How to allow parallel inference requests

[webhaypac]

Hi,
I am running on GPU using the following command.
python3 -m llama_cpp.server --host xx.xx.xxx.xx --port 4444 --model /home/user1/llama.cpp-old/models/codellama-7b-instruct.Q8_0.gguf --n_gpu_layers -1 --n_threads 5 --n_threads_batch 5 --interrupt_requests false

but, I still can't get concurrent inference requests to work. I just need to allow multiple inference requests to return answers in the same time. Currently, all the requests gets queued and handled one at a time.

any idea how to do that?

Thank you

[rookiemann]

I would like to know this too, does llama-cpp-python have a native in-built way to do this?

[RobinBially]

also tried these as described here, doesn't work:

LLAMA_ARG_N_PARALLEL=3 python -m llama_cpp.server --model ./cache/llamacpp/qwen2-0_5b-instruct-q4_0.gguf
python -m llama_cpp.server --model ./cache/llamacpp/qwen2-0_5b-instruct-q4_0.gguf --parallel 3
python -m llama_cpp.server --model ./cache/llamacpp/qwen2-0_5b-instruct-q4_0.gguf -np 3

[javi22020]

Any solution?