mirror of
https://github.com/zebrajr/tensorflow.git
synced 2025-12-06 12:20:11 +01:00
85151a471c
dtensor.run_on() tends to cause confusion since it indicates override semantics (force the functions and ops in the context to run on the supplied mesh), even though the implementation only provides fallback semantics. PiperOrigin-RevId: 513949632
499 lines
16 KiB
Python
499 lines
16 KiB
Python
# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# ==============================================================================
|
|
"""Core DTensor Python API."""
|
|
|
|
import contextlib
|
|
import threading
|
|
from typing import Any, Callable, Optional, Sequence
|
|
|
|
from tensorflow.dtensor.python import dtensor_device
|
|
from tensorflow.dtensor.python import gen_dtensor_ops
|
|
from tensorflow.dtensor.python import layout as layout_lib
|
|
from tensorflow.python.eager import context
|
|
from tensorflow.python.framework import ops
|
|
from tensorflow.python.util import deprecation
|
|
from tensorflow.python.util.tf_export import tf_export
|
|
|
|
_dtensor_singleton = None
|
|
_dtensor_singleton_lock = threading.Lock()
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Main methods to launch DTensor computations.
|
|
|
|
|
|
@tf_export("experimental.dtensor.call_with_layout", v1=[])
|
|
def call_with_layout(fn: Callable[...,
|
|
Any], layout: Optional[layout_lib.Layout],
|
|
*args, **kwargs) -> Any:
|
|
"""Calls a function in the DTensor device scope if `layout` is not None.
|
|
|
|
If `layout` is not None, `fn` consumes DTensor(s) as input and produces a
|
|
DTensor as output; a DTensor is a tf.Tensor with layout-related attributes.
|
|
|
|
If `layout` is None, `fn` consumes and produces regular tf.Tensors.
|
|
|
|
Args:
|
|
fn: A supported TF API function such as tf.zeros.
|
|
layout: Optional, the layout of the output DTensor.
|
|
*args: Arguments given to `fn`.
|
|
**kwargs: Keyword arguments given to `fn`.
|
|
|
|
Returns:
|
|
The return value of `fn` transformed to a DTensor if requested.
|
|
"""
|
|
if layout is not None:
|
|
if not context.executing_eagerly():
|
|
# This is a workaround for b/199324097, where functions such as tf.ones
|
|
# could attach an incorrect layout to the tf.const generated under the
|
|
# hood. The op runs successfully in eager mode, but in graph mode, MLIR
|
|
# passes sometimes attach the default layout to a scalar constant.
|
|
# %cst = tf.Const([1]) -- With the given layout
|
|
# %0 = "tf.DTensorLayout"(%cst). -- Fails in MLIR pass since shape for
|
|
# -- layout could be different than
|
|
# -- shape[0] for %cst.
|
|
# %1 = tf.Fill(%0, 1)
|
|
result = fn(*args, **kwargs)
|
|
return relayout(result, layout)
|
|
else:
|
|
with run_on(layout.mesh):
|
|
with _dtensor_device()._default_layout(layout): # pylint: disable=protected-access
|
|
return fn(*args, **kwargs)
|
|
return fn(*args, **kwargs)
|
|
|
|
|
|
@tf_export("experimental.dtensor.run_on", v1=[])
|
|
@deprecation.deprecated(None, "Use `dtensor.default_mesh` scope instead.")
|
|
@contextlib.contextmanager
|
|
def run_on(mesh: layout_lib.Mesh):
|
|
"""Runs enclosed functions in the DTensor device scope.
|
|
|
|
This function returns a scope. All the ops and tf.functions in this scope will
|
|
run on the DTensor device using the mesh provided.
|
|
This is useful for wrapping any tf.function that doesn't take a DTensor as
|
|
input but would like to produce DTensor as result. The scope will also make
|
|
sure all small constants be replicated as DTensor.
|
|
|
|
Args:
|
|
mesh: A Mesh instance to extract a default mesh from.
|
|
|
|
Yields:
|
|
A context in which all ops and tf.functions will run on the DTensor device.
|
|
"""
|
|
with default_mesh(mesh):
|
|
yield
|
|
|
|
|
|
@tf_export("experimental.dtensor.default_mesh", v1=[])
|
|
@contextlib.contextmanager
|
|
def default_mesh(mesh: layout_lib.Mesh):
|
|
"""Sets the default DTensor device mesh to use for enclosed functions.
|
|
|
|
This function returns a scope. All the ops and tf.functions in this scope will
|
|
default to this DTensor mesh if a mesh cannot be inferred from any of the
|
|
inputs
|
|
This is useful for wrapping any tf.function that doesn't take a DTensor as
|
|
input but would like to produce DTensor as result. The scope will also make
|
|
sure all small constants are replicated as DTensors.
|
|
|
|
Args:
|
|
mesh: A Mesh instance to extract a default mesh from.
|
|
|
|
Yields:
|
|
A context in which all ops and tf.functions will run on the given mesh.
|
|
"""
|
|
if not isinstance(mesh, layout_lib.Mesh):
|
|
raise ValueError(f"Expect `mesh` to be `Mesh`, got {type(mesh)}")
|
|
|
|
with _dtensor_device()._experimental_default_mesh(mesh): # pylint: disable=protected-access
|
|
with ops.device(device_name()):
|
|
yield
|
|
|
|
|
|
@tf_export("experimental.dtensor.device_name", v1=[])
|
|
def device_name() -> str:
|
|
"""Returns the singleton DTensor device's name.
|
|
|
|
This function can be used in the following way:
|
|
|
|
```python
|
|
import tensorflow as tf
|
|
|
|
with tf.device(dtensor.device_name()):
|
|
# ...
|
|
```
|
|
"""
|
|
return _dtensor_device().name
|
|
|
|
|
|
@tf_export("experimental.dtensor.is_dtensor", v1=[])
|
|
def is_dtensor(tensor) -> bool:
|
|
"""Check whether the input tensor is a DTensor.
|
|
|
|
In Python, a DTensor has the same type as a `tf.Tensor`. This method will
|
|
let you check and handle the tensor differently if a tf.Tensor is a DTensor.
|
|
|
|
Args:
|
|
tensor: an object to be checked.
|
|
|
|
Returns:
|
|
bool, True if the given tensor is a DTensor.
|
|
"""
|
|
return _dtensor_device().is_dtensor(tensor)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Data transfer methods.
|
|
|
|
|
|
@tf_export("experimental.dtensor.copy_to_mesh", v1=[])
|
|
def copy_to_mesh(
|
|
tensor: Any,
|
|
layout: layout_lib.Layout,
|
|
source_layout: Optional[layout_lib.Layout] = None) -> ops.Tensor:
|
|
"""Copies a tf.Tensor onto the DTensor device with the given layout.
|
|
|
|
Copies a regular tf.Tensor onto the DTensor device. Use the mesh attached to
|
|
`layout` as target mesh. This method currently only supports replicated
|
|
layouts, or one-to-one copies for sharded layouts.
|
|
|
|
Args:
|
|
tensor: A regular tf.Tensor to be copied as a DTensor.
|
|
layout: Target layout (and mesh) for the result DTensor.
|
|
source_layout: Source layout of the tensor before copy. This argument
|
|
is deprecated.
|
|
|
|
Returns:
|
|
A DTensor on the DTensor device with the given layout.
|
|
"""
|
|
del source_layout
|
|
with run_on(layout.mesh):
|
|
return gen_dtensor_ops.copy_to_mesh(tensor, layout.to_string())
|
|
|
|
|
|
@tf_export("experimental.dtensor.pack", v1=[])
|
|
def pack(tensors: Sequence[Any], layout: layout_lib.Layout) -> Any:
|
|
"""Packs `tf.Tensor` components into a DTensor.
|
|
|
|
Packing and unpacking are inverse operations:
|
|
|
|
```
|
|
* unpack(pack(tensors)) == tensors
|
|
* pack(unpack(dtensor)) == dtensor
|
|
```
|
|
|
|
1. For any DTensor on the mesh, `unpack` returns the raw components placed on
|
|
each underlying device.
|
|
2. Packing these raw components in the same order using `pack` returns a
|
|
DTensor which should be identical to the original DTensor--both the content
|
|
value and the layout.
|
|
|
|
**Shape, Rank, and Scalars**: The rank of the DTensor is the same as the
|
|
rank of its raw components, i.e., rank is preserved. This leads to a
|
|
consistent interpretation for packing scalar values into a DTensor. The only
|
|
valid layout for a scalar value is fully replicated, and the individual
|
|
components must be identical scalars.
|
|
|
|
Each input `tensors[i]` will be copied to `layout.mesh.local_device[i]`
|
|
if not already on the local device. Non-local components should not be passed
|
|
to `pack`; use `copy_to_mesh` and `relayout` to place tensors on all global
|
|
devices on a mesh.
|
|
|
|
It is the caller's responsibility to ensure that the underlying values
|
|
for `pack` adhere to the specified layout, and that only as many values are
|
|
specified as there are local devices. Pack does not move data between clients.
|
|
See examples below for more detail about layouts.
|
|
|
|
For example, assume we have a mesh `[X(2), Y(3)]`, which has in total 6
|
|
underlying devices. Futuremore, assume that the device location mapping is
|
|
the following:
|
|
|
|
```
|
|
device_ID | location X, Y
|
|
0 0, 0
|
|
1 0, 1
|
|
2 0, 2
|
|
3 1, 0
|
|
4 1, 1
|
|
5 1, 2
|
|
```
|
|
|
|
1. For 1-D vector DTensor with shape `[128]` with layout `[mesh.X]` and value
|
|
as `range(128)`, the raw components will have shape `[64]` each, and the
|
|
raw components will be:
|
|
|
|
```
|
|
device_ID | raw component
|
|
0 range(0, 64)
|
|
1 range(0, 64)
|
|
2 range(0, 64)
|
|
3 range(64, 128)
|
|
4 range(64, 128)
|
|
5 range(64, 128)
|
|
```
|
|
|
|
This also means for a 1-D DTensor with shape `[2]` and layout `[mesh.X]`,
|
|
the raw components have shape `[1]` rather than the shape for scalar values
|
|
`[]`.
|
|
|
|
2. For 2-D vector DTensor with shape `[2, 3]` with layout `[mesh.X, mesh.Y]`
|
|
and value as `range(6)`, this is basically a fully-sharded DTensor.
|
|
|
|
From global view, the content looks like
|
|
```
|
|
[
|
|
[0.0, 1.0, 2.0],
|
|
[3.0, 4.0, 5.0],
|
|
]
|
|
```
|
|
|
|
The raw components will have shape `[1, 1]` each, and have the following
|
|
content:
|
|
|
|
```
|
|
device_ID | raw component
|
|
0 [[0.0]]
|
|
1 [[1.0]]
|
|
2 [[2.0]]
|
|
3 [[3.0]]
|
|
4 [[4.0]]
|
|
5 [[5.0]]
|
|
```
|
|
|
|
3. For a scalar value `123.0` DTensor, it can only have one legitimate layout
|
|
`[]` (no dimension, but fully replicated).
|
|
|
|
The raw components will have shape `[]` each, and have the following
|
|
content:
|
|
|
|
```
|
|
device_ID | raw component
|
|
0 123.0
|
|
1 123.0
|
|
2 123.0
|
|
3 123.0
|
|
4 123.0
|
|
5 123.0
|
|
```
|
|
|
|
Again, caller of `pack` is expected to provide 6 identical value raw
|
|
components with scalar shapes.
|
|
|
|
4. For 3-D vector DTensor with shape `[2, 2, 3]` with layout
|
|
`[X, unsharded, unsharded]` and value as `range(12)`,
|
|
|
|
From global view, the content looks like:
|
|
```
|
|
[
|
|
[
|
|
[0.0, 1.0, 2.0],
|
|
[3.0, 4.0, 5.0],
|
|
],
|
|
[
|
|
[6.0, 7.0, 8.0],
|
|
[9.0, 10., 11.],
|
|
],
|
|
]
|
|
```
|
|
|
|
The raw components will have shape `[1, 2, 3]` each, and have the following
|
|
content:
|
|
|
|
```
|
|
device_ID | raw component
|
|
0 range(6).reshape([1, 2, 3])
|
|
1 range(6).reshape([1, 2, 3])
|
|
2 range(6).reshape([1, 2, 3])
|
|
3 range(6, 12).reshape([1, 2, 3])
|
|
4 range(6, 12).reshape([1, 2, 3])
|
|
5 range(6, 12).reshape([1, 2, 3])
|
|
```
|
|
|
|
Args:
|
|
tensors: The list of local tensor components to pack into a DTensor.
|
|
layout: The layout of the DTensor to be created.
|
|
|
|
Returns:
|
|
A DTensor created from the individual component tensors.
|
|
|
|
Raises:
|
|
RuntimeError: When `pack` is not called eagerly.
|
|
"""
|
|
return _dtensor_device().pack(tensors, layout)
|
|
|
|
|
|
@tf_export("experimental.dtensor.unpack", v1=[])
|
|
def unpack(tensor: Any) -> Sequence[Any]:
|
|
"""Unpacks a DTensor into `tf.Tensor` components.
|
|
|
|
Packing and unpacking are inverse operations:
|
|
|
|
```
|
|
* unpack(pack(tensors)) == tensors
|
|
* pack(unpack(dtensor)) == dtensor
|
|
```
|
|
|
|
1. For any DTensor on the mesh, `unpack` returns the raw components placed on
|
|
each underlying device.
|
|
2. Packing these raw components in the same order using `pack` returns a
|
|
DTensor which should be identical to the original DTensor--both the content
|
|
value and the layout.
|
|
|
|
See the documentation for `pack` for more information about how packing and
|
|
unpacking works.
|
|
|
|
Args:
|
|
tensor: The DTensor to unpack.
|
|
|
|
Returns:
|
|
The individual component tensors of the DTensor. This will include only the
|
|
client-local components, i.e. the components placed on the local devices.
|
|
|
|
Raises:
|
|
RuntimeError: When `unpack` is not called eagerly.
|
|
"""
|
|
return _dtensor_device().unpack(tensor)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Layout-related methods.
|
|
|
|
|
|
@tf_export("experimental.dtensor.fetch_layout", v1=[])
|
|
def fetch_layout(tensor: ops.Tensor) -> layout_lib.Layout:
|
|
"""Fetches the layout of a DTensor.
|
|
|
|
Args:
|
|
tensor: The DTensor whose layout is to be fetched.
|
|
|
|
Returns:
|
|
The `Layout` of this DTensor.
|
|
|
|
Raises:
|
|
RuntimeError: When not called eagerly.
|
|
"""
|
|
return _dtensor_device().fetch_layout(tensor)
|
|
|
|
|
|
@tf_export("experimental.dtensor.check_layout", v1=[])
|
|
def check_layout(tensor: ops.Tensor, layout: layout_lib.Layout) -> None:
|
|
"""Asserts that the layout of the DTensor is `layout`.
|
|
|
|
Args:
|
|
tensor: A DTensor whose layout is to be checked.
|
|
layout: The `Layout` to compare against.
|
|
|
|
Raises:
|
|
ValueError: If the layout of `tensor` does not match the supplied `layout`.
|
|
"""
|
|
if fetch_layout(tensor) != layout:
|
|
raise ValueError("Layout of tensor: " + str(fetch_layout(tensor)) +
|
|
", did not match expected layout: " + str(layout))
|
|
|
|
|
|
@tf_export("experimental.dtensor.relayout", v1=[])
|
|
def relayout(tensor: ops.Tensor, layout: layout_lib.Layout) -> ops.Tensor:
|
|
"""Changes the layout of `tensor`.
|
|
|
|
Changes the layout of `tensor` to `layout`. This is used to fine-tune the
|
|
behavior of ops following/connected to `tensor`, such as choosing one SPMD
|
|
expansion pattern over another. This works by forward propagating `layout`
|
|
to connected TensorFlow computation graphs during layout propagation.
|
|
|
|
Currently, only converting layouts from replicated to sharded or sharded to
|
|
replicated per mesh dimension is supported. That is, "x, y" -> "unsharded, y"
|
|
is supported, while "x, y" -> "z, y" is not supported.
|
|
|
|
We also support a special "match" sharding spec, which instructs the relayout
|
|
to act as an identity operation with respect to any sharding on these
|
|
mesh dimensions.
|
|
|
|
Relayout is internally lowered to a set of Split and/or AllToAll ops. When
|
|
tensor layouts are converted from replicated to sharded, the cost is
|
|
comparatively low because we only insert Split ops and no cross-device
|
|
communication is needed. However, when tensor layouts are converted from
|
|
sharded to replicated, cross-device communication may occur, causing potential
|
|
performance impact.
|
|
|
|
Args:
|
|
tensor: A DTensor to specify a new layout for.
|
|
layout: A Layout object specifying a new sharding spec.
|
|
|
|
Returns:
|
|
A DTensor output from the Relayout op.
|
|
"""
|
|
layout_str = layout.to_string()
|
|
with run_on(layout.mesh):
|
|
return gen_dtensor_ops.relayout(tensor, layout_str)
|
|
|
|
|
|
def _set_dtensor_device(device: dtensor_device.DTensorDevice) -> None:
|
|
global _dtensor_singleton
|
|
_dtensor_singleton = device
|
|
|
|
|
|
def _dtensor_device() -> dtensor_device.DTensorDevice:
|
|
with _dtensor_singleton_lock:
|
|
if _dtensor_singleton is None:
|
|
_set_dtensor_device(
|
|
dtensor_device.DTensorDevice(meshes=[], is_async=True))
|
|
return _dtensor_singleton
|
|
|
|
|
|
def _reset() -> None:
|
|
global _dtensor_singleton
|
|
if _dtensor_singleton is not None:
|
|
_dtensor_singleton.clear_tpu_core_ids()
|
|
with _dtensor_singleton_lock:
|
|
_dtensor_singleton = None
|
|
|
|
|
|
# ----------------------------------------------------------------------------
|
|
# Gradients
|
|
|
|
|
|
@ops.RegisterGradient("Relayout")
|
|
def _relayout_gradient(op, grad):
|
|
grad = gen_dtensor_ops.relayout_grad(grad, forward_input=op.inputs[0])
|
|
return grad
|
|
|
|
|
|
@ops.RegisterGradient("RelayoutGrad")
|
|
def _relayout_grad_gradient(op, grad):
|
|
# Gradient of RelayoutGrad is relayout to the original Relayout's output.
|
|
grad = gen_dtensor_ops.relayout_grad(grad, forward_input=op.inputs[0])
|
|
# Return None for forward_input's partial gradient since it is not connected
|
|
# to the target's gradient.
|
|
return grad, None
|
|
|
|
|
|
@ops.RegisterGradient("CopyToMesh")
|
|
def _copy_to_mesh_gradient(op, grad):
|
|
grad = gen_dtensor_ops.copy_to_mesh_grad(
|
|
grad,
|
|
forward_input=op.inputs[0],
|
|
reference_layout=op.get_attr("layout"),
|
|
)
|
|
return grad
|
|
|
|
|
|
@ops.RegisterGradient("CopyToMeshGrad")
|
|
def _copy_to_mesh_grad_gradient(op, grad):
|
|
grad = gen_dtensor_ops.copy_to_mesh_grad(
|
|
grad,
|
|
forward_input=op.inputs[0],
|
|
reference_layout=op.get_attr("reference_layout"),
|
|
)
|
|
return grad, None
|