Add model parallel distribution.

keras_core/distribution/distribution_lib.py

@@ -289,6 +289,102 @@ def get_variable_layout(self, variable):
         return TensorLayout(variable_shard_spec, self.device_mesh)
 
 
+class ModelParallel(Distribution):
+    """Distribution that shard model weights.
+
+    Compare to DataParallel which replicates the weights across all the devices,
+    ModelParallel allows user to shard weights in addition to the input data.
+
+    To construct a ModelParallel distribution, user need to provide device mesh
+    and layout mapping.
+
+    1. `DeviceMesh`contains physcial device information, and the axis names in
+        the mesh will be used to map the weight and data layout.
+    2. `LayoutMap` contains the mapping for the variable path to its
+        corresponding `TensorLayout`.
+
+    Example:
+    ```python
+    devices = list_devices()    # Assume there are 8 devices.
+
+    # Create a mesh with 2 devices on data parallel and 4 devices on weight
+    # parallel.
+    device_mesh = DeviceMesh(shape=(2, 4), axis_names=('batch', 'model'),
+                             devices=devices)
+    # Create a layout map that shard the dense layer and conv2d layer weights
+    # on the last dimension. Based on the device_mesh, this means the weights
+    # will be split across 4 devices. Any other weights that doesn't match for
+    # any key in layout map will get be fully replicated.
+    layout_map = LayoutMap(device_mesh)
+    layout_map['.*dense.*kernel'] = TensorLayout([None, 'model'])
+    layout_map['.*dense.*bias'] = TensorLayout(['model'])
+    layout_map['.*conv2d.*kernel'] = TensorLayout([None, None, None, 'model'])
+    layout_map['.*conv2d.*bias'] = TensorLayout(['model'])
+
+    distribution = ModelParallel(device_mesh=device_mesh,
+                                 layout_map=layout_map,
+                                 batch_dim_name='batch')
+    # Set the global distribution, or via `with distribution.scope():`
+    set_distribution(distribution)
+
+    model = model_creation()
+    model.compile()
+    model.fit(data)
+    ```
+
+    User can quickly update the device mesh shape to change the sharding factor
+    of the weights. E.g.
+    ```
+    # With only the shape change for the device mesh, the weights will be
+    # sharded across 8 devices instead of 4, which further reduce the memory
+    # footprint of weights on each of the device.
+    device_mesh = DeviceMesh(shape=(1, 8), axis_names=('batch', 'model'),
+                             devices=devices)
+    ```
+
+    To figure out a proper layout mapping rule for all the model weights, you
+    can first list out all the model weights path, which will be used as the key
+    to map the weights to `TensorLayout`.
+
+    e.g.
+    ```
+    model = create_model()
+    for w in model.weights:
+        print(w.path)
+    ```
+    """
+
+    def __init__(self, device_mesh, layout_map, batch_dim_name=None):
+        """Initialize the model parallel distribution.
+
+        Args:
+            device_mesh: `DeviceMesh` instance for physical device and its
+                logical mapping.
+            layout_map: `LayoutMap` instance which map the variable path to the
+                corresponding `TensorLayout`. The axis names of the
+                `TensorLayout`s should match to the axis names in the
+                device_mesh, or exception will be raised.
+            batch_dim_name: optional string, the axis name in the device_mesh
+                that will be used to distribute data. The first axis from the
+                device_mesh will be used if user didn't specify any.
+        """
+        super().__init__(device_mesh)
+        self._layout_map = layout_map
+        self._batch_dim_name = batch_dim_name or self.device_mesh.axis_names[0]
+
+    def get_data_layout(self, data_shape):
+        data_shard_spec = [None] * len(data_shape)
+        data_shard_spec[0] = self._batch_dim_name  # Shard on the first dim
+        return TensorLayout(data_shard_spec, self.device_mesh)
+
+    def get_variable_layout(self, variable):
+        variable_layout = self._layout_map[variable.path]
+        if variable_layout is not None:
+            return variable_layout
+        variable_shard_spec = [None] * len(variable.shape)
+        return TensorLayout(variable_shard_spec, self.device_mesh)
+
+
 class LayoutMap(collections.abc.MutableMapping):
     """A dict-like object that maps string to `TensorLayout` instances.
 

keras_core/distribution/distribution_lib_test.py

@@ -193,6 +193,53 @@ def test_get_variable_layout(self):
         self.assertEqual(variable_layout.axes, [None])
 
 
+class ModelParallelDistributionTest(testing.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.devices = ["CPU:{i}" for i in range(8)]
+        shape = (2, 4)
+        axis_names = ["data", "model"]
+
+        self.device_mesh = distribution_lib.DeviceMesh(
+            shape, axis_names, self.devices
+        )
+
+    def test_distribute_weights(self):
+        layout_map = distribution_lib.LayoutMap(self.device_mesh)
+        layout_map[".*kernel"] = distribution_lib.TensorLayout([None, "model"])
+        layout_map[".*bias"] = distribution_lib.TensorLayout(["model"])
+
+        distribution = distribution_lib.ModelParallel(
+            self.device_mesh, layout_map, batch_dim_name="data"
+        )
+        kernel = backend.Variable(initializer=np.arange(8, 4), name="kernel")
+        bias = backend.Variable(initializer=np.arange(4), name="bias")
+        rng_seed = backend.Variable(initializer=[0, 1], name="seed")
+
+        kernel_layout = distribution.get_variable_layout(kernel)
+        self.assertIs(kernel_layout.device_mesh, self.device_mesh)
+        self.assertEqual(kernel_layout.axes, [None, "model"])
+
+        bias_layout = distribution.get_variable_layout(bias)
+        self.assertIs(bias_layout.device_mesh, self.device_mesh)
+        self.assertEqual(bias_layout.axes, ["model"])
+
+        rng_seed_layout = distribution.get_variable_layout(rng_seed)
+        self.assertIs(rng_seed_layout.device_mesh, self.device_mesh)
+        self.assertEqual(rng_seed_layout.axes, [None])
+
+    def test_distribute_data(self):
+        layout_map = distribution_lib.LayoutMap(self.device_mesh)
+        distribution = distribution_lib.ModelParallel(
+            self.device_mesh, layout_map, batch_dim_name="data"
+        )
+
+        data = np.arange(16).reshape((4, 2, 2))
+        data_layout = distribution.get_data_layout(data.shape)
+        self.assertIs(data_layout.device_mesh, self.device_mesh)
+        self.assertEqual(data_layout.axes, ["data", None, None])
+
+
 class LayoutMapTest(testing.TestCase):
     def setUp(self):
         super().setUp()
@@ -345,7 +392,7 @@ def test_validation_for_device_mesh(self):
         ):
             backend_dlib.to_jax_layout(layout)
 
-    def test_e2e_model(self):
+    def test_e2e_data_parallel_model(self):
         distribution = distribution_lib.DataParallel(
             devices=backend_dlib.list_devices()
         )
@@ -368,3 +415,42 @@ def test_e2e_model(self):
         with distribution.scope():
             model.compile(loss="mse")
             model.fit(inputs, labels)
+
+    def test_e2e_model_parallel_model(self):
+        shape = (4, 2)
+        axis_names = ["batch", "model"]
+        device_mesh = distribution_lib.DeviceMesh(
+            shape, axis_names, backend_dlib.list_devices()
+        )
+
+        layout_map = distribution_lib.LayoutMap(device_mesh)
+        layout_map[".*dense.*kernel"] = distribution_lib.TensorLayout(
+            [None, "model"]
+        )
+        layout_map[".*dense.*bias"] = distribution_lib.TensorLayout(["model"])
+
+        distribution = distribution_lib.ModelParallel(
+            device_mesh, layout_map, batch_dim_name="batch"
+        )
+        with distribution.scope():
+            inputs = layers.Input(shape=[28, 28, 1])
+            y = layers.Flatten()(inputs)
+            y = layers.Dense(units=200, use_bias=False, activation="relu")(y)
+            y = layers.Dropout(0.4)(y)
+            y = layers.Dense(units=10, activation="softmax")(y)
+            model = models.Model(inputs=inputs, outputs=y)
+
+        for weight in model.weights:
+            if "kernel" in weight.name:
+                self.assertEqual(weight._value.sharding.spec, (None, "model"))
+            elif "bias" in weight.name:
+                self.assertEqual(weight._value.sharding.spec, ("model",))
+            else:
+                self.assertTrue(weight._value.sharding.is_fully_replicated)
+
+        inputs = np.random.normal(size=(32, 28, 28, 1))
+        labels = np.random.normal(size=(32, 10))
+
+        with distribution.scope():
+            model.compile(loss="mse")
+            model.fit(inputs, labels)