mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
39 lines
1.2 KiB
Python
39 lines
1.2 KiB
Python
from tensorrt_llm import SamplingParams
|
|
from tensorrt_llm._torch import LLM
|
|
from tensorrt_llm.executor import LoRARequest
|
|
from tensorrt_llm.lora_manager import LoraConfig
|
|
|
|
|
|
def main():
|
|
lora_config = LoraConfig(lora_dir=[
|
|
"/home/scratch.trt_llm_data/llm-models/llama-models-v2/chinese-llama-2-lora-13b"
|
|
],
|
|
max_lora_rank=64)
|
|
llm = LLM(
|
|
model=
|
|
"/home/scratch.trt_llm_data/llm-models/llama-models-v2/llama-v2-13b-hf",
|
|
lora_config=lora_config,
|
|
)
|
|
prompts = [
|
|
"今天天气很好，我到公园的时候，",
|
|
]
|
|
|
|
sampling_params = SamplingParams(max_tokens=20, add_special_tokens=False)
|
|
lora_req_2 = LoRARequest(
|
|
"task-0", 0,
|
|
"/home/scratch.trt_llm_data/llm-models/llama-models-v2/chinese-llama-2-lora-13b"
|
|
)
|
|
lora_request = [lora_req_2]
|
|
|
|
outputs = llm.generate(prompts, sampling_params, lora_request=lora_request)
|
|
|
|
# Print the outputs.
|
|
for output in outputs:
|
|
prompt = output.prompt
|
|
generated_text = output.outputs[0].text
|
|
print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|