[Misc]feat: adapt to vLLM main (0d4d334e...b2c58ee9)

[Meihan-chen]

Main2Main Summary

Status: completed
Upstream range: 0d4d334eaa583b9c09aa4eb7538c22db99fd84b3..b2c58ee9427f15563210e184c57a6e530f37e464
Reached upstream commit: b2c58ee9427f15563210e184c57a6e530f37e464
Steps: 6/6
CI suite: e2e-singlecard-light

Result

The target vLLM commit b2c58ee9427f15563210e184c57a6e530f37e464 was fully reached. All 6 planned steps completed successfully with CI passing on each step. No vllm-ascend code adaptation was required - all upstream changes were either hardware-specific optimizations (CPU/RISC-V, ROCm), internal implementation refactors, or new model registrations that flow through to vllm-ascend without override.

Completed Steps

Step	Upstream range	vllm-ascend commit	CI result	Summary
step-1	0d4d334..27b85d2	d362ebf	passed	commit reference only - CPU/RISC-V kernels, InternS2 model, thinking_budget bugfix
step-2	27b85d2..2676ab1	f2b1c59	passed	commit reference only - deprecation removal for get_tokenizer, resolve_hf_chat_template
step-3	2676ab1..4b364f8	0359b16	passed	commit reference only - LoRA OOT support (backward compatible), KV cache alignment_tokens (optional param)
step-4	4b364f8..75fd68c	b764824	passed	commit reference only - PoolingOfflineMixin refactor (API unchanged via mixin inheritance)
step-5	75fd68c..491e8d8	22338a1	passed	commit reference only - DeepSeekV4 cuda graph, IPC weight sync, kv_connector order fix
step-6	491e8d8..b2c58ee	7bd484a	passed	commit reference only - SM12x kernels, FlashAttn FP8 dtype handling, IR op set_default

Changes Made

• Updated vLLM commit reference from 0d4d334eaa583b9c09aa4eb7538c22db99fd84b3 to b2c58ee9427f15563210e184c57a6e530f37e464 (30 commits absorbed).
• Updated workflow files (Dockerfile.lint, pr_test_full.yaml, pr_test_light.yaml) with new commit reference.
• Updated docs/source/conf.py main_vllm_commit to target commit.
• No vllm-ascend code adaptation required - all upstream changes were compatible with existing vllm-ascend implementations.

CI Verification

• Passed: step-1 through step-6 (all 6 steps)
• Treated as env flakes: none
• Last successful step: step-6 (reached target commit)

Commits Created in vllm-ascend

1. d362ebfed81775e03eb531b6f3155fcb82428e63 - step-1
2. f2b1c59a3050602e1cd73182b902d881ebffdd9d - step-2
3. 0359b1676dc403219092763950417d16de5eb834 - step-3
4. b7648245a2bf2c26794918f062c75b92f50bfc5a - step-4
5. 22338a1eb16a12d42d30d9b11f9cdb8040f5619c - step-5
6. 7bd484ab1632c6cf663a1b07b619fd8f0fed037f - step-6

Key Upstream Changes Absorbed

• New Models: InternS2PreviewForConditionalGeneration
• Bug Fixes: thinking_token_budget condition, FlashAttn FP8 dtype handling, kv_connector order, SM121 exclusion
• Refactors: PoolingOfflineMixin, LoRA OOT layer support, IR op priority management
• Optimizations: DeepSeekV4 cuda graph, MLA attention optimization, IPC weight sync

No Push/PR

As instructed, no git push or PR creation was performed. The commits are local only.

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