TensorRT-LLMs/examples/llm-api/llm_multilora.py
2025-09-16 16:35:13 +08:00

90 lines
4.7 KiB
Python

### :section Customization
### :title Generate text with multiple LoRA adapters
### :order 5
import argparse
from typing import Optional
from huggingface_hub import snapshot_download
from tensorrt_llm import LLM
from tensorrt_llm.executor import LoRARequest
from tensorrt_llm.lora_helper import LoraConfig
def main(chatbot_lora_dir: Optional[str], mental_health_lora_dir: Optional[str],
tarot_lora_dir: Optional[str]):
# Download the LoRA adapters from huggingface hub, if not provided via command line args.
if chatbot_lora_dir is None:
chatbot_lora_dir = snapshot_download(
repo_id="snshrivas10/sft-tiny-chatbot")
if mental_health_lora_dir is None:
mental_health_lora_dir = snapshot_download(
repo_id=
"givyboy/TinyLlama-1.1B-Chat-v1.0-mental-health-conversational")
if tarot_lora_dir is None:
tarot_lora_dir = snapshot_download(
repo_id="barissglc/tinyllama-tarot-v1")
# Currently, we need to pass at least one lora_dir to LLM constructor via build_config.lora_config.
# This is necessary because it requires some configuration in the lora_dir to build the engine with LoRA support.
lora_config = LoraConfig(lora_dir=[chatbot_lora_dir],
max_lora_rank=64,
max_loras=3,
max_cpu_loras=3)
llm = LLM(model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
lora_config=lora_config)
# Sample prompts
prompts = [
"Hello, tell me a story: ",
"Hello, tell me a story: ",
"I've noticed you seem a bit down lately. Is there anything you'd like to talk about?",
"I've noticed you seem a bit down lately. Is there anything you'd like to talk about?",
"In this reading, the Justice card represents a situation where",
"In this reading, the Justice card represents a situation where",
]
# At runtime, multiple LoRA adapters can be specified via lora_request; None means no LoRA used.
for output in llm.generate(prompts,
lora_request=[
None,
LoRARequest("chatbot", 1, chatbot_lora_dir),
None,
LoRARequest("mental-health", 2,
mental_health_lora_dir), None,
LoRARequest("tarot", 3, tarot_lora_dir)
]):
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
# Got output like
# Prompt: 'Hello, tell me a story: ', Generated text: '1. Start with a question: "What\'s your favorite color?" 2. Ask a question that leads to a story: "What\'s your'
# Prompt: 'Hello, tell me a story: ', Generated text: '1. A person is walking down the street. 2. A person is sitting on a bench. 3. A person is reading a book.'
# Prompt: "I've noticed you seem a bit down lately. Is there anything you'd like to talk about?", Generated text: "\n\nJASON: (smiling) No, I'm just feeling a bit overwhelmed lately. I've been trying to"
# Prompt: "I've noticed you seem a bit down lately. Is there anything you'd like to talk about?", Generated text: "\n\nJASON: (sighs) Yeah, I've been struggling with some personal issues. I've been feeling like I'm"
# Prompt: 'In this reading, the Justice card represents a situation where', Generated text: 'you are being asked to make a decision that will have a significant impact on your life. The card suggests that you should take the time to consider all the options'
# Prompt: 'In this reading, the Justice card represents a situation where', Generated text: 'you are being asked to make a decision that will have a significant impact on your life. It is important to take the time to consider all the options and make'
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description="Generate text with multiple LoRA adapters")
parser.add_argument('--chatbot_lora_dir',
type=str,
default=None,
help='Path to the chatbot LoRA directory')
parser.add_argument('--mental_health_lora_dir',
type=str,
default=None,
help='Path to the mental health LoRA directory')
parser.add_argument('--tarot_lora_dir',
type=str,
default=None,
help='Path to the tarot LoRA directory')
args = parser.parse_args()
main(args.chatbot_lora_dir, args.mental_health_lora_dir,
args.tarot_lora_dir)