window dimensions, affecting local attention behavior.

Changes

Cohort / File(s)	Summary
Gemma Model Bridges src/megatron/bridge/models/gemma/gemma2_bridge.py, src/megatron/bridge/models/gemma/gemma3_provider.py	Adjusted sliding window size calculations to subtract 1 from configured window dimensions in attention layer configuration.
GPT-OSS Bridge src/megatron/bridge/models/gpt_oss/gpt_oss_bridge.py	Changed window_size configuration from fixed value to dynamically computed as (sliding_window - 1).

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

Suggested labels

r0.3.0

🚥 Pre-merge checks | ✅ 2 | ❌ 2

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 66.67% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.
Test Results For Major Changes	⚠️ Warning	PR contains numerical changes affecting sliding window attention computation but lacks actual test results demonstrating no regression.	Add actual test results or GitHub Actions CI/CD logs showing successful test execution and logit validation.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main changes: fixing an off-by-one error in sliding window size calculations across multiple model bridges (Gemma2, Gemma3, and GPT-OSS).

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

• 📝 Generate docstrings (stacked PR)
• 📝 Generate docstrings (commit on current branch)

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch chcui/fix-sliding-window-off-by-one

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[cuichenx]

/ok to test 237efef

[cuichenx]

/ok to test 871dee9