build: bump vllm 0.20.0 → 0.25.1 - #3280

Merged

terrykong merged 42 commits into

mainfrom

terryk/bump-vllm-0.25.1

Jul 30, 2026

Commits on Jul 29, 2026

build: bump vllm 0.20.0 -> 0.25.1
terrykong
committed
b933510
fix: synchronize packed-broadcast side streams before returning
terrykong
committed
b9807ce
fix: port ModelOpt fakequant to vllm 0.25
terrykong
committed
1e29e51
fix: keep the vllm 0.25 TCPStore port out of the MessageQueue scan range
terrykong
committed
a196a4b
fix: update fp8 linear weights in place on refit to stop GPU memory creep
terrykong
committed
81e94b7
fix: pin fp8 recipe KV cache to 0.20-equivalent size for vllm 0.25 wake-up
terrykong
committed
e238818
test: relax CI metric thresholds for vllm 0.25.1
terrykong
committed
f80137d
fix: port sharded NIXL expert refit to the vllm 0.25 MoE layout
terrykong
committed
c11384f
fix: port MXFP8 rollout refit helpers to vllm 0.25
terrykong
committed
6efd82d
fix: port ModelOpt W4A16 NVFP4 rollout to vllm 0.25
terrykong
committed
eeaf31f
test: adapt real-quant assertions and refit diagnostics to vllm 0.25
terrykong
committed
128568b
fix: load gated ModelOpt MoE refit weights per expert on vllm 0.25
terrykong
committed
155b51a
test: re-enable the Qwen3.5 AutoModel nightlies fixed by vllm 0.25.1
terrykong
committed
478ccde
test: move the Qwen3.5-397B recipe from disabled to the release suite
terrykong
committed
4b19347
fix(vllm): retry MessageQueue remote bind so node-spanning engines start
terrykong
committed
0a67e79
test: make the vllm patch fixtures independent of test order
terrykong
committed
ac79fe8
test: assert the tcpstore invariant instead of one environment's port
terrykong
committed
2f4f631
fix(refit): stream a parameter larger than the IPC staging buffer
terrykong
committed
838a1d3
test: raise the nightly compute cap to 3600 hours after rebasing on main
terrykong
committed
2045f6f
Revert "test: relax CI metric thresholds for vllm 0.25.1"
terrykong
committed
aebd09d
fix(diagnostics): keep script 5 runnable in a standalone bisect env
terrykong
committed
966bce0
fix(modelopt): fail loudly if the triton_moe kernel namespace moves
terrykong
committed
af817d7
fix(refit): fail loudly if vLLM expert params leave .routed_experts
terrykong
committed
ea654be
chore(modelopt): log which QuantModuleRegistry keys the purge drops
terrykong
committed
0e9cec9
chore(fp8): log the MXFP8 CuTe-DSL -> CUTLASS kernel swap
terrykong
committed
5e8a94f
fix(vllm): report whether the _init_workers_ray patch actually applied
terrykong
committed
04b32d3
test: cover the NamespaceTool patch and the Ray env-var merge
terrykong
committed
17e20af
test: pin fp8 in-place weight update across refits
terrykong
committed
6d6c51d
test: pin the .routed_experts. layout guard
terrykong
committed
8e47e59
style: sort the imports added by the review fixes
terrykong
committed
cc34230
test: reach the routed_experts guard instead of the EP ownership check
terrykong
committed
d2be743
test: raise the megatron PP test timeout for vllm 0.25's slower init
terrykong
committed
697953d
test: run the _init_workers_ray guards in the cheap lane
terrykong
committed
cfdf1a3
docs(modelopt): re-sync the vLLM references the bump invalidated
terrykong
committed
839f441
docs(vllm): stop implying the v1 _init_workers_ray patch is what runs
terrykong
committed
d2c5349
test: unblock the bump on the PPO critic grad-norm, tracked in #3412
terrykong
committed
4dba76a
refactor(vllm): drop the legacy v1 Ray executor patch
terrykong
committed
1ca7db7

Commits on Jul 30, 2026