ModelParallel
classkeras.distribution.ModelParallel(layout_map=None, batch_dim_name=None, **kwargs)
Distribution that shards model variables.
Compare to DataParallel
which replicates the variables across all devices,
ModelParallel
allows you to shard variables in addition to the input data.
To construct a ModelParallel
distribution, you need to provide a
DeviceMesh
and a LayoutMap
.
DeviceMesh
contains physical device information. The axis names in
the mesh will be used to map the variable and data layout.LayoutMap
contains the mapping between variable paths to their
corresponding TensorLayout
.Example
devices = list_devices() # Assume there are 8 devices.
# Create a mesh with 2 devices for data parallelism and 4 devices for
# model parallelism.
device_mesh = DeviceMesh(shape=(2, 4), axis_names=('batch', 'model'),
devices=devices)
# Create a layout map that shard the `Dense` layer and `Conv2D`
# layer variables on the last dimension.
# Based on the `device_mesh`, this means the variables
# will be split across 4 devices. Any other variable that doesn't
# match any key in the layout map will be fully replicated.
layout_map = LayoutMap(device_mesh)
layout_map['dense.*kernel'] = (None, 'model')
layout_map['dense.*bias'] = ('model',)
layout_map['conv2d.*kernel'] = (None, None, None, 'model')
layout_map['conv2d.*bias'] = ('model',)
distribution = ModelParallel(
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)
You can quickly update the device mesh shape to change the sharding factor of the variables. E.g.
# With only the shape change for the device mesh, the variables will be
# sharded across 8 devices instead of 4, which further reduces the memory
# footprint of variables 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 variables, you
can first list out all the model variable paths, which will be used as the
key to map the variables to TensorLayout
.
e.g.
model = create_model()
for v in model.variables:
print(v.path)
Arguments
LayoutMap
instance which map the variable path to the
corresponding tensor layout.layout_map
object)
that will be used to distribute data. If unspecified, the
first axis from the device mesh will be used.