[Misc]feat: adapt to vLLM main (c7aa186d...50acdc5b)

[Meihan-chen]

Main2Main Summary

Status: completed
Upstream range: c7aa186d67b6f051680831418e957c67f34ba7a2..50acdc5b5cc00f10408d8f98b21fc97efc615173
Reached upstream commit: 50acdc5b5cc00f10408d8f98b21fc97efc615173
Steps: 9/9
CI suite: e2e-main2main

Result

Successfully adapted vllm-ascend to track 30 upstream vLLM commits across 9 steps.
All steps passed CI verification. One code adaptation was required in step-2 for
spec decode multimodal method rename.

Completed Steps

Step	Upstream range	vllm-ascend commit	CI result	Summary
step-1	c7aa186d67b..16e336491e96	da5c278	passed	commit reference only
step-2	16e336491e96..5d0fd87038b1	447596a	passed	renamed _raise_if_multimodal to _warn_if_multimodal
step-3	5d0fd87038b1..242afc6bf40d	acbd2fe	passed	commit reference only
step-4	242afc6bf40d..df8e63f4edab	f9b6415	passed	commit reference only
step-5	df8e63f4edab..f39bcf1e3069	2cfd454	passed	commit reference only
step-6	f39bcf1e3069..27e0057aeda6	fbe9025	passed	commit reference only (round 2: HCCL env flake)
step-7	27e0057aeda6..38e16678ba7e	fbd1eb7	passed	commit reference only
step-8	38e16678ba7e..ca3e62d3363d	4ce1b69	passed	commit reference only
step-9	ca3e62d3363d..50acdc5b5cc0	e8e4c06	passed	commit reference only

Changes Made

• Updated vLLM commit reference from c7aa186d67b6f051680831418e957c67f34ba7a2 to 50acdc5b5cc00f10408d8f98b21fc97efc615173.
• Renamed _raise_if_multimodal to _warn_if_multimodal in vllm_ascend/spec_decode/dflash_proposer.py (step-2).
• Updated workflow files and docs/source/conf.py with new commit reference at each step.

CI Verification

• Passed: All 9 steps passed CI after verification.
• Treated as env flakes: HCCL communication error in step-6 round 1 (error code 7), resolved in round 2.
• Last successful step: step-9

Key Upstream Changes Absorbed

1. Mistral tokenizer leniency - Internal tokenizer changes, no adaptation needed.
2. Spec decode multimodal warning - Method rename _raise_if_multimodal → _warn_if_multimodal, adapted in vllm-ascend.
3. Gemma4 encoder budget - Model internals, no adaptation needed.
4. nixl refactor - KV transfer connector internals, vllm-ascend uses separate connectors.
5. OpenAI/KV Offload fixes - Compatible with existing vllm-ascend patches.
6. Gemma4 MTP spec decode - New feature flag with default preserving existing behavior.
7. Multiple bugfixes - TRTLLM alignment inherited via class hierarchy.
8. TPU upgrade - Not relevant to Ascend NPU platform.
9. Qwen3 streaming fix - Different methods from vllm-ascend patches.

vllm-ascend Commits Created

1. da5c278
2. 447596a
3. acbd2fe
4. f9b6415
5. 2cfd454
6. fbe9025
7. fbd1eb7
8. 4ce1b69
9. e8e4c06

• vLLM version: v0.20.1
• vLLM main: vllm-project/vllm@c7aa186