What does this PR do?

Type of change: New feature

This PR adds two complementary AutoQuant scoring controls:

• constraints["score_model"]="per_element" normalizes selector coefficients by the number of represented weight elements while preserving the configured effective-bit budget.
• method="group_recon" with score_boundary="group" measures normalized reconstruction error at shared attention or MLP group outputs. This changes where sensitivity is measured without forcing the corresponding projection modules to share a quantization recipe.

It also keeps shared-expert gate/up/down projections in one deployable fused-MoE recipe decision, persists the new scoring metadata in AutoQuant checkpoints, validates restored-state compatibility, and exposes the controls through the HF PTQ CLI.

Existing behavior remains the default: score_model="raw", method="gradient", and score_boundary="local".

Usage

model, search_state = mtq.auto_quantize(
    model,
    constraints={
        "effective_bits": 5.5,
        "cost_model": "active_moe",
        "score_model": "per_element",
    },
    data_loader=calib_dataloader,
    forward_step=forward_step,
    quantization_formats=[mtq.NVFP4_DEFAULT_CFG, mtq.FP8_DEFAULT_CFG],
    method="group_recon",
    score_boundary="group",
)

Equivalent HF PTQ flags:

--auto_quantize_method group_recon \
--auto_quantize_score_model per_element \
--auto_quantize_score_boundary group

Testing

• 72 passed: focused non-distributed AutoQuant unit suite
• 1 passed: distributed AutoQuant checkpoint/search test
• 3 passed: HF PTQ argument tests
• All changed-file pre-commit hooks passed, including Ruff, mypy, RST checks, and Bandit
• bash -n passed for the modified HF PTQ shell entry points
• Qwen3.6-35B end-to-end checkpoint generation and vLLM serving smoke test passed on the implementation-equivalent pre-rename revision; the terminology-only revision is covered by the unit and CLI tests above

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?: ✅
• Did you get Claude approval on this PR?: N/A (draft)

Additional Information

The terms group scoring and recipe grouping are intentionally distinct: group scoring changes the output boundary used to measure a projection's perturbation, while recipe grouping constrains multiple modules to use the same quantization format.