[benchmarks] Fix execution with AMP precision. (#6512)

benchmarks/benchmark_model.py

@@ -1,4 +1,5 @@
 from collections import OrderedDict
+import contextlib
 import logging
 import re
 import torch
@@ -66,6 +67,8 @@ def __init__(self, suite_name, model_name, benchmark_experiment):
     self.suite_name = suite_name
     self.model_name = model_name
     self.benchmark_experiment = benchmark_experiment
+    self.autocast = contextlib.nullcontext
+    self.autocast_kwargs = {}
 
   def set_up(self):
     """Set up module, actual batch_size, example_inputs, and optimizer_class
@@ -125,6 +128,7 @@ def pick_grad(self):
       return torch.no_grad()
     elif self.benchmark_experiment.test == "train":
       return torch.enable_grad()
+    raise NotImplementedError
 
   def _optimizer_zero_grad(self):
     if self.optimizer is not None:
@@ -141,8 +145,9 @@ def compute_loss(self, pred):
 
   def train(self, inputs, collect_full_output=False):
     self._optimizer_zero_grad()
-    pred = self.module(*inputs)
-    loss = self.compute_loss(pred)
+    with self.autocast(**self.autocast_kwargs):
+      pred = self.module(*inputs)
+      loss = self.compute_loss(pred)
     loss.backward()
     self._optimizer_step()
     if collect_full_output:
@@ -152,7 +157,8 @@ def train(self, inputs, collect_full_output=False):
     return None
 
   def eval(self, inputs, collect_full_output=False):
-    pred = self.module(*inputs)
+    with self.autocast(**self.autocast_kwargs):
+      pred = self.module(*inputs)
     return pred
 
   @property

benchmarks/experiment_runner.py

@@ -248,7 +248,7 @@ def run_single_config(self):
 
     # Repeat the experiment and accumulate metrics.
     last_output = None
-    with benchmark_model.pick_context():
+    with benchmark_model.pick_grad():
       accumulated_metrics = OrderedDict()
       for repeat_iteration in range(self._args.repeat):
         metrics, last_output = self.run_once_and_gather_metrics(

benchmarks/torchbench_model.py

@@ -7,11 +7,11 @@
 from os.path import abspath, exists
 import sys
 import torch
+import torch.amp
 import torch.nn as nn
 from torch._dynamo.testing import collect_results, reduce_to_scalar_loss
 from torch._dynamo.utils import clone_inputs
 import torch_xla
-import torch_xla.amp
 import torch_xla.core.xla_model as xm
 import types
 import yaml
@@ -269,11 +269,17 @@ def set_up(self):
 
     This is model suite specific.
     """
+    # Set the optimizer class.
+    # Check if we should use SGD instead of Adam for memory reasons.
     if self.benchmark_experiment.test == "train" and self.model_name in TRAIN_WITH_SGD:
       self.optimizer_class = torch.optim.SGD
     else:
       self.optimizer_class = torch.optim.Adam
 
+    # Setup the autocast environment if we are running on AMP precision.
+    self.autocast, self.autocast_kwargs = self._get_autocast_with_kwargs()
+
+    # Load the actual benchmark instance.
     benchmark = self.load_benchmark()
 
     self.module, self.example_inputs = benchmark.get_module()
@@ -417,26 +423,24 @@ def pick_grad(self):
     # special case
     if self.model_name in ("maml",):
       return torch.enable_grad()
+    return super().pick_grad()
 
-    if self.benchmark_experiment.test == "eval":
-      return torch.no_grad()
-    elif self.benchmark_experiment.test == "train":
-      return torch.enable_grad()
-
-  def pick_amp(self):
+  def _get_autocast_with_kwargs(self):
     if (self.benchmark_experiment.accelerator == "cuda" and
         self.is_cuda_precision_amp()):
+      kwargs = {"dtype": torch.bfloat16}
       if self.benchmark_experiment.xla:
-        return torch_xla.amp.autocast(xm.xla_device())
+        # Should call device specific autocast implementations.
+        # PyTorch/XLA autocast does not run with dynamo, though:
+        # https://github.com/pytorch/xla/issues/6511
+        autocast = torch.amp.autocast
+        kwargs["device_type"] = "xla"
       else:
-        return torch.cuda.amp.autocast()
-    return contextlib.nullcontext()
-
-  def pick_context(self):
-    stack = contextlib.ExitStack()
-    stack.enter_context(self.pick_amp())
-    stack.enter_context(self.pick_grad())
-    return stack
+        autocast = torch.cuda.amp.autocast
+    else:
+      kwargs = {}
+      autocast = contextlib.nullcontext
+    return (autocast, kwargs)
 
   def compute_loss(self, pred):
     """Reduce the output of a model to get scalar loss"""