chore: bump vLLM/vLLM-Omni to 0.22.0 and adapt the worker stack#70

Merged

kaiitunnz merged 6 commits into

mainfrom

kaiitunnz/chore/bump-vllm-0.22

Jun 13, 2026

Commits on Jun 13, 2026

chore: bump vllm/vllm-omni to 0.22.0 across the inference/training stack
kaiitunnz
committed
fix(worker): adapt executors to transformers 5 / trl 0.23 / diffusers 0.38
kaiitunnz
committed
test: default the unit suite to CPU-only
kaiitunnz
committed
refactor(worker): narrow tokenizer.decode() with assert instead of cast
kaiitunnz
committed
fix(worker): let the 4-bit dtype probe fail fast on CUDA init error
kaiitunnz
committed
chore: pin flashinfer to vllm's exact version and bump deepspeed
kaiitunnz
committed