3f9700868da4d09988b7dfbd7d5b3a9d5ea7d401,fairseq/multiprocessing_trainer.py,MultiprocessingTrainer,train_step,#MultiprocessingTrainer#Any#,127

Before Change

        grad_denom = self.criterion.grad_denom(samples)

        // forward pass, backward pass and gradient step
        losses = [
            self.call_async(rank, "_async_train_step", grad_denom=grad_denom)
            for rank in range(self.num_replicas)
        ]

        // aggregate losses and gradient norms
        loss_dicts = Future.gen_list(losses)
        loss_dict = self.criterion.aggregate(loss_dicts)
        loss_dict["gnorm"] = loss_dicts[0]["gnorm"]

        return loss_dict

After Change

        self._scatter_samples(samples, replace_empty_samples=replace_empty_samples)

        // forward pass
        sample_sizes, logging_outputs = Future.gen_tuple_list([
            self.call_async(rank, "_async_forward")
            for rank in range(self.num_replicas)
        ])

        // backward pass, all-reduce gradients and take an optimization step
        grad_denom = self.criterion.__class__.grad_denom(sample_sizes)
        grad_norms = Future.gen_list([

Instances