GGUF: missing split.no metadata

[ngxson]

Related to: ggerganov/llama.cpp#6343 (comment)

Explanation: We recently introduced gguf-split tool to llama.cpp, which allows user to split the model into smaller shards. Each shard has 3 metadata to know its info:

• split.count: Total number of splits
• split.no: The number of the current split
• split.tensors.count: Total number of tensors of the original model (= sum of tensors of all splits)

The split.no is however missing when viewing from GGUF viewer on huggingface. It is still visible when inspecting using gguf-py

This can be reproduce using a smaller model: https://huggingface.co/ngxson/tinyllama_split_test/tree/main?show_tensors=stories15M-q8_0-00001-of-00003.gguf

Here is the command that I used to split the model:

./gguf-split --split-max-size 10M ~/Downloads/stories15M-q8_0.gguf ~/Downloads/stories15M-q8_0

I'd be happy to help you guys with this. Feel free to let me know if you need more info.

[mishig25]

@ngxson you can run the gguf js parser in any node/js env without needing hf frontend, as described in readme https://github.com/huggingface/huggingface.js/blob/main/packages%2Fgguf%2FREADME.md

If you can produce a snippet to reproduce the bug or even better a fix for the gguf js parser, that would be great !

[phymbert]

@julien-c The bug seems easy to fix: do not remove metadata with 0 as value :)
Please include all metadata regardless their value.

[mishig25]

@phymbert thanks a lot for debugging. There was no problem on the js/parser side (i.e. metadata had split.no: 0). However, frontend was treating 0 as falsy, therefore not rendering. Hence, I've submitted a PR that fixes the issue on the frontend

[phymbert]

Cool! thanks @mishig25 for the explanations

[mishig25]

Keeping the issue open until the UI fix is deployed (the fix is already merged)

[mishig25]

The fix is deployed now. You can see results https://huggingface.co/ngxson/tinyllama_split_test/tree/main?show_tensors=stories15M-q8_0-00001-of-00003.gguf