import itertools import warnings from contextlib import contextmanager from typing import (Any, ClassVar, Dict, List, Optional, Sequence, Tuple, Type, Union, cast, overload) from tqdm import tqdm from vllm import envs from vllm.beam_search import (BeamSearchInstance, BeamSearchOutput, BeamSearchSequence, get_beam_search_score) from vllm.engine.arg_utils import (EngineArgs, HfOverrides, PoolerConfig, TaskOption) from vllm.engine.llm_engine import LLMEngine from vllm.entrypoints.chat_utils import (ChatCompletionMessageParam, apply_hf_chat_template, apply_mistral_chat_template, parse_chat_messages) from vllm.inputs import PromptType, TextPrompt, TokensPrompt from vllm.inputs.parse import parse_and_batch_prompt from vllm.logger import init_logger from vllm.lora.request import LoRARequest from vllm.model_executor.guided_decoding.guided_fields import ( GuidedDecodingRequest, LLMGuidedOptions) from vllm.outputs import EmbeddingRequestOutput, RequestOutput from vllm.pooling_params import PoolingParams from vllm.prompt_adapter.request import PromptAdapterRequest from vllm.sampling_params import (BeamSearchParams, GuidedDecodingParams, RequestOutputKind, SamplingParams) from vllm.transformers_utils.tokenizer import (AnyTokenizer, MistralTokenizer, get_cached_tokenizer) from vllm.transformers_utils.tokenizer_group import TokenizerGroup from vllm.usage.usage_lib import UsageContext from vllm.utils import Counter, deprecate_args, deprecate_kwargs, is_list_of logger = init_logger(__name__) class LLM: """An LLM for generating texts from given prompts and sampling parameters. This class includes a tokenizer, a language model (possibly distributed across multiple GPUs), and GPU memory space allocated for intermediate states (aka KV cache). Given a batch of prompts and sampling parameters, this class generates texts from the model, using an intelligent batching mechanism and efficient memory management. Args: model: The name or path of a HuggingFace Transformers model. tokenizer: The name or path of a HuggingFace Transformers tokenizer. tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer if available, and "slow" will always use the slow tokenizer. skip_tokenizer_init: If true, skip initialization of tokenizer and detokenizer. Expect valid prompt_token_ids and None for prompt from the input. trust_remote_code: Trust remote code (e.g., from HuggingFace) when downloading the model and tokenizer. allowed_local_media_path: Allowing API requests to read local images or videos from directories specified by the server file system. This is a security risk. Should only be enabled in trusted environments. tensor_parallel_size: The number of GPUs to use for distributed execution with tensor parallelism. dtype: The data type for the model weights and activations. Currently, we support `float32`, `float16`, and `bfloat16`. If `auto`, we use the `torch_dtype` attribute specified in the model config file. However, if the `torch_dtype` in the config is `float32`, we will use `float16` instead. quantization: The method used to quantize the model weights. Currently, we support "awq", "gptq", "squeezellm", "weightonly", and "fp8" (experimental). If None, we first check the `quantization_config` attribute in the model config file. If that is None, we assume the model weights are not quantized and use `dtype` to determine the data type of the weights. revision: The specific model version to use. It can be a branch name, a tag name, or a commit id. tokenizer_revision: The specific tokenizer version to use. It can be a branch name, a tag name, or a commit id. seed: The seed to initialize the random number generator for sampling. gpu_memory_utilization: The ratio (between 0 and 1) of GPU memory to reserve for the model weights, activations, and KV cache. Higher values will increase the KV cache size and thus improve the model's throughput. However, if the value is too high, it may cause out-of- memory (OOM) errors. swap_space: The size (GiB) of CPU memory per GPU to use as swap space. This can be used for temporarily storing the states of the requests when their `best_of` sampling parameters are larger than 1. If all requests will have `best_of=1`, you can safely set this to 0. Otherwise, too small values may cause out-of-memory (OOM) errors. cpu_offload_gb: The size (GiB) of CPU memory to use for offloading the model weights. This virtually increases the GPU memory space you can use to hold the model weights, at the cost of CPU-GPU data transfer for every forward pass. enforce_eager: Whether to enforce eager execution. If True, we will disable CUDA graph and always execute the model in eager mode. If False, we will use CUDA graph and eager execution in hybrid. max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs. When a sequence has context length larger than this, we fall back to eager mode. Additionally for encoder-decoder models, if the sequence length of the encoder input is larger than this, we fall back to the eager mode. disable_custom_all_reduce: See :class:`~vllm.config.ParallelConfig` disable_async_output_proc: Disable async output processing. This may result in lower performance. hf_overrides: If a dictionary, contains arguments to be forwarded to the HuggingFace config. If a callable, it is called to update the HuggingFace config. **kwargs: Arguments for :class:`~vllm.EngineArgs`. (See :ref:`engine_args`) Note: This class is intended to be used for offline inference. For online serving, use the :class:`~vllm.AsyncLLMEngine` class instead. """ DEPRECATE_LEGACY: ClassVar[bool] = False """A flag to toggle whether to deprecate the legacy generate/encode API.""" DEPRECATE_INIT_POSARGS: ClassVar[bool] = True """ A flag to toggle whether to deprecate positional arguments in :meth:`LLM.__init__`. """ @classmethod @contextmanager def deprecate_legacy_api(cls): cls.DEPRECATE_LEGACY = True yield cls.DEPRECATE_LEGACY = False @deprecate_args( start_index=2, # Ignore self and model is_deprecated=lambda: LLM.DEPRECATE_INIT_POSARGS, additional_message=( "All positional arguments other than `model` will be " "replaced with keyword arguments in an upcoming version."), ) def __init__( self, model: str, tokenizer: Optional[str] = None, tokenizer_mode: str = "auto", skip_tokenizer_init: bool = False, trust_remote_code: bool = False, allowed_local_media_path: str = "", tensor_parallel_size: int = 1, dtype: str = "auto", quantization: Optional[str] = None, revision: Optional[str] = None, tokenizer_revision: Optional[str] = None, seed: int = 0, gpu_memory_utilization: float = 0.9, swap_space: float = 4, cpu_offload_gb: float = 0, enforce_eager: Optional[bool] = None, max_seq_len_to_capture: int = 8192, disable_custom_all_reduce: bool = False, disable_async_output_proc: bool = False, hf_overrides: Optional[HfOverrides] = None, mm_processor_kwargs: Optional[Dict[str, Any]] = None, # After positional args are removed, move this right below `model` task: TaskOption = "auto", override_pooler_config: Optional[PoolerConfig] = None, **kwargs, ) -> None: ''' LLM constructor. Note: if enforce_eager is unset (enforce_eager is None) it defaults to False. ''' if "disable_log_stats" not in kwargs: kwargs["disable_log_stats"] = True engine_args = EngineArgs( model=model, task=task, tokenizer=tokenizer, tokenizer_mode=tokenizer_mode, skip_tokenizer_init=skip_tokenizer_init, trust_remote_code=trust_remote_code, allowed_local_media_path=allowed_local_media_path, tensor_parallel_size=tensor_parallel_size, dtype=dtype, quantization=quantization, revision=revision, tokenizer_revision=tokenizer_revision, seed=seed, gpu_memory_utilization=gpu_memory_utilization, swap_space=swap_space, cpu_offload_gb=cpu_offload_gb, enforce_eager=enforce_eager, max_seq_len_to_capture=max_seq_len_to_capture, disable_custom_all_reduce=disable_custom_all_reduce, disable_async_output_proc=disable_async_output_proc, hf_overrides=hf_overrides, mm_processor_kwargs=mm_processor_kwargs, override_pooler_config=override_pooler_config, **kwargs, ) # Logic to switch between engines is done at runtime instead of import # to avoid import order issues self.engine_class = self.get_engine_class() # TODO(rob): enable mp by default (issue with fork vs spawn) self.llm_engine = self.engine_class.from_engine_args( engine_args, usage_context=UsageContext.LLM_CLASS) self.request_counter = Counter() @staticmethod def get_engine_class() -> Type[LLMEngine]: if envs.VLLM_USE_V1: # Lazy import: the v1 package isn't distributed from vllm.v1.engine.llm_engine import LLMEngine as V1LLMEngine return V1LLMEngine # type: ignore return LLMEngine def get_tokenizer(self) -> AnyTokenizer: return self.llm_engine.get_tokenizer_group(TokenizerGroup).tokenizer def set_tokenizer(self, tokenizer: AnyTokenizer) -> None: tokenizer_group = self.llm_engine.get_tokenizer_group(TokenizerGroup) # While CachedTokenizer is dynamic, have no choice but # compare class name. Misjudgment will arise from # user-defined tokenizer started with 'Cached' if tokenizer.__class__.__name__.startswith("Cached"): tokenizer_group.tokenizer = tokenizer else: tokenizer_group.tokenizer = get_cached_tokenizer(tokenizer) @overload # LEGACY: single (prompt + optional token ids) def generate( self, prompts: str, sampling_params: Optional[Union[SamplingParams, List[SamplingParams]]] = None, prompt_token_ids: Optional[List[int]] = None, use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[RequestOutput]: ... @overload # LEGACY: multi (prompt + optional token ids) def generate( self, prompts: List[str], sampling_params: Optional[Union[SamplingParams, List[SamplingParams]]] = None, prompt_token_ids: Optional[List[List[int]]] = None, use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[RequestOutput]: ... @overload # LEGACY: single (token ids + optional prompt) def generate( self, prompts: Optional[str] = None, sampling_params: Optional[Union[SamplingParams, List[SamplingParams]]] = None, *, prompt_token_ids: List[int], use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[RequestOutput]: ... @overload # LEGACY: multi (token ids + optional prompt) def generate( self, prompts: Optional[List[str]] = None, sampling_params: Optional[Union[SamplingParams, List[SamplingParams]]] = None, *, prompt_token_ids: List[List[int]], use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[RequestOutput]: ... @overload # LEGACY: single or multi token ids [pos-only] def generate( self, prompts: None, sampling_params: None, prompt_token_ids: Union[List[int], List[List[int]]], use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[RequestOutput]: ... @overload def generate( self, prompts: Union[PromptType, Sequence[PromptType]], /, *, sampling_params: Optional[Union[SamplingParams, Sequence[SamplingParams]]] = None, use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[RequestOutput]: ... @deprecate_kwargs( "prompt_token_ids", is_deprecated=lambda: LLM.DEPRECATE_LEGACY, additional_message="Please use the 'prompts' parameter instead.", ) def generate( self, prompts: Union[Union[PromptType, Sequence[PromptType]], Optional[Union[str, List[str]]]] = None, sampling_params: Optional[Union[SamplingParams, Sequence[SamplingParams]]] = None, prompt_token_ids: Optional[Union[List[int], List[List[int]]]] = None, use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, prompt_adapter_request: Optional[PromptAdapterRequest] = None, guided_options_request: Optional[Union[LLMGuidedOptions, GuidedDecodingRequest]] = None, priority: Optional[List[int]] = None, ) -> List[RequestOutput]: """Generates the completions for the input prompts. This class automatically batches the given prompts, considering the memory constraint. For the best performance, put all of your prompts into a single list and pass it to this method. Args: prompts: The prompts to the LLM. You may pass a sequence of prompts for batch inference. See :class:`~vllm.inputs.PromptType` for more details about the format of each prompts. sampling_params: The sampling parameters for text generation. If None, we use the default sampling parameters. When it is a single value, it is applied to every prompt. When it is a list, the list must have the same length as the prompts and it is paired one by one with the prompt. use_tqdm: Whether to use tqdm to display the progress bar. lora_request: LoRA request to use for generation, if any. prompt_adapter_request: Prompt Adapter request to use for generation, if any. priority: The priority of the requests, if any. Only applicable when priority scheduling policy is enabled. Returns: A list of ``RequestOutput`` objects containing the generated completions in the same order as the input prompts. Note: Using ``prompts`` and ``prompt_token_ids`` as keyword parameters is considered legacy and may be deprecated in the future. You should instead pass them via the ``inputs`` parameter. """ task = self.llm_engine.model_config.task if task != "generate": messages = [ "LLM.generate() is only supported for (conditional) generation " "models (XForCausalLM, XForConditionalGeneration).", ] supported_tasks = self.llm_engine.model_config.supported_tasks if "generate" in supported_tasks: messages.append( "Your model supports the 'generate' task, but is " f"currently initialized for the '{task}' task. Please " "initialize the model using `--task generate`.") raise ValueError(" ".join(messages)) if prompt_token_ids is not None: parsed_prompts = self._convert_v1_inputs( prompts=cast(Optional[Union[str, List[str]]], prompts), prompt_token_ids=prompt_token_ids, ) else: parsed_prompts = cast(Union[PromptType, Sequence[PromptType]], prompts) if isinstance(guided_options_request, dict): if len(guided_options_request) > 1: raise ValueError( "You can only use one guided decoding but multiple is " f"specified: {guided_options_request}") guided_options_request = GuidedDecodingRequest( **guided_options_request) if sampling_params is None: # Use default sampling params. sampling_params = SamplingParams() self._validate_and_add_requests( prompts=parsed_prompts, params=sampling_params, lora_request=lora_request, prompt_adapter_request=prompt_adapter_request, guided_options=guided_options_request, priority=priority) outputs = self._run_engine(use_tqdm=use_tqdm) return self.engine_class.validate_outputs(outputs, RequestOutput) def beam_search( self, prompts: List[Union[str, List[int]]], params: BeamSearchParams, ) -> List[BeamSearchOutput]: """ Generate sequences using beam search. Args: prompts: A list of prompts. Each prompt can be a string or a list of token IDs. params: The beam search parameters. TODO: how does beam search work together with length penalty, frequency penalty, and stopping criteria, etc.? """ beam_width = params.beam_width max_tokens = params.max_tokens temperature = params.temperature ignore_eos = params.ignore_eos length_penalty = params.length_penalty def sort_beams_key(x: BeamSearchSequence) -> float: return get_beam_search_score(x.tokens, x.cum_logprob, tokenizer.eos_token_id, length_penalty) tokenizer = self.get_tokenizer() # generate 2 * beam_width candidates at each step # following the huggingface transformers implementation # at https://github.com/huggingface/transformers/blob/e15687fffe5c9d20598a19aeab721ae0a7580f8a/src/transformers/generation/beam_search.py#L534 # noqa beam_search_params = SamplingParams(logprobs=2 * beam_width, max_tokens=1, temperature=temperature) instances: List[BeamSearchInstance] = [] for prompt in prompts: prompt_tokens = prompt if isinstance( prompt, list) else tokenizer.encode(prompt) instances.append(BeamSearchInstance(prompt_tokens)) for _ in range(max_tokens): all_beams: List[BeamSearchSequence] = list( sum((instance.beams for instance in instances), [])) pos = [0] + list( itertools.accumulate( len(instance.beams) for instance in instances)) instance_start_and_end: List[Tuple[int, int]] = list( zip(pos[:-1], pos[1:])) if len(all_beams) == 0: break prompts_batch = [ TokensPrompt(prompt_token_ids=beam.tokens) for beam in all_beams ] # only runs for one step # we don't need to use tqdm here output = self.generate(prompts_batch, sampling_params=beam_search_params, use_tqdm=False) for (start, end), instance in zip(instance_start_and_end, instances): instance_new_beams = [] for i in range(start, end): current_beam = all_beams[i] result = output[i] if result.outputs[0].logprobs is not None: # if `result.outputs[0].logprobs` is None, it means # the sequence is completed because of the max-model-len # or abortion. we don't need to add it to the new beams. logprobs = result.outputs[0].logprobs[0] for token_id, logprob_obj in logprobs.items(): new_beam = BeamSearchSequence( tokens=current_beam.tokens + [token_id], logprobs=current_beam.logprobs + [logprobs], cum_logprob=current_beam.cum_logprob + logprob_obj.logprob) if token_id == tokenizer.eos_token_id and \ not ignore_eos: instance.completed.append(new_beam) else: instance_new_beams.append(new_beam) sorted_beams = sorted(instance_new_beams, key=sort_beams_key, reverse=True) instance.beams = sorted_beams[:beam_width] outputs = [] for instance in instances: instance.completed.extend(instance.beams) sorted_completed = sorted(instance.completed, key=sort_beams_key, reverse=True) best_beams = sorted_completed[:beam_width] for beam in best_beams: beam.text = tokenizer.decode(beam.tokens) outputs.append(BeamSearchOutput(sequences=best_beams)) return outputs def chat( self, messages: Union[List[ChatCompletionMessageParam], List[List[ChatCompletionMessageParam]]], sampling_params: Optional[Union[SamplingParams, List[SamplingParams]]] = None, use_tqdm: bool = True, lora_request: Optional[LoRARequest] = None, chat_template: Optional[str] = None, add_generation_prompt: bool = True, continue_final_message: bool = False, tools: Optional[List[Dict[str, Any]]] = None, mm_processor_kwargs: Optional[Dict[str, Any]] = None, ) -> List[RequestOutput]: """ Generate responses for a chat conversation. The chat conversation is converted into a text prompt using the tokenizer and calls the :meth:`generate` method to generate the responses. Multi-modal inputs can be passed in the same way you would pass them to the OpenAI API. Args: messages: A list of conversations or a single conversation. - Each conversation is represented as a list of messages. - Each message is a dictionary with 'role' and 'content' keys. sampling_params: The sampling parameters for text generation. If None, we use the default sampling parameters. When it is a single value, it is applied to every prompt. When it is a list, the list must have the same length as the prompts and it is paired one by one with the prompt. use_tqdm: Whether to use tqdm to display the progress bar. lora_request: LoRA request to use for generation, if any. chat_template: The template to use for structuring the chat. If not provided, the model's default chat template will be used. add_generation_prompt: If True, adds a generation template to each message. continue_final_message: If True, continues the final message in the conversation instead of starting a new one. Cannot be `True` if `add_generation_prompt` is also `True`. mm_processor_kwargs: Multimodal processor kwarg overrides for this chat request. Only used for offline requests. Returns: A list of ``RequestOutput`` objects containing the generated responses in the same order as the input messages. """ list_of_messages: List[List[ChatCompletionMessageParam]] # Handle multi and single conversations if is_list_of(messages, list): # messages is List[List[...]] list_of_messages = cast(List[List[ChatCompletionMessageParam]], messages) else: # messages is List[...] list_of_messages = [ cast(List[ChatCompletionMessageParam], messages) ] prompts: List[Union[TokensPrompt, TextPrompt]] = [] for msgs in list_of_messages: tokenizer = self.get_tokenizer() model_config = self.llm_engine.get_model_config() # NOTE: _parse_chat_message_content_parts() currently doesn't # handle mm_processor_kwargs, since there is no implementation in # the chat message parsing for it. conversation, mm_data = parse_chat_messages( msgs, model_config, tokenizer) prompt_data: Union[str, List[int]] if isinstance(tokenizer, MistralTokenizer): prompt_data = apply_mistral_chat_template( tokenizer, messages=msgs, chat_template=chat_template, add_generation_prompt=add_generation_prompt, continue_final_message=continue_final_message, tools=tools, ) else: prompt_data = apply_hf_chat_template( tokenizer, conversation=conversation, chat_template=chat_template, add_generation_prompt=add_generation_prompt, continue_final_message=continue_final_message, tools=tools, ) prompt: Union[TokensPrompt, TextPrompt] if is_list_of(prompt_data, int): prompt = TokensPrompt(prompt_token_ids=prompt_data) else: prompt = TextPrompt(prompt=prompt_data) if mm_data is not None: prompt["multi_modal_data"] = mm_data if mm_processor_kwargs is not None: prompt["mm_processor_kwargs"] = mm_processor_kwargs prompts.append(prompt) return self.generate( prompts, sampling_params=sampling_params, use_tqdm=use_tqdm, lora_request=lora_request, ) @overload # LEGACY: single (prompt + optional token ids) def encode( self, prompts: str, pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None, prompt_token_ids: Optional[List[int]] = None, use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[EmbeddingRequestOutput]: ... @overload # LEGACY: multi (prompt + optional token ids) def encode( self, prompts: List[str], pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None, prompt_token_ids: Optional[List[List[int]]] = None, use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[EmbeddingRequestOutput]: ... @overload # LEGACY: single (token ids + optional prompt) def encode( self, prompts: Optional[str] = None, pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None, *, prompt_token_ids: List[int], use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[EmbeddingRequestOutput]: ... @overload # LEGACY: multi (token ids + optional prompt) def encode( self, prompts: Optional[List[str]] = None, pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None, *, prompt_token_ids: List[List[int]], use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[EmbeddingRequestOutput]: ... @overload # LEGACY: single or multi token ids [pos-only] def encode( self, prompts: None, pooling_params: None, prompt_token_ids: Union[List[int], List[List[int]]], use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[EmbeddingRequestOutput]: ... @overload def encode( self, prompts: Union[PromptType, Sequence[PromptType]], /, *, pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None, use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, ) -> List[EmbeddingRequestOutput]: ... @deprecate_kwargs( "prompt_token_ids", is_deprecated=lambda: LLM.DEPRECATE_LEGACY, additional_message="Please use the 'prompts' parameter instead.", ) def encode( self, prompts: Union[Union[PromptType, Sequence[PromptType]], Optional[Union[str, List[str]]]] = None, pooling_params: Optional[Union[PoolingParams, Sequence[PoolingParams]]] = None, prompt_token_ids: Optional[Union[List[int], List[List[int]]]] = None, use_tqdm: bool = True, lora_request: Optional[Union[List[LoRARequest], LoRARequest]] = None, prompt_adapter_request: Optional[PromptAdapterRequest] = None, ) -> List[EmbeddingRequestOutput]: """Generates the completions for the input prompts. This class automatically batches the given prompts, considering the memory constraint. For the best performance, put all of your prompts into a single list and pass it to this method. Args: prompts: The prompts to the LLM. You may pass a sequence of prompts for batch inference. See :class:`~vllm.inputs.PromptType` for more details about the format of each prompts. pooling_params: The pooling parameters for pooling. If None, we use the default pooling parameters. use_tqdm: Whether to use tqdm to display the progress bar. lora_request: LoRA request to use for generation, if any. prompt_adapter_request: Prompt Adapter request to use for generation, if any. Returns: A list of `EmbeddingRequestOutput` objects containing the generated embeddings in the same order as the input prompts. Note: Using ``prompts`` and ``prompt_token_ids`` as keyword parameters is considered legacy and may be deprecated in the future. You should instead pass them via the ``inputs`` parameter. """ task = self.llm_engine.model_config.task if task != "embedding": messages = ["LLM.encode() is only supported for embedding models."] supported_tasks = self.llm_engine.model_config.supported_tasks if "embedding" in supported_tasks: messages.append( "Your model supports the 'embedding' task, but is " f"currently initialized for the '{task}' task. Please " "initialize the model using `--task embedding`.") raise ValueError(" ".join(messages)) if prompt_token_ids is not None: parsed_prompts = self._convert_v1_inputs( prompts=cast(Optional[Union[str, List[str]]], prompts), prompt_token_ids=prompt_token_ids, ) else: parsed_prompts = cast(Union[PromptType, Sequence[PromptType]], prompts) if pooling_params is None: # Use default pooling params. pooling_params = PoolingParams() self._validate_and_add_requests( prompts=parsed_prompts, params=pooling_params, lora_request=lora_request, prompt_adapter_request=prompt_adapter_request, ) outputs = self._run_engine(use_tqdm=use_tqdm) return self.engine_class.validate_outputs(outputs, EmbeddingRequestOutput) def start_profile(self) -> None: self.llm_engine.start_profile() def stop_profile(self) -> None: self.llm_engine.stop_profile() # LEGACY def _convert_v1_inputs( self, prompts: Optional[Union[str, List[str]]], prompt_token_ids: Optional[Union[List[int], List[List[int]]]], ): # skip_tokenizer_init is now checked in engine if prompts is not None: prompts = [p["content"] for p in parse_and_batch_prompt(prompts)] if prompt_token_ids is not None: prompt_token_ids = [ p["content"] for p in parse_and_batch_prompt(prompt_token_ids) ] num_requests = None if prompts is not None: num_requests = len(prompts) if prompt_token_ids is not None: if (num_requests is not None and num_requests != len(prompt_token_ids)): raise ValueError("The lengths of prompts and prompt_token_ids " "must be the same.") num_requests = len(prompt_token_ids) if num_requests is None: raise ValueError("Either prompts or prompt_token_ids must be " "provided.") parsed_prompts: List[PromptType] = [] for i in range(num_requests): item: PromptType if prompts is not None: item = TextPrompt(prompt=prompts[i]) elif prompt_token_ids is not None: item = TokensPrompt(prompt_token_ids=prompt_token_ids[i]) else: raise AssertionError parsed_prompts.append(item) return parsed_prompts def _validate_and_add_requests( self, prompts: Union[PromptType, Sequence[PromptType]], params: Union[SamplingParams, Sequence[SamplingParams], PoolingParams, Sequence[PoolingParams]], lora_request: Optional[Union[Sequence[LoRARequest], LoRARequest]], prompt_adapter_request: Optional[PromptAdapterRequest], guided_options: Optional[GuidedDecodingRequest] = None, priority: Optional[List[int]] = None, ) -> None: if guided_options is not None: warnings.warn( "guided_options_request is deprecated, use " "SamplingParams.guided_decoding instead", DeprecationWarning, stacklevel=2, ) if isinstance(prompts, (str, dict)): # Convert a single prompt to a list. prompts = [prompts] num_requests = len(prompts) if isinstance(params, list) and len(params) != num_requests: raise ValueError("The lengths of prompts and params " "must be the same.") if isinstance(lora_request, list) and len(lora_request) != num_requests: raise ValueError("The lengths of prompts and lora_request " "must be the same.") for sp in params if isinstance(params, list) else (params, ): if isinstance(sp, SamplingParams): self._add_guided_params(sp, guided_options) # We only care about the final output sp.output_kind = RequestOutputKind.FINAL_ONLY # Add requests to the engine. for i, prompt in enumerate(prompts): self._add_request( prompt, params[i] if isinstance(params, Sequence) else params, lora_request=lora_request[i] if isinstance( lora_request, Sequence) else lora_request, prompt_adapter_request=prompt_adapter_request, priority=priority[i] if priority else 0, ) def _add_request( self, prompt: PromptType, params: Union[SamplingParams, PoolingParams], lora_request: Optional[LoRARequest] = None, prompt_adapter_request: Optional[PromptAdapterRequest] = None, priority: int = 0, ) -> None: request_id = str(next(self.request_counter)) self.llm_engine.add_request( request_id, prompt, params, lora_request=lora_request, prompt_adapter_request=prompt_adapter_request, priority=priority, ) def _add_guided_params( self, params: SamplingParams, guided_options: Optional[GuidedDecodingRequest] = None): if guided_options is None: return params if params.guided_decoding is not None: raise ValueError("Cannot set both guided_options_request and" "params.guided_decoding.") params.guided_decoding = GuidedDecodingParams( json=guided_options.guided_json, regex=guided_options.guided_regex, choice=guided_options.guided_choice, grammar=guided_options.guided_grammar, json_object=guided_options.guided_json_object, backend=guided_options.guided_decoding_backend, whitespace_pattern=guided_options.guided_whitespace_pattern) return params def _run_engine( self, *, use_tqdm: bool ) -> List[Union[RequestOutput, EmbeddingRequestOutput]]: # Initialize tqdm. if use_tqdm: num_requests = self.llm_engine.get_num_unfinished_requests() pbar = tqdm( total=num_requests, desc="Processed prompts", dynamic_ncols=True, postfix=(f"est. speed input: {0:.2f} toks/s, " f"output: {0:.2f} toks/s"), ) # Run the engine. outputs: List[Union[RequestOutput, EmbeddingRequestOutput]] = [] total_in_toks = 0 total_out_toks = 0 while self.llm_engine.has_unfinished_requests(): step_outputs = self.llm_engine.step() for output in step_outputs: if output.finished: outputs.append(output) if use_tqdm: if isinstance(output, RequestOutput): # Calculate tokens only for RequestOutput assert output.prompt_token_ids is not None total_in_toks += len(output.prompt_token_ids) in_spd = total_in_toks / pbar.format_dict["elapsed"] total_out_toks += sum( len(stp.token_ids) for stp in output.outputs) out_spd = (total_out_toks / pbar.format_dict["elapsed"]) pbar.postfix = ( f"est. speed input: {in_spd:.2f} toks/s, " f"output: {out_spd:.2f} toks/s") pbar.update(1) if use_tqdm: pbar.close() # Sort the outputs by request ID. # This is necessary because some requests may be finished earlier than # its previous requests. return sorted(outputs, key=lambda x: int(x.request_id))