ngxson's picture
ngxson HF Staff
keep the output tensor in F16, the clef head reads its rows
38bc353 verified
|
Raw History Blame Contribute Delete
1.29 kB
---
license: apache-2.0
base_model: ggml-org-dev/tiny-clef-for-testing
tags:
- gguf
- testing
---
# tiny-clef-for-testing-GGUF
> [!WARNING]
> This is a **modified, damaged copy** of [Cloudflare/clef](https://huggingface.co/Cloudflare/clef), made only to test software. Its output is gibberish. Do not use it for anything else.
GGUF of [ggml-org-dev/tiny-clef-for-testing](https://huggingface.co/ggml-org-dev/tiny-clef-for-testing), converted with `convert_hf_to_gguf.py` of llama.cpp (text model and mmproj in F16, then text model quantized to Q4_0 with `llama-quantize`, token embeddings included, and mmproj quantized to Q8_0). The output tensor stays F16: the clef head also reads its rows with GET_ROWS, which the CPU repacked Q4_0 buffer does not support. The vocab of the text GGUF is also cut with `slice_vocab_gguf.py` (special tokens, single characters, the BPE path of the decision label words and the first 2000 merges are kept, token ids are remapped): it tokenizes text differently from the original model, the HF repo keeps the full tokenizer. Used by the llama.cpp server tests of the `clef` decision type.
```sh
llama-server -hf ggml-org-dev/tiny-clef-for-testing-GGUF:Q4_0
```
## License
Derived from clef by Cloudflare, under the [Apache License 2.0](LICENSE).