mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-02-08 04:01:51 +08:00
75057cd036
* Update TensorRT-LLM --------- Co-authored-by: Puneesh Khanna <puneesh.khanna@tii.ae> Co-authored-by: Ethan Zhang <26497102+ethnzhng@users.noreply.github.com>
32 lines
910 B
Python
32 lines
910 B
Python
#!/usr/bin/env python3
|
|
import os
|
|
|
|
import click
|
|
|
|
from tensorrt_llm.llmapi import LLM, KvCacheConfig, SamplingParams
|
|
|
|
|
|
@click.command()
|
|
@click.option("--model_dir", type=str, required=True)
|
|
@click.option("--tp_size", type=int, required=True)
|
|
@click.option("--engine_dir", type=str, default=None)
|
|
def main(model_dir: str, tp_size: int, engine_dir: str):
|
|
llm = LLM(model_dir,
|
|
tensor_parallel_size=tp_size,
|
|
kv_cache_config=KvCacheConfig(free_gpu_memory_fraction=0.4))
|
|
|
|
if engine_dir is not None and os.path.abspath(
|
|
engine_dir) != os.path.abspath(model_dir):
|
|
llm.save(engine_dir)
|
|
|
|
sampling_params = SamplingParams(max_tokens=10, end_id=-1)
|
|
|
|
prompt_token_ids = [45, 12, 13]
|
|
for output in llm.generate([prompt_token_ids],
|
|
sampling_params=sampling_params):
|
|
print(output)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|