# HFIE stock Python 3.11 container ships torch + transformers + HF # inference toolkit pre-baked. We override transformers to the version # lerobot[pi]==0.5.1 was built against and pin the safe datasets release # (4.8.x imports BucketNotFoundError which only exists on hub `main`). # # lerobot 0.5.1 declares requires-python>=3.12 in its metadata. The # patch_lerobot_for_py311.py script in this repo rewrites the four # PEP 695 sites so the code itself runs on 3.11, but pip's metadata # check still fails by default. Set PIP_IGNORE_REQUIRES_PYTHON=1 in the # endpoint's environment variables (HFIE config) to skip it. # # The shim package drops a .pth file into site-packages that re-exports # the names the HF inference toolkit imports from transformers.file_utils # (`is_tf_available`, `is_torch_available`). transformers 5.x dropped # `is_tf_available` entirely (TF support removed). Without the shim the # toolkit's webservice_starlette fails to start before our handler.py is # ever loaded. # # Note: HFIE's pip install runs from a cwd OTHER than /repository, so # `./shim` doesn't resolve. We use the absolute /repository/shim path # (HFIE-specific: the runtime always mounts the model repo there). /repository/shim transformers==5.3.0 datasets==4.7.0 huggingface_hub>=0.36 lerobot==0.5.1