[Feat]Unquantized Linear to nz and control all nz-cast

[anon189Ty]

What this PR does / why we need it?

Currently, when executing to the Linear layer of models in vLLM-Ascend, the weights format is ND in unquantized case and skipped ascend case.
This PR supplements the execution logic for Linear layer. We use a new global variable: VLLM_ASCEND_ENABLE_NZ. When VLLM_ASCEND_ENABLE_NZ=1 and CANN version is 8.3, the weights of the Linear layer will be converted to FRACTAL_NZ, in both unquantized case and skipped ascend case. We also use VLLM_ASCEND_ENABLE_NZ to control the existing NZ conversion, such as w8a8-quantized case.

Does this PR introduce any user-facing change?

Add a new global variable VLLM_ASCEND_ENABLE_NZ. If you want to use NZ format, you should set VLLM_ASCEND_ENABLE_NZ=1.

How was this patch tested?

• vLLM version: v0.11.0rc3
• vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

[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.

[github-actions]

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

[gemini-code-assist]

Code Review

This pull request introduces a new environment variable VLLM_ASCEND_ENABLE_NZ to control the conversion of weights to the FRACTAL_NZ format, which is a valuable addition for performance tuning on Ascend hardware. The changes are applied consistently across various quantization methods and models. However, I've identified a few critical issues in the test files that would prevent the test suite from running, and a potential logic bug in vllm_ascend/attention/mla_v1.py involving dead code and an incorrect format constant. These issues need to be addressed to ensure the correctness and stability of the codebase.

[gemini-code-assist]

This code is at the class level, which will cause a NameError because self is not defined in this context. This code should be moved inside a test method, for example test_init.

Suggested change

	linear = AscendReplicatedLinear(
	input_size=16,
	output_size=8,
	)
	self.assertTrue(isinstance(linear.quant_method,
	AscendUnquantizedLinearMethod))
	def test_init(self):
	linear = AscendReplicatedLinear(
	input_size=16,
	output_size=8,
	)
	self.assertTrue(isinstance(linear.quant_method,
	AscendUnquantizedLinearMethod))

[gemini-code-assist]

This block of code appears to be dead code. The condition getattr(layer.quant_method, "unquant_to_nz", False) will likely never be true because AscendUnquantizedLinearMethod.process_weights_after_loading sets self.unquant_to_nz = False.

Furthermore, even if this code were to be executed, it casts the weight to ACL_FORMAT_FRACTAL_ND, which is inconsistent with the pull request's goal of converting to FRACTAL_NZ.

If this logic is intended to be used, please correct the condition and the format. Otherwise, it should be removed.

[momo609]

This check can be extracted into a common function.

[weijinqian0]

may can remove this parameter

[realliujiaxu]

why not set NZ by default, but instead add an environment variable for control?

[weijinqian0]

dont need inherit this base class

[github-actions]

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