| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | """This example demonstrates a special case of wrapping a request-level logits |
| 5 | processor, namely the case where it is necessary to utilize engine config or |
| 6 | environment info passed to the constructor. The subclass must override the |
| 7 | wrapper base class `__init__()` method to access the engine config, the device |
| 8 | identifier, or the flag which indicates whether pinned memory is available. |
| 9 | |
| 10 | For demo purposes, a request-level dummy logits processor is employed which |
| 11 | causes the same token (`target_token`) to be decoded in each step. The |
| 12 | request-level dummy logits processor is wrapped to create a batch-level logits |
| 13 | processor, which can apply the logits processor to output logits from all |
| 14 | requests in the persistent batch in a given decode step. |
| 15 | |
| 16 | The wrapped dummy logits processor below models a scenario where we must |
| 17 | disable the logits processor on non-"cuda" platforms. The wrapper base class |
| 18 | `__init__()` is overridden in order to check this condition and set a flag. |
| 19 | |
| 20 | A batch is constructed with `temperature=0.0` and 50% of requests specifying |
| 21 | `target_token`, and for these requests - and *only* these requests - we |
| 22 | expect that on a "cuda" device the output will look something like: |
| 23 | |
| 24 | Generated Outputs: |
| 25 | ------------------------------------------------------------ |
| 26 | Prompt: 'Hello, my name is' |
| 27 | Output: " ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' '" |
| 28 | ------------------------------------------------------------ |
| 29 | Prompt: 'The president of the United States is' |
| 30 | Output: " not a racist. He is a racist.\nHe's a racist because he" |
| 31 | ------------------------------------------------------------ |
| 32 | Prompt: 'The capital of France is' |
| 33 | Output: ' also also also also also also also also also also also also also |
| 34 | also also also' |
| 35 | ------------------------------------------------------------ |
| 36 | Prompt: 'The future of AI is' |
| 37 | Output: ' in the hands of the people.\n\nThe future of AI is in the' |
| 38 | ------------------------------------------------------------ |
| 39 | |
| 40 | which indicates that the logits processor is running. However, on a non-"cuda" |
| 41 | device, the first and third requests would not repeat the same token. |
| 42 | """ |
| 43 | |
| 44 | import torch |
| 45 | |
| 46 | from vllm import LLM, SamplingParams |
| 47 | from vllm.config import VllmConfig |
| 48 | from vllm.logger import init_logger |
| 49 | from vllm.v1.sample.logits_processor import ( |
| 50 | AdapterLogitsProcessor, |
| 51 | RequestLogitsProcessor, |
| 52 | ) |
| 53 | |
| 54 | logger = init_logger(__name__) |
| 55 | |
| 56 | |
| 57 | class DummyPerReqLogitsProcessor: |
| 58 | """The request-level logits processor masks out all logits except the |
| 59 | token id identified by `target_token`""" |
| 60 | |
| 61 | def __init__(self, target_token: int) -> None: |
| 62 | """Specify `target_token`.""" |
| 63 | self.target_token = target_token |
| 64 | |
| 65 | def __call__( |
| 66 | self, |
| 67 | output_ids: list[int], |
| 68 | logits: torch.Tensor, |
| 69 | ) -> torch.Tensor: |
| 70 | val_to_keep = logits[self.target_token].item() |
| 71 | logits[:] = float("-inf") |
| 72 | logits[self.target_token] = val_to_keep |
| 73 | return logits |
| 74 | |
| 75 | |
| 76 | class WrappedPerReqLogitsProcessor(AdapterLogitsProcessor): |
| 77 | """Example of overriding the wrapper class `__init__()` in order to utilize |
| 78 | info about the device type""" |
| 79 | |
| 80 | @classmethod |
| 81 | def validate_params(cls, params: SamplingParams): |
| 82 | target_token = params.extra_args and params.extra_args.get("target_token") |
| 83 | if target_token is not None and not isinstance(target_token, int): |
| 84 | raise ValueError( |
| 85 | f"`target_token` has to be an integer, got {target_token}." |
| 86 | ) |
| 87 | |
| 88 | def __init__( |
| 89 | self, vllm_config: VllmConfig, device: torch.device, is_pin_memory: bool |
| 90 | ): |
| 91 | super().__init__(vllm_config, device, is_pin_memory) |
| 92 | self.is_cuda = device.type == "cuda" |
| 93 | |
| 94 | def is_argmax_invariant(self) -> bool: |
| 95 | return False |
| 96 | |
| 97 | def new_req_logits_processor( |
| 98 | self, |
| 99 | params: SamplingParams, |
| 100 | ) -> RequestLogitsProcessor | None: |
| 101 | """This method returns a new request-level logits processor, customized |
| 102 | to the `target_token` value associated with a particular request. |
| 103 | |
| 104 | Returns None if the logits processor should not be applied to the |
| 105 | particular request. To use the logits processor the request must have |
| 106 | a "target_token" custom argument with an integer value, and the device |
| 107 | must be "cuda"-type |
| 108 | |
| 109 | Args: |
| 110 | params: per-request sampling params |
| 111 | |
| 112 | Returns: |
| 113 | `Callable` request logits processor, or None |
| 114 | |
| 115 | """ |
| 116 | if ( |
| 117 | not self.is_cuda |
| 118 | or ( |
| 119 | target_token := params.extra_args |
| 120 | and params.extra_args.get("target_token") |
| 121 | ) |
| 122 | is None |
| 123 | ): |
| 124 | return None |
| 125 | return DummyPerReqLogitsProcessor(target_token) |
| 126 | |
| 127 | |
| 128 | # Sample prompts. |
| 129 | prompts = [ |
| 130 | "Hello, my name is", |
| 131 | "The president of the United States is", |
| 132 | "The capital of France is", |
| 133 | "The future of AI is", |
| 134 | ] |
| 135 | # Create a mixture of requests which do and don't utilize the dummy logitproc |
| 136 | sampling_params_list = [ |
| 137 | SamplingParams(temperature=0.0, extra_args={"target_token": 128}), |
| 138 | SamplingParams(temperature=0.0), |
| 139 | SamplingParams(temperature=0.0, extra_args={"target_token": 67}), |
| 140 | SamplingParams(temperature=0.0), |
| 141 | ] |
| 142 | |
| 143 | |
| 144 | def main(): |
| 145 | # Create an LLM. |
| 146 | llm = LLM( |
| 147 | model="facebook/opt-125m", |
| 148 | logits_processors=[WrappedPerReqLogitsProcessor], |
| 149 | ) |
| 150 | # Generate texts from the prompts. |
| 151 | # The output is a list of RequestOutput objects |
| 152 | # that contain the prompt, generated text, and other information. |
| 153 | outputs = llm.generate(prompts, sampling_params_list) |
| 154 | # Print the outputs. |
| 155 | print("\nGenerated Outputs:\n" + "-" * 60) |
| 156 | for output in outputs: |
| 157 | prompt = output.prompt |
| 158 | generated_text = output.outputs[0].text |
| 159 | print(f"Prompt: {prompt!r}") |
| 160 | print(f"Output: {generated_text!r}") |
| 161 | print("-" * 60) |
| 162 | |
| 163 | |
| 164 | if __name__ == "__main__": |
| 165 | main() |