Key-value caching

modeling_nort5.py

@@ -1,19 +1,17 @@
-from __future__ import absolute_import, division, print_function, unicode_literals
-
 import math
 from typing import List, Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from torch import _softmax_backward_data as _softmax_backward_data
+from transformers.pytorch_utils import softmax_backward_data
 from torch.utils import checkpoint
 
 from configuration_nort5 import NorT5Config
 from transformers.modeling_utils import PreTrainedModel
 from transformers.activations import gelu_new
 from transformers.modeling_outputs import (
-    Seq2SeqModelOutput, Seq2SeqLMOutput, BaseModelOutput
+    Seq2SeqModelOutput, Seq2SeqLMOutput, BaseModelOutput, BaseModelOutputWithPastAndCrossAttentions
 )
 
 
@@ -58,18 +56,37 @@ class Decoder(nn.Module):
             layer.mlp.mlp[1].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
             layer.mlp.mlp[-2].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
 
-    def forward(self, x, encoder_output, encoder_padding_mask):
+        self.activation_checkpointing = activation_checkpointing
+
+    def forward(self, x, encoder_output, encoder_padding_mask, past_key_values=None):
         self_relative_embedding = self.self_relative_embedding()
         cross_relative_embedding = self.cross_relative_embedding()
 
-        autoreg_mask = torch.triu(
-            torch.full((x.size(0), x.size(0)), True, device=x.device),
-            diagonal=1
-        )
+        if past_key_values is not None:
+            autoreg_mask = torch.triu(
+                torch.full((x.size(0), x.size(0)), True, device=x.device),
+                diagonal=1
+            )
+        else:
+            autoreg_mask = None
 
-        for layer in self.layers:
-            x = layer(x, autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding)
-        return x
+        # initialize past_key_values with `None` if past does not exist
+        if past_key_values is None:
+            past_key_values = [None] * len(self.layers)
+
+        hidden_states, self_attention_probs, cross_attention_probs, key_value_states = [x], [], [], []
+        for layer, past_key_value in zip(self.layers, past_key_values):
+            if self.activation_checkpointing:
+                hidden_state, self_attention_p, cross_attention_p, key_value_state = checkpoint.checkpoint(layer, hidden_states[-1], autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=None)
+            else:
+                hidden_state, self_attention_p, cross_attention_p, key_value_state = layer(hidden_states[-1], autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=past_key_value)
+
+            hidden_states.append(hidden_state)
+            self_attention_probs.append(self_attention_p)
+            cross_attention_probs.append(cross_attention_p)
+            key_value_states.append(key_value_state)
+
+        return hidden_states, self_attention_probs, cross_attention_probs, key_value_states
 
 
 class MaskClassifier(nn.Module):
@@ -95,11 +112,11 @@ class MaskClassifier(nn.Module):
 class EncoderLayer(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.attention = Attention(config)
+        self.attention = Attention(config, is_cross_attention=False)
         self.mlp = FeedForward(config)
 
     def forward(self, x, padding_mask, relative_embedding):
-        attention_output, attention_probs = self.attention(x, x, padding_mask, relative_embedding)
+        attention_output, attention_probs, _ = self.attention(x, x, padding_mask, relative_embedding)
         x = x + attention_output
         x = x + self.mlp(x)
         return x, attention_probs
@@ -108,15 +125,26 @@ class EncoderLayer(nn.Module):
 class DecoderLayer(nn.Module):
     def __init__(self, config):
         super().__init__()
-        self.self_attention = Attention(config)
-        self.cross_attention = Attention(config)
+        self.self_attention = Attention(config, is_cross_attention=False)
+        self.cross_attention = Attention(config, is_cross_attention=True)
         self.mlp = FeedForward(config)
 
-    def forward(self, x, autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding):
-        x = x + self.self_attention(x, x, autoreg_mask, self_relative_embedding)[0]
-        x = x + self.cross_attention(x, encoder_output, encoder_padding_mask, cross_relative_embedding)[0]
+    def forward(self, x, autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=None):
+        query_offset = 0
+        if past_key_value is not None:
+            self_attn_past_key_value = past_key_value[:2]
+            cross_attn_past_key_value = past_key_value[2:]
+            query_offset = self_attn_past_key_value[0].size(1)
+        else:
+            self_attn_past_key_value, cross_attn_past_key_value = None, None
+
+        x_, self_attention_probs, self_key_value_state = self.self_attention(x, x, autoreg_mask, self_relative_embedding, past_key_value=self_attn_past_key_value, query_offset=query_offset)
+        x = x + x_
+        x_, cross_attention_probs, cross_key_value_state = self.cross_attention(x, encoder_output, encoder_padding_mask, cross_relative_embedding, past_key_value=cross_attn_past_key_value, query_offset=query_offset)
+        x = x + x_
         x = x + self.mlp(x)
-        return x
+
+        return x, self_attention_probs, cross_attention_probs, self_key_value_state + cross_key_value_state
 
 
 class GeGLU(nn.Module):
@@ -152,24 +180,27 @@ class MaskedSoftmax(torch.autograd.Function):
     @staticmethod
     def forward(self, x, mask, dim):
         self.dim = dim
-        x.masked_fill_(mask, float('-inf'))
+        if mask is not None:
+            x.masked_fill_(mask, float('-inf'))
         x = torch.softmax(x, self.dim)
-        x.masked_fill_(mask, 0.0)
+        if mask is not None:
+            x.masked_fill_(mask, 0.0)
         self.save_for_backward(x)
         return x
 
     @staticmethod
     def backward(self, grad_output):
         output, = self.saved_tensors
-        inputGrad = _softmax_backward_data(grad_output, output, self.dim, output.dtype)
-        return inputGrad, None, None
+        input_grad = softmax_backward_data(self, grad_output, output, self.dim, output)
+        return input_grad, None, None
 
 
 class Attention(nn.Module):
-    def __init__(self, config):
+    def __init__(self, config, is_cross_attention=False):
         super().__init__()
 
         self.config = config
+        self.is_cross_attention = is_cross_attention
 
         if config.hidden_size % config.num_attention_heads != 0:
             raise ValueError(f"The hidden size {config.hidden_size} is not a multiple of the number of attention heads {config.num_attention_heads}")
@@ -186,9 +217,9 @@ class Attention(nn.Module):
         self.pre_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False)
         self.post_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
 
-        position_indices = torch.arange(config.max_position_embeddings, dtype=torch.long).unsqueeze(1) \
-            - torch.arange(config.max_position_embeddings, dtype=torch.long).unsqueeze(0)
-        position_indices = self.make_log_bucket_position(position_indices, config.position_bucket_size, config.max_position_embeddings)
+        position_indices = torch.arange(512, dtype=torch.long).unsqueeze(1) \
+            - torch.arange(512, dtype=torch.long).unsqueeze(0)
+        position_indices = self.make_log_bucket_position(position_indices, config.position_bucket_size, 512)
         position_indices = config.position_bucket_size - 1 + position_indices
         self.register_buffer("position_indices", position_indices, persistent=True)
 
@@ -215,59 +246,67 @@ class Attention(nn.Module):
         self.in_proj_v.bias.data.zero_()
         self.out_proj.bias.data.zero_()
 
-    def compute_attention_scores(self, q, kv, relative_embedding):
+    def forward(self, q, kv, attention_mask, relative_embedding, past_key_value=None, query_offset=0):
         key_len, batch_size, _ = kv.size()
         query_len, _, _ = q.size()
 
-        if self.position_indices.size(0) < query_len:
-            position_indices = torch.arange(query_len, dtype=torch.long).unsqueeze(1) \
-                - torch.arange(query_len, dtype=torch.long).unsqueeze(0)
+        if not self.is_cross_attention or past_key_value is None or past_key_value[0].size(1) != kv.size(0):
+            kv = self.pre_layer_norm(kv)
+            key = self.in_proj_k(kv)  # shape: [T, B, D]
+            value = self.in_proj_v(kv)  # shape: [T, B, D]
+            key = key.reshape(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)  # shape: [BxH, T, D]
+            value = value.view(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)  # shape: [BxH, T, D]
+
+        if past_key_value is not None:
+            if not self.is_cross_attention:
+                key = torch.cat([past_key_value[0], key], dim=1)
+                value = torch.cat([past_key_value[1], value], dim=1)
+                key_len = key.size(1)
+            elif past_key_value[0].size(1) == kv.size(0):
+                key = past_key_value[0]
+                value = past_key_value[1]
+
+        if self.position_indices.size(0) < max(query_len, key_len):
+            position_indices = torch.arange(max(query_len, key_len), dtype=torch.long).unsqueeze(1) \
+                - torch.arange(max(query_len, key_len), dtype=torch.long).unsqueeze(0)
             position_indices = self.make_log_bucket_position(position_indices, self.config.position_bucket_size, 512)
             position_indices = self.config.position_bucket_size - 1 + position_indices
             self.register_buffer("position_indices", position_indices.to(q.device), persistent=True)
 
-        kv = self.pre_layer_norm(kv)
         q = self.pre_layer_norm(q)
-
         query = self.in_proj_q(q)  # shape: [T, B, D]
-        key = self.in_proj_k(kv)  # shape: [T, B, D]
-        value = self.in_proj_v(kv)  # shape: [T, B, D]
-
-        query_pos = self.in_proj_q(self.dropout(relative_embedding))  # shape: [2T-1, 2D]
-        query_pos = F.embedding(self.position_indices[:query_len, :key_len], query_pos)  # shape: [T, T, 2D]
-        query_pos = query_pos.view(query_len, key_len, self.num_heads, self.head_size)
-
-        key_pos = self.in_proj_k(self.dropout(relative_embedding))  # shape: [2T-1, 2D]
-        key_pos = F.embedding(self.position_indices[:query_len, :key_len], key_pos)  # shape: [T, T, 2D]
-        key_pos = key_pos.view(query_len, key_len, self.num_heads, self.head_size)
-
         query = query.reshape(query_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
-        key = key.reshape(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
-        value = value.view(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
 
         attention_scores = torch.bmm(query, key.transpose(1, 2) * self.scale)
-
-        query = query.view(batch_size, self.num_heads, query_len, self.head_size)
-        key = key.view(batch_size, self.num_heads, key_len, self.head_size)
+        
+        query_pos = self.in_proj_q(self.dropout(relative_embedding))  # shape: [2T-1, D]
+        query_pos = query_pos.view(-1, self.num_heads, self.head_size)  # shape: [2T-1, H, D]
+        key_pos = self.in_proj_k(self.dropout(relative_embedding))  # shape: [2T-1, D]
+        key_pos = key_pos.view(-1, self.num_heads, self.head_size)  # shape: [2T-1, H, D]
+
+        query_ = query.view(batch_size, self.num_heads, query_len, self.head_size)
+        key_ = key.view(batch_size, self.num_heads, key_len, self.head_size)
+        
+        attention_c_p = torch.einsum("bhqd,khd->bhqk", query_, key_pos.squeeze(1) * self.scale)
+        attention_p_c = torch.einsum("bhkd,qhd->bhqk", key_ * self.scale, query_pos.squeeze(1))
+        position_indices = self.position_indices[query_offset:query_offset+query_len, :key_len].expand(batch_size, self.num_heads, -1, -1)
+        attention_c_p = attention_c_p.gather(3, position_indices)
+        attention_p_c = attention_p_c.gather(2, position_indices)
+        
         attention_scores = attention_scores.view(batch_size, self.num_heads, query_len, key_len)
-        attention_scores.add_(torch.einsum("bhqd,qkhd->bhqk", query, key_pos * self.scale))
-        attention_scores.add_(torch.einsum("bhkd,qkhd->bhqk", key * self.scale, query_pos))
+        attention_scores.add_(attention_c_p)
+        attention_scores.add_(attention_p_c)
 
-        return attention_scores, value
+        attention_probs = MaskedSoftmax.apply(attention_scores, attention_mask, -1)
 
-    def compute_output(self, attention_probs, value):
         attention_probs = self.dropout(attention_probs)
         context = torch.bmm(attention_probs.flatten(0, 1), value)  # shape: [B*H, Q, D]
         context = context.transpose(0, 1).reshape(context.size(1), -1, self.hidden_size)  # shape: [Q, B, H*D]
         context = self.out_proj(context)
         context = self.post_layer_norm(context)
         context = self.dropout(context)
-        return context
 
-    def forward(self, q, kv, attention_mask, relative_embedding):
-        attention_scores, value = self.compute_attention_scores(q, kv, relative_embedding)
-        attention_probs = MaskedSoftmax.apply(attention_scores, attention_mask, -1)
-        return self.compute_output(attention_probs, value), attention_probs.detach()
+        return context, attention_probs.detach(), (key.detach(), value.detach())
 
 
 class WordEmbedding(nn.Module):
@@ -348,8 +387,8 @@ class NorT5Model(NorT5PreTrainedModel):
         return self.get_encoder_output
 
     def get_decoder(self):
-        return self.decoder
-    
+        return self.get_decoder_output
+
     def set_decoder_special_tokens(self, target_id):
         target_id.masked_fill_(target_id == self.cls_token_id, self.bos_token_id)
         target_id.masked_fill_(target_id == self.sep_token_id, self.eos_token_id)
@@ -359,12 +398,13 @@ class NorT5Model(NorT5PreTrainedModel):
         shifted_input_ids = input_ids.new_zeros(input_ids.shape)
         shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
         shifted_input_ids[..., 0] = self.bos_token_id
+        shifted_input_ids.masked_fill_(shifted_input_ids == -100, self.pad_token_id)
 
         return shifted_input_ids
 
     def get_encoder_output(
         self,
-        input_ids: Optional[torch.Tensor] = None,
+        input_ids: torch.Tensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         output_hidden_states: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
@@ -394,16 +434,28 @@ class NorT5Model(NorT5PreTrainedModel):
         ]
 
         if not return_dict:
-            return last_layer, contextualized_embeddings, attention_probs
-
+            return (
+                last_layer,
+                *([contextualized_embeddings] if output_hidden_states else []),
+                *([attention_probs] if output_attentions else [])
+            )
+        
         return BaseModelOutput(
             last_hidden_state=last_layer,
-            hidden_states=contextualized_embeddings,
-            attentions=attention_probs
+            hidden_states=contextualized_embeddings if output_hidden_states else None,
+            attentions=attention_probs if output_attentions else None
         )
     
     def get_decoder_output(
-        self, target_ids, encoder_output, attention_mask
+        self,
+        target_ids: torch.Tensor = None,
+        encoder_output: torch.Tensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict = False
     ):
         batch_size, seq_length, _ = encoder_output.shape
         device = target_ids.device
@@ -414,11 +466,37 @@ class NorT5Model(NorT5PreTrainedModel):
             attention_mask = ~attention_mask.bool()
         attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
 
-        return self.decoder(
+        hidden_states, self_attention_p, cross_attention_p, key_value_states = self.decoder(
             self.embedding(target_ids.t()),
             encoder_output.transpose(0, 1),
-            attention_mask
-        ).transpose(0, 1)
+            attention_mask,
+            past_key_values
+        )
+
+        hidden_states = [e.transpose(0, 1) for e in hidden_states]
+        last_layer = hidden_states[-1]
+        hidden_states = [hidden_states[0]] + [
+            hidden_states[i] - hidden_states[i - 1]
+            for i in range(1, len(hidden_states))
+        ]
+
+        if not return_dict:
+            return (
+                last_layer,
+                *([key_value_states] if use_cache else []),
+                *([hidden_states] if output_hidden_states else []),
+                *([self_attention_p] if output_attentions else []),
+                *([cross_attention_p] if output_attentions else []),
+            )
+        
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=last_layer,
+            past_key_values=key_value_states if use_cache else None,
+            hidden_states=hidden_states if output_hidden_states else None,
+            attentions=self_attention_p if output_attentions else None,
+            cross_attentions=cross_attention_p if output_attentions else None
+        )
+
 
     def forward(
         self,
@@ -426,28 +504,45 @@ class NorT5Model(NorT5PreTrainedModel):
         attention_mask: Optional[torch.FloatTensor] = None,
         decoder_input_ids: Optional[torch.LongTensor] = None,
         decoder_attention_mask: Optional[torch.BoolTensor] = None,
-        return_dict: Optional[bool] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None
     ):
 
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         decoder_input_ids = self.set_decoder_special_tokens(decoder_input_ids)
 
-        encoder_outputs, encoder_contextualized_embeddings, encoder_attention_probs = self.get_encoder_output(input_ids, attention_mask)
-        decoder_outputs = self.get_decoder_output(decoder_input_ids, encoder_outputs, attention_mask)
+        if encoder_outputs is None:
+            encoder_outputs = self.get_encoder_output(
+                input_ids, attention_mask, output_hidden_states, output_attentions, return_dict
+            )
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                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,
+            )
+    
+        decoder_outputs = self.get_decoder_output(
+            decoder_input_ids, encoder_outputs[0], attention_mask, past_key_values, use_cache, output_hidden_states, output_attentions, return_dict
+        )
 
         if not return_dict:
-            return (decoder_outputs, encoder_outputs)
-
+            return decoder_outputs + encoder_outputs
+        
         return Seq2SeqModelOutput(
-            last_hidden_state=decoder_outputs,
-            past_key_values=None,
-            decoder_hidden_states=None,
-            decoder_attentions=None,
-            cross_attentions=None,
-            encoder_last_hidden_state=encoder_outputs,
-            encoder_hidden_states=encoder_contextualized_embeddings,
-            encoder_attentions=encoder_attention_probs,
+            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,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
         )
 
 
@@ -475,12 +570,19 @@ class NorT5ForConditionalGeneration(NorT5Model):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
     ):
-
-        use_cache = False
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         if encoder_outputs is None:
-            encoder_outputs = self.get_encoder_output(input_ids, attention_mask, return_dict=True)
+            encoder_outputs = self.get_encoder_output(
+                input_ids, attention_mask, output_hidden_states, output_attentions, return_dict
+            )
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                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,
+            )
 
         if labels is not None:
             labels = self.set_decoder_special_tokens(labels)
@@ -490,24 +592,28 @@ class NorT5ForConditionalGeneration(NorT5Model):
         elif decoder_input_ids is not None:
             decoder_input_ids = self.set_decoder_special_tokens(decoder_input_ids)
 
-        decoder_outputs = self.get_decoder_output(decoder_input_ids, encoder_outputs.last_hidden_state, attention_mask)
-        lm_logits = self.classifier(decoder_outputs)
+        decoder_outputs = self.get_decoder_output(
+            decoder_input_ids, encoder_outputs[0], attention_mask, past_key_values, use_cache, output_hidden_states, output_attentions, return_dict
+        )
+        lm_logits = self.classifier(decoder_outputs[0])
 
         loss = None
         if labels is not None:
-            loss_fct = nn.CrossEntropyLoss(ignore_index=self.pad_token_id)
+            labels.masked_fill_(labels == self.pad_token_id, -100)
+            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
             loss = loss_fct(lm_logits.flatten(0, 1), labels.flatten())
 
         if not return_dict:
-            output = (lm_logits,) + encoder_outputs
+            output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
             return ((loss,) + output) if loss is not None else output
 
         return Seq2SeqLMOutput(
             loss=loss,
             logits=lm_logits,
-            decoder_hidden_states=decoder_outputs,
-            decoder_attentions=None,
-            cross_attentions=None,
+            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,
             encoder_last_hidden_state=encoder_outputs.last_hidden_state,
             encoder_hidden_states=encoder_outputs.hidden_states,
             encoder_attentions=encoder_outputs.attentions,
@@ -525,6 +631,9 @@ class NorT5ForConditionalGeneration(NorT5Model):
         encoder_outputs=None,
         **kwargs,
     ):
+        if past_key_values is not None:
+            input_ids = input_ids[:, -1:]
+
         return {
             "decoder_input_ids": input_ids,
             "past_key_values": past_key_values,
@@ -553,9 +662,10 @@ class NorT5ForConditionalGeneration(NorT5Model):
             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)),
-                )
+                layer_past_state = layer_past_state.unflatten(0, (-1, self.config.num_attention_heads))
+                layer_past_state = layer_past_state.index_select(0, beam_idx.to(layer_past_state.device))
+                layer_past_state = layer_past_state.flatten(0, 1)
+                reordered_layer_past_states = reordered_layer_past_states + (layer_past_state,)
 
             assert reordered_layer_past_states[0].shape == layer_past_states[0].shape
             assert len(reordered_layer_past_states) == len(layer_past_states)
@@ -578,4 +688,6 @@ class NorT5Encoder(NorT5Model):
     ):
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
-        return self.get_encoder_output(input_ids, attention_mask, return_dict=return_dict)
+        return self.get_encoder_output(
+            input_ids, attention_mask, output_hidden_states, output_attentions, return_dict=return_dict
+        )