# This script provides an implementation of GPT2 based mixture-of-experts model. # Most of its functionality is copied from existing GPT2 implementation on huggingface: https://huggingface.co/docs/transformers/v4.20.1/en/model_doc/gpt2 # MoE layers are inspired by Mixtral: https://huggingface.co/docs/transformers/v4.39.1/en/model_doc/mixtral # There are however, slight differences in this implementation to adapt it to behave like DeepSpeed Megatron's GPT2 MoE: https://github.com/microsoft/Megatron-DeepSpeed/blob/main/examples_deepspeed/MoE/ds_pretrain_gpt_1.3B_MoE128.sh # Please note: Most of the the features from DeepSpeed Megatron's GPT MoE are **not** implemented here. import warnings from typing import Optional, Tuple, Union from .configuration_lola_gpt2 import LOLAConfig import torch import torch.utils.checkpoint from torch import nn import torch.nn.functional as F from torch.nn import CrossEntropyLoss from transformers.modeling_outputs import ( BaseModelOutputWithPastAndCrossAttentions, SequenceClassifierOutputWithPast, QuestionAnsweringModelOutput ) from transformers.modeling_utils import SequenceSummary from transformers.pytorch_utils import Conv1D from transformers.utils import ( logging ) from transformers.utils.model_parallel_utils import assert_device_map, get_device_map from transformers.models.gpt2.modeling_gpt2 import GPT2Attention, GPT2MLP, GPT2Block, GPT2PreTrainedModel from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel, GPT2DoubleHeadsModel, GPT2ForSequenceClassification, GPT2ForTokenClassification logger = logging.get_logger(__name__) # LOLA class LOLAModel(GPT2PreTrainedModel): config_class = LOLAConfig def __init__(self, config): super().__init__(config) self.embed_dim = config.hidden_size self.wte = nn.Embedding(config.vocab_size, self.embed_dim) self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim) self.drop = nn.Dropout(config.embd_pdrop) self.h = nn.ModuleList([ GPT2Block(config, layer_idx=i) if i % 2 == 0 else LOLABlock(config, layer_idx=i) for i in range(config.num_hidden_layers) ]) self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon) # Model parallel self.model_parallel = False self.device_map = None self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def parallelize(self, device_map=None): # Check validity of device_map warnings.warn( "`GPT2Model.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your" " model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own" " `device_map` but it needs to be a dictionary module_name to device, so for instance {'h.0': 0, 'h.1': 1," " ...}", FutureWarning, ) self.device_map = ( get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map ) assert_device_map(self.device_map, len(self.h)) self.model_parallel = True self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys())) self.last_device = "cuda:" + str(max(self.device_map.keys())) self.wte = self.wte.to(self.first_device) self.wpe = self.wpe.to(self.first_device) # Load onto devices for k, v in self.device_map.items(): for block in v: cuda_device = "cuda:" + str(k) self.h[block] = self.h[block].to(cuda_device) # ln_f to last self.ln_f = self.ln_f.to(self.last_device) def deparallelize(self): warnings.warn( "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.", FutureWarning, ) self.model_parallel = False self.device_map = None self.first_device = "cpu" self.last_device = "cpu" self.wte = self.wte.to("cpu") self.wpe = self.wpe.to("cpu") for index in range(len(self.h)): self.h[index] = self.h[index].to("cpu") self.ln_f = self.ln_f.to("cpu") torch.cuda.empty_cache() def get_input_embeddings(self): return self.wte def set_input_embeddings(self, new_embeddings): self.wte = new_embeddings def _prune_heads(self, heads_to_prune): """ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} """ for layer, heads in heads_to_prune.items(): self.h[layer].attn.prune_heads(heads) def forward( self, input_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, attention_mask: Optional[torch.FloatTensor] = None, token_type_ids: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, head_mask: Optional[torch.FloatTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, encoder_hidden_states: Optional[torch.Tensor] = None, encoder_attention_mask: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict if input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") elif input_ids is not None: # self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask) input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) batch_size = input_ids.shape[0] elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] batch_size = inputs_embeds.shape[0] else: raise ValueError("You have to specify either input_ids or inputs_embeds") device = input_ids.device if input_ids is not None else inputs_embeds.device if token_type_ids is not None: token_type_ids = token_type_ids.view(-1, input_shape[-1]) if past_key_values is None: past_length = 0 past_key_values = tuple([None] * len(self.h)) else: past_length = past_key_values[0][0].size(-2) if position_ids is None: position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device) position_ids = position_ids.unsqueeze(0) # GPT2Attention mask. if attention_mask is not None: if batch_size <= 0: raise ValueError("batch_size has to be defined and > 0") attention_mask = attention_mask.view(batch_size, -1) # We create a 3D attention mask from a 2D tensor mask. # Sizes are [batch_size, 1, 1, to_seq_length] # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length] # this attention mask is more simple than the triangular masking of causal attention # used in OpenAI GPT, we just need to prepare the broadcast dimension here. attention_mask = attention_mask[:, None, None, :] # Since attention_mask is 1.0 for positions we want to attend and 0.0 for # masked positions, this operation will create a tensor which is 0.0 for # positions we want to attend and the dtype's smallest value for masked positions. # Since we are adding it to the raw scores before the softmax, this is # effectively the same as removing these entirely. attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min # If a 2D or 3D attention mask is provided for the cross-attention # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length] if self.config.add_cross_attention and encoder_hidden_states is not None: encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size() encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length) if encoder_attention_mask is None: encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device) encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask) else: encoder_attention_mask = None # Prepare head mask if needed # 1.0 in head_mask indicate we keep the head # attention_probs has shape bsz x n_heads x N x N # head_mask has shape n_layer x batch x n_heads x N x N head_mask = self.get_head_mask(head_mask, self.config.n_layer) if inputs_embeds is None: inputs_embeds = self.wte(input_ids) position_embeds = self.wpe(position_ids) hidden_states = inputs_embeds + position_embeds if token_type_ids is not None: token_type_embeds = self.wte(token_type_ids) hidden_states = hidden_states + token_type_embeds hidden_states = self.drop(hidden_states) output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),) if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." ) use_cache = False presents = () if use_cache else None all_self_attentions = () if output_attentions else None all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None all_hidden_states = () if output_hidden_states else None for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)): # Model parallel if self.model_parallel: torch.cuda.set_device(hidden_states.device) # Ensure layer_past is on same device as hidden_states (might not be correct) if layer_past is not None: layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past) # Ensure that attention_mask is always on the same device as hidden_states if attention_mask is not None: attention_mask = attention_mask.to(hidden_states.device) if isinstance(head_mask, torch.Tensor): head_mask = head_mask.to(hidden_states.device) if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if self.gradient_checkpointing and self.training: outputs = self._gradient_checkpointing_func( block.__call__, hidden_states, None, attention_mask, head_mask[i], encoder_hidden_states, encoder_attention_mask, use_cache, output_attentions, ) else: outputs = block( hidden_states, layer_past=layer_past, attention_mask=attention_mask, head_mask=head_mask[i], encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_attention_mask, use_cache=use_cache, output_attentions=output_attentions, ) hidden_states = outputs[0] if use_cache is True: presents = presents + (outputs[1],) if output_attentions: all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],) if self.config.add_cross_attention: all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],) # Model Parallel: If it's the last layer for that device, put things on the next device if self.model_parallel: for k, v in self.device_map.items(): if i == v[-1] and "cuda:" + str(k) != self.last_device: hidden_states = hidden_states.to("cuda:" + str(k + 1)) hidden_states = self.ln_f(hidden_states) hidden_states = hidden_states.view(output_shape) # Add last hidden state if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if not return_dict: return tuple( v for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions] if v is not None ) return BaseModelOutputWithPastAndCrossAttentions( last_hidden_state=hidden_states, past_key_values=presents, hidden_states=all_hidden_states, attentions=all_self_attentions, cross_attentions=all_cross_attentions, ) class LOLABlock(nn.Module): def __init__(self, config, layer_idx=None): super().__init__() hidden_size = config.hidden_size inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon) self.attn = GPT2Attention(config, layer_idx=layer_idx) self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon) self.moe = LOLAMOE( hidden_size, inner_dim, config, config.num_experts, k=config.topk, # capacity_factor=1.0, # min_capacity=4, # drop_tokens=False, # use_tutel=False, # enable_expert_tensor_parallelism=False, ) def forward( self, hidden_states: Optional[Tuple[torch.FloatTensor]], layer_past: Optional[Tuple[torch.Tensor]] = None, attention_mask: Optional[torch.FloatTensor] = None, head_mask: Optional[torch.FloatTensor] = None, encoder_hidden_states: Optional[torch.Tensor] = None, encoder_attention_mask: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = False, output_attentions: Optional[bool] = False, ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]: residual = hidden_states hidden_states = self.ln_1(hidden_states) attn_outputs = self.attn( hidden_states, layer_past=layer_past, attention_mask=attention_mask, head_mask=head_mask, use_cache=use_cache, output_attentions=output_attentions, ) attn_output = attn_outputs[0] # output_attn: a, present, (attentions) outputs = attn_outputs[1:] # residual connection hidden_states = attn_output + residual if encoder_hidden_states is not None: # add one self-attention block for cross-attention if not hasattr(self, "crossattention"): raise ValueError( f"If `encoder_hidden_states` are passed, {self} has to be instantiated with " "cross-attention layers by setting `config.add_cross_attention=True`" ) residual = hidden_states hidden_states = self.ln_cross_attn(hidden_states) cross_attn_outputs = self.crossattention( hidden_states, attention_mask=attention_mask, head_mask=head_mask, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_attention_mask, output_attentions=output_attentions, ) attn_output = cross_attn_outputs[0] # residual connection hidden_states = residual + attn_output outputs = outputs + cross_attn_outputs[2:] # add cross attentions if we output attention weights residual = hidden_states hidden_states = self.ln_2(hidden_states) feed_forward_hidden_states, _ = self.moe(hidden_states) # residual connection hidden_states = residual + feed_forward_hidden_states if use_cache: outputs = (hidden_states,) + outputs else: outputs = (hidden_states,) + outputs[1:] return outputs # hidden_states, present, (attentions, cross_attentions) class LOLAMOE(nn.Module): def __init__(self, hidden_size, inner_dim, config, num_experts, k ): super().__init__() self.hidden_dim = hidden_size self.num_experts = num_experts self.top_k = k self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False) self.experts = nn.ModuleList([GPT2MLP(inner_dim, config) for _ in range(self.num_experts)]) def forward(self, hidden_states): # https://github.com/huggingface/transformers/blob/main/src/transformers/models/mixtral/modeling_mixtral.py#L816 # FIXME do it as in top1gating # https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/moe/sharded_moe.py batch_size, sequence_length, hidden_dim = hidden_states.shape hidden_states = hidden_states.view(-1, hidden_dim) router_logits = self.gate(hidden_states) # router_logits = router_logits.squeeze(dim=0) # TODO: fix the weights logic to be the same as Megatron routing_weights = F.softmax(router_logits, dim=1) routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1) # routing_weights /= routing_weights.sum(dim=-1, keepdim=True) # commenting the statement above for LOLA and removing the "/" operator to avoid getting weights as 1 routing_weights = routing_weights.sum(dim=-1, keepdim=True) routing_weights = routing_weights.to(hidden_states.dtype) final_hidden_states = torch.zeros( (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device ) expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0) for expert_idx in range(self.num_experts): expert_layer = self.experts[expert_idx] idx, top_x = torch.where(expert_mask[expert_idx]) if top_x.shape[0] == 0: continue # in torch it is faster to index using lists than torch tensors top_x_list = top_x.tolist() idx_list = idx.tolist() # Index the correct hidden states and compute the expert hidden state for # the current expert. We need to make sure to multiply the output hidden # states by `routing_weights` on the corresponding tokens (top-1 and top-2) current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim) current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None] # However `index_add_` only support torch tensors for indexing so we'll use # the `top_x` tensor here. final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype)) final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim) return final_hidden_states, router_logits class LOLAAttention(GPT2Attention): def __init__(self, config, is_cross_attention=False, layer_idx=None): super(GPT2Attention, SequenceClassifierOutputWithPast).__init__() max_positions = config.max_position_embeddings self.register_buffer( "bias", torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view( 1, 1, max_positions, max_positions ), #persistent=False, ) self.register_buffer("masked_bias", torch.tensor(-1e4), #persistent=False ) self.embed_dim = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.embed_dim // self.num_heads self.split_size = self.embed_dim if self.head_dim * self.num_heads != self.embed_dim: raise ValueError( f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:" f" {self.num_heads})." ) self.scale_attn_weights = config.scale_attn_weights self.is_cross_attention = is_cross_attention # Layer-wise attention scaling, reordering, and upcasting self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx self.layer_idx = layer_idx self.reorder_and_upcast_attn = config.reorder_and_upcast_attn if self.is_cross_attention: self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim) self.q_attn = Conv1D(self.embed_dim, self.embed_dim) else: self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim) self.c_proj = Conv1D(self.embed_dim, self.embed_dim) self.attn_dropout = nn.Dropout(config.attn_pdrop) self.resid_dropout = nn.Dropout(config.resid_pdrop) self.pruned_heads = set() class LOLALMHeadModel(GPT2LMHeadModel): config_class = LOLAConfig def __init__(self, config): # preventing initiation of GPT2LMHeadModel directly super(GPT2LMHeadModel, self).__init__(config) self.transformer = LOLAModel(config) self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False) # Model parallel self.model_parallel = False self.device_map = None # Initialize weights and apply final processing self.post_init() class LOLADoubleHeadsModel(GPT2DoubleHeadsModel): config_class = LOLAConfig def __init__(self, config): super(GPT2DoubleHeadsModel, self).__init__(config) config.num_labels = 1 self.transformer = LOLAModel(config) self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False) self.multiple_choice_head = SequenceSummary(config) # Model parallel self.model_parallel = False self.device_map = None # Initialize weights and apply final processing self.post_init() class LOLAForSequenceClassification(GPT2ForSequenceClassification): config_class = LOLAConfig def __init__(self, config): super(GPT2ForSequenceClassification, self).__init__(config) self.num_labels = config.num_labels self.transformer = LOLAModel(config) self.score = nn.Linear(config.n_embd, self.num_labels, bias=False) # Model parallel self.model_parallel = False self.device_map = None # Initialize weights and apply final processing self.post_init() class LOLAForTokenClassification(GPT2ForTokenClassification): config_class = LOLAConfig def __init__(self, config): super(GPT2ForTokenClassification, self).__init__(config) self.num_labels = config.num_labels self.transformer = LOLAModel(config) if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None: classifier_dropout = config.classifier_dropout elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None: classifier_dropout = config.hidden_dropout else: classifier_dropout = 0.1 self.dropout = nn.Dropout(classifier_dropout) self.classifier = nn.Linear(config.hidden_size, config.num_labels) # Model parallel self.model_parallel = False self.device_map = None # Initialize weights and apply final processing self.post_init() class LOLAForQuestionAnswering(GPT2PreTrainedModel): config_class = LOLAConfig def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.transformer = LOLAModel(config) self.qa_outputs = nn.Linear(config.hidden_size, 2) # Model parallel self.model_parallel = False self.device_map = None # Initialize weights and apply final processing self.post_init() def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.FloatTensor] = None, token_type_ids: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, head_mask: Optional[torch.FloatTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, start_positions: Optional[torch.LongTensor] = None, end_positions: Optional[torch.LongTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, QuestionAnsweringModelOutput]: r""" start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict outputs = self.transformer( input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask, inputs_embeds=inputs_embeds, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) sequence_output = outputs[0] logits = self.qa_outputs(sequence_output) start_logits, end_logits = logits.split(1, dim=-1) start_logits = start_logits.squeeze(-1).contiguous() end_logits = end_logits.squeeze(-1).contiguous() total_loss = None if start_positions is not None and end_positions is not None: # If we are on multi-GPU, split add a dimension if len(start_positions.size()) > 1: start_positions = start_positions.squeeze(-1).to(start_logits.device) if len(end_positions.size()) > 1: end_positions = end_positions.squeeze(-1).to(end_logits.device) # sometimes the start/end positions are outside our model inputs, we ignore these terms ignored_index = start_logits.size(1) start_positions = start_positions.clamp(0, ignored_index) end_positions = end_positions.clamp(0, ignored_index) loss_fct = CrossEntropyLoss(ignore_index=ignored_index) start_loss = loss_fct(start_logits, start_positions) end_loss = loss_fct(end_logits, end_positions) total_loss = (start_loss + end_loss) / 2 if not return_dict: output = (start_logits, end_logits) + outputs[2:] return ((total_loss,) + output) if total_loss is not None else output return QuestionAnsweringModelOutput( loss=total_loss, start_logits=start_logits, end_logits=end_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, )