Running out of memory with TheBloke/CodeLlama-7B-AWQ

[bonuschild]

• Already posted on https://github.com/vllm-project/vllm/issues/1479
• My GPU is RTX 3060 with 12GB VRAM
• My target model isCodeLlama-7B-AWQ, which size is <= 4GB

Looking for help from 2 communities 😄 thx!

[bonuschild]

I've re-tested this on A100 instead of RTX3060, it show that finally it occupy about 20+GB VRAM! Why is that? I use command:

python api_server.py --model path/to/7b-awq/model --port 8000 -q awq --dtype half --trust-remote-code

That was so weired...

[jkrauss82]

I had success running Mistral-7B-v0.1-AWQ and CodeLlama-7B-AWQ of TheBloke on an A6000 with 48G VRAM, restricted to ~8G VRAM with the following parameters:

python api_server.py --model path/to/model --port 8000 --quantization awq --dtype float16 --gpu-memory-utilization 0.167 --max-model-len 4096 --max-num-batched-tokens 4096

nvidia-smi then shows around 8G memory consumed by the python process, should run on the 3060 as well I hope (need to omit the --gpu-memory-utilization obviously).