Cache min / max seq len for torch.nested.as_nested_tensor(t) (pytorch#130766)

For the `torch.nested.as_nested_tensor(t)` constructor, computing min / max seq len is trivial since the sequence lengths are all the same. Might as well cache them during construction.
Pull Request resolved: pytorch#130766
Approved by: https://github.com/YuqingJ, https://github.com/soulitzer

Author: jbschlosser

test/test_nestedtensor.py

@@ -4512,9 +4512,8 @@ def test_as_nested_tensor_from_tensor(
             dim=dim,
             batch_size=expected_batch_size,
             contiguous=True,
-            # TODO: compute min / max during construction for this case since it's easy
-            cached_min_seqlen=None,
-            cached_max_seqlen=None,
+            cached_min_seqlen=expected_seqlen,
+            cached_max_seqlen=expected_seqlen,
         )
 
         if torch.device(device) == t.device and dtype == t.dtype and contiguous:
@@ -4549,9 +4548,8 @@ def test_as_nested_tensor_from_tensor(
             dim=dim,
             batch_size=expected_batch_size,
             contiguous=True,
-            # TODO: compute min / max during construction for this case since it's easy
-            cached_min_seqlen=None,
-            cached_max_seqlen=None,
+            cached_min_seqlen=expected_seqlen,
+            cached_max_seqlen=expected_seqlen,
         )
 
         # we don't support conversion between layouts this way atm

torch/nested/__init__.py

@@ -119,7 +119,9 @@ def as_nested_tensor(
 
             from torch.nested._internal.nested_tensor import nested_view_from_values_offsets
 
-            return nested_view_from_values_offsets(values, offsets)
+            return nested_view_from_values_offsets(
+                values, offsets, min_seqlen=seq_len, max_seqlen=seq_len
+            )
         else:
             from torch.nested._internal.nested_tensor import jagged_from_list