Replace torch.jit.instance with instance to assist torch.compile (#522)

Summary:
Pull Request resolved: #522

torch.jit.instance is mainly needed to refine Generics, but in this case instance is sufficient, instance unlike torch.jit.instance is also playing nicely with torch.compile

Reviewed By: ebsmothers

Differential Revision: D53797660

fbshipit-source-id: b8e4c5542be866bb09dd331ece4725ae8f57dfcf

torchmultimodal/modules/layers/transformer.py

@@ -381,9 +381,8 @@ def _cross_attention_block(
             # TODO: figure out caching for cross-attention
             use_cache=False,
         )
-        assert torch.jit.isinstance(
-            output, Tensor
-        ), "cross-attention output must be Tensor."
+        assert isinstance(output, Tensor), "cross-attention output must be Tensor."
+
         attention_output = self.cross_attention_dropout(output)
         return attention_output