Distributed LLM Generation#

Source NVIDIA/TensorRT-LLM.

 1from tensorrt_llm import LLM, SamplingParams
 2
 3
 4def main():
 5    # model could accept HF model name or a path to local HF model.
 6    llm = LLM(
 7        model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
 8        # Enable 2-way tensor parallelism
 9        tensor_parallel_size=2
10        # Enable 2-way pipeline parallelism if needed
11        # pipeline_parallel_size=2
12        # Enable 2-way expert parallelism for MoE model's expert weights
13        # moe_expert_parallel_size=2
14        # Enable 2-way tensor parallelism for MoE model's expert weights
15        # moe_tensor_parallel_size=2
16    )
17
18    # Sample prompts.
19    prompts = [
20        "Hello, my name is",
21        "The president of the United States is",
22        "The capital of France is",
23        "The future of AI is",
24    ]
25
26    # Create a sampling params.
27    sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
28
29    for output in llm.generate(prompts, sampling_params):
30        print(
31            f"Prompt: {output.prompt!r}, Generated text: {output.outputs[0].text!r}"
32        )
33
34    # Got output like
35    # Prompt: 'Hello, my name is', Generated text: '\n\nJane Smith. I am a student pursuing my degree in Computer Science at [university]. I enjoy learning new things, especially technology and programming'
36    # Prompt: 'The president of the United States is', Generated text: 'likely to nominate a new Supreme Court justice to fill the seat vacated by the death of Antonin Scalia. The Senate should vote to confirm the'
37    # Prompt: 'The capital of France is', Generated text: 'Paris.'
38    # Prompt: 'The future of AI is', Generated text: 'an exciting time for us. We are constantly researching, developing, and improving our platform to create the most advanced and efficient model available. We are'
39
40
41# The entry point of the program need to be protected for spawning processes.
42if __name__ == '__main__':
43    main()