| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """This file demonstrates the usage of text generation with an LLM model, |
| 4 | comparing the performance with and without speculative decoding. |
| 5 | |
| 6 | Note that this example is out of date and not supported in vLLM v1. |
| 7 | """ |
| 8 | |
| 9 | import gc |
| 10 | import time |
| 11 | |
| 12 | from vllm import LLM, SamplingParams |
| 13 | |
| 14 | |
| 15 | def time_generation( |
| 16 | llm: LLM, prompts: list[str], sampling_params: SamplingParams, title: str |
| 17 | ): |
| 18 | # Generate texts from the prompts. The output is a list of RequestOutput |
| 19 | # objects that contain the prompt, generated text, and other information. |
| 20 | # Warmup first |
| 21 | llm.generate(prompts, sampling_params) |
| 22 | llm.generate(prompts, sampling_params) |
| 23 | start = time.time() |
| 24 | outputs = llm.generate(prompts, sampling_params) |
| 25 | end = time.time() |
| 26 | print("-" * 50) |
| 27 | print(title) |
| 28 | print("time: ", (end - start) / sum(len(o.outputs[0].token_ids) for o in outputs)) |
| 29 | # Print the outputs. |
| 30 | for output in outputs: |
| 31 | generated_text = output.outputs[0].text |
| 32 | print(f"text: {generated_text!r}") |
| 33 | print("-" * 50) |
| 34 | |
| 35 | |
| 36 | def main(): |
| 37 | template = ( |
| 38 | "Below is an instruction that describes a task. Write a response " |
| 39 | "that appropriately completes the request.\n\n### Instruction:\n{}" |
| 40 | "\n\n### Response:\n" |
| 41 | ) |
| 42 | |
| 43 | # Sample prompts. |
| 44 | prompts = [ |
| 45 | "Write about the president of the United States.", |
| 46 | ] |
| 47 | prompts = [template.format(prompt) for prompt in prompts] |
| 48 | # Create a sampling params object. |
| 49 | sampling_params = SamplingParams(temperature=0.0, max_tokens=200) |
| 50 | |
| 51 | # Create an LLM without spec decoding |
| 52 | llm = LLM(model="meta-llama/Llama-2-13b-chat-hf") |
| 53 | |
| 54 | time_generation(llm, prompts, sampling_params, "Without speculation") |
| 55 | |
| 56 | del llm |
| 57 | gc.collect() |
| 58 | |
| 59 | # Create an LLM with spec decoding |
| 60 | llm = LLM( |
| 61 | model="meta-llama/Llama-2-13b-chat-hf", |
| 62 | speculative_config={ |
| 63 | "model": "ibm-ai-platform/llama-13b-accelerator", |
| 64 | }, |
| 65 | ) |
| 66 | |
| 67 | time_generation(llm, prompts, sampling_params, "With speculation") |
| 68 | |
| 69 | |
| 70 | if __name__ == "__main__": |
| 71 | main() |