Repository navigation

Move vLLM to 0.29.0 and transformers to v5 - #1127

Closed

tengomucho wants to merge 43 commits into

mainfrom

vllm-update

Closed

Move vLLM to 0.29.0 and transformers to v5#1127
tengomucho wants to merge 43 commits into
mainfrom
vllm-update

Commits on Sep 11, 2026
- fix(vllm): adapt worker and platform to newer vLLM APIs
  
  tengomucho
  and
  claude
  committed
  1945738
- chore(deps): move vllm to 0.23.0
  
  tengomucho
  and
  claude
  committed
  141e93e
- chore(deps): move optimum to 2.3.x
  
  tengomucho
  and
  claude
  committed
  422b248
- fix(nxd): shim transformers.utils.fx removed in transformers v5
  
  tengomucho
  and
  claude
  committed
  15586b9
- fix(utils): replace transformers and hub APIs removed in v5
  
  tengomucho
  and
  claude
  committed
  60b473c
- fix(inference): read rope parameters from config.rope_parameters
  
  tengomucho
  and
  claude
  committed
  ed07809
- fix(seq2seq): drop beam search from the legacy generation mixin
  
  tengomucho
  and
  claude
  committed
  beab702
- fix(vllm): adapt worker to the 0.29 WorkerBase API
  
  tengomucho
  and
  claude
  committed
  59e5687
- chore(deps): move vllm to 0.29.0 and transformers to 5.x
  
  tengomucho
  and
  claude
  committed
  db94d28
- fix(vllm): follow the entrypoint modules moved in vLLM 0.29
  
  tengomucho
  and
  claude
  committed
  1e0b70a
- fix(config): do not fail to load a config when training is unavailable
  
  tengomucho
  and
  claude
  committed
  d88cc72
- fix(diffusion): use the image processor and controlnet module paths
  
  tengomucho
  and
  claude
  committed
  9757599
- fix(inference): do not assume configs carry a pad token id
  
  tengomucho
  and
  claude
  committed
  8794e7b
- test(decoder): export llama4 from a checkpoint with a valid config
  
  tengomucho
  and
  claude
  committed
  59ce7ba
- fix(moe): convert the expert weights fused by transformers v5
  
  tengomucho
  and
  claude
  committed
  7c96f9f
- fix(export): build the encoder-decoder cache with the v5 cache API
  
  tengomucho
  and
  claude
  committed
  ef0981d
- fix(export): trace with eager attention when tracing the attentions
  
  tengomucho
  and
  claude
  committed
  3f96094
- fix(export): select the CLIP text encoder outputs as a tuple
  
  tengomucho
  and
  claude
  committed
  a629ed1
- test(exporters): export the tiny flaubert model
  
  tengomucho
  and
  claude
  committed
  c41b68b
- test: use the tiny flaubert model in the shared model list
  
  tengomucho
  and
  claude
  committed
  402d151
- fix(export): return the traced model outputs as a tuple
  
  tengomucho
  and
  claude
  committed
  fc1aa82
- fix(inference): default the token type ids the model was traced with
  
  tengomucho
  and
  claude
  committed
  4a9113c
- test(inference): use the tiny flaubert model
  
  tengomucho
  and
  claude
  committed
  08ecc41
- fix(pipelines): stop reading the framework removed from pipelines
  
  tengomucho
  and
  claude
  committed
  07e6abd
Commits on Sep 12, 2026
- test(inference): drop the question answering pipeline test
  
  tengomucho
  and
  claude
  committed
  aa31642
- fix(nxd): install the fx shim in the exporters package too
  
  tengomucho
  and
  claude
  committed
  d81e916
Commits on Sep 14, 2026
- fix(nxd): install the fx shim at interpreter startup
  
  tengomucho
  and
  claude
  committed
  04b88bb
- fix(seq2seq): resolve the unset generation parameters
  
  tengomucho
  and
  claude
  committed
  cd0906a
- fix(t5): follow the transformers v5 attention contract
  
  tengomucho
  and
  claude
  committed
  7426845
- test(decoder): record the generations decided by a logit tie
  
  tengomucho
  and
  claude
  committed
  2a7f891
- fix(smolvlm): clamp the vision position coordinates
  
  tengomucho
  and
  claude
  committed
  e62baaa
- test(vlm): compare the tokens the neuron loop actually returns
  
  tengomucho
  and
  claude
  committed
  b4165ef
- fix(vllm): report a KV cache layout to the 0.29 engine core
  
  tengomucho
  and
  claude
  committed
  3dbcc0c
- fix(docker): copy the files the package build needs
  
  tengomucho
  and
  claude
  committed
  e44b726
- fix(serve): accept the underscore spelling of data_parallel_size
  
  tengomucho
  and
  claude
  committed
  0134a53
Commits on Sep 15, 2026
- fix(deps): depend on requests explicitly
  
  tengomucho
  and
  claude
  committed
  3a1ff86
- fix(training): replace the imports removed in transformers v5
  
  tengomucho
  and
  claude
  committed
  810d477
- fix(cache): adapt hub cache sync errors to huggingface_hub v1
  
  tengomucho
  and
  claude
  committed
  1413219
- fix(export): catch the httpx errors raised when the Hub is unreachable
  
  tengomucho
  and
  claude
  committed
  fa03585
Commits on Sep 16, 2026
- fix(deps): keep the CPU builds of torch in the resolver overrides
  
  tengomucho
  and
  claude
  committed
  1f227e2
- test(diffusers): release the exported pipelines before loading another
  
  tengomucho
  and
  claude
  committed
  390cb1f
- fix(moe): accept both fused and per-expert weight checkpoints
  
  tengomucho
  and
  claude
  committed
  b8d9599
- test(vllm): shut the engine core down when an engine test ends
  
  tengomucho
  and
  claude
  committed
  57d186a

Commits

Commits on Sep 11, 2026

Commits on Sep 12, 2026

Commits on Sep 14, 2026

Commits on Sep 15, 2026

Commits on Sep 16, 2026