diff --git a/internlm/core/scheduler/pipeline_scheduler.py b/internlm/core/scheduler/pipeline_scheduler.py index 055bb92..a9955a2 100644 --- a/internlm/core/scheduler/pipeline_scheduler.py +++ b/internlm/core/scheduler/pipeline_scheduler.py @@ -1342,7 +1342,7 @@ class InterleavedPipelineScheduler(PipelineScheduler): accum_moe_loss = self._accum_moe_loss accum_loss = self._accum_loss - if return_loss: + if accum_loss is not None: accum_loss += self._accum_moe_loss self._clear_state()