Add a compile option to nn.value_and_grad

python/mlx/nn/utils.py

@@ -1,11 +1,23 @@
 # Copyright © 2023 Apple Inc.
 
-from typing import Callable
+from contextlib import contextmanager
+from typing import Any, Callable
 
 import mlx.core as mx
 
 
-def value_and_grad(model: "mlx.nn.Module", fn: Callable):
+@contextmanager
+def updated_model(model: "mlx.nn.Module", *parameters: Any):
+    old_state = model.parameters()
+    try:
+        for p in parameters:
+            model.update(p)
+        yield model
+    finally:
+        model.update(old_state)
+
+
+def value_and_grad(model: "mlx.nn.Module", fn: Callable, compile: bool = False):
     """Transform the passed function ``fn`` to a function that computes the
     gradients of ``fn`` wrt the model's trainable parameters and also its
     value.
@@ -14,20 +26,29 @@ def value_and_grad(model: "mlx.nn.Module", fn: Callable):
         model (mlx.nn.Module): The model whose trainable parameters to compute
                                gradients for
         fn (Callable): The scalar function to compute gradients for
+        compile (bool): Whether to "compile" the function before returning it.
+                        Default: ``False``.
 
     Returns:
         A callable that returns the value of ``fn`` and the gradients wrt the
         trainable parameters of ``model``
     """
 
-    def inner_fn(params, *args, **kwargs):
+    def inner_fn(trainable_parameters, other_parameters, *args, **kwargs):
-        model.update(params)
+        with updated_model(model, other_parameters, trainable_parameters):
             return fn(*args, **kwargs)
 
     value_grad_fn = mx.value_and_grad(inner_fn)
+    if compile:
+        value_grad_fn = mx.compile(value_grad_fn)
 
     def wrapped_value_grad_fn(*args, **kwargs):
-        value, grad = value_grad_fn(model.trainable_parameters(), *args, **kwargs)
+        value, grad = value_grad_fn(
+            model.trainable_parameters(),
+            model.non_trainable_parameters(),
+            *args,
+            **kwargs,
+        )
         return value, grad
 
     return wrapped_value_grad_fn