Bug: Missing Sanity Check in convert_hf_to_gguf.py

[mesibo]

What happened?

The script convert_hf_to_gguf.py does not perform a sanity check on the completeness of the model files. We encountered an issue where one of the models we downloaded was incomplete (specifically, a safesensors file was missing). Despite this, convert_hf_to_gguf.py proceeded with the conversion without generating any errors or warnings.

Expected Behavior: The script should validate that all required files are present before converting. Since Hugging Face downloads are organized with sequence and total files (e.g., model-00001-of-00002), it should be easy to validate.

Steps to Reproduce:

1. Download a model from HF and delete one of the safesensors file.
2. Run convert_hf_to_gguf.py on the incomplete set.
3. the script completes without any indication of missing files.

Name and Version

version: 3631 (7d787ed)
built with cc (GCC) 8.5.0 20210514 (Red Hat 8.5.0-22) for x86_64-redhat-linux

What operating system are you seeing the problem on?

Linux

Relevant log output

No response

[compilade]

Steps to Reproduce:

1. Download a model from HF and delete one of the safesensors file.
2. Run convert_hf_to_gguf.py on the incomplete set.
3. the script completes without any indication of missing files.

@mesibo

I tried, but I can't reproduce not getting an error. (tried on version 8f1d81a, but the following sanity check has been in convert_hf_to_gguf.py since #7075, which was merged almost 4 months ago)

I get

ValueError: Mismatch between weight map and model parts for tensor names:

and then a big list of missing tensors.

This comes from

https://github.com/ggerganov/llama.cpp/blob/8f1d81a0b6f50b9bad72db0b6fcd299ad9ecd48c/convert_hf_to_gguf.py#L174-L176

[compilade]

@mesibo

I just now realized that you're right, when there's only one .safetensors detected there is no sanity check, even when there still is a model.safetensors.index.json.

So what you're describing can be reproduced by having only a single .safetensors of the model when it should have been multi-part. In my previous test I used a model with 4 parts and removed one, so it wasn't enough to reproduce the problem.

This could be improved by cross-checking the index files when present even when only a single model file is detected.