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

Speed up expert location update (#6661)

Browse Source
This commit is contained in:
fzyzcjy
2025-05-28 01:00:09 +08:00
committed by GitHub
parent 447be24228
commit 673ff668f7
1 changed files with 2 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
python/sglang/srt/model_executor/expert_location_updater.py
View File

@@ -12,7 +12,6 @@
# limitations under the License. # limitations under the License.
# ============================================================================== # ==============================================================================
import logging import logging
from datetime import timedelta
from typing import Dict, List, Tuple from typing import Dict, List, Tuple
import torch import torch
@@ -348,14 +347,8 @@ def update_expert_weights_single_layer(
return return
reqs = torch.distributed.batch_isend_irecv(p2p_ops) reqs = torch.distributed.batch_isend_irecv(p2p_ops)
try: for req in reqs:
for req in reqs: req.wait()
req.wait(timeout=timedelta(seconds=30))
except RuntimeError:
logger.error(
f"Context: {rank=} {old_physical_to_logical_map=} {new_physical_to_logical_map=} {num_local_physical_experts=} {num_gpu_per_node=}"
)
raise
def _execute_buffer2weight_copies(buffer2weight_copy_infos): def _execute_buffer2weight_copies(buffer2weight_copy_infos):
for ( for (