no causal mask when autoregressively predicting

tft/module.py

@@ -223,7 +223,11 @@ class TemporalFusionDecoder(nn.Module):
         return mask
 
     def forward(
-        self, x: torch.Tensor, static: torch.Tensor, mask: Optional[torch.Tensor] = None
+        self,
+        x: torch.Tensor,
+        static: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        causal: bool = True,
     ) -> torch.Tensor:
         expanded_static = static.expand_as(x)
         # static.repeat((1, self.context_length + self.prediction_length, 1))
@@ -242,7 +246,7 @@ class TemporalFusionDecoder(nn.Module):
             key=query_key_value,
             value=query_key_value,
             # key_padding_mask=key_padding_mask,
-            attn_mask=self.attn_mask,
+            attn_mask=self.attn_mask if causal else None,
         )
         att = self.att_net(attn_output)
 
@@ -571,7 +575,7 @@ class TFTModel(nn.Module):
         states = [c_h.unsqueeze(0), c_c.unsqueeze(0)]
 
         enc_out = self.temporal_encoder(past_selection, tgt_input=None, states=states)
-        dec_output = self.temporal_decoder(enc_out, static_enrichment)
+        dec_output = self.temporal_decoder(enc_out, static_enrichment, causal=False)
         params = self.param_proj(dec_output)
 
         distr = self.output_distribution(params, scale=scale, trailing_n=1)