[fp8] hotfix backward hook (#6053)

* [fp8] hotfix backward hook

* [fp8] hotfix pipeline loss accumulation

colossalai/initialize.py

@@ -9,6 +9,7 @@ import os
 # https://forums.developer.nvidia.com/t/how-many-streams-maximum-number-of-streams/6571/16
 os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "1"
 
+import torch
 import torch.distributed as dist
 
 from colossalai.accelerator import get_accelerator
@@ -64,6 +65,11 @@ def launch(
 
     set_seed(seed)
 
+    try:
+        torch._dynamo.config.optimize_ddp = world_size > 1
+    except AttributeError:
+        pass
+
     if verbose:
         logger = get_dist_logger()
         logger.info(f"Distributed environment is initialized, world size: {dist.get_world_size()}", ranks=[0])