Add LSQ NVFP4 export support

[realAsma]

Summary

Adds LSQ NVFP4 support to unified Hugging Face export.

• Uses the LSQ pre scale to pack FP4 weight codes and serializes the post scale for dequantization.
• Supports quantized or unquantized pre scales, tied or separate pre/post amax values, and the covered learnable-amax modes.
• Keeps this PR scoped to the three export/test files; the LSQ rename and quantizer APIs come from Add LSQ (Learned Scale Quantization) support and recipes #1884.

Testing

• pre-commit run --files modelopt/torch/export/quant_utils.py modelopt/torch/export/unified_export_hf.py tests/unit/torch/quantization/test_nvfp4_static_export_cpu.py
• pytest_pwd tests/unit/torch/quantization/test_lsq.py tests/unit/torch/quantization/test_nvfp4_static_export_cpu.py -q -x — 66 passed
• pytest_pwd tests/unit/torch/quantization/test_nvfp4_static_export_cpu.py -q -x — 18 passed

Checklist

• Backward compatible: ✅
• New dependency or copied code: N/A
• Necessary tests added: ✅
• Changelog update: N/A
• Claude approval: N/A (draft)

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 9b98fd9a-312a-463d-8b65-897a2bc465bf

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch asma/laq-export-support

---

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

[realAsma]

BB: Rebase this PR to the latest main, remove the unnecessary changes.

[realAsma]

🤖 Bot comment.

🐝 I’m on it — I’ll rebase onto the latest main, trim the unnecessary changes, and report back here.

[realAsma]

🤖 Bot comment.

Blocked before rebase: this PR still targets asma/laq-algorithm@6bfefb9, which does not contain latest main@f479e78, and the verified rebase path can only rebase onto the PR’s current base. Please either retarget this PR to main or rebase/update the parent branch onto main; then I can safely rebase and trim the duplicated changes.

[realAsma]

BB: Can you rebase this PR to latest main and only add the changes required to enable export with LSQ? (btw, I have renamed LAQ to LSQ, see #1884)

[realAsma]

BB: Please update PR description and PR title after addressing #1847 (comment)

[realAsma]

🤖 Bot comment.

🐝 I’m on it — I’ll rebase onto the latest main, then narrow the PR to the changes needed for LSQ export.

[realAsma]

🤖 Bot comment.

🐝 Acknowledged this instruction. I’ll update the PR title and description after addressing the linked feedback.

[realAsma]

🤖 Bot comment.

🐝 Got it — I’ll update the PR title and description to match the narrowed LSQ-export scope after the branch changes are prepared.

[realAsma]

🤖 Bot comment.

Blocked for now: the PR still declares asma/laq-algorithm as its base, so the guarded rebase followed that base and paused locally on conflicts instead of rewriting onto main. Nothing was committed or pushed. The title/description update is also deferred until an LSQ-only candidate exists. The next safe step is to retarget this PR to main, clear the wrong-base local rebase, and rerun the LSQ-only preparation.

[realAsma]

🤖 Bot comment.

Blocked outcome for the title/description request: the guarded pass paused because the PR still targets asma/laq-algorithm, not main. No commit, push, title, or description change was made. I’ll update the metadata after a correct-base LSQ-only candidate is prepared; details are in the branch-work outcome.

[realAsma]

BB: Yes please retarget this PR to latest main. then address #1847 (comment)

[realAsma]

🤖 Bot comment.

🐝 Acknowledged — I’ll retarget this PR to the latest main, then address the LSQ export request.

[realAsma]

🤖 Bot comment.

Blocked after the requested retarget attempt: I safely cleared the unpublished wrong-base rebase and restored asma/laq-export-support cleanly at the exact remote head. The guarded GitHub interface available in this run cannot edit a PR’s base branch, so I could not change the live base to main or start the LSQ-only fix pass. Please retarget this PR to main; the active monitor will pick it up and continue automatically.

[realAsma]

BB: Done, I manually rebase to main. do the rest.

[realAsma]

🤖 Bot comment.

Update: the live PR base is now main at 0baad6453a8, and the canonical head remains clean at 30e8b91986. This retarget blocker is cleared. The enabled controller monitor will start the fresh LSQ-only pass under the next dispatch; no code or history was published in this completed one-pass dispatch.

[realAsma]

🤖 Bot comment.

🐝 Got it — I’m taking over from the manual rebase and will finish the LSQ export work, validation, and PR metadata updates.

[realAsma]

🤖 Bot comment.

I verified the PR now targets main, but the branch head itself was not rebased: it is still on the old commit, 195 commits behind, and GitHub reports conflicts. I preserved the clean branch and recorded the exact guarded rebase handoff; the next monitored controller pass will rebase safely, then prepare the LSQ-export-only changes and update the title/description. No code or history was published in this pass.

[realAsma]

🤖 Bot comment.

🐝 Updated this PR to the requested LSQ-export-only scope.

• Rebased cleanly onto current main with one signed and signed-off commit (448e58e3c).
• Live diff is limited to the two export files and one focused CPU test file.
• PR title and description now use LSQ terminology and reference Add LSQ (Learned Scale Quantization) support and recipes #1884 for the LSQ rename/API dependency.
• File-scoped pre-commit passed; focused tests passed (66 passed, including the 18 passed export subset).
• GitHub now reports the PR mergeable. DCO and initial file-change checks pass; the remaining fresh CI jobs are running.

The PR remains draft; no ready-for-review promotion or reviewer notification was made.

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA.github.io/Model-Optimizer/pr-preview/pr-1847/
Built to branch gh-pages at 2026-07-22 23:13 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[codecov]

Codecov Report

❌ Patch coverage is 97.72727% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 75.86%. Comparing base (0baad64) to head (448e58e).

Files with missing lines	Patch %	Lines
modelopt/torch/export/quant_utils.py	96.55%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1847      +/-   ##
==========================================
+ Coverage   75.84%   75.86%   +0.01%     
==========================================
  Files         518      518              
  Lines       58544    58586      +42     
==========================================
+ Hits        44404    44445      +41     
- Misses      14140    14141       +1     

Flag	Coverage Δ
unit	55.09% <97.72%> (+0.04%)	⬆️

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

☔ View full report in Codecov by Harness.
📢 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.