[Refactor][main CI] Refactor code to align with vllm main

[MengqingCao]

What this PR does / why we need it?

1. cherry-pick [CI] Upgrade vllm to newest commit #3423
2. refactor deepseek code
   • seprate modeling code deepseek v3.2 with other deepseek version
   • using the latest MLA code arch in vLLM

How was this patch tested?

Test locally passed with AscendScheduler enabled + torchair/eager mode on DeepSeek-V3.2-Exp-W8A8

• 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.

[gemini-code-assist]

Code Review

This pull request refactors the codebase to align with upstream vllm main, introducing version compatibility checks across multiple files. The changes are extensive and mostly correct. However, I've identified a critical issue in a test case that could break CI for a specific vllm version, and a code duplication issue that should be addressed for better maintainability.

[gemini-code-assist]

The assertion for vllm_config.compilation_config.level is incorrect when vllm_version_is("0.11.0") is true. The test will fail in that case. The assertion should be conditional, similar to other tests in this file, to check for CompilationLevel.NO_COMPILATION for vllm 0.11.0 and CompilationMode.NONE otherwise.

            if vllm_version_is("0.11.0"):
                self.assertEqual(
                    vllm_config.compilation_config.level,
                    CompilationLevel.NO_COMPILATION,
                )
            else:
                self.assertEqual(
                    vllm_config.compilation_config.level,
                    CompilationMode.NONE,
                )

[gemini-code-assist]

This block of code, which pads num_tokens for sequence parallelism, appears to be duplicated. The second block should be removed to avoid redundancy.

Suggested change

	# In multi-DP scenarios, there may be situations where all DP groups are executing dummy runs.
	# If sequence parallelism is enabled, it is essential to ensure that num_tokens is divisible by tp_size.
	if self.use_aclgraph and enable_sp(self.vllm_config):
	tp_size = self.vllm_config.parallel_config.tensor_parallel_size
	num_tokens = math.ceil(num_tokens / tp_size) * tp_size
	# In multi-DP scenarios, there may be situations where all DP groups are executing dummy runs.
	# If sequence parallelism is enabled, it is essential to ensure that num_tokens is divisible by tp_size.
	if self.use_aclgraph and enable_sp(self.vllm_config):
	# In multi-DP scenarios, there may be situations where all DP groups are executing dummy runs.
	# If sequence parallelism is enabled, it is essential to ensure that num_tokens is divisible by tp_size.
	if self.use_aclgraph and enable_sp(self.vllm_config):
	tp_size = self.vllm_config.parallel_config.tensor_parallel_size
	num_tokens = math.ceil(num_tokens / tp_size) * tp_size
	if self.use_aclgraph and enable_sp(self.vllm_config):

[whx-sjtu]

The adaption of MLA part LGTM. It's ok for me to delete deepseek_v2 modeling after this PR merged.

[github-actions]

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

[MengqingCao]

torchair mtp failed in

• https://github.com/vllm-project/vllm-ascend/actions/runs/18611921289/job/53071268132?pr=3504
• https://github.com/vllm-project/vllm-ascend/actions/runs/18626797693/job/53105932398?pr=3504

test_models_distributed_Qwen_Dense_with_flashcomm_v1 failed in https://github.com/vllm-project/vllm-ascend/actions/runs/18611921289/job/53071268128?pr=3504

tp+ep failed due to dispatcher to different communication ops https://github.com/vllm-project/vllm-ascend/actions/runs/18616600778/job/53081964431?pr=3504

[github-actions]

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

[github-actions]

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

[MengqingCao]

e2e tests on singlecard only failed with test_mtp_torchair_correctness_piecewise and test_mtp_torchair_correctness_full

2025-10-18T15:07:07.2022784Z tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_torchair_correctness.py::test_mtp_torchair_correctness_piecewise SKIPPED
2025-10-18T15:07:09.4163583Z tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_torchair_correctness.py::test_mtp_torchair_correctness_full SKIPPED

in https://github.com/vllm-project/vllm-ascend/actions/runs/18616600778/job/53081964430?pr=3504

[github-actions]

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

[MengqingCao]

closing as done in #3612