[Feature]: Add AWQ quantization support for vllm-ascend

[menogrey]

🚀 The feature, motivation and pitch

motivation

AWQ quantization is is a commonly used quantitative method, and there are many quantized models that can be used immediately, such as Qwen. Now vllm-ascend support special quantized model which is quantized by modelslim, but it will take a lot of time to quantize model and we cannot cover all the models if user want to run a quantized model.

implement

Refer to PR: #4316
vllm-ascend: #4316
vllm: v0.11.1rc5 (2918c1b)

validation

Note: You can go to modescope to get the weight with model name as below

• Every model is tested in single-node, and for large model tensor parallel is only used.
• ⚠️WARNING: Use export VLLM_ASCEND_ENABLE_NZ=0 to avoid error when process weights.
• ⚠️WARNING: Not support expert parallel now.

Type	Architecture	Models	Model Name	Aclgraph Mode	Accuracy	Performance	Compare to W8A8
Text-only	DeepseekV3ForCausalLM	DeepSeek-V3	tclf90/DeepSeek-V3.1-AWQ	✅	ceval:0.896
Text-only	DeepseekV3ForCausalLM	DeepSeek-R1	cognitivecomputations/DeepSeek-R1-awq	✅	ceval:0.8923
Text-only	Qwen2ForCausalLM	QwQ, Qwen2	Qwen/Qwen2.5-32B-Instruct-AWQ Qwen/QwQ-32B-AWQ	✅
Text-only	Qwen3ForCausalLM	Qwen3	Qwen/Qwen3-32B-AWQ	✅	ceval:0.85
Text-only	Qwen3MoeForCausalLM	Qwen3MoE	billy800/Qwen3-30B-A3B-Instruct-2507-AWQ swift/Qwen3-235B-A22B-Instruct-2507-AWQ	✅ ❌(accuracy issue)	ceval:0.8403
Multimodal	Qwen2AudioForConditionalGeneration	Qwen2-Audio	No AWQ quantized model provided
Multimodal	Qwen2VLForConditionalGeneration	QVQ, Qwen2-VL	Qwen/Qwen2-VL-7B-Instruct-AWQ	✅
Multimodal	Qwen2_5_VLForConditionalGeneration	Qwen2.5-VL	Qwen/Qwen2.5-VL-32B-Instruct-AWQ	❌(accuracy issue)
Multimodal	Qwen3VLForConditionalGeneration	Qwen3-VL	tclf90/Qwen3-VL-32B-Instruct-AWQ	✅
Multimodal	Qwen3VLMoeForConditionalGeneration	Qwen3-VL-MOE	tclf90/Qwen3-VL-30B-A3B-Instruct-AWQ tclf90/Qwen3-VL-235B-A22B-Instruct-AWQ	✅ ❌(accuracy issue)

Alternatives

No response

Additional context

No response

[menogrey]

PR: #4516 will support w4a16 in ascend quantization. There are some common implements, after this PR, i will refactor AWQ code and will support expert parallel.