/
api.py
484 lines (374 loc) · 15.5 KB
/
api.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
# 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.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:
with default_mesh(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 default_mesh(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 default_mesh(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