fix(export): Split unquantized 3-D dense weights in quant-aware reverse conversion (GLM-5.3-Flash)

[cjluo-nv]

What does this PR do?

Type of change: Bug fix

Unified HF export of GLM-5.3-Flash (glm5_next) writes transformers' in-memory tensor
names instead of the hub layout, and vLLM cannot load the checkpoint:

KeyError: 'layers.0.self_attn.conv1d.weight'

Root cause. transformers fuses GLM-5.3's KDA depthwise convs q_conv1d / k_conv1d /
v_conv1d (each [C, 1, K]) into one 3-D conv1d weight. The quant-aware reverse conversion
gets a dense SplitRule for that fusion. _apply_split_rule treats any 3-D .weight as a
stacked expert tensor and raises QuantConversionUnsupportedError. The reverse is atomic, so
every tensor then keeps its in-memory name (conv1d, forget_gate.A_log, ...). The only
trace is a UserWarning ("Quant-aware reverse weight conversion skipped").

Fix. The guard now rejects a 3-D weight only when it really can be a stacked expert:

• it sits under .experts, or
• it carries quantization state (weight_scale, ...).

Unquantized dense 3-D weights chunk cleanly along the output dim, the same way the 2-D case
already does.

Usage

No API change. For example, this now exports with hub tensor names:

python examples/hf_ptq/hf_ptq.py --pyt_ckpt_path zai-org/GLM-5.3-Flash-BF16 \
    --qformat nvfp4_experts_only --kv_cache_qformat fp8_cast --export_path <out>

Testing

• New unit tests in tests/unit/torch/export/test_quant_aware_conversion.py:
  • An unquantized fused [3C, 1, K] conv weight splits back to exactly the original q/k/v
    tensors.
  • A quantized 3-D weight outside .experts still falls back.
  • The existing test_stacked_3d_expert_raises_unsupported is unchanged and passes.
• tests/unit/torch/export/test_quant_aware_conversion.py: 59 passed with transformers 5.14.1
  (15 passed, 44 skipped with 4.57.6, which lacks core_model_loading).
• tests/unit/torch/export/ without test_export_diffusers.py: 349 passed.
  test_export_diffusers_models_non_quantized[get_tiny_dit] fails in my local env on main as
  well, unrelated to this change.
• End to end (oci-jhb GB300, transformers 5.16.1). Four GLM-5.3-Flash
  nvfp4_experts_only + fp8_cast exports from zai-org/GLM-5.3-Flash-BF16:
  • Without the fix, vLLM failed with the KeyError above.
  • With it, the tensor-name set matches the published nvidia/GLM-5.3-Flash-NVFP4, apart from
    that checkpoint's extra dense-MLP quantization in layers 0-2.
  • The exports serve in vLLM and were evaluated on GPQA-D, AA-LCR, SciCode and Terminal-Bench
    2.1.

Before your PR is "Ready for review"

• Is this change backward compatible?: ✅ (the stacked-expert fallback is unchanged)
• 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?: ✅ (the bug is in 0.46 and 0.47)
• Did you get Claude approval on this PR?: ❌ (draft)

Additional Information

Found while running an NVFP4 calibration-data study on GLM-5.3-Flash. Follows #1833 (quant-aware
reverse weight conversion).

🤖 Generated with Claude Code

[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-2732/
Built to branch gh-pages at 2026-10-11 02:57 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 69.46%. Comparing base (54d4416) to head (5374a5f).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2732      +/-   ##
==========================================
- Coverage   69.47%   69.46%   -0.01%     
==========================================
  Files         646      646              
  Lines       71672    71673       +1     
==========================================
- Hits        49796    49791       -5     
- Misses      21876    21882       +6     

Flag	Coverage Δ
unit	59.88% <100.00%> (-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.