convert : identify missing model files

[compilade]

Should fix #9245.

Previously, when models files (like model-00002-of-00002.safetensors) were missing, a cryptic error message was output, but not when only one model file was left.

convert_hf_to_gguf.py now reads the appropriate (model.safetensors|pytorch_model.bin).index.json when it exists instead of only when there are more than one model file.

I've also made the error message easier to understand by making it include a list of missing tensor files instead of the missing tensor names, when that information can be obtained from the index files.

This can be tested with python3 convert_hf_to_gguf.py --dry-run /path/to/model_dir by renaming the model files of multi-part models so that at least one of them doesn't match model*.safetensors or pytorch_model*.bin.

From my testing this seems to work correctly, and should proceed with conversion as before when all the tensors are present or if the index file is missing.

@mesibo Does this help with the problem you've noticed?

---

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low

[mesibo]

Thanks, it should, I will test it out.

Do you think the script should also validate the part sequence number for continuity, even if an index file is missing? For instance, should it throw an error or exit if seq != prev_seq + 1?