Semble 0.5.0 cannot recover from an incomplete Hugging Face model download

[Szpadel]

After upgrading to Semble 0.5.0, the MCP server started failing when it tried to load the new default model:

ValueError: Could not find expected model files in
~/.cache/huggingface/hub/models--minishlab--potion-code-16M-v2/snapshots/d3daf3e31f36d78f75913030b8bdf4a505d5b833.
Tried model2vec, sentence-transformers, and 0_StaticEmbedding layouts.

I’m running it through Codex as:

uvx --from "semble[mcp]" semble

The Hugging Face snapshot was only partially downloaded. It contained config.json and tokenizer.json, but model.safetensors was missing. The corresponding blob was left as a zero-byte .incomplete file.

Semble loads the model with:

StaticModel.from_pretrained(model_path, force_download=False)

With that option, Model2Vec finds the existing snapshot directory and treats it as cached. It then attempts to load the incomplete directory instead of asking Hugging Face to finish the download. Every subsequent invocation therefore fails with the same error.

Removing the incomplete model cache fixes the problem:

rm -rf ~/.cache/huggingface/hub/models--minishlab--potion-code-16M-v2

The next invocation downloads the model again and can proceed normally.

It would be useful if Semble could recover automatically from an incomplete Hugging Face snapshot.

[Pringled]

Hey @Szpadel, thanks for reporting, this can indeed be handled more gracefully. For context, we use force_download=False to skip HuggingFace checks if the model already exists, which speeds up the loading of the model once it has been download once. However, doing this check would also automatically fix this issue. We'll think about the best solution for fixing this while keeping the speed benefits, either in Semble or Model2vec.