[fp8]update reduce-scatter test (#6002)

* fix

* fix

* fix

* fix

tests/test_fp8/test_fp8_reduce_scatter.py

@@ -13,14 +13,20 @@ from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
 @parameterize("scatter_dim", [0, 1, 2])
 @parameterize("dtype", [torch.bfloat16, torch.float16])
 @parameterize("fp8_format", ["e4m3", "e5m2"])
-def check_4gpu(shape, scatter_dim, dtype, fp8_format):
+@parameterize("async_op", [True, False])
+def check_4gpu(shape, scatter_dim, dtype, fp8_format, async_op):
     x = torch.rand(shape, dtype=dtype, device=get_accelerator().get_current_device())
     input_list = list(torch.chunk(x, dim=scatter_dim, chunks=4))
     input_list = [t.contiguous() for t in input_list]
     output_origin = torch.empty_like(input_list[0])
     output_fp8 = torch.empty_like(input_list[0])
-    reduce_scatter(output_origin, input_list, group=_get_default_group())
+    origin_handle = reduce_scatter(output_origin, input_list, group=_get_default_group(), async_op=async_op)
-    reduce_scatter_fp8(output_fp8, input_list, group=_get_default_group(), fp8_format=fp8_format)
+    fp8_handle = reduce_scatter_fp8(
+        output_fp8, input_list, group=_get_default_group(), fp8_format=fp8_format, async_op=async_op
+    )
+    if async_op:
+        origin_handle.wait()
+        fp8_handle.wait()
     assert_close(output_origin, output_fp8, rtol=0.1, atol=0.1)