| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from vllm import LLM, EngineArgs |
| 5 | from vllm.outputs import RequestOutput |
| 6 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 7 | |
| 8 | |
| 9 | def create_parser(): |
| 10 | parser = FlexibleArgumentParser() |
| 11 | # Add engine args |
| 12 | EngineArgs.add_cli_args(parser) |
| 13 | parser.set_defaults(model="meta-llama/Llama-3.2-1B-Instruct") |
| 14 | # Add sampling params |
| 15 | sampling_group = parser.add_argument_group("Sampling parameters") |
| 16 | sampling_group.add_argument("--max-tokens", type=int) |
| 17 | sampling_group.add_argument("--temperature", type=float) |
| 18 | sampling_group.add_argument("--top-p", type=float) |
| 19 | sampling_group.add_argument("--top-k", type=int) |
| 20 | # Add example params |
| 21 | parser.add_argument("--chat-template-path", type=str) |
| 22 | |
| 23 | return parser |
| 24 | |
| 25 | |
| 26 | def main(args: dict): |
| 27 | # Pop arguments not used by LLM |
| 28 | max_tokens = args.pop("max_tokens") |
| 29 | temperature = args.pop("temperature") |
| 30 | top_p = args.pop("top_p") |
| 31 | top_k = args.pop("top_k") |
| 32 | chat_template_path = args.pop("chat_template_path") |
| 33 | |
| 34 | # Create an LLM |
| 35 | llm = LLM(**args) |
| 36 | |
| 37 | # Create sampling params object |
| 38 | sampling_params = llm.get_default_sampling_params() |
| 39 | if max_tokens is not None: |
| 40 | sampling_params.max_tokens = max_tokens |
| 41 | if temperature is not None: |
| 42 | sampling_params.temperature = temperature |
| 43 | if top_p is not None: |
| 44 | sampling_params.top_p = top_p |
| 45 | if top_k is not None: |
| 46 | sampling_params.top_k = top_k |
| 47 | |
| 48 | def print_outputs(outputs: list[RequestOutput], prompts: list): |
| 49 | assert len(outputs) == len(prompts) |
| 50 | print("\nGenerated Outputs:\n" + "-" * 80) |
| 51 | for i, output in enumerate(outputs): |
| 52 | generated_text = output.outputs[0].text |
| 53 | print(f"Prompt: {prompts[i]!r}\n") |
| 54 | print(f"Generated text: {generated_text!r}") |
| 55 | print("-" * 80) |
| 56 | |
| 57 | print("=" * 80) |
| 58 | |
| 59 | # In this script, we demonstrate how to pass input to the chat method: |
| 60 | conversation = [ |
| 61 | {"role": "system", "content": "You are a helpful assistant"}, |
| 62 | {"role": "user", "content": "Hello"}, |
| 63 | {"role": "assistant", "content": "Hello! How can I assist you today?"}, |
| 64 | { |
| 65 | "role": "user", |
| 66 | "content": "Write an essay about the importance of higher education.", |
| 67 | }, |
| 68 | ] |
| 69 | outputs = llm.chat(conversation, sampling_params, use_tqdm=False) |
| 70 | print_outputs( |
| 71 | outputs, |
| 72 | [ |
| 73 | conversation, |
| 74 | ], |
| 75 | ) |
| 76 | |
| 77 | # You can run batch inference with llm.chat API |
| 78 | conversations = [conversation for _ in range(10)] |
| 79 | |
| 80 | # We turn on tqdm progress bar to verify it's indeed running batch inference |
| 81 | outputs = llm.chat(conversations, sampling_params, use_tqdm=True) |
| 82 | print_outputs(outputs, conversations) |
| 83 | |
| 84 | # A chat template can be optionally supplied. |
| 85 | # If not, the model will use its default chat template. |
| 86 | if chat_template_path is not None: |
| 87 | with open(chat_template_path) as f: |
| 88 | chat_template = f.read() |
| 89 | |
| 90 | outputs = llm.chat( |
| 91 | conversations, |
| 92 | sampling_params, |
| 93 | use_tqdm=False, |
| 94 | chat_template=chat_template, |
| 95 | ) |
| 96 | print_outputs(outputs, conversations) |
| 97 | |
| 98 | |
| 99 | if __name__ == "__main__": |
| 100 | parser = create_parser() |
| 101 | args: dict = vars(parser.parse_args()) |
| 102 | main(args) |