[Quantization] Support compressed tensors moe w8a8 int8 dynamic weight

[LHXuuu]

What this PR does / why we need it?

While using the LLM Compressor quantization tool from the VLLM community to generate quantized weights, the VLLM Ascend engine needs to be adapted to support the compressed tensors quantization format.

1. Support Moe model W8A8 Int8 dynamic weight.
2. Specify W4A16 quantization configuration.

Co-authored-by: menogrey 1299267905@qq.com
Co-authored-by: kunpengW-code 1289706727@qq.com

Does this PR introduce any user-facing change?

No

How was this patch tested?

• vLLM version: v0.13.0
• vLLM main: vllm-project/vllm@2f4e654

[gemini-code-assist]

Code Review

This pull request extends quantization support in the VLLM Ascend engine for compressed tensors, specifically adding handling for Mixture-of-Experts (MoE) models with W8A8 Int8 dynamic weight quantization and specifying W4A16 quantization. The changes involve refactoring how quantization schemes are identified and applied, with new logic for FusedMoE layers.

My review focuses on the correctness of these changes. I've identified a high-severity issue where the new logic for FusedMoE layers is hardcoded to only consider the first expert, which could lead to incorrect quantization for models with multiple experts. The rest of the changes, including refactoring and extending quantization checks, appear to be well-implemented.

[gemini-code-assist]

The logic to determine the quantization scheme for FusedMoE layers is hardcoded to only check the projections of the first expert (expert 0). This is a significant limitation as noted in the TODO on line 179. It can lead to incorrect behavior for models with multiple experts, especially if they use different quantization schemes or if the naming convention for experts differs. The implementation should be generalized to iterate over all experts in the MoE layer to ensure consistent quantization is applied.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[menogrey]

gsm8k accuracy:

ais_bench --models vllm_api_general_chat --datasets gsm8k_gen_0_shot_cot_chat_prompt.py --mode all --dump-eval-details --merge-ds 

Qwen3-235B-A22B

dataset	version	metric	mode	vllm-api-general-chat
gsm8kdataset	-	accuracy	gen	96.44

llmcompressor Qwen3-235B-A22B w8a8 int8 dynamic

dataset	version	metric	mode	vllm-api-general-chat
gsm8kdataset	-	accuracy	gen	96.29

ceval accuracy:

ais_bench --models vllm_api_general_chat --datasets ceval_gen_0_shot_cot_chat_prompt.py --mode all --dump-eval-details --merge-ds 

Qwen3-235B-A22B

dataset	version	metric	mode	vllm-api-general-chat
cevaldataset	-	accuracy	gen	90.86

llmcompressor Qwen3-235B-A22B w8a8 int8 dynamic

dataset	version	metric	mode	vllm-api-general-chat
cevaldataset	-	accuracy	gen	90.27

mmlu accuracy:

ais_bench --models vllm_api_general_chat --datasets mmlu_gen_0_shot_cot_chat_prompt.py --mode all --dump-eval-details --merge-ds

Qwen3-235B-A22B

dataset	version	metric	mode	vllm-api-general-chat
mmludataset	-	accuracy	gen	89.67

llmcompressor Qwen3-235B-A22B w8a8 int8 dynamic

dataset	version	metric	mode	vllm-api-general-chat
mmludataset	-	accuracy	gen	88.83

[kunpengW-code]

mmlu accuracy:

ais_bench --models vllm_api_general --datasets mmlu_gen --merge-ds --debug

Qwen3-30B-A3B

dataset	version	metric	mode	vllm_api_general
mmludataset	-	accuracy	gen	79.58

llmcompressor Qwen3-30B-A3B w8a8 int8 dynamic

dataset	version	metric	mode	vllm_api_general
mmludataset	-	accuracy	gen	79.42

gam8k accuracy:

ais_bench --models vllm_api_general_chat --datasets gsm8k_gen_0_shot_cot_chat_prompt --dump-eval-details --debug

Qwen3-30B-A3B

dataset	version	metric	mode	vllm_api_general
gam8kdataset	-	accuracy	gen	96.44

llmcompressor Qwen3-30B-A3B w8a8 int8 dynamic

dataset	version	metric	mode	vllm_api_general
gam8kdataset	-	accuracy	gen	96.21

ceval accuracy:

ais_bench --models vllm_api_general --datasets ceval_gen --merge-ds --debug --dump-eval-details

Qwen3-30B-A3B

dataset	version	metric	mode	vllm_api_general
cevaldataset	-	accuracy	gen	81.58

llmcompressor Qwen3-30B-A3B w8a8 int8 dynamic

dataset	version	metric	mode	vllm_api_general
cevaldataset	-	accuracy	gen	81.50