| try: |
| from transformers import AutoTokenizer |
| from vllm import LLM, SamplingParams |
| except ImportError as e: |
| |
| pass |
|
|
| from lcb_runner.runner.base_runner import BaseRunner |
|
|
|
|
| class VLLMRunner(BaseRunner): |
| def __init__(self, args, model): |
| super().__init__(args, model) |
| model_tokenizer_path = ( |
| model.model_name if args.local_model_path is None else args.local_model_path |
| ) |
| self.llm = LLM( |
| model=model_tokenizer_path, |
| tokenizer=model_tokenizer_path, |
| tensor_parallel_size=args.tensor_parallel_size, |
| dtype=args.dtype, |
| enforce_eager=True, |
| disable_custom_all_reduce=True, |
| enable_prefix_caching=args.enable_prefix_caching, |
| trust_remote_code=args.trust_remote_code, |
| ) |
| self.sampling_params = SamplingParams( |
| n=self.args.n, |
| max_tokens=self.args.max_tokens, |
| temperature=self.args.temperature, |
| top_p=self.args.top_p, |
| frequency_penalty=0, |
| presence_penalty=0, |
| stop=self.args.stop, |
| ) |
|
|
| def _run_single(self, prompt: str) -> list[str]: |
| pass |
|
|
| def run_batch(self, prompts: list[str]) -> list[list[str]]: |
| outputs = [None for _ in prompts] |
| remaining_prompts = [] |
| remaining_indices = [] |
| for prompt_index, prompt in enumerate(prompts): |
| if self.args.use_cache and prompt in self.cache: |
| if len(self.cache[prompt]) == self.args.n: |
| outputs[prompt_index] = self.cache[prompt] |
| continue |
| remaining_prompts.append(prompt) |
| remaining_indices.append(prompt_index) |
| if remaining_prompts: |
| vllm_outputs = self.llm.generate(remaining_prompts, self.sampling_params) |
| if self.args.use_cache: |
| assert len(remaining_prompts) == len(vllm_outputs) |
| for index, remaining_prompt, vllm_output in zip( |
| remaining_indices, remaining_prompts, vllm_outputs |
| ): |
| self.cache[remaining_prompt] = [o.text for o in vllm_output.outputs] |
| outputs[index] = [o.text for o in vllm_output.outputs] |
| else: |
| for index, vllm_output in zip(remaining_indices, vllm_outputs): |
| outputs[index] = [o.text for o in vllm_output.outputs] |
| return outputs |
|
|
