updating T5 and BART models to support Prefix Tuning (#20601)

* updating T5 and BART models to support Prefix Tuning

* `make fix-copies`

* address comments

* address comments

src/transformers/models/bart/modeling_bart.py

@@ -192,7 +192,14 @@ class BartAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -1266,7 +1266,14 @@ class BigBirdPegasusDecoderAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/blenderbot/modeling_blenderbot.py

@@ -179,7 +179,14 @@ class BlenderbotAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/blenderbot_small/modeling_blenderbot_small.py

@@ -177,7 +177,14 @@ class BlenderbotSmallAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/data2vec/modeling_data2vec_audio.py

@@ -396,7 +396,14 @@ class Data2VecAudioAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/hubert/modeling_hubert.py

@@ -449,7 +449,14 @@ class HubertAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/longt5/modeling_longt5.py

@@ -479,6 +479,12 @@ class LongT5Attention(nn.Module):
                     # self-attn
                     # (batch_size, n_heads, key_length, dim_per_head)
                     hidden_states = torch.cat([past_key_value, hidden_states], dim=2)
+                elif past_key_value.shape[2] != key_value_states.shape[1]:
+                    # checking that the `sequence_length` of the `past_key_value` is the same as
+                    # the provided `key_value_states` to support prefix tuning
+                    # cross-attn
+                    # (batch_size, n_heads, seq_length, dim_per_head)
+                    hidden_states = shape(proj_layer(key_value_states))
                 else:
                     # cross-attn
                     hidden_states = past_key_value

src/transformers/models/m2m_100/modeling_m2m_100.py

@@ -247,7 +247,14 @@ class M2M100Attention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/marian/modeling_marian.py

@@ -194,7 +194,14 @@ class MarianAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/mbart/modeling_mbart.py

@@ -199,7 +199,14 @@ class MBartAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/pegasus/modeling_pegasus.py

@@ -194,7 +194,14 @@ class PegasusAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/pegasus_x/modeling_pegasus_x.py

@@ -203,7 +203,14 @@ class PegasusXAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/plbart/modeling_plbart.py

@@ -196,7 +196,14 @@ class PLBartAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/sew/modeling_sew.py

@@ -450,7 +450,14 @@ class SEWAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/speech_to_text/modeling_speech_to_text.py

@@ -253,7 +253,14 @@ class Speech2TextAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/speech_to_text_2/modeling_speech_to_text_2.py

@@ -199,7 +199,14 @@ class Speech2Text2Attention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/switch_transformers/modeling_switch_transformers.py

@@ -543,6 +543,12 @@ class SwitchTransformersAttention(nn.Module):
                     # self-attn
                     # (batch_size, n_heads, key_length, dim_per_head)
                     hidden_states = torch.cat([past_key_value, hidden_states], dim=2)
+                elif past_key_value.shape[2] != key_value_states.shape[1]:
+                    # checking that the `sequence_length` of the `past_key_value` is the same as
+                    # the provided `key_value_states` to support prefix tuning
+                    # cross-attn
+                    # (batch_size, n_heads, seq_length, dim_per_head)
+                    hidden_states = shape(proj_layer(key_value_states))
                 else:
                     # cross-attn
                     hidden_states = past_key_value

src/transformers/models/t5/modeling_t5.py

@@ -489,6 +489,12 @@ class T5Attention(nn.Module):
                     # self-attn
                     # (batch_size, n_heads, key_length, dim_per_head)
                     hidden_states = torch.cat([past_key_value, hidden_states], dim=2)
+                elif past_key_value.shape[2] != key_value_states.shape[1]:
+                    # checking that the `sequence_length` of the `past_key_value` is the same as
+                    # the provided `key_value_states` to support prefix tuning
+                    # cross-attn
+                    # (batch_size, n_heads, seq_length, dim_per_head)
+                    hidden_states = shape(proj_layer(key_value_states))
                 else:
                     # cross-attn
                     hidden_states = past_key_value

src/transformers/models/time_series_transformer/modeling_time_series_transformer.py

@@ -589,7 +589,14 @@ class TimeSeriesTransformerAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/unispeech/modeling_unispeech.py

@@ -486,7 +486,14 @@ class UniSpeechAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/unispeech_sat/modeling_unispeech_sat.py

@@ -500,7 +500,14 @@ class UniSpeechSatAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/wav2vec2/modeling_wav2vec2.py

@@ -538,7 +538,14 @@ class Wav2Vec2Attention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]

src/transformers/models/whisper/modeling_whisper.py

@@ -160,7 +160,14 @@ class WhisperAttention(nn.Module):
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
         # get key, value proj
-        if is_cross_attention and past_key_value is not None:
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
             # reuse k,v, cross_attentions
             key_states = past_key_value[0]
             value_states = past_key_value[1]