EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[fix][RL] Remove the incorrect barrier in init_weights_update_group (#5914)

Browse Source
This commit is contained in:
Zilin Zhu
2025-05-15 10:15:21 +08:00
committed by GitHub
parent f3bf611054
commit 44a3783d13
2 changed files with 2 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
python/sglang/srt/model_executor/model_runner.py
View File

@@ -631,7 +631,6 @@ class ModelRunner:
rank=rank, rank=rank,
group_name=group_name, group_name=group_name,
) )
dist.barrier(group=self._model_update_group, device_ids=[rank])
return True, "Succeeded to initialize custom process group." return True, "Succeeded to initialize custom process group."
except Exception as e: except Exception as e:
message = f"Failed to initialize custom process group: {e}." message = f"Failed to initialize custom process group: {e}."

5
test/srt/test_update_weights_from_distributed.py
View File

@@ -162,7 +162,6 @@ def init_process_hf(
rank=rank, rank=rank,
group_name="test_parameter_update_group", group_name="test_parameter_update_group",
) )
dist.barrier(group=group, device_ids=[rank])
torch.cuda.synchronize() torch.cuda.synchronize()
time_begin_broadcast = time.perf_counter() time_begin_broadcast = time.perf_counter()
@@ -223,8 +222,8 @@ def init_process_sgl(
if rank == 1: if rank == 1:
url = DEFAULT_URL_FOR_TEST url = DEFAULT_URL_FOR_TEST
else: else:
host, port = DEFAULT_URL_FOR_TEST.split(":") host, _, port = DEFAULT_URL_FOR_TEST.rpartition(":")
url = ":".join(host, str(int(port) + 10000)) url = ":".join([host, str(int(port) + 10000)])
print(f"[sgl] rank {rank} init server on url: {url}") print(f"[sgl] rank {rank} init server on url: {url}")
process = popen_launch_server( process = popen_launch_server(