Skip to content
forked from meta-llama/llama

Quantized inference code for LLaMA models

License

Notifications You must be signed in to change notification settings

tloen/llama-int8

 
 

Repository files navigation

LLaMA: INT8 edition

⚠️ 2023-03-16: LLaMA is now supported in Huggingface transformers, which has out-of-the-box int8 support. I'll keep this repo up as a means of space-efficiently testing LLaMA weights packaged as state_dicts, but for serious inference or training workloads I encourage users to migrate to transformers. Instructions for converting weights can be found here.

This is a fork of the LLaMA code that runs LLaMA-13B comfortably within 24 GiB of RAM. It relies almost entirely on the bitsandbytes and LLM.int8() work of Tim Dettmers. I've tested it on an RTX 4090, and it reportedly works on the 3090. It might also theoretically allow us to run LLaMA-65B on an 80GB A100, but I haven't tried this.

The code contains the following changes:

  • Removes parallelism constructs
  • Quantizes weights on the host machine
  • Loads weights incrementally to avoid severe memory problems
  • Added dependencies on bitsandbytes, tqdm.
  • Repetition penalty settings (--repetition_penalty, default 1.15)

On my Ubuntu machine with 64 GB of RAM and an RTX 4090, it takes about 25 seconds to load in the floats and quantize the model. Users should be ready to expand their swapfiles if they don't have enough RAM. Llamanon has also produced a slightly uncouth user's guide for using this repo, which I won't reproduce here but seems generally trustworthy. You will likely need to build bitsandbytes from source.

If you have interesting ideas for further development, I can be reached at https://twitter.com/ecjwg.

Usage:

python example.py --ckpt_dir [TARGET_DIR]/13B --tokenizer_path [TARGET_DIR]/tokenizer.model --max_batch_size=1


Original README

This repository is intended as a minimal, hackable and readable example to load LLaMA (arXiv) models and run inference. In order to download the checkpoints and tokenizer, fill this google form

Setup

In a conda env with pytorch / cuda available, run

pip install -r requirements.txt

Then in this repository

pip install -e .

Download

Once your request is approved, you will receive links to download the tokenizer and model files. Edit the download.sh script with the signed url provided in the email to download the model weights and tokenizer.

Inference

The provided example.py can be run on a single or multi-gpu node with torchrun and will output completions for two pre-defined prompts. Using TARGET_FOLDER as defined in download.sh:

torchrun --nproc_per_node MP example.py --ckpt_dir $TARGET_FOLDER/model_size --tokenizer_path $TARGET_FOLDER/tokenizer.model

Different models require different MP values:

Model MP
7B 1
13B 2
33B 4
65B 8

FAQ

Model Card

See MODEL_CARD.md

License

See the LICENSE file.

About

Quantized inference code for LLaMA models

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 94.0%
  • Shell 6.0%