Small change to Wav2Vec2 model to support Tensor-Parallelism with Dee…

…pSpeed (#14298)

* minor modification to the wav2vec2 modeling file to support tensor-parallelism with DeepSpeed on this HuggingFace model

* refine the comments

* synch changes

* fix comments

* refine comments

* fix format

src/transformers/models/bart/modeling_bart.py

@@ -171,7 +171,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -257,7 +258,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -1244,7 +1244,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -1330,7 +1331,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -173,7 +173,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -259,7 +260,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -171,7 +171,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -257,7 +258,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -388,7 +388,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -474,7 +475,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -242,7 +242,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -328,7 +329,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -188,7 +188,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -274,7 +275,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -177,7 +177,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -263,7 +264,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -188,7 +188,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -274,7 +275,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -387,7 +387,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -473,7 +474,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -255,7 +255,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -341,7 +342,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -195,7 +195,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -281,7 +282,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -456,7 +456,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -542,7 +543,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -457,7 +457,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -543,7 +544,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        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

@@ -493,7 +493,8 @@ def forward(
         # if key_value_states are provided this layer is used as a cross-attention layer
         # for the decoder
         is_cross_attention = key_value_states is not None
-        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        bsz, tgt_len, _ = hidden_states.size()
 
         # get query proj
         query_states = self.q_proj(hidden_states) * self.scaling
@@ -579,7 +580,10 @@ def forward(
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
-        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
+
+        # 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.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
 
         attn_output = self.out_proj(attn_output)