astronomer/Llama-3-8B-GPTQ-4-Bit

This model is generously created and made open source by Astronomer.

Astronomer is the de facto company for Apache Airflow, the most trusted open-source framework for data orchestration and MLOps.

---

Llama-3-8B-GPTQ-8-Bit

• Original Model creator: Meta Llama from Meta
• Original model: meta-llama/Meta-Llama-3-8B
• Built with Meta Llama 3
• Quantized by David Xue from Astronomer

MUST READ: Very Important!! Note About Untrained Special Tokens in Llama 3 Base (Non-instruct) Models & Fine-tuning Llama 3 Base

• If you intend to fine-tune this model with any added tokens, or fine-tune for instruction following, please use the untrained-special-tokens-fixed branch/revision.
• Special tokens such as the ones used for instruct are undertrained in Llama 3 base models.
• Credits: discovered by Daniel Han https://twitter.com/danielhanchen/status/1781395882925343058
• 

Important Note About Serving with vLLM & oobabooga/text-generation-webui

• For loading this model onto vLLM, make sure all requests have "stop_token_ids":[128001, 128009] to temporarily address the non-stop generation issue.
  • vLLM does not yet respect generation_config.json.
  • vLLM team is working on a a fix for this https://github.com/vllm-project/vllm/issues/4180
• For oobabooga/text-generation-webui
  • Load the model via AutoGPTQ, with no_inject_fused_attention enabled. This is a bug with AutoGPTQ library.
  • Under Parameters -> Generation -> Skip special tokens: turn this off (deselect)
  • Under Parameters -> Generation -> Custom stopping strings: add "<|end_of_text|>","<|eot_id|>" to the field

Description

This repo contains 4 Bit quantized GPTQ model files for meta-llama/Meta-Llama-3-8B.

This model can be loaded with less than 6 GB of VRAM (huge reduction from the original 16.07GB model) and can be served lightning fast with the cheapest Nvidia GPUs possible (Nvidia T4, Nvidia K80, RTX 4070, etc).

The 4 bit GPTQ quant has small quality degradation from the original bfloat16 model but can be served on much smaller GPUs with maximum improvement in latency and throughput.

The untrained-special-tokens-fixed branch is the same model as the main branch but has special tokens and tokens untrained (by finding the tokens where max embedding value of each token in input_embeddings and output_embeddings is 0) and setting them to the average of all trained tokens for each feature. Using this branch is recommended if you plan to do any fine-tuning with your own tokens added or with instruction following.

GPTQ Quantization Method

• This model is quantized by utilizing the AutoGPTQ library, following best practices noted by GPTQ paper
• Quantization is calibrated and aligned with random samples from the specified dataset (wikitext for now) for minimum accuracy loss.

Branch	Bits	Group Size	Act Order	Damp %	GPTQ Dataset	Sequence Length	VRAM Size	ExLlama	Special Tokens Fixed	Description
main	4	128	Yes	0.1	wikitext	8192	5.74 GB	Yes	No	4-bit, with Act Order and group size 128g. Smallest model possible with small accuracy loss
untrained-special-tokens-fixed	4	128	Yes	0.1	wikitext	8192	5.74 GB	Yes	Yes	Same as the main branch. The special tokens that were untrained causing exploding gradients/NaN gradients have had their embedding values set to the average of trained tokens for each feature
More variants to come	TBD	TBD	TBD	TBD	TBD	TBD	TBD	TBD	TBD	May upload additional variants of GPTQ 4 bit models in the future using different parameters such as 128g group size and etc.

Serving this GPTQ model using vLLM

Tested serving this model via vLLM using an Nvidia T4 (16GB VRAM).

Tested with the below command

python -m vllm.entrypoints.openai.api_server --model astronomer-io/Llama-3-8B-GPTQ-4-Bit --max-model-len 8192 --dtype float16

For the non-stop token generation bug, make sure to send requests with stop_token_ids":[128001, 128009] to vLLM endpoint

Contributors

• Quantized by David Xue, Machine Learning Engineer from Astronomer