Audio encodings now match conv2d weight dtype in Gemma3nAudioSSCPConvBlock (#39743)

audio encodings now match conv weight dtype in Gemma3nAudioSSCPConvBlock

Author: Malav-P

src/transformers/models/gemma3n/modeling_gemma3n.py

@@ -695,7 +695,9 @@ def forward(self, audio_encodings: torch.Tensor) -> torch.Tensor:
         # Input audio_encodings is [B, C_in, T_in, F_in] (e.g., C_in=1)
         # manual_padding is (pad_F_left, pad_F_right, pad_T_top, pad_T_bottom)
         # F.pad applies to last two dims: F_in then T_in
-        audio_encodings_padded = F.pad(audio_encodings, self.manual_padding, mode="constant", value=0.0)
+        audio_encodings_padded = F.pad(audio_encodings, self.manual_padding, mode="constant", value=0.0).to(
+            self.conv.weight.dtype
+        )
         # Expected padded shape for F_in, k_w=3, pad_F=(1,1) -> F_padded = F_in+2
         # Expected padded shape for T_in, k_h=3, pad_T=(0,2) -> T_padded = T_in+2
         audio_encodings_conv = self.conv(audio_encodings_padded)

src/transformers/models/gemma3n/modular_gemma3n.py

@@ -1264,7 +1264,9 @@ def forward(self, audio_encodings: torch.Tensor) -> torch.Tensor:
         # Input audio_encodings is [B, C_in, T_in, F_in] (e.g., C_in=1)
         # manual_padding is (pad_F_left, pad_F_right, pad_T_top, pad_T_bottom)
         # F.pad applies to last two dims: F_in then T_in
-        audio_encodings_padded = F.pad(audio_encodings, self.manual_padding, mode="constant", value=0.0)
+        audio_encodings_padded = F.pad(audio_encodings, self.manual_padding, mode="constant", value=0.0).to(
+            self.conv.weight.dtype
+        )
         # Expected padded shape for F_in, k_w=3, pad_F=(1,1) -> F_padded = F_in+2
         # Expected padded shape for T_in, k_h=3, pad_T=(0,2) -> T_padded = T_in+2
         audio_encodings_conv = self.conv(audio_encodings_padded)