meta-llama/Llama-3.1-70B-Instruct · Slow inference/low GPU utilization.

I have noticed low GPU utilization on several LLMs when using the huggingface pipeline api or the AutoModelForCausalLM api

My setup is 8 * H100. During inference, utilization fluctuates between 10% to 15% on each GPU.
How can I improve the same?

import torch

model_id = "meta-llama/Meta-Llama-3.1-70B-Instruct"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto", 
)

reviews = "\n".join(reviews)

messages = [
    {"role": "system", "content": "Your goal is summarize text."},
    {"role": "user", "content": f"Summarize the below reviews. \n {reviews}"  },
]

outputs = pipeline(
    messages,
    max_new_tokens=1256,
)
print(outputs[0]["generated_text"][-1])