[WavLM] Correct position bias computation (#14805)

src/transformers/models/wavlm/modeling_wavlm.py

@@ -394,6 +394,7 @@ class WavLMAttention(nn.Module):
         dropout: float = 0.0,
         num_buckets: int = 320,
         max_distance: int = 800,
+        has_relative_position_bias: bool = True,
     ):
         super().__init__()
         self.embed_dim = embed_dim
@@ -418,6 +419,8 @@ class WavLMAttention(nn.Module):
 
         self.gru_rel_pos_const = nn.Parameter(torch.ones(1, self.num_heads, 1, 1))
         self.gru_rel_pos_linear = nn.Linear(self.head_dim, 8)
+
+        if has_relative_position_bias:
             self.rel_attn_embed = nn.Embedding(self.num_buckets, self.num_heads)
 
     def forward(
@@ -573,7 +576,7 @@ class WavLMFeedForward(nn.Module):
 
 
 class WavLMEncoderLayer(nn.Module):
-    def __init__(self, config):
+    def __init__(self, config: WavLMConfig, has_relative_position_bias: bool = True):
         super().__init__()
         self.attention = WavLMAttention(
             embed_dim=config.hidden_size,
@@ -581,6 +584,7 @@ class WavLMEncoderLayer(nn.Module):
             dropout=config.attention_dropout,
             num_buckets=config.num_buckets,
             max_distance=config.max_bucket_distance,
+            has_relative_position_bias=has_relative_position_bias,
         )
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
@@ -613,7 +617,7 @@ class WavLMEncoderLayer(nn.Module):
 
 
 class WavLMEncoderLayerStableLayerNorm(nn.Module):
-    def __init__(self, config):
+    def __init__(self, config: WavLMConfig, has_relative_position_bias: bool = True):
         super().__init__()
         self.attention = WavLMAttention(
             embed_dim=config.hidden_size,
@@ -621,6 +625,7 @@ class WavLMEncoderLayerStableLayerNorm(nn.Module):
             dropout=config.attention_dropout,
             num_buckets=config.num_buckets,
             max_distance=config.max_bucket_distance,
+            has_relative_position_bias=has_relative_position_bias,
         )
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
@@ -655,7 +660,9 @@ class WavLMEncoder(nn.Module):
         self.pos_conv_embed = WavLMPositionalConvEmbedding(config)
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.hidden_dropout)
-        self.layers = nn.ModuleList([WavLMEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.layers = nn.ModuleList(
+            [WavLMEncoderLayer(config, has_relative_position_bias=(i == 0)) for i in range(config.num_hidden_layers)]
+        )
         self.gradient_checkpointing = False
 
     def forward(
@@ -743,7 +750,10 @@ class WavLMEncoderStableLayerNorm(nn.Module):
         self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.hidden_dropout)
         self.layers = nn.ModuleList(
-            [WavLMEncoderLayerStableLayerNorm(config) for _ in range(config.num_hidden_layers)]
+            [
+                WavLMEncoderLayerStableLayerNorm(config, has_relative_position_bias=(i == 0))
+                for i in range(config.num_hidden_layers)
+            ]
         )
         self.gradient_checkpointing = False