a615533788c1842483a9708787db0d73902dc1ec,fairseq/multiprocessing_trainer.py,MultiprocessingTrainer,_scatter_samples,#MultiprocessingTrainer#Any#Any#,242

Before Change

        events = []
        for d in self.device_ids:
            with torch.cuda.device(d):
                event = torch.cuda.Event(interprocess=True)
                event.record()
                events.append(event)

        return res, events

After Change

        // Pad with None until its size is equal to the number of replicas.
        samples = samples + [None]*(self.num_replicas - len(samples))

        Future.gen_list([
            self.call_async(rank, "_async_prepare_sample", sample=samples[rank], volatile=volatile)
            for rank in range(self.num_replicas)
        ])

    def _async_prepare_sample(self, rank, device_id, sample, volatile):
        if sample is None:

Instances