[Quantization] Support compressed tensors w8a8 static and w8a8 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. Add AscendCompressedTensorsConfig to replace CompressedTensorsConfig in vllm.
2. Support CompressedTensorsW8A8 static weight.
   • weight: per-channel, int8, symmetric; activation: per-tensor, int8, symmetric.
3. Support CompressedTensorsW8A8Dynamic weight.
   • weight: per-channel, int8, symmetric; activation: per-token, int8, symmetric, dynamic.
4. Modify the override_quantization_method in AscendQuantConfig.

Co-authored-by: taoqun110 taoqun@huawei.com
Co-authored-by: chenxi-hh chen464822955@163.com

Does this PR introduce any user-facing change?

No

How was this patch tested?

• vLLM version: v0.11.2
• vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2

[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.

[gemini-code-assist]

Code Review

This pull request adds support for w8a8 static and dynamic quantization using the compressed tensors format on Ascend hardware. The changes include a new AscendCompressedTensorsConfig, corresponding quantization schemes, and integration into the vLLM-Ascend platform and worker.

The implementation looks good overall, but I've found a few issues:

• A critical bug in AscendCompressedTensorsConfig that could lead to a runtime crash due to a missing None check.
• Some robustness issues, such as an unsafe list removal and the use of assert for configuration validation, which could cause crashes.
• A performance issue in the w8a8 static quantization scheme where a transpose operation is inefficiently performed on every forward pass.

I've provided detailed comments and suggestions to address these points.

[gemini-code-assist]

The transpose operation on layer.weight.data is performed on every forward pass for the is_310p() case, which is inefficient. The transposed weight should be computed once and cached to improve performance. A good place for this one-time operation would be in process_weights_after_loading.

        if is_310p():
            # On 300I Duo platform, we need transpose again if
            # using nz. This transpose can be skipped in torchair.
            # The transpose is cached to avoid re-computation on every forward pass.
            if not hasattr(layer, "_weight_transposed_for_310p"):
                layer._weight_transposed_for_310p = layer.weight.data.transpose(1, 0).contiguous()
            output = torch_npu.npu_quant_matmul(
                x,
                layer._weight_transposed_for_310p,
                layer.deq_scale,
                bias=bias,
                output_dtype=layer.params_dtype,
            )

[MengqingCao]

Thanks for this great work! Could you plz add an e2e test of w8a8 static and dynamic quant? And ut is also expected, but we could add ut in the follow-up prs.

And is there any accuracy and performance mertics of your pr?

also cc @wangxiyuan @22dimensions

[MengqingCao]

You can solve the DCO and lint issues by referring to the contributing doc in https://vllm-ascend.readthedocs.io/

[LHXuuu]

Thanks for this great work! Could you plz add an e2e test of w8a8 static and dynamic quant? And ut is also expected, but we could add ut in the follow-up prs.

And is there any accuracy and performance mertics of your pr?

also cc @wangxiyuan @22dimensions

Thanks for your reply. I’m currently running accuracy and performance tests. Once they’re complete, I’ll post them in the comment.

[LHXuuu]

@MengqingCao @wangxiyuan Hi! The precision results are shown in the table below. W8A8 static weights fall back to all down_proj linear, while w8a8 dynamic weights are fully quantized.

Qwen3-32b precision test

	ceval	gsm8k	mmlu
BF16	88.94	95.45	89.27
w8a8 static	89.03	95.45	88.91
w8a8 dynamic	88.72	96.51	89.16

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[LHXuuu]

@MengqingCao @wangxiyuan Hello, this pr is ready to merge.

[Ronald1995]

redundant else, you could remove it.

[Ronald1995]

it's good to return (torch.int8, torch.float16, torch.bfloat16)

[Ronald1995]

it's good to return ()

[Ronald1995]

_is_dynamic_token_w8a8 and _is_static_tensor_w8a8 is duplicated, please extract a method