Inference Speed comparison

[that-rahul-guy]

Hello everyone,

• I am trying to serve TheBloke/Mistral-7B-Instruct-v0.1-GPTQ model.

• I'm currently serving the model via jina and internally I am doing the predictions as:

  input_ids = text_tokenizer.encode(prompt, return_tensors="pt").cuda()
  
  # Generate output with custom configuration
  output_ids = text_model.generate(input_ids, **gen_config)
  
  generated_ids = output_ids[:, num_input_tokens:]
  # Decode only the generated part
  output = text_tokenizer.decode(generated_ids[0], skip_special_tokens=True)

  Here manually we are generating the output tokens and then decoding it. For my test prompt, the performance is as:

  time taken: ~2.5sec token/sec: ~6

• Now, while serving it via OpenLLM using following command: openllm start TheBloke/Mistral-7B-Instruct-v0.1-GPTQ --quantize gptq --backend pt For the same test prompt, the performance is as:

  time taken: ~30sec token/sec: ~0.5

• Env:

  • OS: Ubuntu 20.04
  • GPU: Nvidia T4 (16GiB vRAM)
  • Cuda: 11.8

I want to discuss why is the difference so big. Am I doing something wrong while serving with OpenLLM? Let me know your thoughts.

Thanks

[aarnphm]

gptq is now supported with vLLM and latest openllm version. You can test it with vLLM as I haven't update the pytorch code path for a while now.

You should see signi improve with vLLM.