| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """This script demonstrates how to extend the context length |
| 4 | of a Qwen model using the YARN method (rope_parameters) |
| 5 | and run a simple chat example. |
| 6 | |
| 7 | Usage: |
| 8 | python examples/features/context_extension/context_extension_offline.py |
| 9 | """ |
| 10 | |
| 11 | from vllm import LLM, RequestOutput, SamplingParams |
| 12 | |
| 13 | |
| 14 | def create_llm(): |
| 15 | rope_theta = 1000000 |
| 16 | original_max_position_embeddings = 32768 |
| 17 | factor = 4.0 |
| 18 | |
| 19 | # Use yarn to extend context |
| 20 | hf_overrides = { |
| 21 | "rope_parameters": { |
| 22 | "rope_theta": rope_theta, |
| 23 | "rope_type": "yarn", |
| 24 | "factor": factor, |
| 25 | "original_max_position_embeddings": original_max_position_embeddings, |
| 26 | }, |
| 27 | "max_model_len": int(original_max_position_embeddings * factor), |
| 28 | } |
| 29 | |
| 30 | llm = LLM(model="Qwen/Qwen3-0.6B", hf_overrides=hf_overrides) |
| 31 | return llm |
| 32 | |
| 33 | |
| 34 | def run_llm_chat(llm): |
| 35 | sampling_params = SamplingParams( |
| 36 | temperature=0.8, |
| 37 | top_p=0.95, |
| 38 | max_tokens=128, |
| 39 | ) |
| 40 | |
| 41 | conversation = [ |
| 42 | {"role": "system", "content": "You are a helpful assistant"}, |
| 43 | {"role": "user", "content": "Hello"}, |
| 44 | {"role": "assistant", "content": "Hello! How can I assist you today?"}, |
| 45 | ] |
| 46 | outputs = llm.chat(conversation, sampling_params, use_tqdm=False) |
| 47 | return outputs, [ |
| 48 | conversation, |
| 49 | ] |
| 50 | |
| 51 | |
| 52 | def print_outputs(outputs: list[RequestOutput], conversations: list): |
| 53 | print("\nGenerated Outputs:\n" + "-" * 80) |
| 54 | for i, output in enumerate(outputs): |
| 55 | prompt = conversations[i] |
| 56 | generated_text = output.outputs[0].text |
| 57 | print(f"Prompt: {prompt!r}\n") |
| 58 | print(f"Generated text: {generated_text!r}") |
| 59 | print("-" * 80) |
| 60 | |
| 61 | |
| 62 | def main(): |
| 63 | llm = create_llm() |
| 64 | outputs, conversations = run_llm_chat(llm) |
| 65 | print_outputs(outputs, conversations) |
| 66 | |
| 67 | |
| 68 | if __name__ == "__main__": |
| 69 | main() |