import os import numpy as np import argparse import gradio as gr from typing import Any, Iterator from typing import Iterator, List, Optional, Tuple import filelock import glob import json import time from gradio.routes import Request from gradio.utils import SyncToAsyncIterator, async_iteration from gradio.helpers import special_args import anyio from typing import AsyncGenerator, Callable, Literal, Union, cast from gradio_client.documentation import document, set_documentation_group from typing import List, Optional, Union, Dict, Tuple from tqdm.auto import tqdm from huggingface_hub import snapshot_download from gradio.components import Button from gradio.events import Dependency, EventListenerMethod from .base_engine import BaseEngine # @@ environments ================ from ..configs import ( DTYPE, TENSOR_PARALLEL, MODEL_PATH, QUANTIZATION, MAX_TOKENS, TEMPERATURE, FREQUENCE_PENALTY, PRESENCE_PENALTY, GPU_MEMORY_UTILIZATION, STREAM_CHECK_MULTIPLE, STREAM_YIELD_MULTIPLE, ) llm = None demo = None def vllm_abort(self): sh = self.llm_engine.scheduler for g in (sh.waiting + sh.running + sh.swapped): sh.abort_seq_group(g.request_id) from vllm.sequence import SequenceStatus scheduler = self.llm_engine.scheduler for state_queue in [scheduler.waiting, scheduler.running, scheduler.swapped]: for seq_group in state_queue: # if seq_group.request_id == request_id: # Remove the sequence group from the state queue. state_queue.remove(seq_group) for seq in seq_group.seqs: if seq.is_finished(): continue scheduler.free_seq(seq, SequenceStatus.FINISHED_ABORTED) def _vllm_run_engine(self: Any, use_tqdm: bool = False) -> Dict[str, Any]: from vllm.outputs import RequestOutput # Initialize tqdm. if use_tqdm: num_requests = self.llm_engine.get_num_unfinished_requests() pbar = tqdm(total=num_requests, desc="Processed prompts") # Run the engine. outputs: Dict[str, RequestOutput] = {} while self.llm_engine.has_unfinished_requests(): step_outputs = self.llm_engine.step() for output in step_outputs: outputs[output.request_id] = output if len(outputs) > 0: yield outputs def vllm_generate_stream( self: Any, prompts: Optional[Union[str, List[str]]] = None, sampling_params: Optional[Any] = None, prompt_token_ids: Optional[List[List[int]]] = None, use_tqdm: bool = False, ) -> Dict[str, Any]: """Generates the completions for the input prompts. NOTE: 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: A list of prompts to generate completions for. sampling_params: The sampling parameters for text generation. If None, we use the default sampling parameters. prompt_token_ids: A list of token IDs for the prompts. If None, we use the tokenizer to convert the prompts to token IDs. use_tqdm: Whether to use tqdm to display the progress bar. Returns: A list of `RequestOutput` objects containing the generated completions in the same order as the input prompts. """ from vllm import LLM, SamplingParams if prompts is None and prompt_token_ids is None: raise ValueError("Either prompts or prompt_token_ids must be " "provided.") if isinstance(prompts, str): # Convert a single prompt to a list. prompts = [prompts] if prompts is not None and prompt_token_ids is not None: if len(prompts) != len(prompt_token_ids): raise ValueError("The lengths of prompts and prompt_token_ids " "must be the same.") if sampling_params is None: # Use default sampling params. sampling_params = SamplingParams() # Add requests to the engine. if prompts is not None: num_requests = len(prompts) else: num_requests = len(prompt_token_ids) for i in range(num_requests): prompt = prompts[i] if prompts is not None else None if prompt_token_ids is None: token_ids = None else: token_ids = prompt_token_ids[i] self._add_request(prompt, sampling_params, token_ids) # return self._run_engine(use_tqdm) yield from _vllm_run_engine(self, use_tqdm) class VllmEngine(BaseEngine): def __init__(self, **kwargs) -> None: super().__init__(**kwargs) @property def tokenizer(self): return self._model.get_tokenizer() def load_model(self, ): import torch try: compute_capability = torch.cuda.get_device_capability() print(f'Torch CUDA compute_capability: {compute_capability}') except Exception as e: print(f'Failed to print compute_capability version: {e}') import vllm from vllm import LLM print(f'VLLM: {vllm.__version__=}') if QUANTIZATION == 'awq': print(F'Load model in int4 quantization') llm = LLM( model=MODEL_PATH, dtype="float16", tensor_parallel_size=TENSOR_PARALLEL, gpu_memory_utilization=GPU_MEMORY_UTILIZATION, quantization="awq", max_model_len=MAX_TOKENS ) else: llm = LLM( model=MODEL_PATH, dtype=DTYPE, tensor_parallel_size=TENSOR_PARALLEL, gpu_memory_utilization=GPU_MEMORY_UTILIZATION, max_model_len=MAX_TOKENS ) try: print(llm.llm_engine.workers[0].model) except Exception as e: print(f'Cannot print model worker: {e}') try: llm.llm_engine.scheduler_config.max_model_len = MAX_TOKENS llm.llm_engine.scheduler_config.max_num_batched_tokens = MAX_TOKENS except Exception as e: print(f'Cannot set parameters: {e}') self._model = llm def generate_yield_string(self, prompt, temperature, max_tokens, stop_strings: Optional[Tuple[str]] = None, **kwargs): from vllm import SamplingParams # ! must abort previous ones vllm_abort(llm) sampling_params = SamplingParams( temperature=temperature, max_tokens=max_tokens, # frequency_penalty=frequency_penalty, # presence_penalty=presence_penalty, stop=stop_strings, ) cur_out = None num_tokens = len(self.tokenizer.encode(prompt)) for j, gen in enumerate(vllm_generate_stream(llm, prompt, sampling_params)): if cur_out is not None and (STREAM_YIELD_MULTIPLE < 1 or j % STREAM_YIELD_MULTIPLE == 0) and j > 0: yield cur_out, num_tokens assert len(gen) == 1, f'{gen}' item = next(iter(gen.values())) cur_out = item.outputs[0].text if cur_out is not None: full_text = prompt + cur_out num_tokens = len(self.tokenizer.encode(full_text)) yield cur_out, num_tokens def batch_generate(self, prompts, temperature, max_tokens, stop_strings: Optional[Tuple[str]] = None, **kwargs): """ Only vllm should support this, the other engines is only batch=1 only """ from vllm import SamplingParams # ! must abort previous ones vllm_abort(llm) sampling_params = SamplingParams( temperature=temperature, max_tokens=max_tokens, # frequency_penalty=frequency_penalty, # presence_penalty=presence_penalty, stop=stop_strings, ) generated = llm.generate(prompts, sampling_params, use_tqdm=False) responses = [g.outputs[0].text for g in generated] return responses