[RL] Update worker_process.py (#5651)

This commit is contained in:
Yuanle Liu
2025-12-19 12:07:58 +08:00
committed by GitHub
parent a8fce47195
commit 689f54f671

View File

@@ -421,7 +421,8 @@ class PaddleDisWorkerProc:
if self.fd_config.load_config.dynamic_load_weight:
if self.model_weights_status.value[0] != ModelWeightsStatus.NORMAL:
self.model_weights_signal[0] = int(self.model_weights_status.value[0])
self.model_weights_signal[0] = self._broadcast_model_weights_signal(src=0, group=None)
if self.ranks > 1:
self.model_weights_signal[0] = self._broadcast_model_weights_signal(src=0, group=None)
req_dicts = None
self.worker_healthy_live_signal.value[tp_rank % self.max_chips_per_node] = int(time.time())
@@ -441,7 +442,8 @@ class PaddleDisWorkerProc:
self._tp_barrier_wait() if tp_size > 1 else None
if self.fd_config.load_config.dynamic_load_weight:
paddle.distributed.barrier()
if self.ranks > 1:
paddle.distributed.barrier()
if self.model_weights_signal[0] != ModelWeightsStatus.NORMAL:
logger.info(
f"Rank: {self.local_rank} to update or clear parameters, signal is {self.model_weights_signal[0]}, [-1:clear, 1:update]"