[JAX] Replace uses of jax.devices("cpu") with jax.local_devices(backend="cpu") (#27593)

An upcoming change to JAX will include non-local (addressable) CPU devices in jax.devices() when JAX is used multicontroller-style, where there are multiple Python processes.

This change preserves the current behavior by replacing uses of jax.devices("cpu"), which previously only returned local devices, with jax.local_devices("cpu"), which will return local devices both now and in the future.

This change is always safe (i.e., it should always preserve the previous behavior), but it may sometimes be unnecessary if code is never used in a multicontroller setting.
Co-authored-by: Peter Hawkins <phawkins@google.com>

src/transformers/modeling_flax_utils.py

@@ -915,7 +915,7 @@ class FlaxPreTrainedModel(PushToHubMixin, FlaxGenerationMixin):
                 state = jax.tree_util.tree_map(jnp.array, state)
             else:
                 # keep the params on CPU if we don't want to initialize
-                state = jax.tree_util.tree_map(lambda x: jax.device_put(x, jax.devices("cpu")[0]), state)
+                state = jax.tree_util.tree_map(lambda x: jax.device_put(x, jax.local_devices(backend="cpu")[0]), state)
 
         if "batch_stats" in state:  # if flax model contains batch norm layers
             # if model is base model only use model_prefix key