Instructions to use togethercomputer/LLaMA-2-7B-32K with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use togethercomputer/LLaMA-2-7B-32K with Transformers:

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

pipe = pipeline("text-generation", model="togethercomputer/LLaMA-2-7B-32K")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("togethercomputer/LLaMA-2-7B-32K")
model = AutoModelForCausalLM.from_pretrained("togethercomputer/LLaMA-2-7B-32K")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use togethercomputer/LLaMA-2-7B-32K with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "togethercomputer/LLaMA-2-7B-32K"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "togethercomputer/LLaMA-2-7B-32K",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/togethercomputer/LLaMA-2-7B-32K

SGLang

How to use togethercomputer/LLaMA-2-7B-32K 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 "togethercomputer/LLaMA-2-7B-32K" \
    --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": "togethercomputer/LLaMA-2-7B-32K",
		"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 "togethercomputer/LLaMA-2-7B-32K" \
        --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": "togethercomputer/LLaMA-2-7B-32K",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use togethercomputer/LLaMA-2-7B-32K with Docker Model Runner:
```
docker model run hf.co/togethercomputer/LLaMA-2-7B-32K
```

Training diverges when used with Llama 2 70B and 4-bit QLoRA

#10

by alyssavance - opened Aug 1, 2023

Discussion

alyssavance

Aug 1, 2023

Posted the issue here, but happy to discuss further if anyone can help. The divergence happens after ~20 steps/six hours. Thanks

https://github.com/TimDettmers/bitsandbytes/issues/663

gardner

Aug 1, 2023

Hi @alyssavance , have you read this? https://huggingface.co/togethercomputer/LLaMA-2-7B-32K/discussions/2

since you are doing QLoRA, you might need to set trust_remote_code=False to use HF's llama implementation, flash attention only works for float16 or bfloat16.

alyssavance

Aug 1, 2023

@gardner I did, I had some type problems but fixed them by removing the JIT decorator from rmsnorm. Right now it runs with no type errors, it does inference fine, it just gradually diverges after the first few dozen steps.

mauriceweber

Aug 2, 2023

Hi @alyssavance , did you try a smaller learning rate? Instead of 1e-4, it might be worth to try out 2e-5 (same as in the linear interpolation paper).

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