[NewFeature]custom_allreduce support cudagraph recapture (#4305)

* custom_allreduce support cudagraph recapture * add shut_down/restart default group
2025-10-07 17:41:52 +08:00 · 2025-09-29 15:56:54 +08:00
parent 7c919070f7
commit 81959c7d88
7 changed files with 32 additions and 3 deletions
--- a/fastdeploy/distributed/communication.py
+++ b/fastdeploy/distributed/communication.py
@@ -42,6 +42,12 @@ def use_custom_allreduce(custom_all_reduce_max_bytes: int = 8192 * 1024):
    _TP_AR = CustomAllreduce(model_parallel_group, custom_all_reduce_max_bytes)


+def custom_ar_clear_ipc_handles():
+    global _TP_AR
+    if _TP_AR is not None:
+        _TP_AR.clear_ipc_handles()
+
+
 try:

    @paddle.jit.marker.unified