[6648996] Fix NVFP4 ONNX packed-weight scale rounding

[ajrasane]

What does this PR do?

Type of change: Bug fix

This PR updates nvfp4_exporter.py so NVFP4 ONNX weight compression quantizes FP4 weights with the same FP8-rounded per-block scales serialized in the exported graph. It aligns quant_utils.py scale arithmetic with eager ModelOpt's FP32 operation ordering and applies the same behavior to the deprecated fp4qdq_to_2dq compatibility path in qdq_utils.py.

test_onnx_export_cpu.py adds regression coverage comparing exported FP8 scale bytes and packed FP4 weights against NVFP4QTensor at FP8-rounding, per-tensor arithmetic, and per-block arithmetic boundaries. CHANGELOG.rst documents the fix.

Usage

from modelopt.onnx.export import NVFP4QuantExporter

converted_model = NVFP4QuantExporter.process_model(model)

Testing

Ran:

CUDA_VISIBLE_DEVICES="" python -m pytest -q \
  tests/unit/torch/quantization/test_onnx_export_cpu.py \
  tests/unit/torch/quantization/test_nvfp4_tensor.py \
  tests/unit/onnx/quantization/test_qdq_utils.py \
  tests/unit/onnx/quantization/test_quant_utils.py

All 89 tests passed. All applicable pre-commit hooks passed. Additionally verified exact FP8-scale and packed-FP4 parity against eager quantization for 8,000 randomized tensors.

Before your PR is "Ready for review"

Make sure you read and follow Contributor guidelines and your commits are signed (git commit -s -S).

Make sure you read and follow the Security Best Practices (e.g. avoiding hardcoded trust_remote_code=True, torch.load(..., weights_only=False), pickle, etc.).

• 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?: ✅
• Did you update Changelog?: ✅

🤖 Generated by Codex (AI agent).

[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

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

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

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA.github.io/Model-Optimizer/pr-preview/pr-2243/
Built to branch gh-pages at 2026-08-25 00:13 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[codecov]

Codecov Report

❌ Patch coverage is 95.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 79.01%. Comparing base (73d7784) to head (93174f5).

Files with missing lines	Patch %	Lines
modelopt/onnx/export/nvfp4_exporter.py	92.85%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2243   +/-   ##
=======================================
  Coverage   79.01%   79.01%           
=======================================
  Files         523      523           
  Lines       60695    60705   +10     
=======================================
+ Hits        47960    47969    +9     
- Misses      12735    12736    +1     

Flag	Coverage Δ
unit	55.67% <95.00%> (+<0.01%)	⬆️

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.