Make inputs actually contiguously laid out in memory (#7072)

Author: jackzhxng

examples/models/llama3_2_vision/text_decoder/model.py

@@ -168,11 +168,22 @@ def get_example_inputs(self):
     def get_example_kwarg_inputs(self):
         # For export we must use the prefill versions of the
         # causal mask and input_pos.
+
+        # Make input_pos and mask contiguous in memory.
+        input_pos = self.input_pos[None, : self.n_tokens]
+        mask = self.causal_mask[None, : self.n_tokens]
+        contiguous_input_pos = torch.empty_like(
+            input_pos, memory_format=torch.contiguous_format
+        )
+        contiguous_input_pos.data.copy_(input_pos.data)
+        contiguous_mask = torch.empty_like(mask, memory_format=torch.contiguous_format)
+        contiguous_mask.data.copy_(mask.data)
+
         # Hardcoding # of tiles to be 2. image tokens per tile is 1601.
         if self.use_kv_cache:
             return {
-                "input_pos": self.input_pos[None, : self.n_tokens],
-                "mask": self.causal_mask[None, : self.n_tokens],
+                "input_pos": contiguous_input_pos,
+                "mask": contiguous_mask,
                 "encoder_input": torch.randn(
                     1, self.encoder_max_seq_len, self.model_.dim, dtype=self.dtype
                 ),