Remove the head masking block in some vision models

src/transformers/models/audio_spectrogram_transformer/modeling_audio_spectrogram_transformer.py

@@ -117,10 +117,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/deit/modeling_deit.py

@@ -182,10 +182,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/dinov2/modeling_dinov2.py

@@ -170,10 +170,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/dinov2_with_registers/modeling_dinov2_with_registers.py

@@ -190,10 +190,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/dinov3_vit/modeling_dinov3_vit.py

@@ -208,10 +208,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/dpt/modeling_dpt.py

@@ -288,10 +288,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/ijepa/modeling_ijepa.py

@@ -161,10 +161,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/videomae/modeling_videomae.py

@@ -199,10 +199,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/vit/modeling_vit.py

@@ -187,10 +187,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/vit_mae/modeling_vit_mae.py

@@ -347,10 +347,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/vit_msn/modeling_vit_msn.py

@@ -184,10 +184,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/vitpose_backbone/modeling_vitpose_backbone.py

@@ -116,10 +116,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/vivit/modeling_vivit.py

@@ -177,10 +177,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/vjepa2/modeling_vjepa2.py

@@ -169,10 +169,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()
 

src/transformers/models/yolos/modeling_yolos.py

@@ -232,10 +232,6 @@ def eager_attention_forward(
     # seem a bit unusual, but is taken from the original Transformer paper.
     attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
 
-    # Mask heads if we want to
-    if attention_mask is not None:
-        attn_weights = attn_weights * attention_mask
-
     attn_output = torch.matmul(attn_weights, value)
     attn_output = attn_output.transpose(1, 2).contiguous()