default _use_segment_sum_csr for dynamo (#1798)

Summary:
Pull Request resolved: #1798

Constanting heuristic for dynamo for now.
Dynamo can not pass it without concrete values of the batch compile time.

Reviewed By: joshuadeng

Differential Revision: D54966335

fbshipit-source-id: 1359b1bbc6e50061c65a312df66fc5b2bf01a8b1

Author: Ivan Kobzarev

torchrec/sparse/jagged_tensor.py

@@ -727,6 +727,10 @@ def _use_segment_sum_csr(stride_per_key: List[int]) -> bool:
     per segment that match performance between the kernel and PyTorch solution, to
     determine the threshold of when to use `segment_sum_csr`.
     """
+    if is_torchdynamo_compiling():
+        # dynamo symbolic shapes can not pass this condition without concrete stride values
+        return False
+
     elements_per_segment = sum(stride_per_key) / len(stride_per_key)
     segment_threshold = int(
         1.39771