GPU requirement for hosting this model?

#9
by csgxy2022 - opened

Two A100 GPUs, trying to host this model but got OOM issue.

docker run --runtime nvidia --gpus all -v ~/.cache/huggingface:/root/.cache/huggingface --env "HUGGING_FACE_HUB_TOKEN=xxx" -p 8000:8000 --ipc=host vllm/vllm-openai:latest --model gradientai/Llama-3-8B-Instruct-Gradient-1048k --tensor-parallel-size 2

I got:

torch.cuda.OutOfMemoryError: CUDA out of memory

No problem hosting the original Llma-3-8B-Instruct model

This comment has been hidden
Gradient AI org
edited May 4

The following should do the job for vllm! (A100.x2)

docker run --runtime nvidia --gpus all \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    -p 8000:8000 \
    --shm-size=8g \
    --ipc=host \
    vllm/vllm-openai:latest \
    --model gradientai/Llama-3-8B-Instruct-1048k \
    --tensor-parallel-size 2 \
    --max-model-len 65536

vLLM is not optimal, it would require around ~1000GB vRAM (IIRC) to serve a model with this hidden dim and ctx length.

Sign up or log in to comment