Hello everyone! I'm running LLM models like Starcoder and CodeLLama through BentoML, and they're taking up 60-70 GB of GPU memory. However, when running them without BentoML, it only uses 30-40 GB. The documentation states that runners occupy all available container resources. Can you confirm if this is the case? My service.py:

import bentoml

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

@bentoml.service
class Service:
    def __init__(self) -> None:
        checkpoint = "bigcode/starcoder"
    
        model = AutoModelForCausalLM.from_pretrained(checkpoint)
        tokenizer = AutoTokenizer.from_pretrained(checkpoint)

        self.pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)

    @bentoml.api
    def generate(self, prompt: str):
        result = self.pipe(prompt)
        return result