fix: force synchronization between TP workers when update_weights (#6626)

Co-authored-by: dangkai.dk <dangkai.dk@alibaba-inc.com>

fix: force synchronization between TP workers when update_weights (#6626)
Co-authored-by: dangkai.dk <dangkai.dk@alibaba-inc.com>
bc2e5645 · DangKai · GitHub · 3abc3036 · bc2e5645
Unverified Commit bc2e5645 authored Jun 25, 2025 by DangKai Committed by GitHub Jun 25, 2025
Hide whitespace changes
Inline Side-by-side

Showing with 1 addition and 0 deletions

python/sglang/srt/managers/scheduler.py python/sglang/srt/managers/scheduler.py +1 -0

No files found.
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -2235,6 +2235,7 @@ class Scheduler(
                assert flash_cache_success, "Cache flush failed after updating weights"
        else:
            logger.error(message)
+        barrier(group=self.tp_cpu_group)
        return UpdateWeightsFromTensorReqOutput(success, message)

    def get_weights_by_name(self, recv_req: GetWeightsByNameReqInput):