# coding=utf-8 # Copyright 2022 SwitchTransformers Authors and HuggingFace Inc. team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """PyTorch SwitchTransformers model.""" import copy import math import warnings from typing import Optional, Tuple, Union import torch import torch.nn as nn from torch.nn import CrossEntropyLoss from ...activations import ACT2FN from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache, StaticCache from ...generation import GenerationMixin from ...modeling_attn_mask_utils import AttentionMaskConverter from ...modeling_outputs import ( MoEModelOutput, MoEModelOutputWithPastAndCrossAttentions, Seq2SeqMoEModelOutput, Seq2SeqMoEOutput, ) from ...modeling_utils import PreTrainedModel from ...pytorch_utils import ALL_LAYERNORM_LAYERS, find_pruneable_heads_and_indices, prune_linear_layer from ...utils import ( DUMMY_INPUTS, DUMMY_MASK, add_start_docstrings, add_start_docstrings_to_model_forward, is_torch_fx_proxy, is_torchdynamo_compiling, logging, replace_return_docstrings, ) from .configuration_switch_transformers import SwitchTransformersConfig logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = "SwitchTransformersConfig" _CHECKPOINT_FOR_DOC = "google/switch-base-8" #################################################### # This dict contains ids and associated url # for the pretrained weights provided with the models #################################################### def router_z_loss_func(router_logits: torch.Tensor) -> float: r""" Compute the router z-loss implemented in PyTorch. The router z-loss was introduced in [Designing Effective Sparse Expert Models](https://arxiv.org/abs/2202.08906). It encourages router logits to remain small in an effort to improve stability. Args: router_logits (`float`): Input logits of shape [batch_size, sequence_length, num_experts] Returns: Scalar router z-loss. """ num_groups, tokens_per_group, _ = router_logits.shape log_z = torch.logsumexp(router_logits, dim=-1) z_loss = log_z**2 return torch.sum(z_loss) / (num_groups * tokens_per_group) def load_balancing_loss_func(router_probs: torch.Tensor, expert_indices: torch.Tensor) -> float: r""" Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch. See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between experts is too unbalanced. Args: router_probs (`torch.Tensor`): Probability assigned to each expert per token. Shape: [batch_size, seqeunce_length, num_experts]. expert_indices (`torch.Tensor`): Indices tensor of shape [batch_size, seqeunce_length] identifying the selected expert for a given token. Returns: The auxiliary loss. """ num_experts = router_probs.shape[-1] # cast the expert indices to int64, otherwise one-hot encoding will fail if expert_indices.dtype != torch.int64: expert_indices = expert_indices.to(torch.int64) if len(expert_indices.shape) == 2: expert_indices = expert_indices.unsqueeze(2) expert_mask = torch.nn.functional.one_hot(expert_indices, num_experts) # For a given token, determine if it was routed to a given expert. expert_mask = torch.max(expert_mask, axis=-2).values # cast to float32 otherwise mean will fail expert_mask = expert_mask.to(torch.float32) tokens_per_group_and_expert = torch.mean(expert_mask, axis=-2) router_prob_per_group_and_expert = torch.mean(router_probs, axis=-2) return torch.mean(tokens_per_group_and_expert * router_prob_per_group_and_expert) * (num_experts**2) class SwitchTransformersTop1Router(nn.Module): """ Router using tokens choose top-1 experts assignment. This router uses the same mechanism as in Switch Transformer (https://arxiv.org/abs/2101.03961) and V-MoE (https://arxiv.org/abs/2106.05974): tokens choose their top experts. Items are sorted by router_probs and then routed to their choice of expert until the expert's expert_capacity is reached. **There is no guarantee that each token is processed by an expert**, or that each expert receives at least one token. """ def __init__(self, config: SwitchTransformersConfig): super().__init__() self.num_experts = config.num_experts self.expert_capacity = config.expert_capacity self.classifier = nn.Linear(config.hidden_size, self.num_experts, bias=config.router_bias) self.jitter_noise = config.router_jitter_noise self.ignore_padding_tokens = config.router_ignore_padding_tokens self.dtype = getattr(torch, config.router_dtype) def _compute_router_probabilities(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: r""" Computes router probabilities from input hidden states. Args: hidden_states (`torch.Tensor`): (batch_size, sequence_length, hidden_dim) from which router probabilities are computed. Returns: router_probabilities (`torch.Tensor`): Tensor of shape (batch_size, sequence_length, num_experts) corresponding to the probabilities for each token and expert. Used for routing tokens to experts. router_logits (`torch.Tensor`): Logits tensor of shape (batch_size, sequence_length, num_experts) corresponding to raw router logits. This is used later for computing router z-loss. """ # float32 is used to ensure stability. See the discussion of "selective precision" in # https://arxiv.org/abs/2101.03961. # We also store the previous dtype to cast back the output to the previous dtype self.input_dtype = hidden_states.dtype hidden_states = hidden_states.to(self.dtype) if self.training and self.jitter_noise > 0: # Multiply the token inputs by the uniform distribution - adding some noise hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise) # Shape: [num_groups, tokens_per_group, num_experts] self._cast_classifier() router_logits = self.classifier(hidden_states) # Apply Softmax and cast back to the original `dtype` router_probabilities = nn.functional.softmax(router_logits, dim=-1, dtype=self.dtype).to(self.input_dtype) return router_probabilities, router_logits def _cast_classifier(self): r""" `bitsandbytes` `Linear8bitLt` layers does not support manual casting Therefore we need to check if they are an instance of the `Linear8bitLt` class by checking special attributes. """ if not (hasattr(self.classifier, "SCB") or hasattr(self.classifier, "CB")): self.classifier = self.classifier.to(self.dtype) def forward(self, hidden_states: torch.Tensor) -> Tuple: r""" Generic forward function for every Router class. Each Router expects to have the same input hidden states (`hidden_states`) corresponding to the hidden states for each token, the `expert_capacity` corresponding to the number of tokens the Router will send to each expert, some Routers can send up to few tokens to each expert. Each Router works as the following: it expects the hidden states for each token, gets the `router_probs` and `router_logits` from the `router_weights`. This will assign for each token, the raw probability to be assigned to an expert. Then each Router class will have to define its own `_compute_routing_instructions`. Args: hidden_states (`torch.Tensor`) : [num_groups, tokens_per_group, hidden_dim] inputs to send to experts. Returns: Tuple[`torch.Tensor`, `torch.Tensor`, `torch.Tensor`] Tuple containing the expert index, the router probs and the router logits. The router probabilities and logits are required to compute the loss. """ router_probs, router_logits = self._compute_router_probabilities(hidden_states) expert_index = torch.argmax(router_probs, dim=-1) expert_index = torch.nn.functional.one_hot(expert_index, num_classes=self.num_experts) # Mask tokens outside expert capacity. Sum over each sequence token_priority = torch.cumsum(expert_index, dim=-2) # mask if the token routed to to the expert will overflow expert_capacity_mask = token_priority <= self.expert_capacity expert_index = expert_index * expert_capacity_mask router_probs = torch.max(router_probs, dim=-1).values.unsqueeze(-1) return expert_index, router_probs, router_logits # Copied from transformers.models.t5.modeling_t5.T5LayerNorm with T5->SwitchTransformers class SwitchTransformersLayerNorm(nn.Module): def __init__(self, hidden_size, eps=1e-6): """ Construct a layernorm module in the SwitchTransformers style. No bias and no subtraction of mean. """ super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states): # SwitchTransformers uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for # half-precision inputs is done in fp32 variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) # convert into half-precision if necessary if self.weight.dtype in [torch.float16, torch.bfloat16]: hidden_states = hidden_states.to(self.weight.dtype) return self.weight * hidden_states ALL_LAYERNORM_LAYERS.append(SwitchTransformersLayerNorm) # Copied from transformers.models.t5.modeling_t5.T5DenseActDense with T5->SwitchTransformers class SwitchTransformersDenseActDense(nn.Module): def __init__(self, config: SwitchTransformersConfig): super().__init__() self.wi = nn.Linear(config.d_model, config.d_ff, bias=False) self.wo = nn.Linear(config.d_ff, config.d_model, bias=False) self.dropout = nn.Dropout(config.dropout_rate) self.act = ACT2FN[config.dense_act_fn] def forward(self, hidden_states): hidden_states = self.wi(hidden_states) hidden_states = self.act(hidden_states) hidden_states = self.dropout(hidden_states) if ( isinstance(self.wo.weight, torch.Tensor) and hidden_states.dtype != self.wo.weight.dtype and self.wo.weight.dtype != torch.int8 ): hidden_states = hidden_states.to(self.wo.weight.dtype) hidden_states = self.wo(hidden_states) return hidden_states class SwitchTransformersSparseMLP(nn.Module): r""" Implementation of the Switch Transformers Sparse MLP module. """ def __init__(self, config: SwitchTransformersConfig, expert_class: nn.Module = SwitchTransformersDenseActDense): super().__init__() # Step 1: Get the correct router according to its class self.router = SwitchTransformersTop1Router(config) # Step 2: Get the experts self.experts = nn.ModuleDict() for idx in range(config.num_experts): self.experts[f"expert_{idx}"] = expert_class(config) def forward(self, hidden_states): r""" Hold on, this will be slightly tricky to understand In the correct order, a MoE layer does the following: 1- Gets the `router_mask` from the router. The shape of the mask is `(batch_size, sequence_length, num_expert)` and corresponds to the argmax of the `router_probs`. The probabilities are needed in the computation of the hidden states : they are broadcasted to the hidden states values (can be interpreted as a scaling factor). 2- Dispatch the tokens to its associated experts. We do a classic for loop over the experts and assign for each expert the corresponding hidden states. """ # Step 1: Get the router_mask from the router as wel as the probabilities router_mask, router_probs, router_logits = self.router(hidden_states) expert_index = torch.argmax(router_mask, dim=-1) # The routers introduced might not always map all the tokens, to a router, which means that some hidden states # can be unchanged from one layer to another. That is why the hidden states are cloned before updating only the seleced ones. next_states = hidden_states.clone() router_mask = router_mask.bool() batch_size, seq_len, num_experts = router_mask.shape idx_mask = router_mask.reshape(batch_size * seq_len, num_experts).sum(dim=0) idx_mask = torch.nonzero(idx_mask, as_tuple=True)[ 0 ].tolist() # length: number of "activated" expert / value: index for idx in idx_mask: next_states[router_mask[:, :, idx]] = getattr(self.experts, "expert_{}".format(idx))( hidden_states[router_mask[:, :, idx]] ) hidden_states = router_probs * next_states return hidden_states, (router_logits, expert_index) class SwitchTransformersLayerFF(nn.Module): r""" Switch Transformers Feed Forward layer module. This is a wrapper around the Mixture of Experts module. Parameters: config : ([`SwitchTransformersConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. is_sparse (`bool`): Whether the MLP layer is a `Sparse` layer (contains a Mixture of Experts) or not """ def __init__(self, config: SwitchTransformersConfig, is_sparse=False): super().__init__() self.is_sparse = is_sparse # Check if it is a sparse layer, if not then it is a dense layer if not self.is_sparse: self.mlp = SwitchTransformersDenseActDense(config) else: self.mlp = SwitchTransformersSparseMLP(config) self.layer_norm = SwitchTransformersLayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward(self, hidden_states, output_router_logits): forwarded_states = self.layer_norm(hidden_states) forwarded_states = self.mlp(forwarded_states) if isinstance(forwarded_states, tuple): forwarded_states, router_tuple = forwarded_states else: router_tuple = None output = hidden_states + self.dropout(forwarded_states) if output_router_logits and router_tuple is not None: output = (output, router_tuple) return output # Copied from transformers.models.t5.modeling_t5.T5Attention with T5->SwitchTransformers class SwitchTransformersAttention(nn.Module): def __init__( self, config: SwitchTransformersConfig, has_relative_attention_bias=False, layer_idx: Optional[int] = None, ): super().__init__() self.is_decoder = config.is_decoder self.has_relative_attention_bias = has_relative_attention_bias self.relative_attention_num_buckets = config.relative_attention_num_buckets self.relative_attention_max_distance = config.relative_attention_max_distance self.d_model = config.d_model self.key_value_proj_dim = config.d_kv self.n_heads = config.num_heads self.dropout = config.dropout_rate self.inner_dim = self.n_heads * self.key_value_proj_dim self.layer_idx = layer_idx if layer_idx is None and self.is_decoder: logger.warning_once( f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and " "will to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` " "when creating this class." ) # Mesh TensorFlow initialization to avoid scaling before softmax self.q = nn.Linear(self.d_model, self.inner_dim, bias=False) self.k = nn.Linear(self.d_model, self.inner_dim, bias=False) self.v = nn.Linear(self.d_model, self.inner_dim, bias=False) self.o = nn.Linear(self.inner_dim, self.d_model, bias=False) if self.has_relative_attention_bias: self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads) self.pruned_heads = set() self.gradient_checkpointing = False def prune_heads(self, heads): if len(heads) == 0: return heads, index = find_pruneable_heads_and_indices( heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads ) # Prune linear layers self.q = prune_linear_layer(self.q, index) self.k = prune_linear_layer(self.k, index) self.v = prune_linear_layer(self.v, index) self.o = prune_linear_layer(self.o, index, dim=1) # Update hyper params self.n_heads = self.n_heads - len(heads) self.inner_dim = self.key_value_proj_dim * self.n_heads self.pruned_heads = self.pruned_heads.union(heads) @staticmethod def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128): """ Adapted from Mesh Tensorflow: https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593 Translate relative position to a bucket number for relative attention. The relative position is defined as memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for small absolute relative_position and larger buckets for larger absolute relative_positions. All relative positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket. This should allow for more graceful generalization to longer sequences than the model has been trained on Args: relative_position: an int32 Tensor bidirectional: a boolean - whether the attention is bidirectional num_buckets: an integer max_distance: an integer Returns: a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets) """ relative_buckets = 0 if bidirectional: num_buckets //= 2 relative_buckets += (relative_position > 0).to(torch.long) * num_buckets relative_position = torch.abs(relative_position) else: relative_position = -torch.min(relative_position, torch.zeros_like(relative_position)) # now relative_position is in the range [0, inf) # half of the buckets are for exact increments in positions max_exact = num_buckets // 2 is_small = relative_position < max_exact # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance relative_position_if_large = max_exact + ( torch.log(relative_position.float() / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact) ).to(torch.long) relative_position_if_large = torch.min( relative_position_if_large, torch.full_like(relative_position_if_large, num_buckets - 1) ) relative_buckets += torch.where(is_small, relative_position, relative_position_if_large) return relative_buckets def compute_bias(self, query_length, key_length, device=None, cache_position=None): """Compute binned relative position bias""" if device is None: device = self.relative_attention_bias.weight.device if cache_position is None: context_position = torch.arange(query_length, dtype=torch.long, device=device)[:, None] else: context_position = cache_position[:, None].to(device) memory_position = torch.arange(key_length, dtype=torch.long, device=device)[None, :] relative_position = memory_position - context_position # shape (query_length, key_length) relative_position_bucket = self._relative_position_bucket( relative_position, # shape (query_length, key_length) bidirectional=(not self.is_decoder), num_buckets=self.relative_attention_num_buckets, max_distance=self.relative_attention_max_distance, ) values = self.relative_attention_bias(relative_position_bucket) # shape (query_length, key_length, num_heads) values = values.permute([2, 0, 1]).unsqueeze(0) # shape (1, num_heads, query_length, key_length) return values def forward( self, hidden_states, mask=None, key_value_states=None, position_bias=None, past_key_value=None, layer_head_mask=None, query_length=None, use_cache=False, output_attentions=False, cache_position=None, ): """ Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states). """ # Input is (batch_size, seq_length, dim) # Mask is (batch_size, 1, 1, key_length) (non-causal encoder) or (batch_size, 1, seq_length, key_length) (causal decoder) batch_size, seq_length = hidden_states.shape[:2] # if key_value_states are provided this layer is used as a cross-attention layer for the decoder is_cross_attention = key_value_states is not None query_states = self.q(hidden_states) query_states = query_states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2) if past_key_value is not None: is_updated = past_key_value.is_updated.get(self.layer_idx) if is_cross_attention: # after the first generated id, we can subsequently re-use all key/value_states from cache curr_past_key_value = past_key_value.cross_attention_cache else: curr_past_key_value = past_key_value.self_attention_cache current_states = key_value_states if is_cross_attention else hidden_states if is_cross_attention and past_key_value is not None and is_updated: # reuse k,v, cross_attentions key_states = curr_past_key_value.key_cache[self.layer_idx] value_states = curr_past_key_value.value_cache[self.layer_idx] else: key_states = self.k(current_states) value_states = self.v(current_states) key_states = key_states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2) value_states = value_states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2) if past_key_value is not None: # save all key/value_states to cache to be re-used for fast auto-regressive generation cache_position = cache_position if not is_cross_attention else None key_states, value_states = curr_past_key_value.update( key_states, value_states, self.layer_idx, {"cache_position": cache_position} ) # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls if is_cross_attention: past_key_value.is_updated[self.layer_idx] = True # compute scores, equivalent of torch.einsum("bnqd,bnkd->bnqk", query_states, key_states), compatible with onnx op>9 scores = torch.matmul(query_states, key_states.transpose(3, 2)) if position_bias is None: key_length = key_states.shape[-2] # cache position is 0-indexed so we add 1 to get the real length of queries (aka with past) real_seq_length = query_length if query_length is not None else cache_position[-1] + 1 if not self.has_relative_attention_bias: position_bias = torch.zeros( (1, self.n_heads, seq_length, key_length), device=scores.device, dtype=scores.dtype ) if self.gradient_checkpointing and self.training: position_bias.requires_grad = True else: position_bias = self.compute_bias( real_seq_length, key_length, device=scores.device, cache_position=cache_position ) position_bias = position_bias[:, :, -seq_length:, :] if mask is not None: causal_mask = mask[:, :, :, : key_states.shape[-2]] position_bias = position_bias + causal_mask if self.pruned_heads: mask = torch.ones(position_bias.shape[1]) mask[list(self.pruned_heads)] = 0 position_bias_masked = position_bias[:, mask.bool()] else: position_bias_masked = position_bias scores += position_bias_masked # (batch_size, n_heads, seq_length, key_length) attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(scores) attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training) # Mask heads if we want to if layer_head_mask is not None: attn_weights = attn_weights * layer_head_mask attn_output = torch.matmul(attn_weights, value_states) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.view(batch_size, -1, self.inner_dim) attn_output = self.o(attn_output) outputs = (attn_output, past_key_value, position_bias) if output_attentions: outputs = outputs + (attn_weights,) return outputs # Copied from transformers.models.t5.modeling_t5.T5LayerSelfAttention with T5->SwitchTransformers class SwitchTransformersLayerSelfAttention(nn.Module): def __init__(self, config, has_relative_attention_bias=False, layer_idx: Optional[int] = None): super().__init__() self.SelfAttention = SwitchTransformersAttention( config, has_relative_attention_bias=has_relative_attention_bias, layer_idx=layer_idx ) self.layer_norm = SwitchTransformersLayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward( self, hidden_states, attention_mask=None, position_bias=None, layer_head_mask=None, past_key_value=None, use_cache=False, output_attentions=False, cache_position=None, ): normed_hidden_states = self.layer_norm(hidden_states) attention_output = self.SelfAttention( normed_hidden_states, mask=attention_mask, position_bias=position_bias, layer_head_mask=layer_head_mask, past_key_value=past_key_value, use_cache=use_cache, output_attentions=output_attentions, cache_position=cache_position, ) hidden_states = hidden_states + self.dropout(attention_output[0]) outputs = (hidden_states,) + attention_output[1:] # add attentions if we output them return outputs # Copied from transformers.models.t5.modeling_t5.T5LayerCrossAttention with T5->SwitchTransformers class SwitchTransformersLayerCrossAttention(nn.Module): def __init__(self, config, layer_idx: Optional[int] = None): super().__init__() self.EncDecAttention = SwitchTransformersAttention( config, has_relative_attention_bias=False, layer_idx=layer_idx ) self.layer_norm = SwitchTransformersLayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward( self, hidden_states, key_value_states, attention_mask=None, position_bias=None, layer_head_mask=None, past_key_value=None, use_cache=False, query_length=None, output_attentions=False, cache_position=None, ): normed_hidden_states = self.layer_norm(hidden_states) attention_output = self.EncDecAttention( normed_hidden_states, mask=attention_mask, key_value_states=key_value_states, position_bias=position_bias, layer_head_mask=layer_head_mask, past_key_value=past_key_value, use_cache=use_cache, query_length=query_length, output_attentions=output_attentions, cache_position=cache_position, ) layer_output = hidden_states + self.dropout(attention_output[0]) outputs = (layer_output,) + attention_output[1:] # add attentions if we output them return outputs class SwitchTransformersBlock(nn.Module): def __init__(self, config, has_relative_attention_bias=False, is_sparse=False, layer_idx: Optional[int] = None): super().__init__() self.is_decoder = config.is_decoder self.is_sparse = is_sparse self.layer = nn.ModuleList() self.layer.append( SwitchTransformersLayerSelfAttention( config, has_relative_attention_bias=has_relative_attention_bias, layer_idx=layer_idx ) ) if self.is_decoder: self.layer.append(SwitchTransformersLayerCrossAttention(config, layer_idx=layer_idx)) self.layer.append(SwitchTransformersLayerFF(config, is_sparse=self.is_sparse)) def forward( self, hidden_states, attention_mask=None, position_bias=None, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None, layer_head_mask=None, cross_attn_layer_head_mask=None, past_key_value=None, use_cache=False, output_attentions=False, output_router_logits=True, return_dict=True, cache_position=None, ): self_attention_outputs = self.layer[0]( hidden_states, attention_mask=attention_mask, position_bias=position_bias, layer_head_mask=layer_head_mask, past_key_value=past_key_value, use_cache=use_cache, output_attentions=output_attentions, cache_position=cache_position, ) hidden_states, past_key_value = self_attention_outputs[:2] attention_outputs = self_attention_outputs[2:] # Keep self-attention outputs and relative position weights # clamp inf values to enable fp16 training if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any(): clamp_value = torch.finfo(hidden_states.dtype).max - 1000 hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value) do_cross_attention = self.is_decoder and encoder_hidden_states is not None if do_cross_attention: cross_attention_outputs = self.layer[1]( hidden_states, key_value_states=encoder_hidden_states, attention_mask=encoder_attention_mask, position_bias=encoder_decoder_position_bias, layer_head_mask=cross_attn_layer_head_mask, past_key_value=past_key_value, query_length=cache_position[-1] + 1, use_cache=use_cache, output_attentions=output_attentions, cache_position=cache_position, ) hidden_states, past_key_value = cross_attention_outputs[:2] # clamp inf values to enable fp16 training if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any(): clamp_value = torch.finfo(hidden_states.dtype).max - 1000 hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value) # Keep cross-attention outputs and relative position weights attention_outputs = attention_outputs + cross_attention_outputs[2:] # Apply Feed Forward layer hidden_states = self.layer[-1](hidden_states, output_router_logits) if isinstance(hidden_states, tuple): hidden_states, router_tuple = hidden_states else: router_tuple = (torch.zeros((1,), device=hidden_states.device, dtype=torch.int64),) # clamp inf values to enable fp16 training if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any(): clamp_value = torch.finfo(hidden_states.dtype).max - 1000 hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value) outputs = (hidden_states,) if use_cache: outputs = outputs + (past_key_value,) + attention_outputs + (router_tuple,) else: outputs = outputs + attention_outputs + (router_tuple,) return outputs # hidden-states, past_key_value, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights), (router_tuple) class SwitchTransformersPreTrainedModel(PreTrainedModel): """ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. """ config_class = SwitchTransformersConfig base_model_prefix = "switch_transformers" supports_gradient_checkpointing = True _supports_cache_class = True _supports_static_cache = False _no_split_modules = ["SwitchTransformersBlock"] @property def dummy_inputs(self): input_ids = torch.tensor(DUMMY_INPUTS) input_mask = torch.tensor(DUMMY_MASK) dummy_inputs = { "decoder_input_ids": input_ids, "input_ids": input_ids, "decoder_attention_mask": input_mask, } return dummy_inputs def _init_weights(self, module): """Initialize the weights""" factor = self.config.initializer_factor # Used for testing weights initialization if isinstance(module, SwitchTransformersLayerNorm): module.weight.data.fill_(factor * 1.0) elif isinstance( module, (SwitchTransformersModel, SwitchTransformersForConditionalGeneration, SwitchTransformersEncoderModel), ): # Mesh TensorFlow embeddings initialization # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624 module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0) if hasattr(module, "lm_head") and not self.config.tie_word_embeddings: module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0) elif isinstance(module, SwitchTransformersDenseActDense): # Mesh TensorFlow FF initialization # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56 # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89 module.wi.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5)) if hasattr(module.wi, "bias") and module.wi.bias is not None: module.wi.bias.data.zero_() module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5)) if hasattr(module.wo, "bias") and module.wo.bias is not None: module.wo.bias.data.zero_() elif isinstance(module, SwitchTransformersAttention): # Mesh TensorFlow attention initialization to avoid scaling before softmax # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136 d_model = self.config.d_model key_value_proj_dim = self.config.d_kv n_heads = self.config.num_heads module.q.weight.data.normal_(mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5)) module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5)) module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5)) module.o.weight.data.normal_(mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5)) if module.has_relative_attention_bias: module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor * ((d_model) ** -0.5)) elif isinstance(module, SwitchTransformersSparseMLP): # Mesh TensorFlow attention initialization to avoid scaling before softmax # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136 d_model = self.config.d_model key_value_proj_dim = self.config.d_kv n_heads = self.config.num_heads module.router.classifier.weight.data.normal_(mean=0.0, std=factor * 1) for idx in range(self.config.num_experts): module.experts[f"expert_{idx}"].wi.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5)) module.experts[f"expert_{idx}"].wo.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5)) def _shift_right(self, input_ids): decoder_start_token_id = self.config.decoder_start_token_id pad_token_id = self.config.pad_token_id if decoder_start_token_id is None: raise ValueError( "self.model.config.decoder_start_token_id has to be defined. In SwitchTransformers it is usually set" " to the pad_token_id. See SwitchTransformers docs for more information" ) # shift inputs to the right if is_torch_fx_proxy(input_ids): # Item assignment is not supported natively for proxies. shifted_input_ids = torch.full(input_ids.shape[:-1] + (1,), decoder_start_token_id) shifted_input_ids = torch.cat([shifted_input_ids, input_ids[..., :-1]], dim=-1) else: shifted_input_ids = input_ids.new_zeros(input_ids.shape) shifted_input_ids[..., 1:] = input_ids[..., :-1].clone() shifted_input_ids[..., 0] = decoder_start_token_id if pad_token_id is None: raise ValueError("self.model.config.pad_token_id has to be defined.") # replace possible -100 values in labels by `pad_token_id` shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id) return shifted_input_ids class SwitchTransformersStack(SwitchTransformersPreTrainedModel): def __init__(self, config, embed_tokens=None): super().__init__(config) self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model) if embed_tokens is not None: self.embed_tokens.weight = embed_tokens.weight self.is_decoder = config.is_decoder sparse_step = config.decoder_sparse_step if self.is_decoder else config.encoder_sparse_step config.num_layers = config.num_decoder_layers if self.is_decoder else config.num_layers self.block = nn.ModuleList() for i in range(config.num_layers): is_sparse = (i % sparse_step == 1 or sparse_step == 1) if sparse_step > 0 else False self.block.append( SwitchTransformersBlock( config, has_relative_attention_bias=bool(i == 0), is_sparse=is_sparse, layer_idx=i ) ) self.final_layer_norm = SwitchTransformersLayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) # Initialize weights and apply final processing self.post_init() self.device_map = None self.gradient_checkpointing = False def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, new_embeddings): self.embed_tokens = new_embeddings def forward( self, input_ids=None, attention_mask=None, encoder_hidden_states=None, encoder_attention_mask=None, inputs_embeds=None, head_mask=None, cross_attn_head_mask=None, past_key_values=None, use_cache=None, output_attentions=None, output_hidden_states=None, output_router_logits=True, return_dict=None, cache_position=None, ): use_cache = use_cache if use_cache is not None else self.config.use_cache 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 ) 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: err_msg_prefix = "decoder_" if self.is_decoder else "" raise ValueError( f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time" ) elif input_ids is not None: input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] else: err_msg_prefix = "decoder_" if self.is_decoder else "" raise ValueError(f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds") 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 if inputs_embeds is None: if self.embed_tokens is None: raise ValueError("You have to initialize the model with valid token embeddings") inputs_embeds = self.embed_tokens(input_ids) batch_size, seq_length = input_shape if use_cache is True: if not self.is_decoder: raise ValueError(f"`use_cache` can only be set to `True` if {self} is used as a decoder") # initialize past_key_values return_legacy_cache = False return_self_attention_cache = False if self.is_decoder and (use_cache or past_key_values is not None): if isinstance(past_key_values, Cache) and not isinstance(past_key_values, EncoderDecoderCache): return_self_attention_cache = True past_key_values = EncoderDecoderCache(past_key_values, DynamicCache()) elif not isinstance(past_key_values, EncoderDecoderCache): return_legacy_cache = True logger.warning_once( "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.48.0. " "You should pass an instance of `EncoderDecoderCache` instead, e.g. " "`past_key_values=EncoderDecoderCache.from_legacy_cache(past_key_values)`." ) past_key_values = EncoderDecoderCache.from_legacy_cache(past_key_values) elif past_key_values is None: past_key_values = EncoderDecoderCache(DynamicCache(), DynamicCache()) elif not self.is_decoder: # do not pass cache object down the line for encoder stack # it messes indexing later in decoder-stack because cache object is modified in-place past_key_values = None past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0 if cache_position is None: cache_position = torch.arange( past_key_values_length, past_key_values_length + seq_length, device=inputs_embeds.device ) if attention_mask is None and not is_torchdynamo_compiling(): # required mask seq length can be calculated via length of past cache mask_seq_length = past_key_values_length + seq_length attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device) if self.config.is_decoder: causal_mask = self._update_causal_mask( attention_mask, inputs_embeds, cache_position, past_key_values.self_attention_cache if past_key_values is not None else None, output_attentions, ) else: causal_mask = attention_mask[:, None, None, :] causal_mask = causal_mask.to(dtype=inputs_embeds.dtype) causal_mask = (1.0 - causal_mask) * torch.finfo(inputs_embeds.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.is_decoder 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=inputs_embeds.device) encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask) else: encoder_extended_attention_mask = None # Prepare head mask if needed head_mask = self.get_head_mask(head_mask, self.config.num_layers) cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers) all_hidden_states = () if output_hidden_states else None all_attentions = () if output_attentions else None all_router_probs = () if output_router_logits else None all_cross_attentions = () if (output_attentions and self.is_decoder) else None position_bias = None encoder_decoder_position_bias = None hidden_states = self.dropout(inputs_embeds) for i, layer_module in enumerate(self.block): layer_head_mask = head_mask[i] cross_attn_layer_head_mask = cross_attn_head_mask[i] if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( layer_module.forward, hidden_states, causal_mask, position_bias, encoder_hidden_states, encoder_extended_attention_mask, encoder_decoder_position_bias, layer_head_mask, cross_attn_layer_head_mask, None, # past_key_value is always None with gradient checkpointing use_cache, output_attentions, output_router_logits, return_dict, cache_position, ) else: layer_outputs = layer_module( hidden_states, attention_mask=causal_mask, position_bias=position_bias, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_extended_attention_mask, encoder_decoder_position_bias=encoder_decoder_position_bias, layer_head_mask=layer_head_mask, cross_attn_layer_head_mask=cross_attn_layer_head_mask, past_key_value=past_key_values, use_cache=use_cache, output_attentions=output_attentions, output_router_logits=output_router_logits, return_dict=return_dict, cache_position=cache_position, ) router_probs = layer_outputs[-1] layer_outputs = layer_outputs[:-1] # layer_outputs is a tuple with: # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights) if use_cache is False: layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:] hidden_states, next_decoder_cache = layer_outputs[:2] # We share the position biases between the layers - the first layer store them # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights), # (cross-attention position bias), (cross-attention weights) position_bias = layer_outputs[2] if self.is_decoder and encoder_hidden_states is not None: encoder_decoder_position_bias = layer_outputs[4 if output_attentions else 3] if output_attentions: all_attentions = all_attentions + (layer_outputs[3],) if self.is_decoder: all_cross_attentions = all_cross_attentions + (layer_outputs[5],) if output_router_logits: all_router_probs = all_router_probs + (router_probs,) hidden_states = self.final_layer_norm(hidden_states) hidden_states = self.dropout(hidden_states) # Add last layer if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) next_cache = next_decoder_cache if use_cache else None if return_self_attention_cache: next_cache = past_key_values.self_attention_cache if return_legacy_cache: next_cache = past_key_values.to_legacy_cache() if not return_dict: return tuple( v for v in [ hidden_states, next_cache, all_hidden_states, all_attentions, all_cross_attentions, all_router_probs, ] if v is not None ) return MoEModelOutputWithPastAndCrossAttentions( last_hidden_state=hidden_states, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_attentions, cross_attentions=all_cross_attentions, router_probs=all_router_probs, ) # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask def _update_causal_mask( self, attention_mask: torch.Tensor, input_tensor: torch.Tensor, cache_position: torch.Tensor, past_key_values: Cache, output_attentions: bool, ): if self.config._attn_implementation == "flash_attention_2": if attention_mask is not None and 0.0 in attention_mask: return attention_mask return None # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail # to infer the attention mask. past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 using_static_cache = isinstance(past_key_values, StaticCache) # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: if AttentionMaskConverter._ignore_causal_mask_sdpa( attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens, is_training=self.training, ): return None dtype, device = input_tensor.dtype, input_tensor.device sequence_length = input_tensor.shape[1] if using_static_cache: target_length = past_key_values.get_max_cache_shape() else: target_length = ( attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length + 1 ) # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position( attention_mask, sequence_length=sequence_length, target_length=target_length, dtype=dtype, device=device, cache_position=cache_position, batch_size=input_tensor.shape[0], ) if ( self.config._attn_implementation == "sdpa" and attention_mask is not None and attention_mask.device.type == "cuda" and not output_attentions ): # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. # Details: https://github.com/pytorch/pytorch/issues/110213 min_dtype = torch.finfo(dtype).min causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) return causal_mask @staticmethod # Copied from transformers.models.llama.modeling_llama.LlamaPreTrainedModel._prepare_4d_causal_attention_mask_with_cache_position def _prepare_4d_causal_attention_mask_with_cache_position( attention_mask: torch.Tensor, sequence_length: int, target_length: int, dtype: torch.dtype, device: torch.device, cache_position: torch.Tensor, batch_size: int, **kwargs, ): """ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. device (`torch.device`): The device to plcae the 4D attention mask on. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. """ if attention_mask is not None and attention_mask.dim() == 4: # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. causal_mask = attention_mask else: min_dtype = torch.finfo(dtype).min causal_mask = torch.full( (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device ) if sequence_length != 1: causal_mask = torch.triu(causal_mask, diagonal=1) causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) if attention_mask is not None: causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit mask_length = attention_mask.shape[-1] padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] padding_mask = padding_mask == 0 causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( padding_mask, min_dtype ) return causal_mask SWITCH_TRANSFORMERS_START_DOCSTRING = r""" The SWITCH_TRANSFORMERS model was proposed in [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://arxiv.org/abs/2101.03961) by [William Fedus](https://arxiv.org/search/cs?searchtype=author&query=Fedus%2C+W), [Barret Zoph](https://arxiv.org/search/cs?searchtype=author&query=Zoph%2C+B), and [Noam Shazeer](https://arxiv.org/search/cs?searchtype=author&query=Shazeer%2C+N). It's an encoder-decoder T5-like model with sparse Feed Forward that stands for Mixture of Experts (MoE) architecture. This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`SwitchTransformersConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. """ SWITCH_TRANSFORMERS_INPUTS_DOCSTRING = r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. SWITCH_TRANSFORMERS is a model with relative position embeddings so you should be able to pad the inputs on both the right and the left. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for detail. [What are input IDs?](../glossary#input-ids) To know more on how to prepare `input_ids` for pretraining take a look a [SWITCH_TRANSFORMERS Training](./switch_transformers#training). attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids) SWITCH_TRANSFORMERS uses the `pad_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). To know more on how to prepare `decoder_input_ids` for pretraining take a look at [SWITCH_TRANSFORMERS Training](./switch_transformers#training). decoder_attention_mask (`torch.BoolTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules in the encoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. decoder_head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules in the decoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. cross_attn_head_mask (`torch.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*): Tuple consists of (`last_hidden_state`, `optional`: *hidden_states*, `optional`: *attentions*) `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)` is a sequence of hidden states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be input (see `past_key_values`). This is useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix. If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value of `inputs_embeds`. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. output_router_logits (`bool`, *optional*): Whether or not to return the logits of all the routers. They are useful for computing the router loss, and should not be returned during inference. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. It is used to update the cache in the correct position and to infer the complete sequence length. """ SWITCH_TRANSFORMERS_ENCODER_INPUTS_DOCSTRING = r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. SWITCH_TRANSFORMERS is a model with relative position embeddings so you should be able to pad the inputs on both the right and the left. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for detail. To know more on how to prepare `input_ids` for pretraining take a look a [SWITCH_TRANSFORMERS Training](./switch_transformers#training). attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. output_router_logits (`bool`, *optional*): Whether or not to return the logits of all the routers. They are useful for computing the router loss, and should not be returned during inference. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. """ # Warning message for FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask __HEAD_MASK_WARNING_MSG = """ The input argument `head_mask` was split into two arguments `head_mask` and `decoder_head_mask`. Currently, `decoder_head_mask` is set to copy `head_mask`, but this feature is deprecated and will be removed in future versions. If you do not want to use any `decoder_head_mask` now, please set `decoder_head_mask = torch.ones(num_layers, num_heads)`. """ @add_start_docstrings( "The bare SWITCH_TRANSFORMERS Model transformer outputting raw hidden-states without any specific head on top.", SWITCH_TRANSFORMERS_START_DOCSTRING, ) class SwitchTransformersModel(SwitchTransformersPreTrainedModel): _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"] def __init__(self, config: SwitchTransformersConfig): super().__init__(config) self.shared = nn.Embedding(config.vocab_size, config.d_model) encoder_config = copy.deepcopy(config) encoder_config.is_decoder = False encoder_config.use_cache = False encoder_config.is_encoder_decoder = False self.encoder = SwitchTransformersStack(encoder_config, self.shared) decoder_config = copy.deepcopy(config) decoder_config.is_decoder = True decoder_config.is_encoder_decoder = False self.decoder = SwitchTransformersStack(decoder_config, self.shared) # Initialize weights and apply final processing self.post_init() # Model parallel self.device_map = None def get_input_embeddings(self): return self.shared def set_input_embeddings(self, new_embeddings): self.shared = new_embeddings self.encoder.set_input_embeddings(new_embeddings) self.decoder.set_input_embeddings(new_embeddings) def _tie_weights(self): if self.config.tie_word_embeddings: self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared) self._tie_or_clone_weights(self.decoder.embed_tokens, self.shared) def get_encoder(self): return self.encoder def get_decoder(self): return self.decoder 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} See base class PreTrainedModel """ for layer, heads in heads_to_prune.items(): self.encoder.layer[layer].attention.prune_heads(heads) @add_start_docstrings_to_model_forward(SWITCH_TRANSFORMERS_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=Seq2SeqMoEModelOutput, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.FloatTensor] = None, decoder_input_ids: Optional[torch.LongTensor] = None, decoder_attention_mask: Optional[torch.BoolTensor] = None, head_mask: Optional[torch.FloatTensor] = None, decoder_head_mask: Optional[torch.FloatTensor] = None, cross_attn_head_mask: Optional[torch.Tensor] = None, encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, inputs_embeds: Optional[torch.Tensor] = None, decoder_inputs_embeds: Optional[torch.Tensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple[torch.FloatTensor], Seq2SeqMoEModelOutput]: r""" Returns: Example: ```python >>> from transformers import AutoTokenizer, SwitchTransformersModel >>> tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8") >>> model = SwitchTransformersModel.from_pretrained("google/switch-base-8") >>> input_ids = tokenizer( ... "Studies have been shown that owning a dog is good for you", return_tensors="pt" ... ).input_ids # Batch size 1 >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids # Batch size 1 >>> # preprocess: Prepend decoder_input_ids with start token which is pad token for SwitchTransformersModel. >>> # This is not needed for torch's SwitchTransformersForConditionalGeneration as it does this internally using labels arg. >>> decoder_input_ids = model._shift_right(decoder_input_ids) >>> # forward pass >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids) >>> last_hidden_states = outputs.last_hidden_state ```""" 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 # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask if head_mask is not None and decoder_head_mask is None: if self.config.num_layers == self.config.num_decoder_layers: warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning) decoder_head_mask = head_mask if ( output_router_logits and self.config.num_sparse_encoder_layers == 0 and self.config.num_sparse_encoder_layers == 0 ): raise ValueError( "You asked to return `output_router_logits` but the transformer in dense, and does " " not contain any sparse MLP Layers. Set `output_router_logits = False` and restart" ) # Encode if needed (training, first prediction pass) if encoder_outputs is None: encoder_outputs = self.encoder( input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, head_mask=head_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, output_router_logits=output_router_logits, return_dict=return_dict, ) elif return_dict and not isinstance(encoder_outputs, MoEModelOutput): encoder_outputs = MoEModelOutput( last_hidden_state=encoder_outputs[0], hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None, attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None, router_probs=encoder_outputs[3] if len(encoder_outputs) > 3 else None, ) hidden_states = encoder_outputs[0] # Decode decoder_outputs = self.decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, inputs_embeds=decoder_inputs_embeds, past_key_values=past_key_values, encoder_hidden_states=hidden_states, encoder_attention_mask=attention_mask, head_mask=decoder_head_mask, cross_attn_head_mask=cross_attn_head_mask, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, output_router_logits=output_router_logits, return_dict=return_dict, cache_position=cache_position, ) if not return_dict: return decoder_outputs + encoder_outputs return Seq2SeqMoEModelOutput( last_hidden_state=decoder_outputs.last_hidden_state, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.hidden_states, decoder_attentions=decoder_outputs.attentions, cross_attentions=decoder_outputs.cross_attentions, decoder_router_logits=decoder_outputs.router_probs, encoder_last_hidden_state=encoder_outputs.last_hidden_state, encoder_hidden_states=encoder_outputs.hidden_states, encoder_attentions=encoder_outputs.attentions, encoder_router_logits=encoder_outputs.router_probs, ) @add_start_docstrings( """SWITCH_TRANSFORMERS Model with a `language modeling` head on top.""", SWITCH_TRANSFORMERS_START_DOCSTRING ) class SwitchTransformersForConditionalGeneration(SwitchTransformersPreTrainedModel, GenerationMixin): _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"] def __init__(self, config: SwitchTransformersConfig): super().__init__(config) self.model_dim = config.d_model self.shared = nn.Embedding(config.vocab_size, config.d_model) encoder_config = copy.deepcopy(config) encoder_config.is_decoder = False encoder_config.use_cache = False encoder_config.is_encoder_decoder = False self.encoder = SwitchTransformersStack(encoder_config, self.shared) decoder_config = copy.deepcopy(config) decoder_config.is_decoder = True decoder_config.is_encoder_decoder = False decoder_config.num_layers = config.num_decoder_layers self.decoder = SwitchTransformersStack(decoder_config, self.shared) self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False) self.router_z_loss_coef = config.router_z_loss_coef self.router_aux_loss_coef = config.router_aux_loss_coef # Initialize weights and apply final processing self.post_init() # Model parallel self.device_map = None def get_input_embeddings(self): return self.shared def set_input_embeddings(self, new_embeddings): self.shared = new_embeddings self.encoder.set_input_embeddings(new_embeddings) self.decoder.set_input_embeddings(new_embeddings) def _tie_weights(self): if self.config.tie_word_embeddings: self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared) self._tie_or_clone_weights(self.decoder.embed_tokens, self.shared) def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings def get_output_embeddings(self): return self.lm_head def get_encoder(self): return self.encoder def get_decoder(self): return self.decoder @add_start_docstrings_to_model_forward(SWITCH_TRANSFORMERS_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=Seq2SeqMoEOutput, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.FloatTensor] = None, decoder_input_ids: Optional[torch.LongTensor] = None, decoder_attention_mask: Optional[torch.BoolTensor] = None, head_mask: Optional[torch.FloatTensor] = None, decoder_head_mask: Optional[torch.FloatTensor] = None, cross_attn_head_mask: Optional[torch.Tensor] = None, encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, decoder_inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = True, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple[torch.FloatTensor], Seq2SeqMoEOutput]: r""" labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ..., config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` Returns: Examples: ```python >>> from transformers import AutoTokenizer, SwitchTransformersForConditionalGeneration >>> tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8") >>> model = SwitchTransformersForConditionalGeneration.from_pretrained("google/switch-base-8") >>> # training >>> input_ids = tokenizer("The walks in park", return_tensors="pt").input_ids >>> labels = tokenizer(" cute dog the ", return_tensors="pt").input_ids >>> outputs = model(input_ids=input_ids, labels=labels) >>> loss = outputs.loss >>> logits = outputs.logits >>> # inference >>> input_ids = tokenizer( ... "summarize: studies have shown that owning a dog is good for you", return_tensors="pt" ... ).input_ids # Batch size 1 >>> outputs = model.generate(input_ids) >>> # . To, let’s say you have a dog. To summarize: >>> # Since the model has been trained on MLM, this will output gibberish ```""" 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 # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask if head_mask is not None and decoder_head_mask is None: if self.config.num_layers == self.config.num_decoder_layers: warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning) decoder_head_mask = head_mask # Encode if needed (training, first prediction pass) if encoder_outputs is None: # Convert encoder inputs in embeddings if needed encoder_outputs = self.encoder( input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, head_mask=head_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, output_router_logits=output_router_logits, return_dict=return_dict, ) elif return_dict and not isinstance(encoder_outputs, MoEModelOutput): encoder_outputs = MoEModelOutput( last_hidden_state=encoder_outputs[0], hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None, attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None, router_probs=encoder_outputs[3] if len(encoder_outputs) > 3 else None, ) hidden_states = encoder_outputs[0] if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None: # get decoder inputs from shifting lm labels to the right decoder_input_ids = self._shift_right(labels) # Decode decoder_outputs = self.decoder( input_ids=decoder_input_ids, attention_mask=decoder_attention_mask, inputs_embeds=decoder_inputs_embeds, past_key_values=past_key_values, encoder_hidden_states=hidden_states, encoder_attention_mask=attention_mask, head_mask=decoder_head_mask, cross_attn_head_mask=cross_attn_head_mask, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, output_router_logits=output_router_logits, return_dict=return_dict, cache_position=cache_position, ) sequence_output = decoder_outputs[0] if self.config.tie_word_embeddings: # Rescale output before projecting on vocab # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586 sequence_output = sequence_output * (self.model_dim**-0.5) lm_logits = self.lm_head(sequence_output) loss = None encoder_z_loss = None encoder_aux_loss = None decoder_z_loss = None decoder_aux_loss = None if output_router_logits: # Compute the router loss (z_loss + auxiliary loss) for each router in the encoder and decoder if self.encoder.config.encoder_sparse_step > 1: encoder_router_logits, encoder_expert_indexes = self._unpack_router_logits(encoder_outputs[-1]) encoder_z_loss = router_z_loss_func(encoder_router_logits) encoder_router_probs = nn.Softmax(dim=-1)(encoder_router_logits) encoder_aux_loss = load_balancing_loss_func(encoder_router_probs, encoder_expert_indexes) else: encoder_z_loss = 0 encoder_aux_loss = 0 if self.decoder.config.decoder_sparse_step > 1: decoder_router_logits, decoder_expert_indexes = self._unpack_router_logits(decoder_outputs[-1]) decoder_z_loss = router_z_loss_func(decoder_router_logits) decoder_router_probs = nn.Softmax(dim=-1)(decoder_router_logits) decoder_aux_loss = load_balancing_loss_func(decoder_router_probs, decoder_expert_indexes) else: decoder_z_loss = 0 decoder_aux_loss = 0 if labels is not None: loss_fct = CrossEntropyLoss(ignore_index=-100) # move labels to correct device to enable PP labels = labels.to(lm_logits.device) loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1)) if output_router_logits: z_loss = self.router_z_loss_coef * (encoder_z_loss + decoder_z_loss) aux_loss = self.router_aux_loss_coef * (encoder_aux_loss + decoder_aux_loss) loss = loss + z_loss + aux_loss if not return_dict: output = (lm_logits,) if output_router_logits: output += (encoder_z_loss, encoder_aux_loss, decoder_z_loss, decoder_aux_loss) output += (*decoder_outputs[1:], *encoder_outputs) return ((loss,) + output) if loss is not None else output return Seq2SeqMoEOutput( loss=loss, logits=lm_logits, encoder_z_loss=encoder_z_loss, encoder_aux_loss=encoder_aux_loss, decoder_z_loss=decoder_z_loss, decoder_aux_loss=decoder_aux_loss, past_key_values=decoder_outputs.past_key_values, decoder_hidden_states=decoder_outputs.hidden_states, decoder_attentions=decoder_outputs.attentions, cross_attentions=decoder_outputs.cross_attentions, decoder_router_logits=decoder_outputs.router_probs, encoder_last_hidden_state=encoder_outputs.last_hidden_state, encoder_hidden_states=encoder_outputs.hidden_states, encoder_attentions=encoder_outputs.attentions, encoder_router_logits=encoder_outputs.router_probs, ) def _unpack_router_logits(self, router_outputs): total_router_logits = [] total_expert_indexes = [] for router_output in router_outputs: if len(router_output[0].shape) > 1: router_logits, expert_indexes = router_output total_router_logits.append(router_logits) total_expert_indexes.append(expert_indexes) return torch.cat(total_router_logits, dim=1), torch.cat(total_expert_indexes, dim=1) def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor): return self._shift_right(labels) def _reorder_cache(self, past_key_values, beam_idx): # if decoder past is not included in output # speedy decoding is disabled and no need to reorder if past_key_values is None: logger.warning("You might want to consider setting `use_cache=True` to speed up decoding") return past_key_values reordered_decoder_past = () for layer_past_states in past_key_values: # get the correct batch idx from layer past batch dim # batch dim of `past` is at 2nd position reordered_layer_past_states = () for layer_past_state in layer_past_states: # need to set correct `past` for each of the four key / value states reordered_layer_past_states = reordered_layer_past_states + ( layer_past_state.index_select(0, beam_idx.to(layer_past_state.device)), ) if reordered_layer_past_states[0].shape != layer_past_states[0].shape: raise ValueError( "expected reordered_layer_past_states to have the same shape than layer_past_states, " f"but got {reordered_layer_past_states[0].shape} and {layer_past_states[0].shape}" ) if len(reordered_layer_past_states) != len(layer_past_states): raise ValueError( "expected layer_past_states to have the same length as reordered_layer_past_states, " f"but got {len(layer_past_states)} and {len(reordered_layer_past_states)}" ) reordered_decoder_past = reordered_decoder_past + (reordered_layer_past_states,) return reordered_decoder_past @add_start_docstrings( "The bare SWITCH_TRANSFORMERS Model transformer outputting encoder's raw hidden-states without any specific head" " on top.", SWITCH_TRANSFORMERS_START_DOCSTRING, ) class SwitchTransformersEncoderModel(SwitchTransformersPreTrainedModel): _tied_weights_keys = ["encoder.embed_tokens.weight"] def __init__(self, config: SwitchTransformersConfig): super().__init__(config) self.shared = nn.Embedding(config.vocab_size, config.d_model) encoder_config = copy.deepcopy(config) encoder_config.use_cache = False encoder_config.is_encoder_decoder = False self.encoder = SwitchTransformersStack(encoder_config, self.shared) # Initialize weights and apply final processing self.post_init() # Model parallel self.device_map = None def get_input_embeddings(self): return self.shared def set_input_embeddings(self, new_embeddings): self.shared = new_embeddings self.encoder.set_input_embeddings(new_embeddings) def _tie_weights(self): if self.config.tie_word_embeddings: self._tie_or_clone_weights(self.encoder.embed_tokens, self.shared) def get_encoder(self): return self.encoder 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} See base class PreTrainedModel """ for layer, heads in heads_to_prune.items(): self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads) @add_start_docstrings_to_model_forward(SWITCH_TRANSFORMERS_ENCODER_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=MoEModelOutput, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.FloatTensor] = None, head_mask: Optional[torch.FloatTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = True, return_dict: Optional[bool] = None, ) -> Union[Tuple[torch.FloatTensor], MoEModelOutput]: r""" Returns: Example: ```python >>> from transformers import AutoTokenizer, SwitchTransformersEncoderModel >>> tokenizer = AutoTokenizer.from_pretrained("google/switch-base-8") >>> model = SwitchTransformersEncoderModel.from_pretrained("google/switch-base-8") >>> input_ids = tokenizer( ... "Studies have been shown that owning a dog is good for you", return_tensors="pt" ... ).input_ids # Batch size 1 >>> outputs = model(input_ids=input_ids) >>> last_hidden_states = outputs.last_hidden_state ```""" return_dict = return_dict if return_dict is not None else self.config.use_return_dict encoder_outputs = self.encoder( input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, head_mask=head_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, output_router_logits=output_router_logits, return_dict=return_dict, ) return encoder_outputs