Function _report_progress

train.py:694–697

(self, step)

        if optim_config.get('gradient_release', False):
            # Prevent deepspeed from logging every single param group lr
            def _report_progress(self, step):
                lr = self.get_lr()
                mom = self.get_mom()
                deepspeed.utils.logging.log_dist(f"step={step}, skipped={self.skipped_steps}, lr={lr[0]}, mom={mom[0]}", ranks=[0])
            deepspeed.runtime.engine.DeepSpeedEngine._report_progress = _report_progress

            # Deepspeed executes all the code to reduce grads across data parallel ranks even if the DP world size is 1.