README.md

metadata

license: other
language:
  - en
  - de
  - fr
  - it
  - pt
  - hi
  - es
  - th
library_name: transformers
pipeline_tag: text-generation
tags:
  - llama-3.1
  - meta
  - autoawq

This repository is a community-driven quantized version of the original model meta-llama/Meta-Llama-3.1-405B-Instruct which is the FP16 half-precision official version released by Meta AI.

Model Information

The Meta Llama 3.1 collection of multilingual large language models (LLMs) is a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes (text in/text out). The Llama 3.1 instruction tuned text only models (8B, 70B, 405B) are optimized for multilingual dialogue use cases and outperform many of the available open source and closed chat models on common industry benchmarks.

Model developer: Meta

Model Architecture: Llama 3.1 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety.

	Training Data	Params	Input modalities	Output modalities	Context length	GQA	Token count	Knowledge cutoff
Llama 3.1 (text only)	A new mix of publicly available online data.	8B	Multilingual Text	Multilingual Text and code	128k	Yes	15T+	December 2023
		70B	Multilingual Text	Multilingual Text and code	128k	Yes
		405B	Multilingual Text	Multilingual Text and code	128k	Yes

Supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

Llama 3.1 family of models. Token counts refer to pretraining data only. All model versions use Grouped-Query Attention (GQA) for improved inference scalability.

Model Release Date: July 23, 2024.

Status: This is a static model trained on an offline dataset. Future versions of the tuned models will be released as we improve model safety with community feedback.

License: A custom commercial license, the Llama 3.1 Community License, is available at: https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/LICENSE

Where to send questions or comments about the model Instructions on how to provide feedback or comments on the model can be found in the model README. For more technical information about generation parameters and recipes for how to use Llama 3.1 in applications, please go here.

For more information please refer to the original model card meta-llama/Meta-Llama-3.1-405B-Instruct.

Quantized Model Information

Llama 3.1 405B Instruct has been quantized using AutoAWQ from FP16 down to INT4 using the GEMM kernels performing zero-point quantization with a group size of 128.

Quantized Model Usage

In order to run the inference with Llama 3.1 405B Instruct AWQ in INT4, around 203 GiB of VRAM are needed only for loading the model checkpoint, without including the KV cache or the CUDA graphs, meaning that there should be a bit over that VRAM available.

In order to use the current quantized model, support is offered for different solutions:

🤗 transformers

To run the inference on top of Llama 3.1 405B Instruct AWQ in INT4 precision, the AWQ model can be instantiated as any other causal language modeling model via AutoModelForCausalLM and run the inference normally.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "hugging-quants/Meta-Llama-3.1-405B-Instruct-AWQ-INT4"
prompt = [
  {"role": "system", "content": "You are a helpful assistant, that responds as a pirate."},
  {"role": "user", "content": "What's Deep Learning?"},
]

tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.pad_token_id = tokenizer.eos_token_id
tokenizer.padding_side = "left"

terminators = [
  tokenizer.eos_token_id,
  tokenizer.convert_tokens_to_ids("<|eot_id|>"),
]

inputs = tokenizer.apply_chat_template(prompt, tokenize=True, add_generation_prompt=True, return_tensors="pt").cuda()

model = AutoModelForCausalLM.from_pretrained(
  model_id,
  torch_dtype=torch.float16,
  low_cpu_mem_usage=True,
  device_map="auto",
)

outputs = model.generate(**inputs, do_sample=True, max_new_tokens=256, eos_token_id=terminators)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))

AutoAWQ

Alternatively, one may want to run that via AutoAWQ even though it's built on top of 🤗 transformers, which is the recommended approach instead as described above.

import torch
from autoawq import AutoAWQForCausalLM
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "hugging-quants/Meta-Llama-3.1-405B-Instruct-AWQ-INT4"
prompt = [
  {"role": "system", "content": "You are a helpful assistant, that responds as a pirate."},
  {"role": "user", "content": "What's Deep Learning?"},
]

tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.pad_token_id = tokenizer.eos_token_id
tokenizer.padding_side = "left"

inputs = tokenizer.apply_chat_template(prompt, tokenize=True, add_generation_prompt=True, return_tensors="pt").cuda()

model = AutoAWQForCausalLM.from_pretrained(
  model_id,
  torch_dtype=torch.float16,
  low_cpu_mem_usage=True,
  device_map="auto",
  fuse_layers=True,
)

outputs = model.generate(**inputs, do_sample=True, max_new_tokens=256)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))

The AutoAWQ script has been adapted from AutoAWQ/examples/generate.py.

🤗 Text Generation Inference (TGI)

Coming soon!

Quantization Reproduction

In order to quantize Llama 3.1 405B Instruct using AutoAWQ, you will need to use an instance with at least enough CPU RAM to fit the whole model i.e. ~800GiB, and an NVIDIA GPU with 80GiB of VRAM to quantize it.

In order to quantize Llama 3.1 405B Instruct, first install torch and autoawq as follows:

pip install "torch>=2.2.0,<2.3.0" autoawq --upgrade

Otherwise the quantization may fail, since the AutoAWQ kernels are built with PyTorch 2.2.1, meaning that those will break with PyTorch 2.3.0.

Then install the latest version of transformers as follows:

pip install "transformers>=4.43.0" --upgrade

And then, run the following script, adapted from AutoAWQ/examples/quantize.py as follows:

from awq import AutoAWQForCausalLM
from transformers import AutoTokenizer

model_path = "meta-llama/Meta-Llama-3.1-405B-Instruct"
quant_path = "hugging-quants/Meta-Llama-3.1-405B-Instruct-AWQ-INT4"
quant_config = {
  "zero_point": True,
  "q_group_size": 128,
  "w_bit": 4,
  "version": "GEMM",
}

# Load model
model = AutoAWQForCausalLM.from_pretrained(
  model_path, **{"low_cpu_mem_usage": True, "use_cache": False}
)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# Quantize
model.quantize(tokenizer, quant_config=quant_config)

# Save quantized model
model.save_quantized(quant_path)
tokenizer.save_pretrained(quant_path)

print(f'Model is quantized and saved at "{quant_path}"')