What
KempnerForge loads vision encoders and tokenizers via HuggingFace from_pretrained, which caches downloads in the default HF location (under the user's home directory). On many clusters home storage is limited, so a large download (e.g. a SigLIP2 vision tower) can exceed the available space and the run fails. Let users set the HF cache location once in the per-user cluster config, and have check_env export it as HF_HOME at launch — consistent with the existing HF-token / WANDB checks and the ib_interface → NCCL_SOCKET_IFNAME export pattern. This is convenience over setting the env var by hand, and is complementary to pointing vision_encoder.path at a local directory (which skips downloading entirely).
Scope
configs/cluster/local.toml(.example): an optional cache-dir field (e.g. [paths].hf_cache).scripts/check_env.py: read it and export HF_HOME into the generated launch scripts; optionally warn if the location looks unsuitable.- Docs: a short how-to note (cache location + the local-path alternative).
- Decision: export at launch only, or also set it early in
train.py (before transformers is imported) to cover interactive / direct runs.
Backward compatibility
Fully backward compatible — new optional field; unset = current behavior. No change to the training runtime or model code.
What
KempnerForge loads vision encoders and tokenizers via HuggingFace
from_pretrained, which caches downloads in the default HF location (under the user's home directory). On many clusters home storage is limited, so a large download (e.g. a SigLIP2 vision tower) can exceed the available space and the run fails. Let users set the HF cache location once in the per-user cluster config, and havecheck_envexport it asHF_HOMEat launch — consistent with the existing HF-token / WANDB checks and theib_interface → NCCL_SOCKET_IFNAMEexport pattern. This is convenience over setting the env var by hand, and is complementary to pointingvision_encoder.pathat a local directory (which skips downloading entirely).Scope
configs/cluster/local.toml(.example): an optional cache-dir field (e.g.[paths].hf_cache).scripts/check_env.py: read it and exportHF_HOMEinto the generated launch scripts; optionally warn if the location looks unsuitable.train.py(beforetransformersis imported) to cover interactive / direct runs.Backward compatibility
Fully backward compatible — new optional field; unset = current behavior. No change to the training runtime or model code.