Update model architecture: d_ff=1024, new weights from merged7.pt

modeling_chessbot.py

@@ -418,14 +418,16 @@ class AbsolutePositionalEncoder(nn.Module):
     def __init__(self, d_model):
         super(AbsolutePositionalEncoder, self).__init__()
         self.d_model = d_model
-        self.register_buffer('position', torch.arange(64).unsqueeze(1))
+        # Don't register as buffers since these are computed values
+        position = torch.arange(64).unsqueeze(1).float()
         
         positional_encoding = torch.zeros(1, 64, d_model)
         _2i = torch.arange(0, d_model, step=2).float()
-        positional_encoding[:, :, 0::2] = torch.sin(self.position / (10000 ** (_2i / d_model)))
-        positional_encoding[:, :, 1::2] = torch.cos(self.position / (10000 ** (_2i / d_model)))
+        positional_encoding[:, :, 0::2] = torch.sin(position / (10000 ** (_2i / d_model)))
+        positional_encoding[:, :, 1::2] = torch.cos(position / (10000 ** (_2i / d_model)))
         
-        self.register_buffer('positional_encoding', positional_encoding)
+        # Register as non-persistent buffer (won't be saved/loaded)
+        self.register_buffer('positional_encoding', positional_encoding, persistent=False)
     
     def forward(self, x):
         batch_size, _, _ = x.size()