Instructions to use dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8")
model = AutoModelForCausalLM.from_pretrained("dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8

SGLang

How to use dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8 with Docker Model Runner:
```
docker model run hf.co/dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8
```

dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8

This repository contains a Hugging Face export of Llama-3.1-8B-Instruct quantized with AQLM using the 8-bit 8x8 scheme.

Base model

meta-llama/Llama-3.1-8B-Instruct

Quantization

Method: AQLM
Scheme: 8x8
Effective label: 8-bit
Source checkpoint: /work/bduan1/quantized_models/Llama-3.1-8B-Instruct-AQLM-8bit-8x8-n4096

Conversion

This repo was produced with convert_to_hf.py from the AQLM project, then exported with --save_safetensors and --save_tokenizer.

Inference requires a Transformers version with AQLM quantization support and the aqlm package installed in the runtime environment.

Local Evaluation

Wikitext2 perplexity: 6.5135
C4 perplexity: 8.0386
Calibration data: 4096 RedPajama samples tokenized with the meta-llama/Llama-3.1-8B-Instruct tokenizer at sequence length 8192
AQLM settings: num_codebooks=8, nbits_per_codebook=8, in_group_size=8, out_group_size=1

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

Downloads last month: 112

Safetensors

Model size

8B params

Tensor type

F16

Model tree for dbw6/Llama-3.1-8B-Instruct-AQLM-8bit-8x8

Base model

meta-llama/Llama-3.1-8B

Finetuned

meta-llama/Llama-3.1-8B-Instruct

Quantized

(882)

this model