Np parall fix

examples/torch_native_parallelism/nd_parallel.py

@@ -31,6 +31,7 @@
     PerformanceTracker,
     create_collate_fn,
     get_dataset,
+    get_model_flops_per_token,
     setup_tokenizer,
 )
 
@@ -73,7 +74,7 @@ def forward(model, batch, optimizer, accelerator: Accelerator):
         loss = outputs.loss
         accelerator.backward(loss)
         optimizer.step()
-        optimizer.zero_grad()
+        optimizer.zero_grad(set_to_none=False)
         dist.all_reduce(loss, op=dist.ReduceOp.AVG, group=loss_reduce_grp)
 
     return loss
@@ -123,6 +124,7 @@ def train(args):
 
     total_num_steps = min(args.num_steps, len(dataloader))
     performance_tracker = PerformanceTracker(warmup_steps=5)
+    model_flops_per_token = get_model_flops_per_token(model, args.sequence_length)
 
     accelerator.print("Starting training...")
     for step, batch in enumerate(dataloader):
@@ -132,7 +134,9 @@ def train(args):
         loss = forward(model, batch, optimizer, accelerator)
 
         # We report TPS per device, so we divide by the number of devices in the non-data parallel dimension
-        metrics = performance_tracker.step(batch["input_ids"].shape[1] / parallelism_config.non_data_parallel_size)
+        metrics = performance_tracker.step(
+            batch["input_ids"].shape[1] / parallelism_config.non_data_parallel_size, model_flops_per_token
+        )
 
         print_msg = f"Step {step}/{total_num_steps}, Loss: {loss.item():.4f}"
         if "warmup_completed" in metrics:

src/accelerate/accelerator.py

@@ -1623,7 +1623,14 @@ def _get_tensor_address(p):
                     # so that the optimizer can correctly update the model parameters.
                     param_group["params"] = [mapping[_get_tensor_address(p)] for p in param_group["params"]]
 
-        return args
+        for item in result:
+            if any(
+                item in container
+                for container in (self._dataloaders, self._models, self._optimizers, self._schedulers)
+            ):
+                item._is_accelerate_prepared = True
+
+        return result
 
     def _prepare_cp(self, *args):
         from torch.distributed.tensor.experimental import context_parallel