NNX migration: modify the print_shardings_params to support NNX

Author: xibinliu

src/maxtext/utils/maxtext_utils.py

@@ -1653,26 +1653,41 @@ def print_shardings_params(params, params_sharding, mesh, logical_annotations=No
   """
   Print state shardings comparing Logical Definition vs Physical Result.
   """
-  if not hasattr(params, "params"):
-    params = {"params": params}
-  if not hasattr(params_sharding, "params"):
-    params_sharding = {"params": params_sharding}
-  if logical_annotations and not hasattr(logical_annotations, "params"):
-    logical_annotations = {"params": logical_annotations}
+  if not isinstance(params, nnx.State):
+    if not hasattr(params, "params"):
+      params = {"params": params}
+    if not hasattr(params_sharding, "params"):
+      params_sharding = {"params": params_sharding}
+    if logical_annotations and not hasattr(logical_annotations, "params"):
+      logical_annotations = {"params": logical_annotations}
 
   leaves_params, _ = jax.tree_util.tree_flatten_with_path(params)
   leaves_sharding, _ = jax.tree_util.tree_flatten_with_path(params_sharding)
-  leaves_logical, _ = jax.tree_util.tree_flatten_with_path(logical_annotations)
 
-  for (path, leaf_val), (_, leaf_sharding), (_, leaf_logical_val) in zip(leaves_params, leaves_sharding, leaves_logical):
-    path_str = "/".join(str(p.key if hasattr(p, "key") else p.name) for p in path)
-    shape = jax.typeof(leaf_val)
-    pspec = sharding.remove_size_one_mesh_axis(leaf_sharding.spec, mesh)
-    pspec_str = str(tuple(pspec))
-    logical_str = str(leaf_logical_val)
-
-    message = f" {path_str}\n" f"    Shape:     {shape}\n" f"    Logical:   {logical_str}\n" f"    Physical:  {pspec_str}"
-    max_logging.info(message)
+  if logical_annotations is not None:
+    leaves_logical, _ = jax.tree_util.tree_flatten_with_path(logical_annotations)
+    for (path, leaf_val), (_, leaf_sharding), (_, leaf_logical_val) in zip(
+        leaves_params, leaves_sharding, leaves_logical
+    ):
+      path_str = "/".join(str(p.key if hasattr(p, "key") else p.name) for p in path)
+      shape = jax.typeof(leaf_val)
+      pspec = sharding.remove_size_one_mesh_axis(leaf_sharding.spec, mesh)
+      pspec_str = str(tuple(pspec))
+      logical_str = str(leaf_logical_val)
+
+      message = (
+          f" {path_str}\n" f"    Shape:     {shape}\n" f"    Logical:   {logical_str}\n" f"    Physical:  {pspec_str}"
+      )
+      max_logging.info(message)
+  else:
+    for (path, leaf_val), (_, leaf_sharding) in zip(leaves_params, leaves_sharding):
+      path_str = "/".join(str(p.key if hasattr(p, "key") else p.name) for p in path)
+      shape = jax.typeof(leaf_val)
+      pspec = sharding.remove_size_one_mesh_axis(leaf_sharding.spec, mesh)
+      pspec_str = str(tuple(pspec))
+
+      message = f" {path_str}\n" f"    Shape:     {shape}\n" f"    Physical:  {pspec_str}"
+      max_logging.info(message)
 
   print(flush=True)
 

tests/unit/maxtext_utils_test.py

@@ -180,7 +180,9 @@ def setUp(self):
         },
         "decoder": {"gate": {"bias": jnp.array([0.5, 0.5])}},
     }
-    self.state = train_state.TrainState(step=0, apply_fn=self.model.apply, params=self.initial_params, tx=None, opt_state={})
+    self.state = train_state.TrainState(
+        step=0, apply_fn=self.model.apply, params=self.initial_params, tx=None, opt_state={}
+    )
 
   def test_update_mode_add(self):
     target_path = ("decoder", "gate", "bias")
@@ -721,7 +723,9 @@ def test_low_temperature_is_greedy(self):
     rngs = jax.random.split(self.rng, 10)
 
     for r in rngs:
-      token = inference_utils.sample_topk_topp_weighted(self.logits, topk=10, nucleus_topp=1.0, temperature=low_temp, rng=r)
+      token = inference_utils.sample_topk_topp_weighted(
+          self.logits, topk=10, nucleus_topp=1.0, temperature=low_temp, rng=r
+      )
       self.assertEqual(token.item(), greedy_token_index)
 
   def test_invalid_args_raise_error(self):