internal

init2winit/main.py

@@ -153,6 +153,11 @@
     'optax_training_algorithm',
     'Name of the training algorithm to use.',
 )
+flags.DEFINE_boolean(
+    'compile_init_on_cpu',
+    False,
+    'Whether or not to compile the init function on CPU or on device.',
+)
 
 FLAGS = flags.FLAGS
 
@@ -208,6 +213,7 @@ def _run(
     callback_configs,
     external_checkpoint_path,
     training_algorithm_name,
+    compile_init_on_cpu,
 ):
   """Function that runs a Jax experiment. See flag definitions for args."""
   model_cls = models.get_model(model_name)
@@ -246,7 +252,13 @@ def _run(
   rng = jax.random.PRNGKey(rng_seed)
 
   # Build the loss_fn, metrics_bundle, and flax_module.
-  model = model_cls(merged_hps, dataset_meta_data, loss_name, metrics_name)
+  model = model_cls(
+      merged_hps,
+      dataset_meta_data,
+      loss_name,
+      metrics_name,
+      compile_init_on_cpu=compile_init_on_cpu,
+  )
   trial_dir = os.path.join(experiment_dir, str(worker_id))
   meta_data_path = os.path.join(trial_dir, 'meta_data.json')
   meta_data = {'worker_id': worker_id, 'status': 'incomplete'}
@@ -373,6 +385,7 @@ def main(unused_argv):
         callback_configs=callback_configs,
         external_checkpoint_path=FLAGS.external_checkpoint_path,
         training_algorithm_name=FLAGS.training_algorithm,
+        compile_init_on_cpu=FLAGS.compile_init_on_cpu,
     )
 
 

init2winit/main_config_flags.py

@@ -130,6 +130,7 @@ def _run(
     external_checkpoint_path,
     training_algorithm_name,
     checkpoint_ttl,
+    compile_init_on_cpu,
 ):
   """Function that runs a Jax experiment. See flag definitions for args."""
   model_cls = models.get_model(model_name)
@@ -168,7 +169,13 @@ def _run(
   rng = jax.random.PRNGKey(rng_seed)
 
   # Build the loss_fn, metrics_bundle, and flax_module.
-  model = model_cls(merged_hps, dataset_meta_data, loss_name, metrics_name)
+  model = model_cls(
+      merged_hps,
+      dataset_meta_data,
+      loss_name,
+      metrics_name,
+      compile_init_on_cpu=compile_init_on_cpu,
+  )
   trial_dir = os.path.join(experiment_dir, str(worker_id))
   meta_data_path = os.path.join(trial_dir, 'meta_data.json')
   meta_data = {'worker_id': worker_id, 'status': 'incomplete'}
@@ -307,6 +314,7 @@ def main(unused_argv):
         external_checkpoint_path=config.external_checkpoint_path,
         training_algorithm_name=config.training_algorithm,
         checkpoint_ttl=config.ttl,
+        compile_init_on_cpu=config.compile_init_on_cpu,
     )
 
 

init2winit/model_lib/base_model.py

@@ -142,13 +142,21 @@ class BaseModel(object):
   https://flax.readthedocs.io/en/latest/flax.linen.html#flax.linen.Module.apply.
   """
 
-  def __init__(self, hps, dataset_meta_data, loss_name, metrics_name):
+  def __init__(
+      self,
+      hps,
+      dataset_meta_data,
+      loss_name,
+      metrics_name,
+      compile_init_on_cpu=False,
+  ):
     self.hps = hps
     self.dataset_meta_data = dataset_meta_data
     self._loss_name = loss_name
     self.loss_fn = losses.get_loss_fn(loss_name, hps)
     self.output_activation_fn = losses.get_output_activation_fn(loss_name)
     self.metrics_bundle = metrics.get_metrics(metrics_name, hps)
+    self._compile_init_on_cpu = compile_init_on_cpu
     self.flax_module = self.build_flax_module()
 
   def initialize(self, initializer, hps, rng, metrics_logger):
@@ -200,10 +208,17 @@ def initialize(self, initializer, hps, rng, metrics_logger):
     # construction.
     # We initialize model params on host to avoid memory issues.
 
+    jit_kwargs = {}
+    if self._compile_init_on_cpu:
+      jit_kwargs['backend'] = 'cpu'
+    logging.info(
+        'Compiling model init on %s.',
+        'cpu' if self._compile_init_on_cpu else 'device',
+    )
     start_time = time.time()
     model_init_fn = jax.jit(
-        functools.partial(self.flax_module.init, train=False),
-        backend='cpu')
+        functools.partial(self.flax_module.init, train=False), **jit_kwargs
+    )
 
     init_dict = model_init_fn({'params': params_rng, 'dropout': dropout_rng},
                               *fake_input_batch)

init2winit/model_lib/nqm.py

@@ -177,15 +177,21 @@ class NQM(base_model.BaseModel):
   generates isotropic Gaussian noise.
   """
 
-  def __init__(self, hps, dataset_meta_data, loss_name, metrics_name):
+  def __init__(
+      self,
+      hps,
+      dataset_meta_data,
+      loss_name,
+      metrics_name,
+      compile_init_on_cpu=False,
+  ):
     del loss_name
-
     # This is ignored, but is needed to satisfy the initializer API.
     self.loss_fn = None
     self.metrics_name = metrics_name
-
     self.hps = hps
     self.dataset_meta_data = dataset_meta_data
+    self._compile_init_on_cpu = compile_init_on_cpu
     self.flax_module = self.build_flax_module()
 
   def evaluate_batch(self, params, batch_stats, batch):