Spaces:
Running
Running
File size: 2,780 Bytes
939262b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 |
GGUF and interaction with Transformers
The GGUF file format is used to store models for inference with GGML and other
libraries that depend on it, like the very popular llama.cpp or
whisper.cpp.
It is a file format supported by the Hugging Face Hub with features
allowing for quick inspection of tensors and metadata within the file.
This file format is designed as a "single-file-format" where a single file usually contains both the configuration
attributes, the tokenizer vocabulary and other attributes, as well as all tensors to be loaded in the model. These
files come in different formats according to the quantization type of the file. We briefly go over some of them
here.
Support within Transformers
We have added the ability to load gguf files within transformers in order to offer further training/fine-tuning
capabilities to gguf models, before converting back those models to gguf to use within the ggml ecosystem. When
loading a model, we first dequantize it to fp32, before loading the weights to be used in PyTorch.
[!NOTE]
The support is still very exploratory and we welcome contributions in order to solidify it across quantization types
and model architectures.
For now, here are the supported model architectures and quantization types:
Supported quantization types
The initial supported quantization types are decided according to the popular quantized files that have been shared
on the Hub.
F32
Q2_K
Q3_K
Q4_0
Q4_K
Q5_K
Q6_K
Q8_0
We take example from the excellent 99991/pygguf Python parser to dequantize the
weights.
Supported model architectures
For now the supported model architectures are the architectures that have been very popular on the Hub, namely:
LLaMa
Mistral
Qwen2
Example usage
In order to load gguf files in transformers, you should specify the gguf_file argument to the from_pretrained
methods of both tokenizers and models. Here is how one would load a tokenizer and a model, which can be loaded
from the exact same file:
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF"
filename = "tinyllama-1.1b-chat-v1.0.Q6_K.gguf"
tokenizer = AutoTokenizer.from_pretrained(model_id, gguf_file=filename)
model = AutoModelForCausalLM.from_pretrained(model_id, gguf_file=filename)
Now you have access to the full, unquantized version of the model in the PyTorch ecosystem, where you can combine it
with a plethora of other tools.
In order to convert back to a gguf file, we recommend using the
convert-hf-to-gguf.py file from llama.cpp.
Here's how you would complete the script above to save the model and export it back to gguf:
tokenizer.save_pretrained('directory')
model.save_pretrained('directory')
!python ${path_to_llama_cpp}/convert-hf-to-gguf.py ${directory}
``` |