| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """This file demonstrates preempt requests when using the `LLMEngine` |
| 4 | for processing prompts with various sampling parameters. |
| 5 | """ |
| 6 | |
| 7 | import argparse |
| 8 | |
| 9 | from vllm import EngineArgs, LLMEngine, RequestOutput, SamplingParams |
| 10 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 11 | |
| 12 | |
| 13 | def create_test_prompts() -> list[tuple[str, SamplingParams]]: |
| 14 | """Create a list of test prompts with their sampling parameters.""" |
| 15 | return [ |
| 16 | ( |
| 17 | "A robot may not injure a human being " * 50, |
| 18 | SamplingParams( |
| 19 | temperature=0.0, logprobs=1, prompt_logprobs=1, max_tokens=16 |
| 20 | ), |
| 21 | ), |
| 22 | ( |
| 23 | "A robot may not injure a human being " * 50, |
| 24 | SamplingParams( |
| 25 | temperature=0.0, logprobs=1, prompt_logprobs=1, max_tokens=16 |
| 26 | ), |
| 27 | ), |
| 28 | ( |
| 29 | "To be or not to be,", |
| 30 | SamplingParams( |
| 31 | temperature=0.8, top_k=5, presence_penalty=0.2, max_tokens=128 |
| 32 | ), |
| 33 | ), |
| 34 | ( |
| 35 | "What is the meaning of life?", |
| 36 | SamplingParams( |
| 37 | n=2, temperature=0.8, top_p=0.95, frequency_penalty=0.1, max_tokens=128 |
| 38 | ), |
| 39 | ), |
| 40 | ] |
| 41 | |
| 42 | |
| 43 | def process_requests(engine: LLMEngine, test_prompts: list[tuple[str, SamplingParams]]): |
| 44 | """Continuously process a list of prompts and handle the outputs.""" |
| 45 | request_id = 0 |
| 46 | |
| 47 | print("-" * 50) |
| 48 | step_id = 0 |
| 49 | while test_prompts or engine.has_unfinished_requests(): |
| 50 | print("-" * 50) |
| 51 | import os |
| 52 | |
| 53 | print(f"Step {step_id} (pid={os.getpid()})") |
| 54 | |
| 55 | if test_prompts: |
| 56 | prompt, sampling_params = test_prompts.pop(0) |
| 57 | engine.add_request(str(request_id), prompt, sampling_params) |
| 58 | request_id += 1 |
| 59 | |
| 60 | if step_id == 10: |
| 61 | print(f"Resetting prefix cache at {step_id}") |
| 62 | engine.reset_prefix_cache(reset_running_requests=True) |
| 63 | |
| 64 | request_outputs: list[RequestOutput] = engine.step() |
| 65 | |
| 66 | for request_output in request_outputs: |
| 67 | if request_output.finished: |
| 68 | print("-" * 50) |
| 69 | print(request_output) |
| 70 | print("-" * 50) |
| 71 | step_id += 1 |
| 72 | |
| 73 | |
| 74 | def initialize_engine(args: argparse.Namespace) -> LLMEngine: |
| 75 | """Initialize the LLMEngine from the command line arguments.""" |
| 76 | engine_args = EngineArgs.from_cli_args(args) |
| 77 | return LLMEngine.from_engine_args(engine_args) |
| 78 | |
| 79 | |
| 80 | def parse_args(): |
| 81 | parser = FlexibleArgumentParser( |
| 82 | description="Demo on using the LLMEngine class directly" |
| 83 | ) |
| 84 | parser = EngineArgs.add_cli_args(parser) |
| 85 | return parser.parse_args() |
| 86 | |
| 87 | |
| 88 | def main(args: argparse.Namespace): |
| 89 | """Main function that sets up and runs the prompt processing.""" |
| 90 | engine = initialize_engine(args) |
| 91 | test_prompts = create_test_prompts() |
| 92 | process_requests(engine, test_prompts) |
| 93 | |
| 94 | |
| 95 | if __name__ == "__main__": |
| 96 | args = parse_args() |
| 97 | main(args) |