Add `accelerate` support for BART-like models (#19927)

* forward contrib credits from suggestion

* add `accelerate` support for BART-like models
Co-authored-by: sgugger <sgugger@users.noreply.github.com>

src/transformers/models/bart/modeling_bart.py

@@ -500,6 +500,7 @@ class BartPretrainedModel(PreTrainedModel):
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
     _keys_to_ignore_on_load_unexpected = [r"encoder.version", r"decoder.version"]
+    _no_split_modules = [r"BartEncoderLayer", r"BartDecoderLayer"]
 
     def _init_weights(self, module):
         std = self.config.init_std
@@ -712,11 +713,11 @@ class BartEncoder(BartPretrainedModel):
         self.max_source_positions = config.max_position_embeddings
         self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
 
-        if embed_tokens is not None:
-            self.embed_tokens = embed_tokens
-        else:
         self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
 
+        if embed_tokens is not None:
+            self.embed_tokens.weight = embed_tokens.weight
+
         self.embed_positions = BartLearnedPositionalEmbedding(
             config.max_position_embeddings,
             embed_dim,
@@ -801,6 +802,7 @@ class BartEncoder(BartPretrainedModel):
             inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
 
         embed_pos = self.embed_positions(input)
+        embed_pos = embed_pos.to(inputs_embeds.device)
 
         hidden_states = inputs_embeds + embed_pos
         hidden_states = self.layernorm_embedding(hidden_states)
@@ -884,11 +886,11 @@ class BartDecoder(BartPretrainedModel):
         self.max_target_positions = config.max_position_embeddings
         self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
 
-        if embed_tokens is not None:
-            self.embed_tokens = embed_tokens
-        else:
         self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
 
+        if embed_tokens is not None:
+            self.embed_tokens.weight = embed_tokens.weight
+
         self.embed_positions = BartLearnedPositionalEmbedding(
             config.max_position_embeddings,
             config.d_model,
@@ -1043,6 +1045,7 @@ class BartDecoder(BartPretrainedModel):
 
         # embed positions
         positions = self.embed_positions(input, past_key_values_length)
+        positions = positions.to(inputs_embeds.device)
 
         hidden_states = inputs_embeds + positions
         hidden_states = self.layernorm_embedding(hidden_states)
@@ -1373,7 +1376,9 @@ class BartForConditionalGeneration(BartPretrainedModel):
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
         )
-        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+
+        lm_logits = self.lm_head(outputs[0])
+        lm_logits = lm_logits + self.final_logits_bias.to(lm_logits.device)
 
         masked_lm_loss = None
         if labels is not None:

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -1595,6 +1595,7 @@ class BigBirdPegasusPreTrainedModel(PreTrainedModel):
     config_class = BigBirdPegasusConfig
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
+    _no_split_modules = ["BigBirdPegasusEncoderLayer", "BigBirdPegasusDecoderLayer"]
 
     def _init_weights(self, module):
         std = self.config.init_std
@@ -1788,11 +1789,11 @@ class BigBirdPegasusEncoder(BigBirdPegasusPreTrainedModel):
         self.max_source_positions = config.max_position_embeddings
         self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
 
-        if embed_tokens is not None:
-            self.embed_tokens = embed_tokens
-        else:
         self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
 
+        if embed_tokens is not None:
+            self.embed_tokens.weight = embed_tokens.weight
+
         self.embed_positions = BigBirdPegasusLearnedPositionalEmbedding(
             config.max_position_embeddings,
             embed_dim,
@@ -2082,11 +2083,11 @@ class BigBirdPegasusDecoder(BigBirdPegasusPreTrainedModel):
         self.max_target_positions = config.max_position_embeddings
         self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
 
-        if embed_tokens is not None:
-            self.embed_tokens = embed_tokens
-        else:
         self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
 
+        if embed_tokens is not None:
+            self.embed_tokens.weight = embed_tokens.weight
+
         self.embed_positions = BigBirdPegasusLearnedPositionalEmbedding(
             config.max_position_embeddings,
             config.d_model,
@@ -2240,6 +2241,7 @@ class BigBirdPegasusDecoder(BigBirdPegasusPreTrainedModel):
 
         # embed positions
         positions = self.embed_positions(input_shape, past_key_values_length)
+        positions = positions.to(inputs_embeds.device)
 
         hidden_states = inputs_embeds + positions
 
@@ -2573,7 +2575,9 @@ class BigBirdPegasusForConditionalGeneration(BigBirdPegasusPreTrainedModel):
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
         )
-        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+
+        lm_logits = self.lm_head(outputs[0])
+        lm_logits = lm_logits + self.final_logits_bias.to(lm_logits.device)
 
         masked_lm_loss = None
         if labels is not None:

src/transformers/models/plbart/modeling_plbart.py

@@ -506,6 +506,7 @@ class PLBartPreTrainedModel(PreTrainedModel):
     config_class = PLBartConfig
     base_model_prefix = "model"
     supports_gradient_checkpointing = True
+    _no_split_modules = ["PLBartDecoderLayer", "PLBartEncoderLayer"]
 
     def _init_weights(self, module):
         std = self.config.init_std
@@ -683,11 +684,11 @@ class PLBartEncoder(PLBartPreTrainedModel):
         self.max_source_positions = config.max_position_embeddings
         self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
 
-        if embed_tokens is not None:
-            self.embed_tokens = embed_tokens
-        else:
         self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)
 
+        if embed_tokens is not None:
+            self.embed_tokens.weight = embed_tokens.weight
+
         self.embed_positions = PLBartLearnedPositionalEmbedding(
             config.max_position_embeddings,
             embed_dim,
@@ -772,6 +773,7 @@ class PLBartEncoder(PLBartPreTrainedModel):
             inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
 
         embed_pos = self.embed_positions(input)
+        embed_pos = embed_pos.to(inputs_embeds.device)
 
         hidden_states = inputs_embeds + embed_pos
         hidden_states = self.layernorm_embedding(hidden_states)
@@ -856,11 +858,11 @@ class PLBartDecoder(PLBartPreTrainedModel):
         self.max_target_positions = config.max_position_embeddings
         self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
 
-        if embed_tokens is not None:
-            self.embed_tokens = embed_tokens
-        else:
         self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
 
+        if embed_tokens is not None:
+            self.embed_tokens.weight = embed_tokens.weight
+
         self.embed_positions = PLBartLearnedPositionalEmbedding(
             config.max_position_embeddings,
             config.d_model,
@@ -1015,6 +1017,7 @@ class PLBartDecoder(PLBartPreTrainedModel):
 
         # embed positions
         positions = self.embed_positions(input, past_key_values_length)
+        positions = positions.to(inputs_embeds.device)
 
         hidden_states = inputs_embeds + positions
         hidden_states = self.layernorm_embedding(hidden_states)
@@ -1334,7 +1337,8 @@ class PLBartForConditionalGeneration(PLBartPreTrainedModel):
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
         )
-        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias
+        lm_logits = self.lm_head(outputs[0])
+        lm_logits = lm_logits + self.final_logits_bias.to(lm_logits.device)
 
         masked_lm_loss = None
         if labels is not None: