[Misc]feat: adapt to vLLM main (c7aa186d...2a16ece2)

[Meihan-chen]

Automated adaptation to upstream vLLM main branch changes.
Commit range: c7aa186d67b6f051680831418e957c67f34ba7a2...2a16ece2d342c0c154a4949ad317b521f8c04ec4

Created Commits

• 59076f5d1ca9b73af617ab34d2e458080268f529 main2main: update vLLM commit reference to 16e33649
• 917370e89373ea551fcebadd3d8d1359c02a6e8e main2main: update vLLM commit reference to 5d0fd87
• cebe1f2e9855c2b2300a3626f765c37fe3733672 main2main: update vLLM commit reference to 242afc6
• 580e7b0bf4ef575938358bcb30fd6b10ea6f466c main2main: update vLLM commit reference to df8e63f
• afca80f716b4a63278bc89c440e22383ed1cbc7c main2main: update vLLM commit reference to f39bcf1
• fd4140f85413df0f0626ceaf583476910bc54477 main2main: update vLLM commit reference to 27e0057
• 4a7858b883429f9197a0b6678aec790ccc2bd79e main2main: update vLLM commit reference to 38e1667
• d0a03e142cb62a5926b534979aaa88e8f7b73424 main2main: update vLLM commit reference to ca3e62d
• 09e37f9d1a74895b0612d0be8994889a86e929e4 main2main: update vLLM commit reference to 20cac26
• 64317d20e0ec625cf7418deee3ba8ee6385a09c4 main2main: update vLLM commit reference to 51f22dc
• 19d4702441c486368529a16e65e3f5b4e2b57248 main2main: update vLLM commit reference to 9c0812f
• bd4b616fcade0f876626c177b3eef4f235ef22d3 main2main: update vLLM commit reference to ffee741
• d775b2abf216bfeee9afebfc2bcd54a1baefb4d4 main2main: update vLLM commit reference to b3945cc
• 341a48e46bb18a5b9c4c0f6b9a15c0acad71810e main2main: update vLLM commit reference to 2a16ece (target)

main2main Summary

Main2Main Summary

Status: completed
Upstream range: c7aa186d67b6f051680831418e957c67f34ba7a2..2a16ece2d342c0c154a4949ad317b521f8c04ec4
Reached upstream commit: 2a16ece2d342c0c154a4949ad317b521f8c04ec4
Steps: 14/14
CI suite: e2e-main2main (not run - Ascend NPU hardware unavailable)

Result

Successfully absorbed 50 upstream commits across 14 steps. The target commit was fully reached with minimal code adaptations required. Most upstream changes were platform-specific (CPU, ROCm, XPU) or feature additions that did not affect vllm-ascend's Ascend-specific implementation.

Completed Steps

Step	Upstream range	vllm-ascend commit	CI result	Summary
step-1	c7aa186..16e33649	59076f5	import/syntax pass	Mistral tokenizer refactoring - commit reference only
step-2	16e33649..5d0fd87	917370e	import/syntax pass	PP bugfix + spec_decode multimodal warning
step-3	5d0fd87..242afc6	cebe1f2	import/syntax pass	Gemma4 encoder budget - commit reference only
step-4	242afc6..df8e63f	580e7b0	import/syntax pass	nixl refactor - commit reference only
step-5	df8e63f..f39bcf1	afca80f	import/syntax pass	KV offload + forced tool choice - commit reference only
step-6	f39bcf1..27e0057	fd4140f	import/syntax pass	Gemma4 MTP spec decode - commit reference only
step-7	27e0057..38e1667	4a7858b	import/syntax pass	Various bugfixes - commit reference only
step-8	38e1667..ca3e62d	d0a03e1	import/syntax pass	TPU requirements - commit reference only
step-9	ca3e62d..20cac26	09e37f9	import/syntax pass	DeepSeekV3/V4 tool parser - commit reference only
step-10	20cac26..51f22dc	64317d2	import/syntax pass	CPU Gated DeltaNet - commit reference only
step-11	51f22dc..9c0812f	19d4702	import/syntax pass	CPU MoE + LoRA fixes - commit reference only
step-12	9c0812f..ffee741	bd4b616	import/syntax pass	AXK1 model - commit reference only
step-13	ffee741..b3945cc	d775b2a	import/syntax pass	CPU kernels bump - commit reference only
step-14	b3945cc..2a16ece	341a48e	import/syntax pass	fastokens tokenizer - commit reference only

Changes Made

• Updated vLLM commit reference from c7aa186d67b6f051680831418e957c67f34ba7a2 to 2a16ece2d342c0c154a4949ad317b521f8c04ec4
• model_runner_v1.py: Fixed PP sampled-token receive condition (world_size > 1 → not is_last_rank)
• dflash_proposer.py: Renamed _raise_if_multimodal to _warn_if_multimodal to match upstream base class

CI Verification

• Import checks passed for all steps
• Syntax checks passed for all steps
• Full CI (e2e-main2main suite) not run - Ascend NPU hardware unavailable in this environment

Adaptation Summary

Only 2 of 14 steps required code adaptation:

1. Step-2: Bugfix for PP sampled-token receive condition in model_runner_v1.py. The upstream changed get_pp_group().world_size > 1 to not get_pp_group().is_last_rank to correctly receive sampled tokens only on non-last ranks.

2. Step-2: Speculative decoding method rename in dflash_proposer.py. The upstream changed _raise_if_multimodal to _warn_if_multimodal to allow multimodal models with a warning instead of raising an error.

All other steps were commit reference updates only because:

• CPU/RISC-V specific kernel changes (Ascend uses its own kernels)
• ROCm/XPU specific changes (not relevant to Ascend platform)
• New feature additions (nixl refactor, Gemma4 MTP, fastokens) that don't change existing interfaces
• Internal implementation changes (detokenizer import path) that don't affect vllm-ascend

Pre-Completion Checklist

• No temp files in the repo
• All commits signed (git commit -s)
• All intermediate files in /tmp/main2main/, not in the repo
• conf.py main_vllm_commit updated to target commit
• conf.py main_vllm_tag unchanged (v0.20.1 - no new tag released)
• Import/syntax verification performed for each step
• Each commit message includes the upstream commit range
• Final summary written to /tmp/main2main/final-summary.md

Notes

• This run was executed in an environment without Ascend NPU hardware, so full CI verification could not be performed.

• The pipeline should be verified by running e2e-main2main suite on Ascend hardware before merging to main.

• No pushes or PRs were created as per the task requirements.

• vLLM version: v0.20.1

• vLLM main: vllm-project/vllm@c7aa186