Don't assert for symbolic stride in dim_order_from_stride() (#15472)

Summary:
Curently dim_order_for_stride() checks if any stride is 0, and fails if
so. N7613577 shows a min repro from factorized joiner usecase with
symbolic stride, where this check fails: P2015309933
This diff skips the assert for symbolic strides.

Differential Revision: D85875885

Author: pssrawat

exir/tensor.py

@@ -67,12 +67,15 @@ def dim_order_from_stride(stride: Tuple[int]) -> Tuple[bytes]:
     Another example is: sizes = (1, 3, 1, 1) with strides = (3, 1, 3, 3), returned
     value is (0, 2, 3, 1)
     """
+    from torch.fx.experimental.symbolic_shapes import (
+        guard_or_false,
+        guard_size_oblivious,
+    )
+
     for _, s in enumerate(stride):
-        if s == 0:
+        if guard_or_false(s == 0):
             raise ValueError("0 in strides is not supported for ExecuTorch.")
 
-    from torch.fx.experimental.symbolic_shapes import guard_size_oblivious
-
     class K(NamedTuple):
         stride: int
 

exir/tests/TARGETS

@@ -385,6 +385,7 @@ python_unittest(
     deps = [
         "//caffe2:torch",
         "//executorch/exir:dim_order_utils",
+        "//executorch/exir:lib",
     ],
 )
 

exir/tests/test_dim_order_utils.py

@@ -8,6 +8,7 @@
 import unittest
 
 import torch
+from executorch.exir import to_edge_transform_and_lower
 from executorch.exir.dim_order_utils import get_dim_order, get_memory_format
 
 
@@ -27,3 +28,22 @@ def test_get_dim_order(self) -> None:
                 list(range(ndim)), get_dim_order(torch.contiguous_format, ndim)
             )
         self.assertEqual([0, 2, 3, 1], get_dim_order(torch.channels_last, 4))
+
+    def test_dim_order_from_stride(self):
+        class Test(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, t1, t2):
+                idx = torch.nonzero(t1).reshape(-1)
+                y = torch.index_select(t2, 0, idx)
+                return y
+
+        M = Test()
+        x = torch.tensor([0, 1, 1, 0, 1], dtype=torch.bool)
+        y = torch.randn(5, 6)
+        M(x, y)
+
+        expo_prog = torch.export.export_for_training(M, (x, y))
+        edge_prog = to_edge_transform_and_lower(expo_prog)
+        edge_prog.to_executorch()