| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | """This example demonstrates instantiating vLLM with a custom logits processor |
| 5 | class object. |
| 6 | |
| 7 | For a basic example of implementing a custom logits processor, see |
| 8 | the `DummyLogitsProcessor` implementation in `vllm/test_utils.py`. |
| 9 | |
| 10 | For testing purposes, a dummy logits processor is employed which, if |
| 11 | `target_token` is passed as a keyword argument to `SamplingParams.extra_args`, |
| 12 | will mask out all tokens except `target_token`. |
| 13 | |
| 14 | A batch is constructed with `temperature=0.0` and 50% of requests specifying |
| 15 | `target_token`, and for these requests - and *only* these requests - we |
| 16 | expect the `target_token` to be decoded in each step, yielding an output |
| 17 | similar to that shown below: |
| 18 | |
| 19 | Generated Outputs: |
| 20 | ------------------------------------------------------------ |
| 21 | Prompt: 'Hello, my name is' |
| 22 | Output: " ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' '" |
| 23 | ------------------------------------------------------------ |
| 24 | Prompt: 'The president of the United States is' |
| 25 | Output: " not a racist. He is a racist.\nHe's a racist because he" |
| 26 | ------------------------------------------------------------ |
| 27 | Prompt: 'The capital of France is' |
| 28 | Output: ' also also also also also also also also also also also also also |
| 29 | also also also' |
| 30 | ------------------------------------------------------------ |
| 31 | Prompt: 'The future of AI is' |
| 32 | Output: ' in the hands of the people.\n\nThe future of AI is in the' |
| 33 | ------------------------------------------------------------ |
| 34 | """ |
| 35 | |
| 36 | from typing import Any |
| 37 | |
| 38 | import torch |
| 39 | |
| 40 | from vllm import LLM, SamplingParams |
| 41 | from vllm.config import VllmConfig |
| 42 | from vllm.v1.sample.logits_processor import ( |
| 43 | BatchUpdate, |
| 44 | LogitsProcessor, |
| 45 | ) |
| 46 | from vllm.v1.sample.logits_processor.builtin import process_dict_updates |
| 47 | |
| 48 | |
| 49 | # Hypothetical custom logits processor |
| 50 | class DummyLogitsProcessor(LogitsProcessor): |
| 51 | """Fake logit processor to support unit testing and examples.""" |
| 52 | |
| 53 | @classmethod |
| 54 | def validate_params(cls, params: SamplingParams): |
| 55 | target_token: Any | None = params.extra_args and params.extra_args.get( |
| 56 | "target_token" |
| 57 | ) |
| 58 | if target_token is not None and not isinstance(target_token, int): |
| 59 | raise ValueError( |
| 60 | f"target_token value {target_token} {type(target_token)} is not int" |
| 61 | ) |
| 62 | |
| 63 | def __init__( |
| 64 | self, vllm_config: VllmConfig, device: torch.device, is_pin_memory: bool |
| 65 | ): |
| 66 | self.req_info: dict[int, int] = {} |
| 67 | |
| 68 | def is_argmax_invariant(self) -> bool: |
| 69 | return False |
| 70 | |
| 71 | def update_state(self, batch_update: BatchUpdate | None): |
| 72 | def extract_extra_arg(params: SamplingParams) -> int | None: |
| 73 | self.validate_params(params) |
| 74 | return params.extra_args and params.extra_args.get("target_token") |
| 75 | |
| 76 | process_dict_updates( |
| 77 | self.req_info, |
| 78 | batch_update, |
| 79 | # This function returns the LP's per-request state based on the |
| 80 | # request details, or None if this LP does not apply to the |
| 81 | # request. |
| 82 | lambda params, _, __: extract_extra_arg(params), |
| 83 | ) |
| 84 | |
| 85 | def apply(self, logits: torch.Tensor) -> torch.Tensor: |
| 86 | if not self.req_info: |
| 87 | return logits |
| 88 | |
| 89 | # Save target values before modification |
| 90 | cols = torch.tensor( |
| 91 | list(self.req_info.values()), dtype=torch.long, device=logits.device |
| 92 | ) |
| 93 | rows = torch.tensor( |
| 94 | list(self.req_info.keys()), dtype=torch.long, device=logits.device |
| 95 | ) |
| 96 | values_to_keep = logits[rows, cols].clone() |
| 97 | |
| 98 | # Mask all but target tokens |
| 99 | logits[rows] = float("-inf") |
| 100 | logits[rows, cols] = values_to_keep |
| 101 | |
| 102 | return logits |
| 103 | |
| 104 | |
| 105 | # Sample prompts. |
| 106 | prompts = [ |
| 107 | "Hello, my name is", |
| 108 | "The president of the United States is", |
| 109 | "The capital of France is", |
| 110 | "The future of AI is", |
| 111 | ] |
| 112 | # Create a mixture of requests which do and don't utilize the dummy logitproc |
| 113 | sampling_params_list = [ |
| 114 | SamplingParams(temperature=0.0, extra_args={"target_token": 128}), |
| 115 | SamplingParams(temperature=0.0), |
| 116 | SamplingParams(temperature=0.0, extra_args={"target_token": 67}), |
| 117 | SamplingParams(temperature=0.0), |
| 118 | ] |
| 119 | |
| 120 | |
| 121 | def main(): |
| 122 | # Create an LLM. |
| 123 | llm = LLM( |
| 124 | model="facebook/opt-125m", |
| 125 | logits_processors=[DummyLogitsProcessor], |
| 126 | ) |
| 127 | # Generate texts from the prompts. |
| 128 | # The output is a list of RequestOutput objects |
| 129 | # that contain the prompt, generated text, and other information. |
| 130 | outputs = llm.generate(prompts, sampling_params_list) |
| 131 | # Print the outputs. |
| 132 | print("\nGenerated Outputs:\n" + "-" * 60) |
| 133 | for output in outputs: |
| 134 | prompt = output.prompt |
| 135 | generated_text = output.outputs[0].text |
| 136 | print(f"Prompt: {prompt!r}") |
| 137 | print(f"Output: {generated_text!r}") |
| 138 | print("-" * 60) |
| 139 | |
| 140 | |
| 141 | if __name__ == "__main__": |
| 142 | main() |