Update model.py

model.py

@@ -31,7 +31,8 @@ class CustomWhisperConfig(WhisperConfig):
                     "kernel_size": 3,
                     "stride": 1,
                     "padding": 1,
-                    "activation": "gelu"
+                    "activation": "gelu",
+                    "bias": True
                 },
                 {
                     "in_channels": self.d_model,
@@ -39,7 +40,8 @@ class CustomWhisperConfig(WhisperConfig):
                     "kernel_size": 3,
                     "stride": 2,
                     "padding": 1,
-                    "activation": "gelu"
+                    "activation": "gelu",
+                    "bias": True
                 }
             ]
             
@@ -996,16 +998,21 @@ class WhisperEncoder(WhisperPreTrainedModel):
         # CUSTOM
         # Create conv layers dynamically based on config
         self.conv_layers = nn.ModuleList()
+        self.conv_layers = nn.ModuleList()
         for layer_config in config.conv_preprocessing_layers:
-            conv_layer = nn.Conv1d(
-                layer_config.in_channels,
-                layer_config.out_channels,
-                kernel_size=layer_config.kernel_size,
-                stride=layer_config.stride,
-                padding=layer_config.padding,
-                bias=config.conv_bias
-        )
-        self.conv_layers.append(conv_layer)
+            # Create sequential module for each conv+activation pair
+            conv_sequence = nn.Sequential(
+                nn.Conv1d(
+                    layer_config["in_channels"],
+                    layer_config["out_channels"],
+                    kernel_size=layer_config["kernel_size"],
+                    stride=layer_config["stride"],
+                    padding=layer_config["padding"],
+                    bias=True
+                ),
+                nn.GELU() if layer_config["activation"] == "gelu" else nn.ReLU()
+            )
+            self.conv_layers.append(conv_sequence)
 
         self.embed_positions = nn.Embedding(self.max_source_positions, embed_dim)
         self.embed_positions.requires_grad_(False)