SwitchTransformers
Overview
The SwitchTransformers model was proposed in Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity by William Fedus, Barret Zoph, Noam Shazeer.
The Switch Transformer model uses a sparse T5 encoder-decoder architecture, where the MLP are replaced by a Mixture of Experts (MoE). A routing mechanism (top 1 in this case) associates each token to one of the expert, where each expert is a dense MLP. While switch transformers have a lot more weights than their equivalent dense models, the sparsity allows better scaling and better finetuning performance at scale. During a forward pass, only a fraction of the weights are used. The routing mechanism allows the model to select relevant weights on the fly which increases the model capacity without increasing the number of operations.
The abstract from the paper is the following:
In deep learning, models typically reuse the same parameters for all inputs. Mixture of Experts (MoE) defies this and instead selects different parameters for each incoming example. The result is a sparsely-activated model — with outrageous numbers of parameters — but a constant computational cost. However, despite several notable successes of MoE, widespread adoption has been hindered by complexity, communication costs and training instability — we address these with the Switch Transformer. We simplify the MoE routing algorithm and design intuitive improved models with reduced communication and computational costs. Our proposed training techniques help wrangle the instabilities and we show large sparse models may be trained, for the first time, with lower precision (bfloat16) formats. We design models based off T5-Base and T5-Large to obtain up to 7x increases in pre-training speed with the same computational resources. These improvements extend into multilingual settings where we measure gains over the mT5-Base version across all 101 languages. Finally, we advance the current scale of language models by pre-training up to trillion parameter models on the “Colossal Clean Crawled Corpus” and achieve a 4x speedup over the T5-XXL model.
This model was contributed by Younes Belkada and Arthur Zucker. The original code can be found here.
Usage tips
- SwitchTransformers uses the T5Tokenizer, which can be loaded directly from each model’s repository.
- The released weights are pretrained on English Masked Language Modeling task, and should be finetuned.
Resources
SwitchTransformersConfig
class transformers.SwitchTransformersConfig
< source >( vocab_size = 32128 d_model = 768 d_kv = 64 d_ff = 2048 expert_capacity = 64 num_layers = 12 num_sparse_encoder_layers = 3 num_decoder_layers = 12 num_sparse_decoder_layers = 3 num_heads = 12 num_experts = 8 router_bias = False router_jitter_noise = 0.01 router_dtype = 'float32' router_ignore_padding_tokens = False relative_attention_num_buckets = 32 relative_attention_max_distance = 128 dropout_rate = 0.1 layer_norm_epsilon = 1e-06 router_z_loss_coef = 0.001 router_aux_loss_coef = 0.001 initializer_factor = 1.0 dense_act_fn = 'relu' is_encoder_decoder = True add_router_probs = False use_cache = True pad_token_id = 0 eos_token_id = 1 **kwargs )
Parameters
- vocab_size (
int, optional, defaults to 32128) — Vocabulary size of the SwitchTransformers model. Defines the number of different tokens that can be represented by theinputs_idspassed when calling SwitchTransformersModel. - d_model (
int, optional, defaults to 768) — Size of the encoder layers and the pooler layer. - d_kv (
int, optional, defaults to 64) — Size of the key, query, value projections per attention head.d_kvhas to be equal tod_model // num_heads. - d_ff (
int, optional, defaults to 2048) — Size of the intermediate feed forward layer in eachSwitchTransformersBlock. - expert_capacity (
int, optional, defaults to 64) — Number of tokens that can be stored in each expert. If set to 1, the model will behave like a regular Transformer. - num_layers (
int, optional, defaults to 12) — Number of dense hidden layers in the Transformer encoder layer. - num_sparse_encoder_layers (
int, optional, defaults to 3) — Number of sparse (MoE) dense hidden layers in the Transformer encoder layer. - num_decoder_layers (
int, optional, defaults to 12) — Number of hidden layers in the Transformer decoder. Will use the same value asnum_layersif not set. - num_sparse_decoder_layers (
int, optional, defaults to 3) — Number of sparse (MoE) dense hidden layers in the Transformer decoder layer. - num_heads (
int, optional, defaults to 12) — Number of attention heads for each attention layer in the Transformer encoder. - num_experts (
int, optional, defaults to 8) — Number of experts for each SwitchTransformer layer. - router_bias (
bool, optional, defaults toFalse) — Whether to add a bias to the router. - router_jitter_noise (
float, optional, defaults to 0.01) — Amount of noise to add to the router. - router_dtype (
str, optional, default to"float32") — Thedtypeused for the routers. It is preferable to keep thedtypeto"float32"as specified in the selective precision discussion in the paper. - router_ignore_padding_tokens (
bool, optional, defaults toFalse) — Whether to ignore padding tokens when routing. - relative_attention_num_buckets (
int, optional, defaults to 32) — The number of buckets to use for each attention layer. - relative_attention_max_distance (
int, optional, defaults to 128) — The maximum distance of the longer sequences for the bucket separation. - dropout_rate (
float, optional, defaults to 0.1) — The ratio for all dropout layers. - layer_norm_eps (
float, optional, defaults to 1e-6) — The epsilon used by the layer normalization layers. - router_z_loss_coef (
float, optional, defaults to 0.001) — The z loss factor for the total loss. - router_aux_loss_coef (
float, optional, defaults to 0.001) — The aux loss factor for the total loss. - initializer_factor (
float, optional, defaults to 1.0) — A factor for initializing all weight matrices (should be kept to 1, used internally for initialization testing). - dense_act_fn (
string, optional, defaults to"relu") — Type of feed forward layer to be used. Should be one of"relu"or"gated-gelu". SwitchTransformersv1.1 uses the"gated-gelu"feed forward projection. Original SwitchTransformers uses"relu". - add_router_probs (
bool, optional, defaults toFalse) — Whether to output router probabilities to compute router auxiliary loss. - use_cache (
bool, optional, defaults toTrue) — Whether or not the model should return the last key/values attentions (not used by all models).
This is the configuration class to store the configuration of a SwitchTransformersModel. It is used to instantiate a SwitchTransformers model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the SwitchTransformers google/switch-base-8 architecture.
Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.
SwitchTransformersTop1Router
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.
_compute_router_probabilities
< source >( hidden_states: Tensor ) → router_probabilities (torch.Tensor)
Parameters
- 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.
Computes router probabilities from input hidden states.
forward
< source >( hidden_states: Tensor )
Parameters
- hidden_states (
torch.Tensor) — [num_groups, tokens_per_group, hidden_dim] inputs to send to experts.
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.
SwitchTransformersSparseMLP
class transformers.SwitchTransformersSparseMLP
< source >( config: SwitchTransformersConfig expert_class: Module = <class 'transformers.models.switch_transformers.modeling_switch_transformers.SwitchTransformersDenseActDense'> )
Implementation of the Switch Transformers Sparse MLP module.
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.
SwitchTransformersModel
class transformers.SwitchTransformersModel
< source >( config: SwitchTransformersConfig )
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 from_pretrained() method to load the model weights.
The bare SWITCH_TRANSFORMERS Model transformer outputting raw hidden-states without any specific head on top.
The SWITCH_TRANSFORMERS model was proposed in Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity by William Fedus, Barret Zoph, and Noam Shazeer. 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 subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
forward
< source >( input_ids: typing.Optional[torch.LongTensor] = None attention_mask: typing.Optional[torch.FloatTensor] = None decoder_input_ids: typing.Optional[torch.LongTensor] = None decoder_attention_mask: typing.Optional[torch.BoolTensor] = None head_mask: typing.Optional[torch.FloatTensor] = None decoder_head_mask: typing.Optional[torch.FloatTensor] = None cross_attn_head_mask: typing.Optional[torch.Tensor] = None encoder_outputs: typing.Optional[typing.Tuple[typing.Tuple[torch.FloatTensor]]] = None past_key_values: typing.Optional[typing.Tuple[typing.Tuple[torch.FloatTensor]]] = None inputs_embeds: typing.Optional[torch.Tensor] = None decoder_inputs_embeds: typing.Optional[torch.Tensor] = None use_cache: typing.Optional[bool] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None output_router_logits: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None cache_position: typing.Optional[torch.LongTensor] = None ) → transformers.modeling_outputs.Seq2SeqMoEModelOutput or tuple(torch.FloatTensor)
Parameters
- input_ids (
torch.LongTensorof 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_idsfor pretraining take a look a SWITCH_TRANSFORMERS Training. - attention_mask (
torch.FloatTensorof 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.
- decoder_input_ids (
torch.LongTensorof 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.
SWITCH_TRANSFORMERS uses the
pad_token_idas the starting token fordecoder_input_idsgeneration. Ifpast_key_valuesis used, optionally only the lastdecoder_input_idshave to be input (seepast_key_values).To know more on how to prepare
decoder_input_idsfor pretraining take a look at SWITCH_TRANSFORMERS Training. - decoder_attention_mask (
torch.BoolTensorof shape(batch_size, target_sequence_length), optional) — Default behavior: generate a tensor that ignores pad tokens indecoder_input_ids. Causal mask will also be used by default. - head_mask (
torch.FloatTensorof 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.FloatTensorof 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.Tensorof 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_stateof 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 lengthconfig.n_layerswith 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_valuesare used, the user can optionally input only the lastdecoder_input_ids(those that don’t have their past key value states given to this model) of shape(batch_size, 1)instead of alldecoder_input_idsof shape(batch_size, sequence_length). - inputs_embeds (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passinginput_idsyou can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_idsindices into associated vectors than the model’s internal embedding lookup matrix. - decoder_inputs_embeds (
torch.FloatTensorof shape(batch_size, target_sequence_length, hidden_size), optional) — Optionally, instead of passingdecoder_input_idsyou can choose to directly pass an embedded representation. Ifpast_key_valuesis used, optionally only the lastdecoder_inputs_embedshave to be input (seepast_key_values). This is useful if you want more control over how to convertdecoder_input_idsindices into associated vectors than the model’s internal embedding lookup matrix.If
decoder_input_idsanddecoder_inputs_embedsare both unset,decoder_inputs_embedstakes the value ofinputs_embeds. - use_cache (
bool, optional) — If set toTrue,past_key_valueskey value states are returned and can be used to speed up decoding (seepast_key_values). - output_attentions (
bool, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentionsunder returned tensors for more detail. - output_hidden_states (
bool, optional) — Whether or not to return the hidden states of all layers. Seehidden_statesunder 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 ModelOutput instead of a plain tuple. - cache_position (
torch.LongTensorof 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.
Returns
transformers.modeling_outputs.Seq2SeqMoEModelOutput or tuple(torch.FloatTensor)
A transformers.modeling_outputs.Seq2SeqMoEModelOutput or a tuple of
torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various
elements depending on the configuration (SwitchTransformersConfig) and inputs.
-
last_hidden_state (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the decoder of the model.If
past_key_valuesis used only the last hidden-state of the sequences of shape(batch_size, 1, hidden_size)is output. -
past_key_values (
tuple(tuple(torch.FloatTensor)), optional, returned whenuse_cache=Trueis passed or whenconfig.use_cache=True) — Tuple oftuple(torch.FloatTensor)of lengthconfig.n_layers, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)) and 2 additional tensors of shape(batch_size, num_heads, encoder_sequence_length, embed_size_per_head).Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see
past_key_valuesinput) to speed up sequential decoding. -
decoder_hidden_states (
tuple(torch.FloatTensor), optional, returned whenoutput_hidden_states=Trueis passed or whenconfig.output_hidden_states=True) — Tuple oftorch.FloatTensor(one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape(batch_size, sequence_length, hidden_size).Hidden-states of the decoder at the output of each layer plus the optional initial embedding outputs.
-
decoder_attentions (
tuple(torch.FloatTensor), optional, returned whenoutput_attentions=Trueis passed or whenconfig.output_attentions=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length).Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
-
decoder_router_logits (
tuple(torch.FloatTensor), optional, returned whenoutput_router_logits=Trueis passed or whenconfig.add_router_probs=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, sequence_length, num_experts).Router logits of the decoder model, useful to compute the auxiliary loss for Mixture of Experts models.
-
cross_attentions (
tuple(torch.FloatTensor), optional, returned whenoutput_attentions=Trueis passed or whenconfig.output_attentions=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length).Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.
-
encoder_last_hidden_state (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model. -
encoder_hidden_states (
tuple(torch.FloatTensor), optional, returned whenoutput_hidden_states=Trueis passed or whenconfig.output_hidden_states=True) — Tuple oftorch.FloatTensor(one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape(batch_size, sequence_length, hidden_size).Hidden-states of the encoder at the output of each layer plus the optional initial embedding outputs.
-
encoder_attentions (
tuple(torch.FloatTensor), optional, returned whenoutput_attentions=Trueis passed or whenconfig.output_attentions=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length).Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
-
encoder_router_logits (
tuple(torch.FloatTensor), optional, returned whenoutput_router_logits=Trueis passed or whenconfig.add_router_probs=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, sequence_length, num_experts).Router logits of the encoder model, useful to compute the auxiliary loss and the z_loss for the sparse modules.
The SwitchTransformersModel forward method, overrides the __call__ special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
Example:
>>> 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_stateSwitchTransformersForConditionalGeneration
class transformers.SwitchTransformersForConditionalGeneration
< source >( config: SwitchTransformersConfig )
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 from_pretrained() method to load the model weights.
SWITCH_TRANSFORMERS Model with a language modeling head on top.
The SWITCH_TRANSFORMERS model was proposed in Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity by William Fedus, Barret Zoph, and Noam Shazeer. 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 subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
forward
< source >( input_ids: typing.Optional[torch.LongTensor] = None attention_mask: typing.Optional[torch.FloatTensor] = None decoder_input_ids: typing.Optional[torch.LongTensor] = None decoder_attention_mask: typing.Optional[torch.BoolTensor] = None head_mask: typing.Optional[torch.FloatTensor] = None decoder_head_mask: typing.Optional[torch.FloatTensor] = None cross_attn_head_mask: typing.Optional[torch.Tensor] = None encoder_outputs: typing.Optional[typing.Tuple[typing.Tuple[torch.Tensor]]] = None past_key_values: typing.Optional[typing.Tuple[typing.Tuple[torch.Tensor]]] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None decoder_inputs_embeds: typing.Optional[torch.FloatTensor] = None labels: typing.Optional[torch.LongTensor] = None use_cache: typing.Optional[bool] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None output_router_logits: typing.Optional[bool] = True return_dict: typing.Optional[bool] = None cache_position: typing.Optional[torch.LongTensor] = None ) → transformers.modeling_outputs.Seq2SeqMoEOutput or tuple(torch.FloatTensor)
Parameters
- input_ids (
torch.LongTensorof 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_idsfor pretraining take a look a SWITCH_TRANSFORMERS Training. - attention_mask (
torch.FloatTensorof 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.
- decoder_input_ids (
torch.LongTensorof 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.
SWITCH_TRANSFORMERS uses the
pad_token_idas the starting token fordecoder_input_idsgeneration. Ifpast_key_valuesis used, optionally only the lastdecoder_input_idshave to be input (seepast_key_values).To know more on how to prepare
decoder_input_idsfor pretraining take a look at SWITCH_TRANSFORMERS Training. - decoder_attention_mask (
torch.BoolTensorof shape(batch_size, target_sequence_length), optional) — Default behavior: generate a tensor that ignores pad tokens indecoder_input_ids. Causal mask will also be used by default. - head_mask (
torch.FloatTensorof 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.FloatTensorof 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.Tensorof 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_stateof 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 lengthconfig.n_layerswith 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_valuesare used, the user can optionally input only the lastdecoder_input_ids(those that don’t have their past key value states given to this model) of shape(batch_size, 1)instead of alldecoder_input_idsof shape(batch_size, sequence_length). - inputs_embeds (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passinginput_idsyou can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_idsindices into associated vectors than the model’s internal embedding lookup matrix. - decoder_inputs_embeds (
torch.FloatTensorof shape(batch_size, target_sequence_length, hidden_size), optional) — Optionally, instead of passingdecoder_input_idsyou can choose to directly pass an embedded representation. Ifpast_key_valuesis used, optionally only the lastdecoder_inputs_embedshave to be input (seepast_key_values). This is useful if you want more control over how to convertdecoder_input_idsindices into associated vectors than the model’s internal embedding lookup matrix.If
decoder_input_idsanddecoder_inputs_embedsare both unset,decoder_inputs_embedstakes the value ofinputs_embeds. - use_cache (
bool, optional) — If set toTrue,past_key_valueskey value states are returned and can be used to speed up decoding (seepast_key_values). - output_attentions (
bool, optional) — Whether or not to return the attentions tensors of all attention layers. Seeattentionsunder returned tensors for more detail. - output_hidden_states (
bool, optional) — Whether or not to return the hidden states of all layers. Seehidden_statesunder 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 ModelOutput instead of a plain tuple. - cache_position (
torch.LongTensorof 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. - labels (
torch.LongTensorof 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-100are ignored (masked), the loss is only computed for labels in[0, ..., config.vocab_size]
Returns
transformers.modeling_outputs.Seq2SeqMoEOutput or tuple(torch.FloatTensor)
A transformers.modeling_outputs.Seq2SeqMoEOutput or a tuple of
torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various
elements depending on the configuration (SwitchTransformersConfig) and inputs.
-
loss (
torch.FloatTensorof shape(1,), optional, returned whenlabelsis provided) — Language modeling loss. -
logits (
torch.FloatTensorof shape(batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). -
past_key_values (
tuple(tuple(torch.FloatTensor)), optional, returned whenuse_cache=Trueis passed or whenconfig.use_cache=True) — Tuple oftuple(torch.FloatTensor)of lengthconfig.n_layers, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)) and 2 additional tensors of shape(batch_size, num_heads, encoder_sequence_length, embed_size_per_head).Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see
past_key_valuesinput) to speed up sequential decoding. -
decoder_hidden_states (
tuple(torch.FloatTensor), optional, returned whenoutput_hidden_states=Trueis passed or whenconfig.output_hidden_states=True) — Tuple oftorch.FloatTensor(one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape(batch_size, sequence_length, hidden_size).Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
-
decoder_attentions (
tuple(torch.FloatTensor), optional, returned whenoutput_attentions=Trueis passed or whenconfig.output_attentions=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length).Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
-
decoder_router_logits (
tuple(torch.FloatTensor), optional, returned whenoutput_router_logits=Trueis passed or whenconfig.add_router_probs=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, sequence_length, num_experts).Router logits of the decoder model, useful to compute the auxiliary loss for Mixture of Experts models.
-
cross_attentions (
tuple(torch.FloatTensor), optional, returned whenoutput_attentions=Trueis passed or whenconfig.output_attentions=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length).Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.
-
encoder_last_hidden_state (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model. -
encoder_hidden_states (
tuple(torch.FloatTensor), optional, returned whenoutput_hidden_states=Trueis passed or whenconfig.output_hidden_states=True) — Tuple oftorch.FloatTensor(one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape(batch_size, sequence_length, hidden_size).Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
-
encoder_attentions (
tuple(torch.FloatTensor), optional, returned whenoutput_attentions=Trueis passed or whenconfig.output_attentions=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length).Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.
-
encoder_router_logits (
tuple(torch.FloatTensor), optional, returned whenoutput_router_logits=Trueis passed or whenconfig.add_router_probs=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, sequence_length, num_experts).Router logits of the encoder model, useful to compute the auxiliary loss and z_loss for Mixture of Experts models.
The SwitchTransformersForConditionalGeneration forward method, overrides the __call__ special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
Examples:
>>> 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 <extra_id_0> walks in <extra_id_1> park", return_tensors="pt").input_ids
>>> labels = tokenizer("<extra_id_0> cute dog <extra_id_1> the <extra_id_2>", 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 gibberishSwitchTransformersEncoderModel
class transformers.SwitchTransformersEncoderModel
< source >( config: SwitchTransformersConfig )
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 from_pretrained() method to load the model weights.
The bare SWITCH_TRANSFORMERS Model transformer outputting encoder’s raw hidden-states without any specific head on top.
The SWITCH_TRANSFORMERS model was proposed in Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity by William Fedus, Barret Zoph, and Noam Shazeer. 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 subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.
forward
< source >( input_ids: typing.Optional[torch.LongTensor] = None attention_mask: typing.Optional[torch.FloatTensor] = None head_mask: typing.Optional[torch.FloatTensor] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None output_router_logits: typing.Optional[bool] = True return_dict: typing.Optional[bool] = None ) → transformers.modeling_outputs.MoEModelOutput or tuple(torch.FloatTensor)
Parameters
- input_ids (
torch.LongTensorof 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_idsfor pretraining take a look a SWITCH_TRANSFORMERS Training. - attention_mask (
torch.FloatTensorof 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.
- head_mask (
torch.FloatTensorof 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.FloatTensorof shape(batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passinginput_idsyou can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_idsindices 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. Seeattentionsunder returned tensors for more detail. - output_hidden_states (
bool, optional) — Whether or not to return the hidden states of all layers. Seehidden_statesunder 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 ModelOutput instead of a plain tuple.
Returns
transformers.modeling_outputs.MoEModelOutput or tuple(torch.FloatTensor)
A transformers.modeling_outputs.MoEModelOutput or a tuple of
torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various
elements depending on the configuration (SwitchTransformersConfig) and inputs.
-
last_hidden_state (
torch.FloatTensorof shape(batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the model. -
hidden_states (
tuple(torch.FloatTensor), optional, returned whenoutput_hidden_states=Trueis passed or whenconfig.output_hidden_states=True) — Tuple oftorch.FloatTensor(one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape(batch_size, sequence_length, hidden_size).Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
-
attentions (
tuple(torch.FloatTensor), optional, returned whenoutput_attentions=Trueis passed or whenconfig.output_attentions=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, num_heads, sequence_length, sequence_length).Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
-
router_probs (
tuple(torch.FloatTensor), optional, returned whenoutput_router_probs=Trueandconfig.add_router_probs=Trueis passed or whenconfig.output_router_probs=True) — Tuple oftorch.FloatTensor(one for each layer) of shape(batch_size, sequence_length, num_experts).Raw router probabilities that are computed by MoE routers, these terms are used to compute the auxiliary loss and the z_loss for Mixture of Experts models.
The SwitchTransformersEncoderModel forward method, overrides the __call__ special method.
Although the recipe for forward pass needs to be defined within this function, one should call the Module
instance afterwards instead of this since the former takes care of running the pre and post processing steps while
the latter silently ignores them.
Example:
>>> 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