5028ed1b6bedd526dee27ea731284f43e87303f0,fairseq/trainer.py,Trainer,train_step,#Trainer#Any#Any#,271

Before Change

        metrics.log_scalar("oom", ooms, len(samples), priority=600, round=3)
        if ooms == self.args.distributed_world_size * len(samples):
            logger.warning("OOM in all workers, skipping update")
            self.zero_grad()
            return None

        try:

After Change

        if is_dummy_batch:
            sample_size *= 0.  // multiply by 0 to preserve device
        if torch.is_tensor(sample_size):
            sample_size = sample_size.float()
        else:
            sample_size = float(sample_size)

        // gather logging outputs from all replicas
        if self._sync_stats():
            logging_outputs, (sample_size, ooms) = self._aggregate_logging_outputs(
                logging_outputs, sample_size, ooms, ignore=is_dummy_batch,
            )

Instances