| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | """Example offline usage of sequence reward models. |
| 5 | |
| 6 | The key distinction between sequence classification and token classification |
| 7 | lies in their output granularity: sequence classification produces a single |
| 8 | result for an entire input sequence, whereas token classification yields a |
| 9 | result for each individual token within the sequence. |
| 10 | """ |
| 11 | |
| 12 | from argparse import Namespace |
| 13 | |
| 14 | from vllm import LLM, EngineArgs |
| 15 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 16 | from vllm.utils.print_utils import print_embeddings |
| 17 | |
| 18 | |
| 19 | def parse_args(): |
| 20 | parser = FlexibleArgumentParser() |
| 21 | parser = EngineArgs.add_cli_args(parser) |
| 22 | # Set example specific arguments |
| 23 | parser.set_defaults( |
| 24 | model="Skywork/Skywork-Reward-V2-Qwen3-0.6B", |
| 25 | runner="pooling", |
| 26 | enforce_eager=True, |
| 27 | max_model_len=1024, |
| 28 | trust_remote_code=True, |
| 29 | ) |
| 30 | return parser.parse_args() |
| 31 | |
| 32 | |
| 33 | def main(args: Namespace): |
| 34 | # Sample prompts. |
| 35 | prompts = [ |
| 36 | "Hello, my name is", |
| 37 | "The president of the United States is", |
| 38 | "The capital of France is", |
| 39 | "The future of AI is", |
| 40 | ] |
| 41 | |
| 42 | # Create an LLM. |
| 43 | # You should pass runner="pooling" for reward models |
| 44 | llm = LLM(**vars(args)) |
| 45 | |
| 46 | # Generate rewards. The output is a list of PoolingRequestOutput. |
| 47 | # Use pooling_task="classify" for sequence reward models. |
| 48 | outputs = llm.encode(prompts, pooling_task="classify") |
| 49 | |
| 50 | # Print the outputs. |
| 51 | print("\nGenerated Outputs:\n" + "-" * 60) |
| 52 | for prompt, output in zip(prompts, outputs): |
| 53 | rewards = output.outputs.data |
| 54 | print(f"Prompt: {prompt!r}") |
| 55 | print_embeddings(rewards.tolist(), prefix="Reward") |
| 56 | print("-" * 60) |
| 57 | |
| 58 | |
| 59 | if __name__ == "__main__": |
| 60 | args = parse_args() |
| 61 | main(args) |