AWQModifier fast resolve mappings, better logging, MoE support

[brian-dellabetta]

SUMMARY:
In AWQ, resolving mappings can take a while because it is traversing the entire model tree, rather than just the parent, to find the balance layers. This scopes the search to just the parent module. For MoE models, the previous implementation only found a single layer for each regex string provided in mappings. This updates that to find as many as it can, which is necessary for mappings like

AWQMapping(
    "re:.*post_attention_layernorm$",
    ["re:.*mlp.experts.*.gate_proj$", "re:.*mlp.experts.*.up_proj$"],
)

which have multiple gate_proj and up_proj layers, one for each expert.

gsm8k results with Qwen/Qwen3-30B-A3B MoE model after AWQ W4A16 Symmetric:

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.3813	±	0.0134
		strict-match	5	exact_match	↑	0.8810	±	0.0089

TEST PLAN:

• working with Qwen/Qwen3-30B-A3B with same set of mappings used in AutoAWQ. Example included in this PR in examples/awq/qwen3_moe_example.py. Ran successfully in ~2 hours on a single H100 with ~70GB of 80GB used (additional memory needed during saving)
• Same wikitext PPL (14.0814) as on main for meta-llama/Llama-3.2-3B-Instruct

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[kylesayrs]

Many nits, otherwise looks good

[Chao-Xue]

https://github.com/vllm-project/llm-compressor/blob/ceffa644072b1d440df3d99b0f98f6416a05bf2f/examples/awq/qwen3_moe_example.py#L55~L56
It looks like re:.*mlp.shared_expert_gate$ is included in the ignore list, but Qwen3-MoE doesn’t seem to have a shared_expert_gate module. Probably not an issue tho

[brian-dellabetta]

@Chao-Xue , yes that is for other qwen MoE architectures where there is a shared expert that we ignore. not all qwen MoE models have that though