import os import sys import os.path as osp import torch from ..smp import * def get_gpu_num(model_name): model_name = model_name.lower() kws = { 8: ['65b', '70b'], 4: ['30b', '33b', '35b', '40b'], 2: ['13b', '14b', '20b', '8b'], 1: ['6b', '7b', 'moss'], } for k in [8, 4, 2, 1]: for keyword in kws[k]: if keyword in model_name: return k return 8 validated_llms = [ 'internlm/internlm-chat-7b', 'internlm/internlm-chat-7b-8k', 'internlm/internlm-chat-20b', 'Qwen/Qwen-7B-Chat', 'Qwen/Qwen-14B-Chat', 'THUDM/chatglm2-6b', 'THUDM/chatglm2-6b-32k', 'THUDM/chatglm3-6b', 'THUDM/chatglm3-6b-32k', 'baichuan-inc/Baichuan2-7B-Chat', 'baichuan-inc/Baichuan2-13B-Chat', 'lmsys/vicuna-7b-v1.5', 'lmsys/vicuna-13b-v1.5', 'meta-llama/Llama-2-7b-chat-hf', 'meta-llama/Llama-3.1-8B-Instruct' ] Auto_model = ['chatglm'] class HFChatModel: def _get_context_length(self, model, model_path): # By default, we use model.config.seq_length model_path = model_path.lower() if 'baichuan' in model_path: context_window = model.config.model_max_length elif 'internlm' in model_path or 'llama' in model_path: context_window = model.config.max_position_embeddings elif 'vicuna' in model_path: context_window = model.generation_config.max_length else: # chatglm & qwen context_window = model.config.seq_length return context_window def _get_context_length_robust(self, model, model_path): try: context_window = self._get_context_length(model, model_path) return context_window except Exception as err: self.logger.critical(f'{type(err)}: {err}') self.logger.critical( 'Failed to extract context_window information from config / generation_config. ' 'Please read the above code and check if the logic works for you model path' ) raise NotImplementedError def __init__(self, model_path, system_prompt: str = None, **kwargs): self.logger = get_logger('HFChatModel') if 'vicuna' in model_path.lower() or 'llama' in model_path.lower(): try: from fastchat.model import get_conversation_template except Exception as err: self.logger.critical('Please install fastchat first to use vicuna. ') raise err self.explicit_device = kwargs.pop('device', None) if self.explicit_device is None: # If CUDA_VISIBLE_DEVICES is not properly set if 'CUDA_VISIBLE_DEVICES' not in os.environ or os.environ['CUDA_VISIBLE_DEVICES'] == '0,1,2,3,4,5,6,7': num_gpu = get_gpu_num(model_path) gpu_offset = kwargs.pop('gpu_offset', 0) cuda_visible_devices = ','.join([str(i) for i in range(gpu_offset, gpu_offset + num_gpu)]) os.environ['CUDA_VISIBLE_DEVICES'] = cuda_visible_devices from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModel from transformers.generation import GenerationConfig if model_path not in validated_llms: self.logger.warning(f'{model_path} not in validated LLMs, may have inference troubles. ') self.model_path = model_path if listinstr(Auto_model, model_path): LoadModel = AutoModel else: LoadModel = AutoModelForCausalLM assert osp.exists(model_path) or len(model_path.split('/')) == 2 device = self.explicit_device if self.explicit_device else 'auto' precision = {} if 'internlm-chat-7b' in model_path: precision = {'torch_dtype': torch.float16} elif 'internlm-chat-20b' in model_path: precision = {'torch_dtype': torch.bfloat16} self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) cuda_devices = os.environ.get('CUDA_VISIBLE_DEVICES', '0') if ',' in cuda_devices: device_ids = [int(x) for x in cuda_devices.split(',')] device_map = {i: i for i in range(len(device_ids))} else: device_map = {'': 0} if 'llama' in self.model_path.lower(): from lmdeploy import pipeline, GenerationConfig, TurbomindEngineConfig print(f"Loading model {model_path} with {num_gpu} GPUs") backend_config = TurbomindEngineConfig(tp=num_gpu) self.gen_config = GenerationConfig(max_new_tokens=256) model = pipeline(model_path, backend_config=backend_config) else: model = LoadModel.from_pretrained(model_path, trust_remote_code=True, device_map='cpu', **precision) model = model.eval() if device != 'cpu': model = model.to(f'cuda:{device}' if isinstance(device, int) else 'cuda') try: model.generation_config = GenerationConfig.from_pretrained( model_path, trust_remote_code=True, device_map=device) except Exception as err: self.logger.warning(f'{type(err)}: {err}') self.context_length = self._get_context_length_robust(model=model, model_path=model_path) torch.cuda.empty_cache() self.model = model self.answer_buffer = 192 self.system_prompt = system_prompt for k, v in kwargs.items(): self.logger.info(f'Following args will be used for generation (If not set specifically), {k}: {v}. ') self.kwargs = kwargs def generate_str(self, input, **kwargs): if 'baichuan' in self.model_path.lower(): messages = [] messages.append({'role': 'user', 'content': input}) resp = self.model.chat(self.tokenizer, messages, **kwargs) elif 'vicuna' in self.model_path.lower(): from fastchat.model import get_conversation_template conv = get_conversation_template('vicuna') conv.append_message(conv.roles[0], input) conv.append_message(conv.roles[1], None) prompt = conv.get_prompt() inputs = self.tokenizer([prompt], return_tensors='pt') if torch.cuda.is_available(): for k in inputs: inputs[k] = inputs[k].cuda() params = dict(do_sample=True, temperature=0.7, repetition_penalty=1.0, max_new_tokens=512) params.update(self.kwargs) params.update(kwargs) outputs = self.model.generate(**inputs, **params) resp = self.tokenizer.decode( outputs[0][len(inputs['input_ids'][0]):], skip_special_tokens=True, spaces_between_special_tokens=False) elif 'llama' in self.model_path.lower(): prompt = [{'role': 'system', 'content': self.system_prompt}, {'role': 'user', 'content': input}] resp = self.model(prompt, gen_config=self.gen_config).text else: params = self.kwargs params.update(kwargs) resp, _ = self.model.chat(self.tokenizer, input, history=[], **params) return resp def length_ok(self, inputs): tot = len(self.tokenizer.encode(self.system_prompt)) if self.system_prompt is not None else 0 for s in inputs: tot += len(self.tokenizer.encode(s)) return tot + self.answer_buffer < self.context_length def generate_list(self, full_inputs, offset=0, **kwargs): assert isinstance(full_inputs, list) inputs = full_inputs[offset:] if not self.length_ok(inputs): return self.chat(full_inputs, offset + 1) model_path = self.model_path.lower() if sum([x in model_path for x in ['baichuan']]): input_msgs = [] if self.system_prompt is not None: input_msgs.append(dict(role='user', content=self.system_prompt)) if len(inputs): assert isinstance(inputs, list) and isinstance(inputs[0], str) roles = ['user', 'assistant'] if len(inputs) % 2 == 1 else ['assistant', 'user'] roles = roles * len(inputs) for role, msg in zip(roles, inputs): input_msgs.append(dict(role=role, content=msg)) response = self.model.chat(self.tokenizer, input_msgs) elif sum([x in model_path for x in ['vicuna']]): from fastchat.model import get_conversation_template conv = get_conversation_template('vicuna') assert isinstance(inputs, list) and isinstance(inputs[0], str) if len(inputs) % 2 == 1: if self.system_prompt is not None: conv.append_message(conv.roles[0], self.system_prompt) for i in range(len(inputs) // 2): conv.append_message(conv.roles[0], inputs[2 * i]) conv.append_message(conv.roles[1], inputs[2 * i + 1]) else: assert self.system_prompt is not None conv.append_message(conv.roles[0], self.system_prompt) conv.append_message(conv.roles[1], inputs[0]) for i in range(len(inputs) // 2 - 1): conv.append_message(conv.roles[0], inputs[2 * i + 1]) conv.append_message(conv.roles[1], inputs[2 * i + 2]) conv.append_message(conv.roles[0], inputs[-1]) conv.append_message(conv.roles[1], None) prompt = conv.get_prompt() inputs = self.tokenizer([prompt], return_tensors='pt') if torch.cuda.is_available(): for k in inputs: inputs[k] = inputs[k].cuda() params = dict(do_sample=True, temperature=0.7, repetition_penalty=1.0, max_new_tokens=512) params.update(self.kwargs) params.update(kwargs) outputs = self.model.generate(**inputs, **params) response = self.tokenizer.decode( outputs[0][len(inputs['input_ids'][0]):], skip_special_tokens=True, spaces_between_special_tokens=False) response = response.lstrip('\n') else: # The default option, support internlm, chatglm, qwen history, msg = [], None if len(inputs) % 2 == 1: if self.system_prompt is not None: history = [(self.system_prompt, '')] for i in range(len(inputs) // 2): history.append((inputs[2 * i], inputs[2 * i + 1])) else: assert self.system_prompt is not None history = [(self.system_prompt, inputs[0])] for i in range(len(inputs) // 2 - 1): history.append((inputs[2 * i + 1], inputs[2 * i + 2])) msg = inputs[-1] params = self.kwargs params.update(kwargs) response, _ = self.model.chat(self.tokenizer, msg, history=history, **params) return response, offset def generate(self, inputs, **kwargs): if isinstance(inputs, str): return self.generate_str(inputs, **kwargs) elif isinstance(inputs, list): return self.generate_list(inputs, **kwargs)