Allow configuring and testing XLA in model export.

PiperOrigin-RevId: 723135244
google-research · Feb 4, 2025 · 827f0e1 · 827f0e1
1 parent ac6e047
commit 827f0e1
Show file tree

Hide file tree

Showing 3 changed files with 22 additions and 5 deletions.
diff --git a/chirp/train/classifier.py b/chirp/train/classifier.py
@@ -400,6 +400,7 @@ def export_tf_model(
     tf_lite_dtype: str = "float16",
     tf_lite_select_ops: bool = True,
     export_dir: str | None = None,
+    enable_xla: bool = False,
 ):
   """Export SavedModel and TFLite."""
   # Get model_ouput keys from output_head_metadatas and add the 'embedding' key
@@ -427,7 +428,7 @@ def infer_fn(audio_batch, variables):
     else:
       shape = (1,) + input_shape
     converted_model = export_utils.Jax2TfModelWrapper(
-        infer_fn, variables, shape, False
+        infer_fn, variables, shape, enable_xla=enable_xla
     )
     class_lists = {
         md.key: md.class_list for md in model_bundle.output_head_metadatas

diff --git a/chirp/train_tests/frontend_test.py b/chirp/train_tests/frontend_test.py
@@ -132,6 +132,7 @@ def test_inverse(self, module_type, inverse_module_type, module_kwargs):
               "freq_range": (60, 10_000),
           },
           "atol": 1e-4,
+          "enable_xla": False,
       },
       {
           "module_type": frontend.SimpleMelspec,
@@ -169,7 +170,12 @@ def test_inverse(self, module_type, inverse_module_type, module_kwargs):
       },
   )
   def test_tflite_stft_export(
-      self, module_type, module_kwargs, signal_shape=None, atol=1e-6
+      self,
+      module_type,
+      module_kwargs,
+      signal_shape=None,
+      atol=1e-6,
+      enable_xla=False,
   ):
     # Note that the TFLite stft requires power-of-two nfft, given by:
     # nfft = 2 * (features - 1).
@@ -182,7 +188,8 @@ def test_tflite_stft_export(
 
     tf_predict = tf.function(
         jax2tf.convert(
-            lambda signal: fe.apply(params, signal), enable_xla=False
+            lambda signal: fe.apply(params, signal),
+            enable_xla=enable_xla,
         ),
         input_signature=[
             tf.TensorSpec(shape=signal.shape, dtype=tf.float32, name="input")

diff --git a/chirp/train_tests/train_test.py b/chirp/train_tests/train_test.py
@@ -157,12 +157,20 @@ def test_config_structure(self):
         jax.tree_util.tree_structure(test_config.to_dict()),
     )
 
-  def test_export_model(self):
+  @parameterized.named_parameters(
+      # Note that b0 tests tend to timeout.
+      # ("xla", True, False),
+      ("no_xla", False, False),
+  )
+  def test_export_model(self, enable_xla, test_b0):
     # NOTE: This test might fail when run on a machine that has a GPU but when
     # CUDA is not linked (JAX will detect the GPU so jax2tf will try to create
     # a TF graph on the GPU and fail)
     config = self._get_test_config()
-    config = self._add_const_model_config(config)
+    if test_b0:
+      config = self._add_b0_model_config(config)
+    else:
+      config = self._add_const_model_config(config)
     config = self._add_pcen_melspec_frontend(config)
 
     model_bundle, train_state = classifier.initialize_model(
@@ -177,6 +185,7 @@ def test_export_model(self):
         config.init_config.input_shape,
         num_train_steps=0,
         eval_sleep_s=0,
+        enable_xla=enable_xla,
     )
     self.assertTrue(
         tf.io.gfile.exists(os.path.join(self.train_dir, "model.tflite"))