from typing import Optional, Tuple, Union, List from transformers.models.mpt.modeling_mpt import MptBlock, build_mpt_alibi_tensor from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask import torch import torch.nn as nn from torch.nn import CrossEntropyLoss, LayerNorm from transformers.models.mpt.modeling_mpt import MptBlock, build_mpt_alibi_tensor from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions, CausalLMOutputWithPast, \ BaseModelOutputWithPastAndCrossAttentions, BaseModelOutputWithPast # from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig, MptForCausalLM, MptModel from transformers import PreTrainedTokenizerFast import os import torch.nn.functional as F from mpt_7b.modeling_mpt import MPTModel, MPTForCausalLM, gen_attention_mask_in_length from mpt_7b.configuration_mpt import MPTConfig from mpt_7b.blocks import MPTBlock from mpt_7b.norm import NORM_CLASS_REGISTRY from mpt_7b.custom_embedding import SharedEmbedding from mpt_7b.attention import ATTN_CLASS_REGISTRY, attn_bias_shape, build_attn_bias, gen_slopes import logging log = logging.getLogger(__name__) class Custom_MPTConfig(MPTConfig): def __init__(self): super().__init__() class Custom_MptModel(MPTModel): # MptModel def __init__(self, config: MPTConfig, modality0_dim=128, modality2_dim=1536): config._validate_config() super().__init__(config) self.attn_impl = config.attn_config['attn_impl'] self.prefix_lm = config.attn_config['prefix_lm'] self.attn_uses_sequence_id = config.attn_config['attn_uses_sequence_id'] self.alibi = config.attn_config['alibi'] self.alibi_bias_max = config.attn_config['alibi_bias_max'] self.learned_pos_emb = config.learned_pos_emb if config.init_device == 'mixed': if dist.get_local_rank() == 0: config.init_device = 'cpu' else: config.init_device = 'meta' if config.norm_type.lower() not in NORM_CLASS_REGISTRY.keys(): norm_options = ' | '.join(NORM_CLASS_REGISTRY.keys()) raise NotImplementedError(f'Requested norm type ({config.norm_type}) is not implemented within this repo (Options: {norm_options}).') norm_class = NORM_CLASS_REGISTRY[config.norm_type.lower()] self.embedding_fraction = config.embedding_fraction self.wte = SharedEmbedding(config.vocab_size, config.d_model, device=config.init_device) if self.learned_pos_emb: self.wpe = torch.nn.Embedding(config.max_seq_len, config.d_model, device=config.init_device) self.emb_drop = nn.Dropout(config.emb_pdrop) self.blocks = nn.ModuleList([MPTBlock(device=config.init_device, **config.to_dict()) for _ in range(config.n_layers)]) self.norm_f = norm_class(config.d_model, device=config.init_device) ### Added for P3GPT - START # Freeze all parameters except the projection layer for param in self.wte.parameters(): param.requires_grad = False for param in self.blocks.parameters(): param.requires_grad = False # Add a projection layer for the custom embedding # torch.set_default_dtype(torch.bfloat16) self.modality0_embedding_projection = nn.ModuleList([nn.Linear(modality0_dim, config.d_model), # nn.BatchNorm1d(config.d_model), nn.ReLU(), nn.Linear(config.d_model, config.d_model), # nn.BatchNorm1d(config.d_model), nn.ReLU(), nn.Linear(config.d_model, config.d_model)])# nn.Linear(modality0_dim, self.hidden_size) self.modality2_embedding_projection = nn.ModuleList([nn.Linear(modality2_dim, config.d_model), # nn.BatchNorm1d(config.d_model), nn.ReLU(), nn.Linear(config.d_model, config.d_model), # nn.BatchNorm1d(config.d_model), nn.ReLU(), nn.Linear(config.d_model, config.d_model)])# nn.Linear(modality0_dim, self.hidden_size) ### Added for P3GPT - FINISH self.rope = config.attn_config['rope'] self.rope_impl = None if self.rope: self.rope_impl = config.attn_config['rope_impl'] self.rotary_embedding = gen_rotary_embedding(rope_head_dim=config.d_model // config.n_heads, rope_impl=self.rope_impl, rope_theta=config.attn_config['rope_theta'], rope_dail_config=config.attn_config['rope_dail_config'], rope_hf_config=config.attn_config['rope_hf_config'], max_seq_len=self.config.max_seq_len) if config.init_device != 'meta': log.info(f'We recommend using config.init_device="meta" with Composer + FSDP for faster initialization.') self.apply(self.param_init_fn) self.is_causal = not self.prefix_lm self._attn_bias_initialized = False self.attn_bias = None 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) if config.no_bias: for module in self.modules(): if hasattr(module, 'bias') and isinstance(module.bias, nn.Parameter): log.info(f'Removing bias from module={module!r}.') module.register_parameter('bias', None) if hasattr(module, 'use_bias'): log.info(f'Setting use_bias=False for module={module!r}.') module.use_bias = False log.debug(self) log.debug(f"Using {self.config.init_config['name']} initialization.") # Initialize weights and apply final processing # self.post_init() def get_input_embeddings(self): return self.wte def set_input_embeddings(self, new_embeddings): # self.wte = new_embeddings self.wte.weight = new_embeddings def forward(self, input_ids: Optional[torch.LongTensor]=None, 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, inputs_embeds: Optional[torch.Tensor]=None, modality0_emb: Optional[bool] = None, modality0_token_id: Optional[bool] = None, modality1_emb: Optional[bool] = None, modality1_token_id: Optional[bool] = None, modality2_emb: Optional[bool] = None, modality2_token_id: Optional[bool] = None, modality3_emb: Optional[bool] = None, modality3_token_id: Optional[bool] = None,) -> BaseModelOutputWithPast: return_dict = return_dict if return_dict is not None else self.config.return_dict use_cache = use_cache if use_cache is not None else self.config.use_cache if attention_mask is not None: attention_mask = attention_mask.bool() if prefix_mask is not None: prefix_mask = prefix_mask.bool() if not return_dict: raise NotImplementedError('return_dict False is not implemented yet for MPT') if output_attentions: if self.attn_impl != 'torch': raise NotImplementedError('output_attentions is not implemented for MPT when using attn_impl `flash` or `triton`.') if self.training and attention_mask is not None and (attention_mask[:, 0].sum() != attention_mask.shape[0]): raise NotImplementedError('MPT does not support training with left padding.') if self.prefix_lm and prefix_mask is None: raise ValueError('prefix_mask is a required argument when MPT is configured with prefix_lm=True.') if self.training: if self.attn_uses_sequence_id and sequence_id is None: 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.') elif self.attn_uses_sequence_id is False and sequence_id is not None: 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.') ### ADDED FOR P3 - START if modality0_emb is not None: modality0_emb = torch.tensor(modality0_emb, dtype=torch.bfloat16) hidden_states = self.wte.weight.detach() for layer in self.modality0_embedding_projection: modality0_emb = layer(modality0_emb) proj_modality0_emb = modality0_emb # Replace the original embedding for the custom token with the custom embedding hidden_states[modality0_token_id, :] = torch.mean(torch.squeeze(proj_modality0_emb, 1), dim=0) self.set_input_embeddings(torch.nn.Parameter(hidden_states)) if modality1_emb is not None: modality1_emb = torch.tensor(modality1_emb, dtype=torch.bfloat16) hidden_states = self.wte.weight.detach() for layer in self.modality0_embedding_projection: modality1_emb = layer(modality1_emb) proj_modality1_emb = modality1_emb # Replace the original embedding for the custom token with the custom embedding hidden_states[modality1_token_id, :] = torch.mean(torch.squeeze(proj_modality1_emb, 1), dim=0) self.set_input_embeddings(torch.nn.Parameter(hidden_states)) if modality2_emb is not None: modality2_emb = torch.tensor(modality2_emb, dtype=torch.bfloat16) hidden_states = self.wte.weight.detach() for layer in self.modality2_embedding_projection: modality2_emb = layer(modality2_emb) proj_modality2_emb = modality2_emb # Replace the original embedding for the custom token with the custom embedding hidden_states[modality2_token_id, :] = torch.mean(torch.squeeze(proj_modality2_emb, 1), dim=0) self.set_input_embeddings(torch.nn.Parameter(hidden_states)) if modality3_emb is not None: modality3_emb = torch.tensor(modality3_emb, dtype=torch.bfloat16) hidden_states = self.wte.weight.detach() for layer in self.modality2_embedding_projection: modality3_emb = layer(modality3_emb) proj_modality3_emb = modality3_emb # Replace the original embedding for the custom token with the custom embedding hidden_states[modality3_token_id, :] = torch.mean(torch.squeeze(proj_modality3_emb, 1), dim=0) self.set_input_embeddings(torch.nn.Parameter(hidden_states)) ### ADDED FOR P3 - END if input_ids is not None and inputs_embeds is not None: raise ValueError('You cannot specify both input_ids and inputs_embeds.') elif input_ids is not None: bsz = input_ids.size(0) S = input_ids.size(1) x = self.wte(input_ids) input_device = input_ids.device elif inputs_embeds is not None: bsz = inputs_embeds.size(0) S = inputs_embeds.size(1) x = inputs_embeds input_device = inputs_embeds.device else: raise ValueError('You must specify input_ids or inputs_embeds') 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}' rotary_emb_w_meta_info = None past_position = 0 if past_key_values is not None: if len(past_key_values) != self.config.n_layers: 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}).') past_position = past_key_values[0][0].size(1) if self.attn_impl == 'torch': past_position = past_key_values[0][0].size(3) if self.learned_pos_emb or self.rope: if self.learned_pos_emb and S + past_position > self.config.max_seq_len: raise ValueError(f'Cannot forward input with past sequence length {past_position} and current sequence length ' + f'{S + 1}, this model only supports total sequence length <= {self.config.max_seq_len}.') if self.learned_pos_emb or (self.rope and self.rope_impl == 'hf'): pos = torch.arange(past_position, S + past_position, dtype=torch.long, device=input_device).unsqueeze(0) if attention_mask is not None: pos = torch.clamp(pos - torch.cumsum((~attention_mask).to(torch.int32), dim=1)[:, past_position:], min=0) if self.learned_pos_emb: x = x + self.wpe(pos) elif self.rope and self.rope_impl == 'hf': rotary_emb_w_meta_info = {'impl': self.rope_impl, 'rotary_emb': self.rotary_embedding, 'offset_info': pos, 'seq_len': S + past_position} elif self.rope and self.rope_impl == 'dail': rotary_emb_w_meta_info = {'impl': self.rope_impl, 'rotary_emb': self.rotary_embedding, 'offset_info': past_position, 'seq_len': S + past_position} if self.embedding_fraction == 1: x = self.emb_drop(x) else: x_shrunk = x * self.embedding_fraction + x.detach() * (1 - self.embedding_fraction) assert isinstance(self.emb_drop, nn.Module) x = self.emb_drop(x_shrunk) (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) attention_mask_in_length = gen_attention_mask_in_length(sequence_id=sequence_id, S=S, attn_uses_sequence_id=self.attn_uses_sequence_id, attn_impl=self.attn_impl, attention_mask=attention_mask) alibi_slopes = None if self.alibi and self.attn_impl == 'flash': alibi_slopes = gen_slopes(n_heads=self.config.n_heads, alibi_bias_max=self.alibi_bias_max, device=x.device, return_1d=True) presents = () if use_cache else None if use_cache and past_key_values is None: past_key_values = [() for _ in range(self.config.n_layers)] all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None flash_attn_padding_info = {} if self.attn_impl == 'flash': flash_attn_padding_info = gen_flash_attn_padding_info(bsz, S, past_position, x.device, attention_mask_in_length, attention_mask) for (b_idx, block) in enumerate(self.blocks): if output_hidden_states: assert all_hidden_states is not None all_hidden_states = all_hidden_states + (x,) past_key_value = past_key_values[b_idx] if past_key_values is not None else None (x, attn_weights, present) = block(x, past_key_value=past_key_value, attn_bias=attn_bias, rotary_emb_w_meta_info=rotary_emb_w_meta_info, attention_mask=attention_mask, is_causal=self.is_causal, output_attentions=bool(output_attentions), alibi_slopes=alibi_slopes, flash_attn_padding_info=flash_attn_padding_info) if presents is not None: presents += (present,) if output_attentions: assert all_self_attns is not None all_self_attns = all_self_attns + (attn_weights,) x = self.norm_f(x) if output_hidden_states: assert all_hidden_states is not None all_hidden_states = all_hidden_states + (x,) return BaseModelOutputWithPast(last_hidden_state=x, past_key_values=presents, hidden_states=all_hidden_states, attentions=all_self_attns) class Custom_MPTForCausalLM(MPTForCausalLM): def __init__(self, config: MPTConfig): super().__init__(config) # log.info(f'Instantiating an MPTForCausalLM model from {__file__}') self.transformer: MPTModel = Custom_MptModel(config) self.lm_head = None if not config.tie_word_embeddings: self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False, device=config.init_device) self.lm_head._fsdp_wrap = True for child in self.transformer.children(): if isinstance(child, torch.nn.ModuleList): continue if isinstance(child, torch.nn.Module): child._fsdp_wrap = True self.logit_scale = None if config.logit_scale is not None: logit_scale = config.logit_scale if isinstance(logit_scale, str): if logit_scale == 'inv_sqrt_d_model': logit_scale = 1 / math.sqrt(config.d_model) else: raise ValueError(f"logit_scale={logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'.") self.logit_scale = logit_scale def forward(self, input_ids: Optional[torch.LongTensor]=None, 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, inputs_embeds: Optional[torch.FloatTensor]=None, modality0_emb: Optional[bool] = None, modality0_token_id: Optional[bool] = None, modality1_emb: Optional[bool] = None, modality1_token_id: Optional[bool] = None, modality2_emb: Optional[bool] = None, modality2_token_id: Optional[bool] = None, modality3_emb: Optional[bool] = None, modality3_token_id: Optional[bool] = None) -> CausalLMOutputWithPast: return_dict = return_dict if return_dict is not None else self.config.return_dict use_cache = use_cache if use_cache is not None else self.config.use_cache 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, inputs_embeds=inputs_embeds, modality0_emb=modality0_emb, modality0_token_id=modality0_token_id, modality1_emb=modality1_emb, modality1_token_id=modality1_token_id, modality2_emb=modality2_emb, modality2_token_id=modality2_token_id, modality3_emb=modality3_emb, modality3_token_id=modality3_token_id ) if self.lm_head is not None: logits = self.lm_head(outputs.last_hidden_state) else: out = outputs.last_hidden_state out = out.to(self.transformer.wte.weight.device) logits = self.transformer.wte(out, True) if self.logit_scale is not None: if self.logit_scale == 0: warnings.warn(f'Multiplying logits by self.logit_scale={self.logit_scale!r}. This will produce uniform (uninformative) outputs.') logits *= self.logit_scale loss = None if labels is not None: _labels = torch.roll(labels, shifts=-1) _labels[:, -1] = -100 loss = F.cross_entropy(logits.view(-1, logits.size(-1)), _labels.to(logits.device).view(-1)) return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions)