Multi-Domain-Expert-Learning/given-mpt-7b
053
1"""A simple, flexible implementation of a GPT model.2 3Inspired by https://github.com/karpathy/minGPT/blob/master/mingpt/model.py4"""5import math6import warnings7from typing import List, Optional, Tuple, Union8import torch9import torch.nn as nn10import torch.nn.functional as F11from transformers import PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast12from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast13from .attention import attn_bias_shape, build_attn_bias14from .blocks import MPTBlock15from .custom_embedding import SharedEmbedding16from .norm import NORM_CLASS_REGISTRY17from .configuration_mpt import MPTConfig18from .adapt_tokenizer import AutoTokenizerForMOD, adapt_tokenizer_for_denoising19from .hf_prefixlm_converter import add_bidirectional_mask_if_missing, convert_hf_causal_lm_to_prefix_lm20from .meta_init_context import init_empty_weights21from .param_init_fns import MODEL_INIT_REGISTRY, generic_param_init_fn_22try:23 from .flash_attn_triton import flash_attn_func24except:25 pass26Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]27 28class MPTPreTrainedModel(PreTrainedModel):29 config_class = MPTConfig30 base_model_prefix = 'model'31 _no_split_modules = ['MPTBlock']32 33class MPTModel(MPTPreTrainedModel):34 35 def __init__(self, config: MPTConfig):36 config._validate_config()37 super().__init__(config)38 self.attn_impl = config.attn_config['attn_impl']39 self.prefix_lm = config.attn_config['prefix_lm']40 self.attn_uses_sequence_id = config.attn_config['attn_uses_sequence_id']41 self.alibi = config.attn_config['alibi']42 self.alibi_bias_max = config.attn_config['alibi_bias_max']43 if config.init_device == 'mixed':44 if dist.get_local_rank() == 0:45 config.init_device = 'cpu'46 else:47 config.init_device = 'meta'48 if config.norm_type.lower() not in NORM_CLASS_REGISTRY.keys():49 norm_options = ' | '.join(NORM_CLASS_REGISTRY.keys())50 raise NotImplementedError(f'Requested norm type ({config.norm_type}) is not implemented within this repo (Options: {norm_options}).')51 norm_class = NORM_CLASS_REGISTRY[config.norm_type.lower()]52 self.embedding_fraction = config.embedding_fraction53 self.wte = SharedEmbedding(config.vocab_size, config.d_model, device=config.init_device)54 if not self.alibi:55 self.wpe = torch.nn.Embedding(config.max_seq_len, config.d_model, device=config.init_device)56 self.emb_drop = nn.Dropout(config.emb_pdrop)57 self.blocks = nn.ModuleList([MPTBlock(device=config.init_device, **config.to_dict()) for _ in range(config.n_layers)])58 self.norm_f = norm_class(config.d_model, device=config.init_device)59 if config.init_device != 'meta':60 print(f'You are using config.init_device={config.init_device!r}, but you can also use config.init_device="meta" with Composer + FSDP for fast initialization.')61 self.apply(self.param_init_fn)62 self.is_causal = not self.prefix_lm63 self._attn_bias_initialized = False64 self.attn_bias = None65 self.attn_bias_shape = attn_bias_shape(self.attn_impl, config.n_heads, config.max_seq_len, self.alibi, prefix_lm=self.prefix_lm, causal=self.is_causal, use_sequence_id=self.attn_uses_sequence_id)66 if config.no_bias:67 for module in self.modules():68 if hasattr(module, 'bias') and isinstance(module.bias, nn.Parameter):69 if config.verbose:70 warnings.warn(f'Removing bias ({module.bias}) from {module}.')71 module.register_parameter('bias', None)72 if config.verbose and config.verbose > 2:73 print(self)74 if 'verbose' not in self.config.init_config:75 self.config.init_config['verbose'] = self.config.verbose76 if self.config.init_config['verbose'] > 1:77 init_fn_name = self.config.init_config['name']78 warnings.warn(f'Using {init_fn_name} initialization.')79 80 def get_input_embeddings(self):81 return self.wte82 83 def set_input_embeddings(self, value):84 self.wte = value85 86 @torch.no_grad()87 def _attn_bias(self, device, dtype, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None):88 if not self._attn_bias_initialized:89 if self.attn_bias_shape:90 self.attn_bias = torch.zeros(self.attn_bias_shape, device=device, dtype=dtype)91 self.attn_bias = build_attn_bias(self.attn_impl, self.attn_bias, self.config.n_heads, self.config.max_seq_len, causal=self.is_causal, alibi=self.alibi, alibi_bias_max=self.alibi_bias_max)92 self._attn_bias_initialized = True93 if self.attn_impl == 'flash':94 return (self.attn_bias, attention_mask)95 if self.attn_bias is not None:96 self.attn_bias = self.attn_bias.to(dtype=dtype, device=device)97 attn_bias = self.attn_bias98 if self.prefix_lm:99 assert isinstance(attn_bias, torch.Tensor)100 assert isinstance(prefix_mask, torch.Tensor)101 attn_bias = self._apply_prefix_mask(attn_bias, prefix_mask)102 if self.attn_uses_sequence_id and sequence_id is not None:103 assert isinstance(attn_bias, torch.Tensor)104 attn_bias = self._apply_sequence_id(attn_bias, sequence_id)105 if attention_mask is not None:106 s_k = attention_mask.shape[-1]107 if attn_bias is None:108 attn_bias = torch.zeros((1, 1, 1, s_k), device=device, dtype=dtype)109 else:110 _s_k = max(0, attn_bias.size(-1) - s_k)111 attn_bias = attn_bias[:, :, :, _s_k:]112 if prefix_mask is not None and attention_mask.shape != prefix_mask.shape:113 raise ValueError(f'attention_mask shape={attention_mask.shape} ' + f'and prefix_mask shape={prefix_mask.shape} are not equal.')114 min_val = torch.finfo(attn_bias.dtype).min115 attn_bias = attn_bias.masked_fill(~attention_mask.view(-1, 1, 1, s_k), min_val)116 return (attn_bias, None)117 118 def _apply_prefix_mask(self, attn_bias: torch.Tensor, prefix_mask: torch.Tensor):119 (s_k, s_q) = attn_bias.shape[-2:]120 if s_k != self.config.max_seq_len or s_q != self.config.max_seq_len:121 raise ValueError('attn_bias does not match the expected shape. ' + f'The last two dimensions should both be {self.config.max_length} ' + f'but are {s_k} and {s_q}.')122 seq_len = prefix_mask.shape[-1]123 if seq_len > self.config.max_seq_len:124 raise ValueError(f'prefix_mask sequence length cannot exceed max_seq_len={self.config.max_seq_len}')125 attn_bias = attn_bias[..., :seq_len, :seq_len]126 causal = torch.tril(torch.ones((seq_len, seq_len), dtype=torch.bool, device=prefix_mask.device)).view(1, 1, seq_len, seq_len)127 prefix = prefix_mask.view(-1, 1, 1, seq_len)128 cannot_attend = ~torch.logical_or(causal, prefix.bool())129 min_val = torch.finfo(attn_bias.dtype).min130 attn_bias = attn_bias.masked_fill(cannot_attend, min_val)131 return attn_bias132 133 def _apply_sequence_id(self, attn_bias: torch.Tensor, sequence_id: torch.LongTensor):134 seq_len = sequence_id.shape[-1]135 if seq_len > self.config.max_seq_len:136 raise ValueError(f'sequence_id sequence length cannot exceed max_seq_len={self.config.max_seq_len}')137 attn_bias = attn_bias[..., :seq_len, :seq_len]138 cannot_attend = torch.logical_not(torch.eq(sequence_id.view(-1, seq_len, 1), sequence_id.view(-1, 1, seq_len))).unsqueeze(1)139 min_val = torch.finfo(attn_bias.dtype).min140 attn_bias = attn_bias.masked_fill(cannot_attend, min_val)141 return attn_bias142 143 def forward(self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple[torch.FloatTensor]]]=None, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None, return_dict: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, use_cache: Optional[bool]=None):144 return_dict = return_dict if return_dict is not None else self.config.return_dict145 use_cache = use_cache if use_cache is not None else self.config.use_cache146 if attention_mask is not None:147 attention_mask = attention_mask.bool()148 if prefix_mask is not None:149 prefix_mask = prefix_mask.bool()150 if not return_dict:151 raise NotImplementedError('return_dict False is not implemented yet for MPT')152 if output_attentions:153 if self.attn_impl != 'torch':154 raise NotImplementedError('output_attentions is not implemented for MPT when using attn_impl `flash` or `triton`.')155 if attention_mask is not None and attention_mask[:, 0].sum() != attention_mask.shape[0] and self.training:156 raise NotImplementedError('MPT does not support training with left padding.')157 if self.prefix_lm and prefix_mask is None:158 raise ValueError('prefix_mask is a required argument when MPT is configured with prefix_lm=True.')159 if self.training:160 if self.attn_uses_sequence_id and sequence_id is None:161 raise ValueError('sequence_id is a required argument when MPT is configured with attn_uses_sequence_id=True ' + 'and the model is in train mode.')162 elif self.attn_uses_sequence_id is False and sequence_id is not None:163 warnings.warn('MPT received non-None input for `sequence_id` but is configured with attn_uses_sequence_id=False. ' + 'This input will be ignored. If you want the model to use `sequence_id`, set attn_uses_sequence_id to True.')164 S = input_ids.size(1)165 assert S <= self.config.max_seq_len, f'Cannot forward input with seq_len={S}, this model only supports seq_len<={self.config.max_seq_len}'166 tok_emb = self.wte(input_ids)167 if self.alibi:168 x = tok_emb169 else:170 past_position = 0171 if past_key_values is not None:172 if len(past_key_values) != self.config.n_layers:173 raise ValueError(f'past_key_values must provide a past_key_value for each attention ' + f'layer in the network (len(past_key_values)={len(past_key_values)!r}; self.config.n_layers={self.config.n_layers!r}).')174 past_position = past_key_values[0][0].size(1)175 if self.attn_impl == 'torch':176 past_position = past_key_values[0][0].size(3)177 if S + past_position > self.config.max_seq_len:178 raise ValueError(f'Cannot forward input with past sequence length {past_position} and current sequence length {S + 1}, this model only supports total sequence length <= {self.config.max_seq_len}.')179 pos = torch.arange(past_position, S + past_position, dtype=torch.long, device=input_ids.device).unsqueeze(0)180 if attention_mask is not None:181 pos = torch.clamp(pos - torch.cumsum((~attention_mask).to(torch.int32), dim=1)[:, past_position:], min=0)182 pos_emb = self.wpe(pos)183 x = tok_emb + pos_emb184 if self.embedding_fraction == 1:185 x = self.emb_drop(x)186 else:187 x_shrunk = x * self.embedding_fraction + x.detach() * (1 - self.embedding_fraction)188 assert isinstance(self.emb_drop, nn.Module)189 x = self.emb_drop(x_shrunk)190 (attn_bias, attention_mask) = self._attn_bias(device=x.device, dtype=torch.float32, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id)191 if use_cache and past_key_values is None:192 past_key_values = [() for _ in range(self.config.n_layers)]193 all_hidden_states = () if output_hidden_states else None194 all_self_attns = () if output_attentions else None195 for (b_idx, block) in enumerate(self.blocks):196 if output_hidden_states:197 assert all_hidden_states is not None198 all_hidden_states = all_hidden_states + (x,)199 past_key_value = past_key_values[b_idx] if past_key_values is not None else None200 (x, attn_weights, past_key_value) = block(x, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=self.is_causal)201 if past_key_values is not None:202 past_key_values[b_idx] = past_key_value203 if output_attentions:204 assert all_self_attns is not None205 all_self_attns = all_self_attns + (attn_weights,)206 x = self.norm_f(x)207 if output_hidden_states:208 assert all_hidden_states is not None209 all_hidden_states = all_hidden_states + (x,)210 return BaseModelOutputWithPast(last_hidden_state=x, past_key_values=past_key_values, hidden_states=all_hidden_states, attentions=all_self_attns)211 212 def param_init_fn(self, module):213 init_fn_name = self.config.init_config['name']214 MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)215 216 def fsdp_wrap_fn(self, module):217 return isinstance(module, MPTBlock)218 219 def activation_checkpointing_fn(self, module):220 return isinstance(module, MPTBlock)221 222class MPTForCausalLM(MPTPreTrainedModel):223 224 def __init__(self, config: MPTConfig):225 super().__init__(config)226 if not config.tie_word_embeddings:227 raise ValueError('MPTForCausalLM only supports tied word embeddings')228 self.transformer = MPTModel(config)229 for child in self.transformer.children():230 if isinstance(child, torch.nn.ModuleList):231 continue232 if isinstance(child, torch.nn.Module):233 child._fsdp_wrap = True234 self.logit_scale = None235 if config.logit_scale is not None:236 logit_scale = config.logit_scale237 if isinstance(logit_scale, str):238 if logit_scale == 'inv_sqrt_d_model':239 logit_scale = 1 / math.sqrt(config.d_model)240 else:241 raise ValueError(f"logit_scale={logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'.")242 self.logit_scale = logit_scale243 244 def get_input_embeddings(self):245 return self.transformer.wte246 247 def set_input_embeddings(self, value):248 self.transformer.wte = value249 250 def get_output_embeddings(self):251 return self.transformer.wte252 253 def set_output_embeddings(self, new_embeddings):254 self.transformer.wte = new_embeddings255 256 def set_decoder(self, decoder):257 self.transformer = decoder258 259 def get_decoder(self):260 return self.transformer261 262 def forward(self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple[torch.FloatTensor]]]=None, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None, labels: Optional[torch.LongTensor]=None, return_dict: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, use_cache: Optional[bool]=None):263 return_dict = return_dict if return_dict is not None else self.config.return_dict264 use_cache = use_cache if use_cache is not None else self.config.use_cache265 outputs = self.transformer(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id, return_dict=return_dict, output_attentions=output_attentions, output_hidden_states=output_hidden_states, use_cache=use_cache)266 logits = self.transformer.wte(outputs.last_hidden_state.to(self.transformer.wte.weight.device), True)267 if self.logit_scale is not None:268 if self.logit_scale == 0:269 warnings.warn(f'Multiplying logits by self.logit_scale={self.logit_scale!r}. This will produce uniform (uninformative) outputs.')270 logits *= self.logit_scale271 loss = None272 if labels is not None:273 labels = torch.roll(labels, shifts=-1)274 labels[:, -1] = -100275 loss = F.cross_entropy(logits.view(-1, logits.size(-1)), labels.to(logits.device).view(-1))276 return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions)277 278 def param_init_fn(self, module):279 init_fn_name = self.config.init_config['name']280 MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)281 282 def fsdp_wrap_fn(self, module):283 return isinstance(module, MPTBlock)284 285 def activation_checkpointing_fn(self, module):286 return isinstance(module, MPTBlock)287 288 def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):289 if inputs_embeds is not None:290 raise NotImplementedError('inputs_embeds is not implemented for MPT yet')291 attention_mask = kwargs['attention_mask'].bool()292 if attention_mask[:, -1].sum() != attention_mask.shape[0]:293 raise NotImplementedError('MPT does not support generation with right padding.')294 if self.transformer.attn_uses_sequence_id and self.training:295 sequence_id = torch.zeros_like(input_ids[:1])296 else:297 sequence_id = None298 if past_key_values is not None:299 input_ids = input_ids[:, -1].unsqueeze(-1)300 if self.transformer.prefix_lm:301 prefix_mask = torch.ones_like(attention_mask)302 if kwargs.get('use_cache') == False:303 raise NotImplementedError('MPT with prefix_lm=True does not support use_cache=False.')304 else:305 prefix_mask = None306 return {'input_ids': input_ids, 'attention_mask': attention_mask, 'prefix_mask': prefix_mask, 'sequence_id': sequence_id, 'past_key_values': past_key_values, 'use_cache': kwargs.get('use_cache', True)}307 308 @staticmethod309 def _reorder_cache(past_key_values, beam_idx):310 """Used by HuggingFace generate when using beam search with kv-caching.311 312 See https://github.com/huggingface/transformers/blob/3ec7a47664ebe40c40f4b722f6bb1cd30c3821ec/src/transformers/models/gpt2/modeling_gpt2.py#L1122-L1133313 for an example in transformers.314 """315 reordered_past = []316 for layer_past in past_key_values:317 reordered_past += [tuple((past_state.index_select(0, beam_idx) for past_state in layer_past))]318 return reordered_past