Instructions to use relaxml/Llama-2-70b-E8P-2Bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use relaxml/Llama-2-70b-E8P-2Bit with Transformers:

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

pipe = pipeline("text-generation", model="relaxml/Llama-2-70b-E8P-2Bit")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("relaxml/Llama-2-70b-E8P-2Bit")
model = AutoModelForCausalLM.from_pretrained("relaxml/Llama-2-70b-E8P-2Bit")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use relaxml/Llama-2-70b-E8P-2Bit with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "relaxml/Llama-2-70b-E8P-2Bit"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "relaxml/Llama-2-70b-E8P-2Bit",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/relaxml/Llama-2-70b-E8P-2Bit

SGLang

How to use relaxml/Llama-2-70b-E8P-2Bit 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 "relaxml/Llama-2-70b-E8P-2Bit" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "relaxml/Llama-2-70b-E8P-2Bit",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "relaxml/Llama-2-70b-E8P-2Bit" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "relaxml/Llama-2-70b-E8P-2Bit",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use relaxml/Llama-2-70b-E8P-2Bit with Docker Model Runner:
```
docker model run hf.co/relaxml/Llama-2-70b-E8P-2Bit
```

Out of CPU memory in pipeline

by dmiche - opened Dec 13, 2023

Discussion

dmiche

Dec 13, 2023

Hi!
The model files itself are small and should fit in 24Gb of GPU.
But if I try the suggested pipeline:

pipe = pipeline("text-generation", model="relaxml/Llama-2-70b-E8P-2Bit")

it start to grow the CPU memory till 128Gb and then be killed on OOM.
Can I avoid this memory allocation?

at676

RelaxML org Dec 13, 2023

I'm confused, what is this suggested pipeline? I don't think we have any code in our codebase that uses a pipeline() call.

dmiche

Dec 13, 2023

Hm... "Model card" tab, most right button above "Downloads" chart "Use in Transformers" :)

Do you offer a working example of code?

at676

RelaxML org Dec 13, 2023

Yes, in our github repo https://github.com/Cornell-RelaxML/quip-sharp. We use a modified version of the modeling_llama.py file to handle our quantized linear layers, which is why calling the default "pipeline" command without using our repo will not work.

dmiche

Dec 13, 2023

Thank You!

dmiche changed discussion status to closed Dec 13, 2023

eramax

Dec 15, 2023

@at676 , I would appriate if you provide us with a colab link for QUIP inference

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment