Repository navigation

[Rebase] Rebase to vLLM 0.29.0 - #7230

Merged

Gaohan123 merged 31 commits into

mainfrom

dev/vllm-align

Sep 10, 2026

Commits on Sep 9, 2026
- rebase: initialize vLLM v0.29 alignment
  tzhouam
  committed
  6a63f5c
- rebase: adapt to vLLM v0.29 APIs
  tzhouam
  committed
  87a1fd8
- chore: satisfy changed-file policy
  tzhouam
  committed
  f987923
- rebase: migrate diffusion KV cache to the vLLM 0.29 layout model
  
  tzhouam
  and
  claude
  committed
  c306445
- rebase: adapt to vLLM 0.29 callback, embedding and video decoder changes
  
  tzhouam
  and
  claude
  committed
  08884c4
- chore: satisfy changed-file policy
  
  tzhouam
  and
  claude
  committed
  9a399ad
- rebase: close the remaining vLLM 0.29 drift found by CI 3009
  
  tzhouam
  and
  claude
  committed
  c1e32fc
- test: guard prompt-update helpers against upstream removal
  
  tzhouam
  and
  claude
  committed
  85d1f92
- rebase: rename Qwen3-Omni audio attention projection to qkv_proj
  
  tzhouam
  and
  claude
  committed
  b5146fa
- rebase: keep the omni scheduler on the v1 model runner
  
  tzhouam
  and
  claude
  committed
  94914a1
- rebase: align vllm-omni with vLLM d2906cc19586
  tzhouam
  committed
  e4c47a2
- rebase: align vllm-omni with vLLM d2906cc19586
  tzhouam
  committed
  9af0331
- [CI] Preserve secret-backed HF token in scheduled GPU jobs (#7118)
  tzhouam
  committed
  90b7b6c
- rebase: remove speech import superseded by main adapter migration
  tzhouam
  committed
  51c5a28
- test: measure dots TTS PCM quality at its native 48 kHz rate
  tzhouam
  committed
  4653b49
- fix: keep frontend log-stats flag out of stage ownership validation
  tzhouam
  committed
  b799118
- fix: preserve Qwen Image rotary frequency precision on CUDA
  tzhouam
  committed
  b751b6e
- test: align SANA SP CPU fixtures with vLLM 0.29 GEMM dispatch
  tzhouam
  committed
  26a37ab
- ci: advance vLLM 0.29 release pin to 98dff2a81d74
  tzhouam
  committed
  eef8e2b
- refactor: reduce copied vLLM 0.29 compatibility code
  tzhouam
  committed
  efda936
- refactor(processors): drop _hf_processor_applies_updates overrides removed in vLLM 0.29
  
  tzhouam
  and
  claude
  committed
  188d021
- refactor(processors): drop _call_hf_processor overrides upstream no longer calls
  
  tzhouam
  and
  claude
  committed
  1439847
- refactor(voxtral_tts): drop pass-through _cached_apply_hf_processor
  
  tzhouam
  and
  claude
  committed
  ec3880d
- refactor(qwen3_omni): inherit _maybe_apply_prompt_updates from the Qwen2.5-Omni processor
  
  tzhouam
  and
  claude
  committed
  2f45482
- perf(omnivoice): tokenize the prompt once per request
  
  tzhouam
  and
  claude
  committed
  a58825e
- refactor(inputs): build the Omni renderer as a subclass instead of a proxy
  
  tzhouam
  and
  claude
  committed
  fb94b87
- test(minicpmo): teach the audio_embeds tokenizer stub the 0.29 token-id contract
  
  tzhouam
  and
  claude
  committed
  f270ff7
- test(qwen3_tts): accept the mapper argument in the code2wav weight-loader double
  
  tzhouam
  and
  claude
  committed
  c4b3747
- fix(glm_image): build the target-shape scaffold without HF's private formatter
  
  tzhouam
  and
  claude
  committed
  121d0b9
Commits on Sep 10, 2026
- ci: use vLLM 0.29 release image and align installation docs
  tzhouam
  committed
  80a0d8d
- Merge branch 'main' into dev/vllm-align
  Gaohan123
  authored
  0ae9bdd

Navigation Menu

Repository navigation

[Rebase] Rebase to vLLM 0.29.0 - #7230

[Rebase] Rebase to vLLM 0.29.0#7230
Gaohan123 merged 31 commits into
mainfrom
dev/vllm-align

Commits on Sep 9, 2026

rebase: initialize vLLM v0.29 alignment

rebase: adapt to vLLM v0.29 APIs

chore: satisfy changed-file policy

rebase: migrate diffusion KV cache to the vLLM 0.29 layout model

rebase: adapt to vLLM 0.29 callback, embedding and video decoder changes

chore: satisfy changed-file policy

rebase: close the remaining vLLM 0.29 drift found by CI 3009

test: guard prompt-update helpers against upstream removal

rebase: rename Qwen3-Omni audio attention projection to qkv_proj

rebase: keep the omni scheduler on the v1 model runner

rebase: align vllm-omni with vLLM d2906cc19586

rebase: align vllm-omni with vLLM d2906cc19586

[CI] Preserve secret-backed HF token in scheduled GPU jobs (#7118)

rebase: remove speech import superseded by main adapter migration

test: measure dots TTS PCM quality at its native 48 kHz rate

fix: keep frontend log-stats flag out of stage ownership validation

fix: preserve Qwen Image rotary frequency precision on CUDA

test: align SANA SP CPU fixtures with vLLM 0.29 GEMM dispatch

ci: advance vLLM 0.29 release pin to 98dff2a81d74

refactor: reduce copied vLLM 0.29 compatibility code

refactor(processors): drop _hf_processor_applies_updates overrides removed in vLLM 0.29

refactor(processors): drop _call_hf_processor overrides upstream no longer calls

refactor(voxtral_tts): drop pass-through _cached_apply_hf_processor

refactor(qwen3_omni): inherit _maybe_apply_prompt_updates from the Qwen2.5-Omni processor

perf(omnivoice): tokenize the prompt once per request

refactor(inputs): build the Omni renderer as a subclass instead of a proxy

test(minicpmo): teach the audio_embeds tokenizer stub the 0.29 token-id contract

test(qwen3_tts): accept the mapper argument in the code2wav weight-loader double

fix(glm_image): build the target-shape scaffold without HF's private formatter

Commits on Sep 10, 2026

ci: use vLLM 0.29 release image and align installation docs

Merge branch 'main' into dev/vllm-align