Repository navigation

[Rebase] Rebase to vllm 0.27.0 - #5976

Merged

Gaohan123 merged 48 commits into

mainfrom

dev/vllm-align

Aug 11, 2026

Merged

[Rebase] Rebase to vllm 0.27.0#5976
Gaohan123 merged 48 commits into
mainfrom
dev/vllm-align

Commits on Aug 8, 2026
- docker: install vLLM from the precompiled wheel at the rebase baseline
  
  tzhouam
  and
  claude
  committed
  a0cf8ea
- ci: sync Dockerfile.ci wheel pin/deps to vLLM 4b25076c8186
  tzhouam
  committed
  f80609c
Commits on Aug 9, 2026
- rebase: adapt omni modules to vLLM 0.27 upstream API changes
  
  tzhouam
  and
  claude
  committed
  bae7df6
- rebase: Phase 3 repairs verified green by their own tests
  
  tzhouam
  and
  claude
  committed
  a9da981
- sched: track discarded async tokens via num_stale_output_tokens
  
  tzhouam
  and
  claude
  committed
  59b9e71
- tests: mock FusedMoEFactory, the name vLLM 0.27 actually exports
  
  tzhouam
  and
  claude
  committed
  efa11ca
- tests: defer the scenarios import so --help does not load the engine
  
  tzhouam
  and
  claude
  committed
  b06e58c
- tests: give _HashableRequest the 0.27 stale-output counter
  
  tzhouam
  and
  claude
  committed
  d43ca5a
- tests: point the comfyui api_server at a writable storage dir
  
  tzhouam
  and
  claude
  committed
  319f457
- tests: mark mimo-audio and voxtral-tts expansion tests for their CI jobs
  
  tzhouam
  and
  claude
  committed
  4c9b8f3
Commits on Aug 10, 2026
- engine: register omni archs unconditionally so omni classes win the registry
  
  tzhouam
  and
  claude
  committed
  029a839
- tests: escalate voxtral-tts transcript grading to whisper large-v3
  
  tzhouam
  and
  claude
  committed
  a4a0497
- Revert test edits that were made to satisfy drifted harness config
  
  tzhouam
  and
  claude
  committed
  2bb5574
- docker: pin torch before flashinfer, not after
  
  tzhouam
  and
  claude
  committed
  8b2e581
- Merge branch 'main' into dev/vllm-align
  tzhouam
  authored
  ae32ef1
Commits on Aug 11, 2026
- ci: give the Custom Pipeline step 8gb of /dev/shm
  
  tzhouam
  and
  claude
  committed
  0dbcbbb
- fix: tolerate absent model_config.model_stage in non-staged runs
  
  tzhouam
  and
  claude
  committed
  e64c054
- fix: unwrap MediaWithBytes before handing videos to the HF processor
  
  tzhouam
  and
  claude
  committed
  0518f05
- ci: target the released vLLM 0.27.0 base image
  
  tzhouam
  and
  claude
  committed
  78028e1
- Revert "synthetic media cache dir write-probe fallback"
  
  tzhouam
  and
  claude
  committed
  d109df4
- fix: accept randomize_inputs in the omni _dummy_run overrides
  
  tzhouam
  and
  claude
  committed
  701bcc6
- fix: return a bare tensor from the qwen3-omni thinker when no capture is requested
  
  tzhouam
  and
  claude
  committed
  e3efeaa
- fix: honor size/width/height from extra_args in lingbot dimension resolution
  
  tzhouam
  and
  claude
  committed
  2ebda68
- fix: coerce numpy arrays to PIL in the text-to-image example output path
  
  tzhouam
  and
  claude
  committed
  06e36eb
- fix: fall back to matched torch SDPA when the FA3 hub kernel has no build
  
  tzhouam
  and
  claude
  committed
  213c149
- fix: never randomize generation-stage dummy ids during warmup
  
  tzhouam
  and
  claude
  committed
  dedda28
- fix: create the diffusion EP group with use_all2all so vLLM 0.27 MoE loads
  
  tzhouam
  and
  claude
  committed
  c382e80
- fix: match the diffusers reference backend to the adapter's fallback chain
  
  tzhouam
  and
  claude
  committed
  b6956a9
- test: report all bagel SHM probe pixels on mismatch, not just the first
  
  tzhouam
  and
  claude
  committed
  64e8cd3
- fix: fall back to eager when a regionally compiled block hits a compiler bug
  
  tzhouam
  and
  claude
  committed
  e3b4285
- fix: keep the eager-fallback wrapper signature-transparent for cache-dit
  
  tzhouam
  and
  claude
  committed
  30603e4
- test: re-golden bagel SHM img2img reference pixels to the 0.27-aligned output
  
  tzhouam
  and
  claude
  committed
  5c1fe35
- test: looser SSIM gate for the SDPA-fallback accuracy path only
  
  tzhouam
  and
  claude
  committed
  6638b5d
- fix: isolate the compile cache per additional stage replica at spawn
  
  tzhouam
  and
  claude
  committed
  9c31928
- test: run both quantization-quality arms eager so LPIPS measures quant only
  
  tzhouam
  and
  claude
  committed
  e39f260
- fix: prove factorable symbolic divisibility that torch 2.13's inductor lost
  
  tzhouam
  and
  claude
  committed
  acf245b
- revert: drop the compiled-block eager fallback; the inductor bug is fixed at root
  
  tzhouam
  and
  claude
  committed
  de3502c
- Revert "fix: coerce numpy arrays to PIL in the text-to-image example output path"
  tzhouam
  committed
  c59d1fc
- Merge branch 'main' into dev/vllm-align
  tzhouam
  committed
  6965a5a
- test: pin the stale-drain defects behind the seed-TTS WER regression
  
  tzhouam
  and
  claude
  committed
  e21bffa
- fix: seed the stale-output drain in scheduled-token units
  
  tzhouam
  and
  claude
  committed
  c2b6523
- fix: read duplex metrics tolerantly for plain-vLLM benchmark backends
  
  tzhouam
  and
  claude
  committed
  2dc5e51
- test: PSNR fallback gate for the SDPA-matched accuracy path
  
  tzhouam
  and
  claude
  committed
  9b4eccc
- Revert "test: PSNR fallback gate for the SDPA-matched accuracy path"
  tzhouam
  committed
  b26b908
- ci: restore h100_2 for the MiniCPM-o 4.5 merge job
  
  tzhouam
  and
  claude
  committed
  86bee6d
- refactor: address PR 5976 review comments from Gaohan123
  
  tzhouam
  and
  claude
  committed
  a1e8834
- chore: DCO remediation for prior dev/vllm-align commits
  tzhouam
  committed
  81aecaa
- style: apply ruff format to gpu_model_runner
  
  tzhouam
  and
  claude
  committed
  25b80bb

Commits

Commits on Aug 8, 2026

Commits on Aug 9, 2026

Commits on Aug 10, 2026

Commits on Aug 11, 2026