| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Example of offline classification with a LoRA adapter.""" |
| 4 | |
| 5 | from vllm import LLM |
| 6 | from vllm.lora.request import LoRARequest |
| 7 | from vllm.transformers_utils.repo_utils import hf_api |
| 8 | |
| 9 | |
| 10 | def main(): |
| 11 | lora_path = hf_api().snapshot_download( |
| 12 | repo_id="AmirMohseni/skywork-qwen3-0.6b-reward-lora" |
| 13 | ) |
| 14 | llm = LLM( |
| 15 | model="Skywork/Skywork-Reward-V2-Qwen3-0.6B", |
| 16 | runner="pooling", |
| 17 | enable_lora=True, |
| 18 | max_lora_rank=8, |
| 19 | max_model_len=512, |
| 20 | ) |
| 21 | |
| 22 | prompt = "Which response is more helpful and correct?" |
| 23 | (output,) = llm.classify( |
| 24 | prompt, |
| 25 | lora_request=LoRARequest("reward-lora", 1, lora_path), |
| 26 | ) |
| 27 | print(f"Prompt: {prompt!r}") |
| 28 | print(f"Class Probabilities: {output.outputs.probs}") |
| 29 | |
| 30 | |
| 31 | if __name__ == "__main__": |
| 32 | main() |