Jj/cumsum nvfuserex opinfo tolerance

thunder/tests/opinfos.py

@@ -6229,6 +6229,20 @@ def cumsum_sample_generator(op, device, dtype, requires_grad, **kwargs):
             dtypes=(datatypes.float16,),
             devicetypes=(devices.DeviceType.CPU,),
         ),
+        # nvfuserex follows pytorch convention to run cumsum in reduced
+        # precision, this causes opinfo tests numerical mismatch for bf16/fp16
+        # NOTE: Even though both nvfuserex and torch uses reduced precision
+        # math, because the reduction order is not the same due to
+        # implementation, error would accumulate.
+        DecorateInfo(
+            custom_comparator(partial(assert_close, atol=1e-1, rtol=1e-1)),
+            "test_core_vs_torch_consistency",
+            dtypes=(
+                datatypes.bfloat16,
+                datatypes.float16,
+            ),
+            executors=("nvfuser",),
+        ),
     ),
 )
 reduction_ops.append(cumsum_opinfo)