File size: 10,752 Bytes
1ce5e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# coding=utf-8
# Copyright 2018 The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  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.
"""
    Benchmarking the library on inference and training in PyTorch.
"""


import timeit
from typing import Callable, Optional

from ..configuration_utils import PretrainedConfig
from ..models.auto.modeling_auto import MODEL_MAPPING, MODEL_WITH_LM_HEAD_MAPPING
from ..utils import is_py3nvml_available, is_torch_available, logging
from .benchmark_utils import (
    Benchmark,
    Memory,
    MemorySummary,
    measure_peak_memory_cpu,
    start_memory_tracing,
    stop_memory_tracing,
)


if is_torch_available():
    import torch

    from .benchmark_args import PyTorchBenchmarkArguments


if is_py3nvml_available():
    import py3nvml.py3nvml as nvml


logger = logging.get_logger(__name__)


class PyTorchBenchmark(Benchmark):
    args: PyTorchBenchmarkArguments
    configs: PretrainedConfig
    framework: str = "PyTorch"

    @property
    def framework_version(self):
        return torch.__version__

    def _inference_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
        _inference = self._prepare_inference_func(model_name, batch_size, sequence_length)
        return self._measure_speed(_inference)

    def _inference_memory(
        self, model_name: str, batch_size: int, sequence_length: int
    ) -> [Memory, Optional[MemorySummary]]:
        _inference = self._prepare_inference_func(model_name, batch_size, sequence_length)
        return self._measure_memory(_inference)

    def _train_speed(self, model_name: str, batch_size: int, sequence_length: int) -> float:
        _train = self._prepare_train_func(model_name, batch_size, sequence_length)
        return self._measure_speed(_train)

    def _train_memory(
        self, model_name: str, batch_size: int, sequence_length: int
    ) -> [Memory, Optional[MemorySummary]]:
        _train = self._prepare_train_func(model_name, batch_size, sequence_length)
        return self._measure_memory(_train)

    def _prepare_inference_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
        config = self.config_dict[model_name]

        if self.args.torchscript:
            config.torchscript = True

        has_model_class_in_config = (
            hasattr(config, "architectures")
            and isinstance(config.architectures, list)
            and len(config.architectures) > 0
        )
        if not self.args.only_pretrain_model and has_model_class_in_config:
            try:
                model_class = config.architectures[0]
                transformers_module = __import__("transformers", fromlist=[model_class])
                model_cls = getattr(transformers_module, model_class)
                model = model_cls(config)
            except ImportError:
                raise ImportError(
                    f"{model_class} does not exist. If you just want to test the pretrained model, you might want to"
                    " set `--only_pretrain_model` or `args.only_pretrain_model=True`."
                )
        else:
            model = MODEL_MAPPING[config.__class__](config)

        model.eval()
        model.to(self.args.device)

        # encoder-decoder has vocab size saved differently
        vocab_size = config.vocab_size if hasattr(config, "vocab_size") else config.encoder.vocab_size
        input_ids = torch.randint(vocab_size, (batch_size, sequence_length), dtype=torch.long, device=self.args.device)

        if self.args.fp16:
            logger.info("Running training in Mixed Precision...")
            if not self.args.is_gpu:
                raise ValueError("Mixed precision is possible only for GPU.")
            # amp seems to have memory leaks so that memory usage
            # is measured using .half() for now https://github.com/NVIDIA/apex/issues/439
            model.half()

        if self.args.torchscript:
            with torch.no_grad():
                inference_model = torch.jit.trace(model, input_ids)
        else:
            inference_model = model

        def encoder_decoder_forward():
            with torch.no_grad():
                outputs = inference_model(input_ids, decoder_input_ids=input_ids)
            return outputs

        def encoder_forward():
            with torch.no_grad():
                outputs = inference_model(input_ids)
            return outputs

        _forward = encoder_decoder_forward if config.is_encoder_decoder else encoder_forward
        return _forward

    def _prepare_train_func(self, model_name: str, batch_size: int, sequence_length: int) -> Callable[[], None]:
        config = self.config_dict[model_name]

        has_model_class_in_config = (
            hasattr(config, "architectures")
            and isinstance(config.architectures, list)
            and len(config.architectures) > 0
        )
        if not self.args.only_pretrain_model and has_model_class_in_config:
            try:
                model_class = config.architectures[0]
                transformers_module = __import__("transformers", fromlist=[model_class])
                model_cls = getattr(transformers_module, model_class)
                model = model_cls(config)
            except ImportError:
                raise ImportError(
                    f"{model_class} does not exist. If you just want to test the pretrained model, you might want to"
                    " set `--only_pretrain_model` or `args.only_pretrain_model=True`."
                )
        else:
            model = MODEL_WITH_LM_HEAD_MAPPING[config.__class__](config)

        if self.args.torchscript:
            raise NotImplementedError("Training for torchscript is currently not implemented")
        else:
            train_model = model

        model.train()
        model.to(self.args.device)

        # encoder-decoder has vocab size saved differently
        vocab_size = config.vocab_size if hasattr(config, "vocab_size") else config.encoder.vocab_size
        input_ids = torch.randint(vocab_size, (batch_size, sequence_length), dtype=torch.long, device=self.args.device)

        if self.args.fp16:
            logger.info("Running training in Mixed Precision...")
            if not self.args.is_gpu:
                raise ValueError("Mixed precision is possible only for GPU.")

            # amp seems to have memory leaks so that memory usage
            # is measured using .half() for now https://github.com/NVIDIA/apex/issues/439
            model.half()

        def compute_loss_and_backprob_encoder():
            loss = train_model(input_ids, labels=input_ids)[0]
            loss.backward()
            return loss

        def compute_loss_and_backprob_encoder_decoder():
            loss = train_model(input_ids, decoder_input_ids=input_ids, labels=input_ids)[0]
            loss.backward()
            return loss

        _train = (
            compute_loss_and_backprob_encoder_decoder
            if config.is_encoder_decoder
            else compute_loss_and_backprob_encoder
        )
        return _train

    def _measure_speed(self, func) -> float:
        try:
            if self.args.is_tpu or self.args.torchscript:
                # run additional 10 times to stabilize compilation for tpu and torchscript
                logger.info("Do inference on TPU or torchscript. Running model 5 times to stabilize compilation")
                timeit.repeat(
                    func,
                    repeat=1,
                    number=5,
                )

            # as written in https://docs.python.org/2/library/timeit.html#timeit.Timer.repeat, min should be taken rather than the average
            runtimes = timeit.repeat(
                func,
                repeat=self.args.repeat,
                number=10,
            )

            if self.args.is_tpu and self.args.torch_xla_tpu_print_metrics:
                import torch_xla.debug.metrics as met

                self.print_fn(met.metrics_report())

            return min(runtimes) / 10.0
        except RuntimeError as e:
            self.print_fn(f"Doesn't fit on GPU. {e}")
            return "N/A"

    def _measure_memory(self, func: Callable[[], None]) -> [Memory, MemorySummary]:
        try:
            if self.args.trace_memory_line_by_line:
                trace = start_memory_tracing("transformers")

            if self.args.is_tpu:
                # tpu
                raise NotImplementedError(
                    "Memory Benchmarking is currently not implemented for TPU. Please disable memory benchmarking with"
                    " `--no-memory` or `args.memory=False`"
                )
            elif self.args.is_gpu:
                if not is_py3nvml_available():
                    logger.warning(
                        "py3nvml not installed, we won't log GPU memory usage. "
                        "Install py3nvml (pip install py3nvml) to log information about GPU."
                    )
                    memory = "N/A"
                else:
                    logger.info(
                        "Measuring total GPU usage on GPU device. Make sure to not have additional processes running"
                        " on the same GPU."
                    )
                    # init nvml
                    nvml.nvmlInit()
                    func()
                    handle = nvml.nvmlDeviceGetHandleByIndex(self.args.device_idx)
                    meminfo = nvml.nvmlDeviceGetMemoryInfo(handle)
                    max_bytes_in_use = meminfo.used
                    memory = Memory(max_bytes_in_use)
                    # shutdown nvml
                    nvml.nvmlShutdown()
            else:
                # cpu
                memory_bytes = measure_peak_memory_cpu(func)
                memory = Memory(memory_bytes) if isinstance(memory_bytes, int) else memory_bytes

            if self.args.trace_memory_line_by_line:
                summary = stop_memory_tracing(trace)
            else:
                summary = None

            return memory, summary
        except RuntimeError as e:
            self.print_fn(f"Doesn't fit on GPU. {e}")
            return "N/A", None