[Bugfix] Enhance extra_config handling for layer name suffix matching

[xin3he]

This pull request improves how the configuration is selected for layers with extra configuration by adding support for suffix matching. This addresses situations where the keys in extra_config are short names (like "lm_head"), but the actual layer_name is fully qualified (like "model.language_model.lm_head"), which can happen due to model nesting.

Configuration handling improvements:

• Added suffix matching for layer_name when looking up keys in extra_config, so that short names in the config will correctly apply to fully qualified layer names. This ensures that extra configuration is applied even when model nesting changes the layer naming.

Purpose

Support MLLM quantized lm_head in inc/auto-round format, related issue: intel/auto-round#1709
@Yi4Liu @wenhuach21

Test Plan

• test_inc_config_parser_suffix_match_for_lm_head
• test_inc_get_quant_method_lm_head_uses_suffix_match

Real model: quantize and load Qwen/Qwen3.5-9B.
Quantization conmmand: auto-round /models/Qwen3.5-9B/ --quant_lm_head --iters 0 --disable_opt_rtn
Serve command: vllm serve tmp_autoround/ --port 8000 --tensor-parallel-size 1 --max-model-len 2048 --reasoning-parser qwen3 --served-model-name qwen

Test Result

UT passed.
Model loaded successfully.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[cloudintheskyfield]

This PR remains the minimal fix for #49137: it correctly handles the reported direct lm_head short-key / nested-runtime-name case.

A deeper audit found several adjacent AutoRound name-resolution paths that are outside this PR current code and tests. I prepared and locally validated a stacked follow-up that:

• centralizes exact, longest component-boundary suffix, bounded literal (for example .experts.), and regex matching;
• applies the same lookup to packed constituent names such as q/k/v and to RoutedExperts descendant configs;
• preserves exact-key priority over the model. fallback in get_quant_method;
• replaces the sentinel lm_head dispatch assertion with an observable ParallelLMHead parameter-registration test (qweight rather than weight).

Against commit 44cf86362, seven added cases fail: longest-suffix selection, exact-vs-model fallback, packed short-name resolution/consistency, FusedMoE short descendants/consistency, and the AutoRound bounded .experts. pattern. On the stacked patch:

• AutoRound non-model unit tests: 37 passed, 2 deselected
• pre-commit on the three changed files: all passed
• a 1,920-key / 10-layer FusedMoE matcher check completes in about 0.10 s after removing an O(N^2) scan found during independent review

@xin3he, would you like this as a stacked PR targeting your fix_inc branch, or would you prefer to keep #48589 minimal and handle the broader resolver separately? I will not publish the patch until the submitting human has completed the repository-required line-by-line review and local test run.

[cloudintheskyfield]

@xin3he The stacked patch is now published after the repository-required human line-by-line review and personal test run:

• commit: cloudintheskyfield@c85dfb9
• branch: https://github.com/cloudintheskyfield/vllm/tree/agent/pr48589-unify-inc-extra-config-matching
• comparison from your current fix_inc head: xin3he/vllm@fix_inc...cloudintheskyfield:agent/pr48589-unify-inc-extra-config-matching

The commit is based directly on 44cf8636213adb424eb1e357d298d1545ed5c0ec; no competing upstream PR has been opened. Validation remains:

• AutoRound non-model tests: 37 passed, 2 deselected
• all pre-commit hooks for the three touched files passed
• the seven broader resolver cases fail on 44cf86362 and pass on the stacked commit

Could you either run your Qwen3.5-9B --quant_lm_head load against this commit, or let me know whether you prefer a draft stacked PR targeting xin3he/vllm:fix_inc? I have also asked the #49137 reporter to validate the 27B checkpoint on 2×RTX 3090.

[cloudintheskyfield]

Update: the stacked follow-up is now available as a draft PR:

• [Test] Cover nested AutoRound lm_head registration xin3he/vllm#1

It targets xin3he/vllm:fix_inc, is exactly one commit ahead of the current 44cf86362 head, and remains Draft while real-model revalidation is pending. @xin3he, feedback on the resolver scope and the Qwen3.5-9B revalidation can go directly on the stacked PR.

[xin3he]

@jikunshang Could you please help review this PR?

[yiliu30]

LGTM, please add model level test, thanks!

[xin3he]

Model loaded successfully.

The model level test is executed locally as described above, it's a bug fix so I only updated the UT to cover.

[yiliu30]

Model loaded successfully.

The model level test is executed locally as described above, it's a bug fix so I only updated the UT to cover.

Great, we already have a model test in our unit tests. Please check the test_auto_round_model.
We can extend it to include a tiny model with a quantized LM head.

[mergify]

Hi @xin3he, the pre-commit checks have failed. Please run:

uv pip install pre-commit>=4.5.1
pre-commit install
pre-commit run --all-files

Then, commit the changes and push to your branch.

For future commits, pre-commit will run automatically on changed files before each commit.

[xin3he]

Model loaded successfully.

The model level test is executed locally as described above, it's a bug fix so I only updated the UT to cover.

Great, we already have a model test in our unit tests. Please check the test_auto_round_model. We can extend it to include a tiny model with a quantized LM head.

lm_head is usually tied and cannot be quantized for tiny model, so I will skip it.

[xin3he]

Hi @xin3he, the pre-commit checks have failed. Please run:

uv pip install pre-commit>=4.5.1
pre-commit install
pre-commit run --all-files

Then, commit the changes and push to your branch.

For future commits, pre-commit will run automatically on changed files before each commit.

I ran these commands, but nothing changed.