Repository navigation
fix: propagate revision/code_revision pins to all artifact boundaries - #42616
Conversation
There was a problem hiding this comment.
Code Review
This pull request updates several model loaders and configurations to consistently use the revision and code_revision attributes from the model configuration. Specifically, it modifies the GGUF loader, Kimi Audio, Kimi K25, and RoBERTa models, as well as the model registry, to ensure that specific model versions are correctly identified during download and initialization. I have no feedback to provide.
|
Hi @jperezdealgaba, the pre-commit checks have failed. Please run: uv pip install pre-commit>=4.5.1
pre-commit install
pre-commit run --all-filesThen, commit the changes and push to your branch. For future commits, Tip Is
|
When an operator supplies --revision or --code-revision, vLLM did not consistently forward these pins to all artifact load paths. This patch closes six pin-decay gaps so that every code, config, processor, and weight artifact loaded for a model resolves under the operator's configured revision. Fixes: - registry.py: forward code_revision to try_get_class_from_dynamic_module - gguf_loader.py: pass revision to hf_hub_download for repo/file.gguf - roberta.py: use model_config.revision for BGE-M3 secondary weights - kimi_k25.py: pass revision to cached_get_image_processor - kimi_audio.py: pass revision to WhisperConfig.from_pretrained and secondary whisper-large-v3 weight source Signed-off-by: jperezde <jperezde@redhat.com>
37d5df0 to
54d1835
Compare
|
@DarkLight1337 Do you mind checking this? I think the errors are not caused by my changes |
|
That failure should be fixed on main already, let's rebase and see if the tests pass now |
…vllm-project#42616) Signed-off-by: jperezde <jperezde@redhat.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
…vllm-project#42616) Signed-off-by: jperezde <jperezde@redhat.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
…vllm-project#42616) Signed-off-by: jperezde <jperezde@redhat.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk> Signed-off-by: Liuweixiong0118 <lwx34158427@gmail.com>
…vllm-project#42616) Signed-off-by: jperezde <jperezde@redhat.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Summary
When an operator supplies
--revisionor--code-revision, vLLM did not consistently forward these pins to all artifact load paths. This patch closes six pin-decay gaps so that every code, config, processor, and weight artifact loaded for a model resolves under the operator's configured revision.Fixes
code_revisiontotry_get_class_from_dynamic_modulerevisiontohf_hub_downloadforrepo/file.ggufmodel_config.revisionfor BGE-M3 secondary weightsrevisiontocached_get_image_processorrevisiontoWhisperConfig.from_pretrainedand secondary whisper-large-v3 weight sourceTest plan
test_gguf_download.pyto verifyrevisionandcache_dirare passed through in_prepare_weights--revisionis suppliedSigned-off-by: jperezde jperezde@redhat.com