Fix hf_ptq mllama path not honoring --calib_seq flag

[cjluo-nv]

What does this PR do?

Type of change: Bug fix

examples/llm_ptq/hf_ptq.py accepts a --calib_seq argument to control the max calibration sample length, but the mllama branch of make_calib_dataloader never forwarded it to get_vlm_dataset_dataloader. As a result, mllama VLM calibration silently used the default max_length regardless of what the user passed on the command line.

This complements #1311, which landed the analogous fix for the LLM (get_dataset_dataloader) branch. After this PR, every branch of make_calib_dataloader that has a length concept honors --calib_seq:

Branch	How calib_seq flows
specdec_offline_dataset	EagleOfflineDataCollator(train_len=args.calib_seq) (already correct)
calib_with_images (VLM)	get_vlm_dataset_dataloader(..., max_length=args.calib_seq) (already correct)
mllama (VLM)	get_vlm_dataset_dataloader(..., max_length=args.calib_seq) ← this PR
whisper (speech)	N/A — get_speech_dataset_dataloader has no length parameter; audio is fixed-shape 80×3000 mel features
default (LLM)	get_dataset_dataloader(..., max_sample_length=args.calib_seq) (fixed by #1311)

Usage

python examples/llm_ptq/hf_ptq.py \
    --pyt_ckpt_path <mllama-model> \
    --qformat fp8 \
    --calib_seq 2048

Testing

Validated on Qwen3-8B using a two-part script:

1. Library contract: directly called get_dataset_dataloader with max_sample_length ∈ {64, 128, 512} and confirmed every batch had input_ids.shape[1] ≤ max_sample_length.
2. hf_ptq.py plumbing: invoked make_calib_dataloader with args.calib_seq ∈ {128, 1024} and a mocked inner dataloader — confirmed the spy received max_sample_length=args.calib_seq (LLM path) and max_length=args.calib_seq (VLM path) verbatim.

Before your PR is "Ready for review"

• Is this change backward compatible?: ✅
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: N/A
• Did you write any new necessary tests?: ❌ (single-line fix to an example script; no unit test harness for this path)
• Did you update Changelog?: N/A

Additional Information

Single-line fix in examples/llm_ptq/hf_ptq.py::make_calib_dataloader (mllama branch). Builds on #1311.

Summary by CodeRabbit

• Bug Fixes
  • Calibration dataloader now correctly respects the configured sequence length during calibration runs, ensuring input sequences are truncated/padded to the specified length. This improves consistency and reliability of calibration and quantization results for applicable model types.

[coderabbitai]

Caution

Review failed

Pull request was closed or merged during review

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 9a8b59a1-524b-458e-beb7-135dacb2be94

📥 Commits

Reviewing files that changed from the base of the PR and between ef33b9c and 48489e8.

📒 Files selected for processing (1)

• examples/llm_ptq/hf_ptq.py

---

📝 Walkthrough

Walkthrough

The mllama calibration branch in examples/llm_ptq/hf_ptq.py now passes max_length=args.calib_seq to get_vlm_dataset_dataloader, explicitly constraining text sequence/input length for the mllama calibration dataloader.

Changes

Cohort / File(s)	Summary
mllama calibration dataloader examples/llm_ptq/hf_ptq.py	Added max_length=args.calib_seq argument to get_vlm_dataset_dataloader call in the mllama branch to enforce calibration sequence length limiting.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

🚥 Pre-merge checks | ✅ 5 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and accurately summarizes the main fix: the mllama path in hf_ptq now honors the --calib_seq flag by passing max_length=args.calib_seq to the dataloader.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Security Anti-Patterns	✅ Passed	The PR adds a benign max_length parameter to pass args.calib_seq, with no security anti-patterns or unsafe patterns introduced.

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

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch chenjiel/fix_seq

Warning

Review ran into problems

🔥 Problems

Git: Failed to clone repository. Please run the @coderabbitai full review command to re-trigger a full review. If the issue persists, set path_filters to include or exclude specific files.

---

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.19%. Comparing base (e4e3508) to head (48489e8).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1329      +/-   ##
==========================================
+ Coverage   74.67%   76.19%   +1.52%     
==========================================
  Files         468      468              
  Lines       50374    50374              
==========================================
+ Hits        37615    38383     +768     
+ Misses      12759    11991     -768     

Flag	Coverage Δ
examples	41.58% <ø> (+5.97%)	⬆️
unit	52.53% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[sugunav14]

@cjluo-nv Do we also want to update the speech and vlm cases?

[kevalmorabia97]

Duplicate of #1311

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-04-23 16:10 UTC