Replace deprecated jax.tree_* functions with jax.tree.*

The top-level `jax.tree_*` aliases have long been deprecated, and will soon be removed. Alternate APIs are in `jax.tree_util`, with shorter aliases in the `jax.tree` submodule, added in JAX version 0.4.25.

PiperOrigin-RevId: 634095268

praxis/layers/adapters_test.py

@@ -87,7 +87,7 @@ def test_residual_adapter_tf_equivalent(self):
     theta = initial_vars['params'].copy()
     theta['layer_norm'] = theta['norm']
     del theta['norm']
-    theta = jax.tree_map(np.array, theta)
+    theta = jax.tree.map(np.array, theta)
     theta = py_utils.NestedMap.FromNestedDict(theta)
     theta.down_w = tf.convert_to_tensor(theta.down_w)
     theta.up_w = tf.convert_to_tensor(theta.up_w)

praxis/layers/attentions.py

@@ -2247,12 +2247,12 @@ def _vmap_on_broadcast_prefixes(
 
     # Wraps fn with slicing on args_to_slice and broadcast_args_to_slice.
     def _sliced_fn(layer, args, args_to_slice, broadcast_args_to_slice, states):
-      sliced = jax.tree_map(
+      sliced = jax.tree.map(
           lambda x, d: self._slice_decode_chunk(x, chunk_id, d),
           args_to_slice,
           args_time_dims,
       )
-      broadcast_sliced = jax.tree_map(
+      broadcast_sliced = jax.tree.map(
           lambda x, d: self._slice_decode_chunk(x, chunk_id, d),
           broadcast_args_to_slice,
           broadcast_args_time_dims,

praxis/layers/convolutions_test.py

@@ -141,7 +141,7 @@ def test_causal_conv2d_layer(self):
 
     prng_key = jax.random.PRNGKey(seed=123)
     initial_vars = conv_layer.init(prng_key, inputs)
-    initial_vars = jax.tree_map(jnp.ones_like, initial_vars)
+    initial_vars = jax.tree.map(jnp.ones_like, initial_vars)
 
     # Test odd length sequence.
     output = conv_layer.apply(initial_vars, inputs)
@@ -156,7 +156,7 @@ def test_causal_conv2d_layer(self):
 
     prng_key = jax.random.PRNGKey(seed=123)
     initial_vars = conv_layer.init(prng_key, inputs)
-    initial_vars = jax.tree_map(jnp.ones_like, initial_vars)
+    initial_vars = jax.tree.map(jnp.ones_like, initial_vars)
 
     output = conv_layer.apply(initial_vars, inputs)
     np_output = np.array(

praxis/layers/flax_adapter_test.py

@@ -149,14 +149,14 @@ def test_mix_layer(self):
       def assert_learnable(x):
         assert not x.collections
 
-      jax.tree_map(assert_learnable, init_var_meta['params'])
+      jax.tree.map(assert_learnable, init_var_meta['params'])
 
       def assert_non_learnable(x):
         assert WeightHParamsCollection.NON_TRAINABLE in x.collections
         assert WeightHParamsCollection.REQUIRES_MEAN_SYNC in x.collections
 
-      jax.tree_map(assert_non_learnable, init_var_meta['batch_stats'])
-      jax.tree_map(assert_non_learnable, init_var_meta['non_trainable'])
+      jax.tree.map(assert_non_learnable, init_var_meta['batch_stats'])
+      jax.tree.map(assert_non_learnable, init_var_meta['non_trainable'])
       init_vars = test_layer.init(prng_key, input_x)
       _ = test_layer.apply(init_vars, input_x, mutable=True)
       _ = test_layer.apply(

praxis/layers/frnn.py

@@ -78,7 +78,7 @@ def reset_mask(
 
 
 def _sum_aux_loss(tree):
-  return jax.tree_map(jnp.sum, tree)
+  return jax.tree.map(jnp.sum, tree)
 
 
 class FRnn(base_layer.BaseLayer):
@@ -143,7 +143,7 @@ def __call__(
       state: Final state.
     """
     # Make a copy of the input structure to avoid side-effect.
-    inputs = jax.tree_map(lambda x: x, inputs)
+    inputs = jax.tree.map(lambda x: x, inputs)
     assert hasattr(inputs, 'act')
     assert hasattr(inputs, 'padding')
     assert isinstance(self.cell, rnn_cell.BaseRnnCell)
@@ -159,7 +159,7 @@ def __call__(
       inputs.reset_mask = jnp.ones_like(inputs.padding, dtype=self.fprop_dtype)
 
     if self.reverse:
-      inputs = jax.tree_map(lambda x: jnp.flip(x, axis=[1]), inputs)
+      inputs = jax.tree.map(lambda x: jnp.flip(x, axis=[1]), inputs)
 
     if not state0:
       batch_size = inputs.padding.shape[0]
@@ -176,7 +176,7 @@ def body_fn(sub, state0, inputs):
     if self.is_initializing():
       # inputs has shape [b, t, dim] or [b, t, 1]
       # sliced_inputs has shape [b, dim] or [b, 1].
-      sliced_inputs = jax.tree_map(lambda x: x[:, 1], inputs)
+      sliced_inputs = jax.tree.map(lambda x: x[:, 1], inputs)
       _ = body_fn(self.cell, state0, sliced_inputs)
 
     # NON_TRAINABLE variables are carried over from one iteration to another.
@@ -248,7 +248,7 @@ def init_states(self, batch_size: int) -> list[NestedMap]:
   def extend_step(
       self, inputs: NestedMap, state: list[NestedMap]
   ) -> tuple[list[NestedMap], JTensor]:
-    inputs = jax.tree_map(lambda x: x, inputs)
+    inputs = jax.tree.map(lambda x: x, inputs)
     new_states = []
     for i in range(self.num_layers):
       new_state, act_i = self.frnn[i].extend_step(inputs, state[i])
@@ -275,7 +275,7 @@ def __call__(
       act: A tensor of [batch, time, dims]. The output.
       state: Final state.
     """
-    inputs = jax.tree_map(lambda x: x, inputs)
+    inputs = jax.tree.map(lambda x: x, inputs)
 
     if not state0:
       batch_size = inputs.padding.shape[0]
@@ -375,7 +375,7 @@ def __call__(
       state: Final state - a list of NestedMap of fwd and bwd states.
     """
     # This is to create a copy.
-    inputs = jax.tree_map(lambda x: x, inputs)
+    inputs = jax.tree.map(lambda x: x, inputs)
 
     if not state0:
       batch_size = inputs.padding.shape[0]
@@ -428,7 +428,7 @@ def __call__(
       state: Final state.
     """
     # Make a copy of the input structure to avoid side-effect.
-    inputs = jax.tree_map(lambda x: x, inputs)
+    inputs = jax.tree.map(lambda x: x, inputs)
     assert hasattr(inputs, 'act')
     assert hasattr(inputs, 'padding')
     assert isinstance(self.cell, rnn_cell.BaseRnnCell)
@@ -444,7 +444,7 @@ def __call__(
       inputs.reset_mask = jnp.ones_like(inputs.padding, dtype=self.fprop_dtype)
 
     if self.reverse:
-      inputs = jax.tree_map(lambda x: jnp.flip(x, axis=[1]), inputs)
+      inputs = jax.tree.map(lambda x: jnp.flip(x, axis=[1]), inputs)
 
     if not state0:
       batch_size = inputs.padding.shape[0]
@@ -466,7 +466,7 @@ def body_fn(sub, state0, inputs):
     if self.is_initializing():
       # inputs has shape [b, t, dim] or [b, t, 1]
       # sliced_inputs has shape [b, dim] or [b, 1].
-      sliced_inputs = jax.tree_map(lambda x: x[:, 1], inputs)
+      sliced_inputs = jax.tree.map(lambda x: x[:, 1], inputs)
       # `body_fn` is sufficient to trigger PARAMS initialization.
       _ = body_fn(self.cell, state0, sliced_inputs)
 

praxis/layers/frnn_test.py

@@ -206,7 +206,7 @@ def test_frnn_lstm_cell(self, jax_cell_class, output_nonlinearity):
 
     rnn_theta = {'params': theta['params']['cell']}
     ys = []
-    cell_state = jax.tree_map(lambda x: x, state0)
+    cell_state = jax.tree.map(lambda x: x, state0)
     for t in range(act_in.shape[1]):
       with base_layer.JaxContext.new_context():
         inputs_t = NestedMap(act=act_in[:, t], padding=padding[:, t])
@@ -414,7 +414,7 @@ def test_frnn_reset_cell_state(
 
     rnn_theta = {'params': theta['params']['cell']}
     ys = []
-    cell_state = jax.tree_map(lambda x: x, state0)
+    cell_state = jax.tree.map(lambda x: x, state0)
     for t in range(act_in.shape[1]):
       with base_layer.JaxContext.new_context():
         inputs_t = NestedMap(

praxis/layers/multi_query_attention.py

@@ -1319,12 +1319,16 @@ def _vmap_on_broadcast_prefixes(self, fn: attentions.FnOnDecodeStateChunk,
 
     # Wraps fn with slicing on args_to_slice and broadcast_args_to_slice.
     def _sliced_fn(layer, args, args_to_slice, broadcast_args_to_slice, states):
-      sliced = jax.tree_map(
-          lambda x, d: self._slice_decode_chunk(x, chunk_id, d), args_to_slice,
-          args_time_dims)
-      broadcast_sliced = jax.tree_map(
+      sliced = jax.tree.map(
           lambda x, d: self._slice_decode_chunk(x, chunk_id, d),
-          broadcast_args_to_slice, broadcast_args_time_dims)
+          args_to_slice,
+          args_time_dims,
+      )
+      broadcast_sliced = jax.tree.map(
+          lambda x, d: self._slice_decode_chunk(x, chunk_id, d),
+          broadcast_args_to_slice,
+          broadcast_args_time_dims,
+      )
       return fn(layer, args, sliced, broadcast_sliced, states)
 
     broadcast_dim_sizes = self.get_decode_state(