fix typo in Bart's attention (#21898)

src/transformers/models/bart/modeling_bart.py

@@ -276,7 +276,7 @@ class BartAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -284,7 +284,7 @@ class BartAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -1335,7 +1335,7 @@ class BigBirdPegasusDecoderAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -1343,7 +1343,7 @@ class BigBirdPegasusDecoderAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/biogpt/modeling_biogpt.py

@@ -237,7 +237,7 @@ class BioGptAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -245,7 +245,7 @@ class BioGptAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/blenderbot/modeling_blenderbot.py

@@ -263,7 +263,7 @@ class BlenderbotAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -271,7 +271,7 @@ class BlenderbotAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/blenderbot_small/modeling_blenderbot_small.py

@@ -260,7 +260,7 @@ class BlenderbotSmallAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -268,7 +268,7 @@ class BlenderbotSmallAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/data2vec/modeling_data2vec_audio.py

@@ -466,7 +466,7 @@ class Data2VecAudioAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -474,7 +474,7 @@ class Data2VecAudioAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/gptsan_japanese/modeling_gptsan_japanese.py

@@ -498,7 +498,7 @@ class GPTSanJapaneseAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -506,7 +506,7 @@ class GPTSanJapaneseAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/hubert/modeling_hubert.py

@@ -528,7 +528,7 @@ class HubertAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -536,7 +536,7 @@ class HubertAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/m2m_100/modeling_m2m_100.py

@@ -331,7 +331,7 @@ class M2M100Attention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -339,7 +339,7 @@ class M2M100Attention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/marian/modeling_marian.py

@@ -278,7 +278,7 @@ class MarianAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -286,7 +286,7 @@ class MarianAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/mbart/modeling_mbart.py

@@ -272,7 +272,7 @@ class MBartAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -280,7 +280,7 @@ class MBartAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/pegasus/modeling_pegasus.py

@@ -278,7 +278,7 @@ class PegasusAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -286,7 +286,7 @@ class PegasusAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/pegasus_x/modeling_pegasus_x.py

@@ -287,7 +287,7 @@ class PegasusXAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -295,7 +295,7 @@ class PegasusXAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/plbart/modeling_plbart.py

@@ -271,7 +271,7 @@ class PLBartAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -279,7 +279,7 @@ class PLBartAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/sew/modeling_sew.py

@@ -528,7 +528,7 @@ class SEWAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -536,7 +536,7 @@ class SEWAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/speech_to_text/modeling_speech_to_text.py

@@ -338,7 +338,7 @@ class Speech2TextAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -346,7 +346,7 @@ class Speech2TextAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/speech_to_text_2/modeling_speech_to_text_2.py

@@ -284,7 +284,7 @@ class Speech2Text2Attention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -292,7 +292,7 @@ class Speech2Text2Attention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/time_series_transformer/modeling_time_series_transformer.py

@@ -764,7 +764,7 @@ class TimeSeriesTransformerAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -772,7 +772,7 @@ class TimeSeriesTransformerAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/unispeech/modeling_unispeech.py

@@ -564,7 +564,7 @@ class UniSpeechAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -572,7 +572,7 @@ class UniSpeechAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/unispeech_sat/modeling_unispeech_sat.py

@@ -578,7 +578,7 @@ class UniSpeechSatAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -586,7 +586,7 @@ class UniSpeechSatAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/wav2vec2/modeling_wav2vec2.py

@@ -621,7 +621,7 @@ class Wav2Vec2Attention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -629,7 +629,7 @@ class Wav2Vec2Attention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)

src/transformers/models/whisper/modeling_whisper.py

@@ -366,7 +366,7 @@ class WhisperAttention(nn.Module):
 
         if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
             raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                 f" {attn_output.size()}"
             )
 
@@ -374,7 +374,7 @@ class WhisperAttention(nn.Module):
         attn_output = attn_output.transpose(1, 2)
 
         # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
-        # partitioned aross GPUs when using tensor-parallelism.
+        # partitioned across GPUs when using tensor-parallelism.
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)