build: bump vllm to 0.29.0 - #4080

Draft

yfw wants to merge 31 commits into

mainfrom

Commits on Sep 10, 2026

build: bump vllm to 0.28.0
yfw
committed
83fa485
fix: use supported PyTorch all-gather API
yfw
committed
3558f0a
build: regenerate lockfile for Docker uv compatibility
yfw
committed
2e0bdb8
build: fix the vLLM 0.28 dependency set exposed by the nightlies

yfw
and
claude
committed
9d8f808
fix(vllm): port fp8 MoE kernel setup and KV-cache scale processing to vLLM 0.28

yfw
and
claude
committed
1734158
fix(vllm): keep multimodal rescale/normalize on the CPU processor

yfw
and
claude
committed
f6993c8

Commits on Sep 14, 2026

build: pin ModelOpt to 5dde396bd, not main

yfw
and
claude
committed
e991336
fix(modelopt): route per-expert fakequant amax around vLLM 0.28's expert loader

yfw
and
claude
committed
ee35b44
build: bump vllm 0.28.0 → 0.29.0

yfw
and
claude
committed
e179799
fix(vllm): recognise vLLM 0.29's upstream TCPStore fix in the port patch

yfw
and
claude
committed
a419679
fix(vllm): stop shadowing vLLM 0.29's Worker.synchronize_device in the extension

yfw
and
claude
committed
ae4ebcf
build(docker): keep the TRT-LLM wheel in the image so cached rebuilds find it

yfw
and
claude
committed
40f5bcc
fix(vllm): import ErrorResponse from its vLLM 0.29 location

yfw
and
claude
committed
ae6507c
fix(vllm): drop tied-embedding aliases before vLLM 0.29's load_weights

yfw
and
claude
committed
e1c16e8
fix(modelopt): emit non-gated real-quant w13 experts per expert for vLLM >= 0.28

yfw
and
claude
committed
a3aa6cd
fix(vllm): proxy check_admission through the async worker's engine client

yfw
and
claude
committed
710d278
test(nvfp4): accept the cached-checkpoint restore wording in the qwen3 nvfp4 recipe

yfw
and
claude
committed
3e3dea6
build(audio): pair torchcodec with torch 2.13 in the audio-deps installer

yfw
and
claude
committed
2c0ad3c
docs(audio): state the torchcodec pin the audio-deps installer applies

yfw
and
claude
committed
72b3b35
fix(modelopt): map checkpoint names to vLLM names before routing expert amax

yfw
and
claude
committed
f3fea21
fix(vllm): mark the tied-alias log flag as an implicitly defined attribute

yfw
and
claude
committed
6ce1256
Merge remote-tracking branch 'origin/main' into yifu/vllm0_29
yfw
committed
6ec3dd6
build(lint): whitelist vllm_quant_moe_amax.py in pyrefly.toml

yfw
and
claude
committed
d8e1ab6

Commits on Sep 17, 2026

build: pin torch per backend extra instead of globally

yfw
and
claude
committed
237a52c
build: bound the base torch range so the driver cannot drift on a relock

yfw
and
claude
committed
04b9dc9
fix(refit): let torch-2.11 inference venvs open CUDA IPC handles from torch 2.13

yfw
and
claude
committed
94dd20b
Merge remote-tracking branch 'origin/main' into vllm029-merge-prep
yfw
committed
be83d77
build(docker): drop the release-stage TRT-LLM wheel re-seed superseded by #4153

yfw
and
claude
committed
fe6945a

Commits on Sep 18, 2026

fix(vllm): defer the DeepSeek V4 post-load hook during layerwise refit

yfw
and
claude
committed
10e1a86
Merge remote-tracking branch 'origin/main' into yifu/vllm0_29
yfw
committed
92a1e7c
Merge remote-tracking branch 'origin/main' into yifu/vllm0_29
yfw
committed
ebc8f0c