feat(export): add functional quantized weight export

[mxinO]

Summary

Adds a functional per-weight export API for online real-quant refit.

• ModelOpt resolves the exact weight quantizer and captures current quantizer state without mutating the module.
• Opaque state supports topology-aligned merge and select operations.
• ModelOpt emits canonical Hugging Face packed weights, sidecars, and quantization_config.
• Reusable export specs contain metadata only; mutable amax and scale tensors are recaptured for each export stream.

The API supports non-integer FP8, MXFP8, MXFP4, NVFP4 variants, and mixed per-layer configurations. Ordinary offline export remains on its existing path. AWQ, SmoothQuant, SVDQuant, and integer weight formats are outside this PR.

Usage

spec = get_quantized_weight_export_spec(module, "weight")
state = capture_quantized_weight_export_state(module, "weight")
exported = export_quantized_weight_tensors(weight, state, dtype)
config = build_hf_quantization_config({"model.layers.0.mlp.weight": spec})

Validation

• 24 focused functional-export tests passed.
• 32 existing offline-export tests passed across Transformers, Diffusers, streaming/offload, and AWQ fallback coverage.
• ModelOpt pre-commit hooks passed.
• Downstream two-step W4A16/NVFP4 QA-GRPO completed two refits with finite loss and KL.

Dependency

Megatron-Bridge integration: NVIDIA-NeMo/Megatron-Bridge#5812

Checklist

• Added focused regression tests.
• Preserved ordinary offline-export behavior.
• No new dependency.
• Changelog update, if required before review.

[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-2251/
Built to branch gh-pages at 2026-08-29 07:47 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[codecov]

Codecov Report

❌ Patch coverage is 85.33333% with 55 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.91%. Comparing base (022767c) to head (8802ee7).

Files with missing lines	Patch %	Lines
modelopt/torch/export/quant_utils.py	85.09%	55 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2251      +/-   ##
==========================================
- Coverage   78.95%   78.91%   -0.05%     
==========================================
  Files         524      524              
  Lines       60866    61186     +320     
==========================================
+ Hits        48058    48284     +226     
- Misses      12808    12902      +94     

Flag	Coverage Δ
unit	56.01% <85.33%> (+0.21%)	⬆️

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.