getting a couple of errors " No model.safetensors.index.json found in remote" and "assert hidden_states.dtype == torch.bfloat16, "hidden_states must be bfloat16"

#1
by bluenevus - opened

getting a couple of errors " No model.safetensors.index.json found in remote" and "assert hidden_states.dtype == torch.bfloat16, "hidden_states must be bfloat16"

Symptoms:

Assertion failures and hidden-state mismatches when loading models with recent versions of the transformers library.
Root Cause: A guard statement introduced upstream now enforces stricter type alignment for tensor dtypes (bfloat16 → fp16).

Resolution:

Two mutually-exclusive upgrade paths are available:

  1. Full Re-download (Recommended)
    Delete the local model directory.
    Re-pull the latest weights from the Hugging Face repository (main branch).
    No further action required.

  2. In-place Patch (Lightweight fallback)
    Run python fp16.py located in the repository root.
    The script performs an in-place dtype cast from bfloat16 to fp16 using sequential on-loading, eliminating the need for additional GPU memory or large RAM buffers.

Verify SHA-256 checksums after completion to ensure integrity.
Both paths resolve the compatibility issue; choose whichever option best aligns with your bandwidth and compute constraints.

twhitworth changed discussion status to closed

Sign up or log in to comment