Don't carry source tensors a dequantizing load already folded into the weights

[shengliangxu]

What does this PR do?

Type of change: Bug fix

The export copies every source tensor unplaced_source_keys reports into the checkpoint verbatim. That pass replayed only the model's stock conversion mapping, so it missed whatever loading added on top, and those additions consume tensors the stock mapping never mentions:

• Dequantize-on-load quantizers. Mxfp4Config(dequantize=True), which examples/hf_ptq uses for gpt-oss, folds each expert's *_blocks/*_scales into a bf16 gate_up_proj/down_proj. FineGrainedFP8Config(dequantize=True) folds every FP8 weight_scale_inv into its weight.
• A patched mapping. An on-the-fly compressed-tensors decompressor feeds weight_packed/weight_scale experts in.

Each such tensor read as unplaced, so the export carried it in beside the weight it had become: stale pre-quantization copies under the source names, plus matching exclude_modules entries.

• Where it was found: an expert-parallel Kimi-K3 export through such a decompressor carried 494,592 packed expert tensors. The checkpoint grew from 1.5 TB to 2.8 TB, and rank 0 spent 21 of the run's 37 minutes writing them.
• Where it hits on main today: gpt-oss. A tiny gpt-oss checkpoint loaded the way hf_ptq loads it reports all four expert block/scale tensors as unplaced.

Fix: from_pretrained records the conversions it actually ran on the model (model._weight_conversions, the list transformers' own save path replays, and the one our quant-aware reverse conversion in quant_aware_conversion.py already prefers), with quantizer and patch additions included. A key now counts as placed when either that record or the stock mapping resolves it to a parameter. The record keeps only conversions some key used, so the stock mapping still answers for weights one of our own loaders placed (the FSDP2 meta-device load).

Usage

N/A: internal to the export's carry-over of unplaced weights.

Testing

• New tests/unit/torch/utils/test_unplaced_source_keys.py. It runs a real from_pretrained on CPU of a tiny FP8-block Llama (FineGrainedFP8Config(dequantize=True)) and a tiny gpt-oss (Mxfp4Config(dequantize=True)). Each checkpoint also holds one genuinely stray tensor.
  • Before the fix: every scale (7) and every expert block/scale tensor (4) is reported unplaced, alongside the stray.
  • After the fix: only the stray.
• Decompressor case, simulated with an on-the-fly decompressor's converter rewrite on a tiny Qwen3-MoE: 24 of 24 packed expert keys were unplaced before the fix, none after.
• Existing suites: tests/unit/torch/utils/test_model_load_utils.py, tests/examples/hf_ptq/test_carry_over_layouts.py, tests/examples/hf_ptq/test_example_utils.py and tests/unit/torch/export/{test_unified_export_hf,test_vllm_fakequant_hf,test_get_quantization}.py: 161 passed.
• mypy and the rest of pre-commit: clean.
• Kimi-K3 expert-parallel export (16 nodes, EP64) through such a decompressor: with the fix, no "Carrying N source weight(s)" line appears, and the export has the expected 995,715 keys (1.5 TB, down from 2.8 TB; the difference is exactly the 494,592 packed expert tensors), matching a reference export key for key. In the unfixed run the carry pass alone took 20:53 of the 37:16 wall time, per the log timestamps; the fixed run took 14:57. The two runs also differed in branch changes unrelated to this PR, so the time saved is read from the log rather than from the run-to-run difference.

Before your PR is "Ready for review"

• Is this change backward compatible?: ✅ Exports stop carrying tensors that were already folded into weights.
• 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?: N/A. The carry-over of unplaced weights is new in the unreleased 0.48.0 cycle.
• Did you get Claude approval on this PR?: ❌ Not yet.

Additional Information

Prerequisite of the HF expert-parallel PTQ (hfep) stack, which is rebased on top of it.

[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

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{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-2733/
Built to branch gh-pages at 2026-10-11 07:26 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[codecov]

Codecov Report

❌ Patch coverage is 92.85714% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 69.47%. Comparing base (54d4416) to head (2a49c89).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
modelopt/torch/utils/plugins/model_load_utils.py	92.85%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2733      +/-   ##
==========================================
- Coverage   69.47%   69.47%   -0.01%     
==========================================
  Files         646      646              
  Lines       71672    71681       +9     
==========================================
+ Hits        49796    49798       +2     
- Misses      21876    21883       +7     

Flag	Coverage Δ
unit	59.89% <92.85%> (-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.