File size: 16,598 Bytes
9c6594c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
# Copyright 2022 The HuggingFace Team. 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.

import importlib
import importlib.metadata
import os
import warnings
from functools import lru_cache, wraps

import torch
from packaging import version
from packaging.version import parse

from .environment import parse_flag_from_env, patch_environment, str_to_bool
from .versions import compare_versions, is_torch_version


# Try to run Torch native job in an environment with TorchXLA installed by setting this value to 0.
USE_TORCH_XLA = parse_flag_from_env("USE_TORCH_XLA", default=True)

_torch_xla_available = False
if USE_TORCH_XLA:
    try:
        import torch_xla.core.xla_model as xm  # noqa: F401
        import torch_xla.runtime

        _torch_xla_available = True
    except ImportError:
        pass

# Keep it for is_tpu_available. It will be removed along with is_tpu_available.
_tpu_available = _torch_xla_available

# Cache this result has it's a C FFI call which can be pretty time-consuming
_torch_distributed_available = torch.distributed.is_available()


def _is_package_available(pkg_name, metadata_name=None):
    # Check we're not importing a "pkg_name" directory somewhere but the actual library by trying to grab the version
    package_exists = importlib.util.find_spec(pkg_name) is not None
    if package_exists:
        try:
            # Some libraries have different names in the metadata
            _ = importlib.metadata.metadata(pkg_name if metadata_name is None else metadata_name)
            return True
        except importlib.metadata.PackageNotFoundError:
            return False


def is_torch_distributed_available() -> bool:
    return _torch_distributed_available


def is_xccl_available():
    if is_torch_version(">=", "2.7.0"):
        return torch.distributed.distributed_c10d.is_xccl_available()
    if is_ipex_available():
        return False
    return False


def is_ccl_available():
    try:
        pass
    except ImportError:
        print(
            "Intel(R) oneCCL Bindings for PyTorch* is required to run DDP on Intel(R) XPUs, but it is not"
            " detected. If you see \"ValueError: Invalid backend: 'ccl'\" error, please install Intel(R) oneCCL"
            " Bindings for PyTorch*."
        )
    return importlib.util.find_spec("oneccl_bindings_for_pytorch") is not None


def get_ccl_version():
    return importlib.metadata.version("oneccl_bind_pt")


def is_import_timer_available():
    return _is_package_available("import_timer")


def is_pynvml_available():
    return _is_package_available("pynvml") or _is_package_available("pynvml", "nvidia-ml-py")


def is_pytest_available():
    return _is_package_available("pytest")


def is_msamp_available():
    return _is_package_available("msamp", "ms-amp")


def is_schedulefree_available():
    return _is_package_available("schedulefree")


def is_transformer_engine_available():
    if is_hpu_available():
        return _is_package_available("intel_transformer_engine", "intel-transformer-engine")
    else:
        return _is_package_available("transformer_engine", "transformer-engine")


def is_lomo_available():
    return _is_package_available("lomo_optim")


def is_cuda_available():
    """
    Checks if `cuda` is available via an `nvml-based` check which won't trigger the drivers and leave cuda
    uninitialized.
    """
    with patch_environment(PYTORCH_NVML_BASED_CUDA_CHECK="1"):
        available = torch.cuda.is_available()

    return available


@lru_cache
def is_torch_xla_available(check_is_tpu=False, check_is_gpu=False):
    """
    Check if `torch_xla` is available. To train a native pytorch job in an environment with torch xla installed, set
    the USE_TORCH_XLA to false.
    """
    assert not (check_is_tpu and check_is_gpu), "The check_is_tpu and check_is_gpu cannot both be true."

    if not _torch_xla_available:
        return False
    elif check_is_gpu:
        return torch_xla.runtime.device_type() in ["GPU", "CUDA"]
    elif check_is_tpu:
        return torch_xla.runtime.device_type() == "TPU"

    return True


def is_torchao_available():
    package_exists = _is_package_available("torchao")
    if package_exists:
        torchao_version = version.parse(importlib.metadata.version("torchao"))
        return compare_versions(torchao_version, ">=", "0.6.1")
    return False


def is_deepspeed_available():
    return _is_package_available("deepspeed")


def is_pippy_available():
    return is_torch_version(">=", "2.4.0")


def is_bf16_available(ignore_tpu=False):
    "Checks if bf16 is supported, optionally ignoring the TPU"
    if is_torch_xla_available(check_is_tpu=True):
        return not ignore_tpu
    if is_cuda_available():
        return torch.cuda.is_bf16_supported()
    if is_mlu_available():
        return torch.mlu.is_bf16_supported()
    if is_xpu_available():
        return torch.xpu.is_bf16_supported()
    if is_mps_available():
        return False
    return True


def is_fp16_available():
    "Checks if fp16 is supported"
    if is_habana_gaudi1():
        return False

    return True


def is_fp8_available():
    "Checks if fp8 is supported"
    return is_msamp_available() or is_transformer_engine_available() or is_torchao_available()


def is_4bit_bnb_available():
    package_exists = _is_package_available("bitsandbytes")
    if package_exists:
        bnb_version = version.parse(importlib.metadata.version("bitsandbytes"))
        return compare_versions(bnb_version, ">=", "0.39.0")
    return False


def is_8bit_bnb_available():
    package_exists = _is_package_available("bitsandbytes")
    if package_exists:
        bnb_version = version.parse(importlib.metadata.version("bitsandbytes"))
        return compare_versions(bnb_version, ">=", "0.37.2")
    return False


def is_bnb_available(min_version=None):
    package_exists = _is_package_available("bitsandbytes")
    if package_exists and min_version is not None:
        bnb_version = version.parse(importlib.metadata.version("bitsandbytes"))
        return compare_versions(bnb_version, ">=", min_version)
    else:
        return package_exists


def is_bitsandbytes_multi_backend_available():
    if not is_bnb_available():
        return False
    import bitsandbytes as bnb

    return "multi_backend" in getattr(bnb, "features", set())


def is_torchvision_available():
    return _is_package_available("torchvision")


def is_megatron_lm_available():
    if str_to_bool(os.environ.get("ACCELERATE_USE_MEGATRON_LM", "False")) == 1:
        if importlib.util.find_spec("megatron") is not None:
            try:
                megatron_version = parse(importlib.metadata.version("megatron-core"))
                if compare_versions(megatron_version, ">=", "0.8.0"):
                    return importlib.util.find_spec(".training", "megatron")
            except Exception as e:
                warnings.warn(f"Parse Megatron version failed. Exception:{e}")
                return False


def is_transformers_available():
    return _is_package_available("transformers")


def is_datasets_available():
    return _is_package_available("datasets")


def is_peft_available():
    return _is_package_available("peft")


def is_timm_available():
    return _is_package_available("timm")


def is_triton_available():
    if is_xpu_available():
        return _is_package_available("triton", "pytorch-triton-xpu")
    return _is_package_available("triton")


def is_aim_available():
    package_exists = _is_package_available("aim")
    if package_exists:
        aim_version = version.parse(importlib.metadata.version("aim"))
        return compare_versions(aim_version, "<", "4.0.0")
    return False


def is_tensorboard_available():
    return _is_package_available("tensorboard") or _is_package_available("tensorboardX")


def is_wandb_available():
    return _is_package_available("wandb")


def is_comet_ml_available():
    return _is_package_available("comet_ml")


def is_swanlab_available():
    return _is_package_available("swanlab")


def is_boto3_available():
    return _is_package_available("boto3")


def is_rich_available():
    if _is_package_available("rich"):
        return parse_flag_from_env("ACCELERATE_ENABLE_RICH", False)
    return False


def is_sagemaker_available():
    return _is_package_available("sagemaker")


def is_tqdm_available():
    return _is_package_available("tqdm")


def is_clearml_available():
    return _is_package_available("clearml")


def is_pandas_available():
    return _is_package_available("pandas")


def is_matplotlib_available():
    return _is_package_available("matplotlib")


def is_mlflow_available():
    if _is_package_available("mlflow"):
        return True

    if importlib.util.find_spec("mlflow") is not None:
        try:
            _ = importlib.metadata.metadata("mlflow-skinny")
            return True
        except importlib.metadata.PackageNotFoundError:
            return False
    return False


def is_mps_available(min_version="1.12"):
    "Checks if MPS device is available. The minimum version required is 1.12."
    # With torch 1.12, you can use torch.backends.mps
    # With torch 2.0.0, you can use torch.mps
    return is_torch_version(">=", min_version) and torch.backends.mps.is_available() and torch.backends.mps.is_built()


def is_ipex_available():
    "Checks if ipex is installed."

    def get_major_and_minor_from_version(full_version):
        return str(version.parse(full_version).major) + "." + str(version.parse(full_version).minor)

    _torch_version = importlib.metadata.version("torch")
    if importlib.util.find_spec("intel_extension_for_pytorch") is None:
        return False
    _ipex_version = "N/A"
    try:
        _ipex_version = importlib.metadata.version("intel_extension_for_pytorch")
    except importlib.metadata.PackageNotFoundError:
        return False
    torch_major_and_minor = get_major_and_minor_from_version(_torch_version)
    ipex_major_and_minor = get_major_and_minor_from_version(_ipex_version)
    if torch_major_and_minor != ipex_major_and_minor:
        warnings.warn(
            f"Intel Extension for PyTorch {ipex_major_and_minor} needs to work with PyTorch {ipex_major_and_minor}.*,"
            f" but PyTorch {_torch_version} is found. Please switch to the matching version and run again."
        )
        return False
    return True


@lru_cache
def is_mlu_available(check_device=False):
    """
    Checks if `mlu` is available via an `cndev-based` check which won't trigger the drivers and leave mlu
    uninitialized.
    """
    if importlib.util.find_spec("torch_mlu") is None:
        return False

    import torch_mlu  # noqa: F401

    with patch_environment(PYTORCH_CNDEV_BASED_MLU_CHECK="1"):
        available = torch.mlu.is_available()

    return available


@lru_cache
def is_musa_available(check_device=False):
    "Checks if `torch_musa` is installed and potentially if a MUSA is in the environment"
    if importlib.util.find_spec("torch_musa") is None:
        return False

    import torch_musa  # noqa: F401

    if check_device:
        try:
            # Will raise a RuntimeError if no MUSA is found
            _ = torch.musa.device_count()
            return torch.musa.is_available()
        except RuntimeError:
            return False
    return hasattr(torch, "musa") and torch.musa.is_available()


@lru_cache
def is_npu_available(check_device=False):
    "Checks if `torch_npu` is installed and potentially if a NPU is in the environment"
    if importlib.util.find_spec("torch_npu") is None:
        return False

    import torch_npu  # noqa: F401

    if check_device:
        try:
            # Will raise a RuntimeError if no NPU is found
            _ = torch.npu.device_count()
            return torch.npu.is_available()
        except RuntimeError:
            return False
    return hasattr(torch, "npu") and torch.npu.is_available()


@lru_cache
def is_sdaa_available(check_device=False):
    "Checks if `torch_sdaa` is installed and potentially if a SDAA is in the environment"
    if importlib.util.find_spec("torch_sdaa") is None:
        return False

    import torch_sdaa  # noqa: F401

    if check_device:
        try:
            # Will raise a RuntimeError if no NPU is found
            _ = torch.sdaa.device_count()
            return torch.sdaa.is_available()
        except RuntimeError:
            return False
    return hasattr(torch, "sdaa") and torch.sdaa.is_available()


@lru_cache
def is_hpu_available(init_hccl=False):
    "Checks if `torch.hpu` is installed and potentially if a HPU is in the environment"
    if (
        importlib.util.find_spec("habana_frameworks") is None
        or importlib.util.find_spec("habana_frameworks.torch") is None
    ):
        return False

    import habana_frameworks.torch  # noqa: F401

    if init_hccl:
        import habana_frameworks.torch.distributed.hccl as hccl  # noqa: F401

    return hasattr(torch, "hpu") and torch.hpu.is_available()


def is_habana_gaudi1():
    if is_hpu_available():
        import habana_frameworks.torch.utils.experimental as htexp  # noqa: F401

        if htexp._get_device_type() == htexp.synDeviceType.synDeviceGaudi:
            return True

    return False


@lru_cache
def is_xpu_available(check_device=False):
    """
    Checks if XPU acceleration is available either via `intel_extension_for_pytorch` or via stock PyTorch (>=2.4) and
    potentially if a XPU is in the environment
    """

    if is_ipex_available():
        import intel_extension_for_pytorch  # noqa: F401
    else:
        if is_torch_version("<=", "2.3"):
            return False

    if check_device:
        try:
            # Will raise a RuntimeError if no XPU  is found
            _ = torch.xpu.device_count()
            return torch.xpu.is_available()
        except RuntimeError:
            return False
    return hasattr(torch, "xpu") and torch.xpu.is_available()


def is_dvclive_available():
    return _is_package_available("dvclive")


def is_torchdata_available():
    return _is_package_available("torchdata")


# TODO: Remove this function once stateful_dataloader is a stable feature in torchdata.
def is_torchdata_stateful_dataloader_available():
    package_exists = _is_package_available("torchdata")
    if package_exists:
        torchdata_version = version.parse(importlib.metadata.version("torchdata"))
        return compare_versions(torchdata_version, ">=", "0.8.0")
    return False


def torchao_required(func):
    """
    A decorator that ensures the decorated function is only called when torchao is available.
    """

    @wraps(func)
    def wrapper(*args, **kwargs):
        if not is_torchao_available():
            raise ImportError(
                "`torchao` is not available, please install it before calling this function via `pip install torchao`."
            )
        return func(*args, **kwargs)

    return wrapper


# TODO: Rework this into `utils.deepspeed` and migrate the "core" chunks into `accelerate.deepspeed`
def deepspeed_required(func):
    """
    A decorator that ensures the decorated function is only called when deepspeed is enabled.
    """

    @wraps(func)
    def wrapper(*args, **kwargs):
        from accelerate.state import AcceleratorState
        from accelerate.utils.dataclasses import DistributedType

        if AcceleratorState._shared_state != {} and AcceleratorState().distributed_type != DistributedType.DEEPSPEED:
            raise ValueError(
                "DeepSpeed is not enabled, please make sure that an `Accelerator` is configured for `deepspeed` "
                "before calling this function."
            )
        return func(*args, **kwargs)

    return wrapper


def is_weights_only_available():
    # Weights only with allowlist was added in 2.4.0
    # ref: https://github.com/pytorch/pytorch/pull/124331
    return is_torch_version(">=", "2.4.0")


def is_numpy_available(min_version="1.25.0"):
    numpy_version = parse(importlib.metadata.version("numpy"))
    return compare_versions(numpy_version, ">=", min_version)