forked from pytorch/torchtitan
-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_runner.py
executable file
·423 lines (401 loc) · 14.1 KB
/
test_runner.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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
import argparse
import logging
import os
import subprocess
from collections import defaultdict
from dataclasses import dataclass
from typing import Sequence
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
try:
import tomllib
except ModuleNotFoundError:
import tomli as tomllib
@dataclass
class OverrideDefinitions:
"""
This class is used to define the override definitions for the integration tests.
"""
override_args: Sequence[Sequence[str]] = tuple(tuple(" "))
test_descr: str = "default"
test_name: str = "default"
requires_seed_checkpoint: bool = False
ngpu: int = 4
model_flavor: str = "debugmodel"
def __repr__(self):
return self.test_descr
def build_test_list():
"""
key is the config file name and value is a list of OverrideDefinitions
that is used to generate variations of integration tests based on the
same root config file.
"""
integration_tests_flavors = defaultdict(list)
integration_tests_flavors["debug_model.toml"] = [
OverrideDefinitions(
[
[],
],
"default",
"default",
),
OverrideDefinitions(
[
[
"--training.compile",
],
],
"1D compile",
"1d_compile",
),
OverrideDefinitions(
[
[
"--training.compile",
"--activation_checkpoint.mode selective",
"--activation_checkpoint.selective_ac_option op",
],
],
"1D compile with selective op AC",
"1d_compile_sac_op",
),
OverrideDefinitions(
[
[
"--training.tensor_parallel_degree 2",
],
],
"2D eager",
"2d_eager",
),
OverrideDefinitions(
[
[
"--training.compile",
"--training.tensor_parallel_degree 2",
],
],
"2D compile",
"2d_compile",
),
OverrideDefinitions(
[
[
"--training.tensor_parallel_degree 2",
"--model.norm_type=fused_rmsnorm",
],
],
"2D eager with fused_rmsnorm",
"2d_eager_fused_rmsnorm",
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
],
[
"--checkpoint.enable_checkpoint",
"--training.steps 20",
],
],
"Checkpoint Integration Test - Save Load Full Checkpoint",
"full_checkpoint",
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--checkpoint.model_weights_only",
],
],
"Checkpoint Integration Test - Save Model Weights Only fp32",
"model_weights_only_fp32",
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--checkpoint.model_weights_only",
"--checkpoint.export_dtype bfloat16",
],
],
"Checkpoint Integration Test - Save Model Weights Only bf16",
"model_weights_only_bf16",
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 4",
"--experimental.pipeline_parallel_split_points layers.1,layers.2,layers.3,layers.4,layers.5,layers.6,layers.7",
"--experimental.pipeline_parallel_schedule flexible_interleaved_1f1b",
],
],
"PP looped flexible 1f1b test",
"pp_looped_flexible_1f1b",
requires_seed_checkpoint=True,
ngpu=4,
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 2",
"--experimental.pipeline_parallel_split_points layers.4",
"--experimental.pipeline_parallel_schedule 1f1b",
"--training.data_parallel_shard_degree 1",
],
],
"PP 1D test 1f1b",
"pp_1f1b",
requires_seed_checkpoint=True,
ngpu=2,
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 2",
"--experimental.pipeline_parallel_split_points layers.4",
"--experimental.pipeline_parallel_schedule gpipe",
"--training.data_parallel_shard_degree 1",
],
],
"PP 1D test gpipe",
"pp_gpipe",
requires_seed_checkpoint=True,
ngpu=2,
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 2",
"--experimental.pipeline_parallel_split_points layers.4",
"--experimental.pipeline_parallel_schedule 1f1b",
"--training.data_parallel_shard_degree 2",
],
],
"PP+DP 1f1b 2D test",
"pp_dp_1f1b",
requires_seed_checkpoint=True,
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 2",
"--experimental.pipeline_parallel_split_points layers.4",
"--experimental.pipeline_parallel_schedule gpipe",
"--training.data_parallel_shard_degree 2",
],
],
"PP+DP gpipe 2D test",
"pp_dp_gpipe",
requires_seed_checkpoint=True,
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 2",
"--experimental.pipeline_parallel_split_points layers.4",
"--training.tensor_parallel_degree 2",
],
],
"PP+TP 2D test",
"pp_tp",
requires_seed_checkpoint=True,
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 2",
"--experimental.pipeline_parallel_split_points layers.4",
"--training.data_parallel_shard_degree 2",
"--training.tensor_parallel_degree 2",
],
[
"--training.steps 20",
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 2",
"--experimental.pipeline_parallel_split_points layers.4",
"--training.data_parallel_shard_degree 2",
"--training.tensor_parallel_degree 2",
],
],
"PP+DP+TP 3D test with save/load resume ckpt",
"pp_dp_tp",
requires_seed_checkpoint=True,
ngpu=8,
),
OverrideDefinitions(
[
[
"--experimental.pipeline_parallel_degree 2",
"--experimental.pipeline_parallel_split_points layers.4",
"--training.data_parallel_shard_degree 2",
"--training.tensor_parallel_degree 2",
"--training.compile",
],
],
"PP+DP+TP 3D test with torch.compile",
"3d_compile",
requires_seed_checkpoint=True,
ngpu=8,
),
OverrideDefinitions(
[
[
"--checkpoint.enable_checkpoint",
"--experimental.pipeline_parallel_degree 4",
"--experimental.pipeline_parallel_split_points layers.1,layers.2,layers.3,layers.4,layers.5,layers.6,layers.7",
"--experimental.pipeline_parallel_schedule interleaved_1f1b",
],
],
"PP looped 1f1b test",
"pp_looped_1f1b",
requires_seed_checkpoint=True,
ngpu=4,
),
OverrideDefinitions(
[
[
"--optimizer.name Adam --optimizer.fused",
"--optimizer.name AdamW --optimizer.fused",
]
],
"Fused Optimizer Test",
),
OverrideDefinitions(
[
[
"--training.data_parallel_shard_degree=1",
"--training.data_parallel_replicate_degree=4",
]
],
"DDP",
"ddp",
ngpu=4,
),
OverrideDefinitions(
[
[
"--training.data_parallel_shard_degree=2",
"--training.data_parallel_replicate_degree=2",
]
],
"HSDP",
"hsdp",
ngpu=4,
),
OverrideDefinitions(
[
[
"--training.data_parallel_shard_degree=2",
"--training.data_parallel_replicate_degree=2",
"--training.tensor_parallel_degree=2",
]
],
"HSDP+TP",
"hsdp+tp",
ngpu=8,
),
OverrideDefinitions(
[
[
"--memory_estimation.enabled",
]
],
"FSDP2 Memory Tracking and Estimation",
"fsdp2_mem_tracker",
ngpu=4,
),
]
return integration_tests_flavors
def _run_cmd(cmd):
return subprocess.run(
[cmd],
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True,
shell=True,
)
def run_test(test_flavor: OverrideDefinitions, full_path: str, output_dir: str):
# run_test supports sequence of tests.
test_name = test_flavor.test_name
dump_folder_arg = f"--job.dump_folder {output_dir}/{test_name}"
model_flavor_arg = f"--model.flavor {test_flavor.model_flavor}"
all_ranks = ",".join(map(str, range(test_flavor.ngpu)))
if test_flavor.requires_seed_checkpoint:
cmd = f"CONFIG_FILE={full_path} ./create_seed_checkpoint.sh {dump_folder_arg} {model_flavor_arg}"
logger.info(
f"=====Integration test, flavor : {test_flavor.test_descr}, command : {cmd}====="
)
result = _run_cmd(cmd)
logger.info(result.stdout)
for override_arg in test_flavor.override_args:
cmd = f"CONFIG_FILE={full_path} NGPU={test_flavor.ngpu} LOG_RANK={all_ranks} ./run_llama_train.sh"
if test_name == "fsdp2_mem_tracker":
cmd = f"CONFIG_FILE={full_path} NGPU={test_flavor.ngpu} LOG_RANK={all_ranks} ./run_memory_estimation.sh"
cmd += " " + dump_folder_arg
cmd += " " + model_flavor_arg
if override_arg:
cmd += " " + " ".join(override_arg)
logger.info(
f"=====Integration test, flavor : {test_flavor.test_descr}, command : {cmd}====="
)
result = _run_cmd(cmd)
logger.info(result.stdout)
if result.returncode != 0:
raise Exception(
f"Integration test failed, flavor : {test_flavor.test_descr}, command : {cmd}"
)
def run_tests(args):
integration_tests_flavors = build_test_list()
for config_file in os.listdir(args.config_dir):
if config_file.endswith(".toml"):
full_path = os.path.join(args.config_dir, config_file)
with open(full_path, "rb") as f:
config = tomllib.load(f)
is_integration_test = config["job"].get(
"use_for_integration_test", False
)
if is_integration_test:
for test_flavor in integration_tests_flavors[config_file]:
if args.test == "all" or test_flavor.test_name == args.test:
if args.ngpu < test_flavor.ngpu:
logger.info(
f"Skipping test {test_flavor.test_name} that requires {test_flavor.ngpu} gpus,"
f" because --ngpu arg is {args.ngpu}"
)
elif args.ngpu == 8 and test_flavor.ngpu != 8:
logger.info(
f"Skipping non-8gpu test {test_flavor.test_name} on 8-gpu runner"
)
else:
run_test(test_flavor, full_path, args.output_dir)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("output_dir")
parser.add_argument("--config_dir", default="./train_configs")
parser.add_argument(
"--test",
default="all",
help="test to run, acceptable values: `test_name` in `build_test_list` (default: all)",
)
parser.add_argument("--ngpu", default=4, type=int)
args = parser.parse_args()
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if os.listdir(args.output_dir):
raise RuntimeError("Please provide an empty output directory.")
run_tests(args)
if __name__ == "__main__":
main()