Avoid reuse of polymorphic variable names in tf2jax internal.

PiperOrigin-RevId: 563039856
google-deepmind · Sep 7, 2023 · 0f0eb93 · 0f0eb93
1 parent 749b21b
commit 0f0eb93
Show file tree

Hide file tree

Showing 2 changed files with 25 additions and 7 deletions.
diff --git a/tf2jax/_src/roundtrip_test.py b/tf2jax/_src/roundtrip_test.py
@@ -744,6 +744,13 @@ def forward(x, w):
     tf_outputs2 = concrete_tf_fn2(x, w)
     self.assertAllClose(expected_outputs, tf_outputs2)
 
+    # JAX -> TF -> JAX -> TF -> SavedModel
+    module = tf.Module()
+    module.fn = tf_fn2
+    tmp_dir = self.create_tempdir()
+    options = tf.saved_model.SaveOptions(experimental_custom_gradients=True)
+    tf.saved_model.save(module, tmp_dir.full_path, options=options)
+
   @chex.variants(with_jit=True)
   @parameterized.named_parameters(
       chex.params_product(
@@ -796,6 +803,13 @@ def forward(x, y):
     tf_outputs2 = concrete_tf_fn2(x, y)
     self.assertAllClose(expected_outputs, tf_outputs2)
 
+    # JAX -> TF -> JAX -> TF -> SavedModel
+    module = tf.Module()
+    module.fn = tf_fn2
+    tmp_dir = self.create_tempdir()
+    options = tf.saved_model.SaveOptions(experimental_custom_gradients=True)
+    tf.saved_model.save(module, tmp_dir.full_path, options=options)
+
   @chex.variants(with_jit=True)
   @parameterized.named_parameters(
       chex.params_product(

diff --git a/tf2jax/experimental/mhlo.py b/tf2jax/experimental/mhlo.py
@@ -96,23 +96,24 @@ def mhlo_apply_abstract_eval(
     symtab = ir.SymbolTable(mhlo_module.operation)
 
     # Check we are not reusing existing dimension vars.
+    dynamic_count = 0
     has_polymorphic = False
     for val in in_avals:
       for dim in val.shape:
         if not isinstance(dim, int):
           has_polymorphic = True
           if any(x.startswith(_UKNOWN_DIM_PREFIX) for x in dim.get_vars()):
-            raise ValueError(
-                "Polymorphic variable name that start with"
-                f" `{_UKNOWN_DIM_PREFIX}` are reserved for use by tf2jax"
-                f" internal for outputs: `{val.shape}`"
-            )
+            for dim in dim.get_vars():
+              if dim.startswith(_UKNOWN_DIM_PREFIX):
+                dynamic_count = max(
+                    dynamic_count,
+                    (int(dim.removeprefix(_UKNOWN_DIM_PREFIX + "_"))),
+                )
 
     # Map each `dynamic`` dimension to a unique dimension variable because we
     # do not have the information from the avals of the original JAX function.
     # In practice, the output shapes may actually be much more constrained, but
     # the information is not available here.
-    dynamic_count = 0
     output_specs = []
     for res in symtab["main"].type.results:
       if any(dim == res.get_dynamic_size() for dim in res.shape):
@@ -124,7 +125,10 @@ def mhlo_apply_abstract_eval(
         )
 
         assert has_polymorphic, has_polymorphic
-        from jax.experimental.jax2tf import shape_poly  # pylint: disable=g-import-not-at-top  # pytype: disable=import-error
+        if jax.__version_info__ <= (0, 4, 14):
+          from jax.experimental.jax2tf import shape_poly  # pylint: disable=g-import-not-at-top  # pytype: disable=import-error
+        else:
+          from jax.experimental.export import shape_poly  # pylint: disable=g-import-not-at-top  # pytype: disable=import-error
         out_shape = shape_poly._parse_spec(out_shape, res.shape)  # pylint: disable=protected-access
       else:
         out_shape = res.shape