[Misc]feat: adapt to vLLM main (0d4d334e...46a95815)

[Meihan-chen]

main2main Upgrade Complete

Summary

Successfully upgraded vllm-ascend from vLLM commit 0d4d334eaa583b9c09aa4eb7538c22db99fd84b3 to target commit 46a95815d344e43ebd6b4757632c55d03e90983a across 8 steps covering 23 upstream commits.

Upgrade Details

Step	Upstream Range	Commits	vllm-ascend Changes	CI Status
1	0d4d334 → 27b85d2	6	No-op (CPU/RISC-V kernels, InternS2 model)	passed
2	27b85d2 → 2676ab1	1	No-op (deprecation removals)	passed
3	2676ab1 → 4b364f8	2	lora/utils.py (maybe_get_oot_by_class)	passed
4	4b364f8 → 75fd68c	1	No-op (PoolingOfflineMixin refactoring)	passed
5	75fd68c → 0fe7550	6	spec_decode/eagle_proposer.py (dtype=np.int32)	passed
6	0fe7550 → e0a45f1	1	No-op (IPC weight sync - CUDA specific)	passed
7	e0a45f1 → fb5bd03	5	No-op (IR Op Priority - inherited via super().init)	passed
8	fb5bd03 → 46a9581	1	No-op (ROCm MLA FP8 - AMD specific)	passed

vllm-ascend Code Changes

Files Modified

1. vllm_ascend/lora/utils.py (Step-3)

   • Updated 4 can_replace_layer methods to use maybe_get_oot_by_class(AscendQKVParallelLinear) for consistency with upstream's out-of-tree layer handling pattern
   • Added import: from vllm.model_executor.custom_op import maybe_get_oot_by_class

2. vllm_ascend/spec_decode/eagle_proposer.py (Step-5)

   • Changed self.token_arange_np = np.arange(self.max_num_tokens + 1) to self.token_arange_np = np.arange(self.max_num_tokens + 1, dtype=np.int32) for consistency with upstream dtype fix

No-Op Steps

Steps with no vllm-ascend code changes required:

• Step-1: CPU/RISC-V RVV kernels (not relevant to Ascend NPU), InternS2 Preview model inherits from already-patched Qwen3_5Moe
• Step-2: Deprecation removals - vllm-ascend doesn't import deprecated functions
• Step-4: PoolingOfflineMixin refactoring - API unchanged, LLM still works via inheritance
• Step-6: IPC weight sync - CUDA-specific, vllm-ascend has its own eplb weight transfer for NPU
• Step-7: IR Op Priority optimization - NPUWorker inherits via super().__init__() call
• Step-8: ROCm MLA FP8 - AMD GPU-specific, Ascend uses different MLA backend

Commits Created

All 8 commits signed and verified:

Step	Commit SHA	Message
1	bc0c59df9e7d07ad5be43547742f6bd4bc71b043	main2main step-1: absorb vLLM commits 0d4d334..27b85d2
2	6fa51f1d7c1f3a1235c6105d2e2d2a3236fc5b69	main2main step-2: absorb vLLM commit 27b85d2..2676ab1
3	36b99ffd5a8d5145b6bc047e05478c59d50210c6	main2main step-3: absorb vLLM commits 2676ab1..4b364f8
4	063e33acea8def1641b80a8046885227185cb194	main2main step-4: absorb vLLM commit 4b364f8..75fd68c
5	025382e396be0ce93cb02a9e6394cc1c1ee6ccc6	main2main step-5: absorb vLLM commits 75fd68c..0fe7550
6	d55688ecd22925e89dfe2c5cba424c864a316693	main2main step-6: absorb vLLM commit 0fe7550..e0a45f1
7	5fa0dd85f558fd5deec622ba686d5eeeae3b3de9	main2main step-7: absorb vLLM commits e0a45f1..fb5bd03
8	aeabe986b3d0eab9eb96fcdea714608f8c7b06fa	main2main step-8: absorb vLLM commit fb5bd03..46a9581

CI Verification

All steps verified via e2e-singlecard-light suite:

• 8 rounds executed
• All passed with exit code 0
• No code bugs or env flakes detected

Pre-Completion Checklist

• ✅ No temp files in the repo
• ✅ All commits signed (git commit -s)
• ✅ All intermediate files in /tmp/main2main/
• ✅ conf.py main_vllm_commit updated at each step
• ✅ Every step ran CI after commit reference update
• ✅ New vllm_version_is() guards not needed (no version-specific code paths)
• ✅ Each commit message includes upstream commit range
• ✅ Each step has summary in /tmp/main2main/steps/<step-id>/summary.md
• ✅ Final summary written to /tmp/main2main/final-summary.md

Target Status

✅ TARGET REACHED: vLLM commit 46a95815d344e43ebd6b4757632c55d03e90983a

The vllm-ascend repository is now compatible with the requested vLLM upstream commit.

• vLLM version: v0.20.2
• vLLM main: vllm-project/vllm@0d4d334