File size: 15,251 Bytes
8c9ba62 | 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 | # -*- coding: utf-8 -*-
"""Test cases for Synchronizer modules."""
import asyncio
import multiprocessing
import os
import shutil
import time
import unittest
from copy import deepcopy
from datetime import datetime
from typing import Dict, List
import ray
from parameterized import parameterized_class
from tests.tools import (
TensorBoardParser,
get_checkpoint_path,
get_model_path,
get_template_config,
get_unittest_dataset_config,
)
from trinity.algorithm import ALGORITHM_TYPE
from trinity.cli.launcher import both, explore, train
from trinity.common.config import Config, ExperienceBufferConfig
from trinity.common.constants import StorageType, SyncMethod, SyncStyle
from trinity.explorer.explorer import Explorer
from trinity.trainer.trainer import Trainer
from trinity.utils.log import get_logger
logger = get_logger(__name__)
CHECKPOINT_ROOT_DIR = os.path.join(os.path.dirname(__file__), "temp_checkpoint_dir")
def trainer_monkey_patch(config: Config, max_steps: int, intervals: List[int]):
async def new_sample_data(self):
self.logger.info(f"Sample data for step {self.train_step_num + 1} started.")
await asyncio.sleep(0.1)
time.sleep(intervals[self.engine.global_steps - 1])
self.logger.info(f"Sample data for step {self.train_step_num + 1} finished.")
return [], {}, []
async def new_train_step(self, exps) -> Dict:
self.engine.algorithm = ALGORITHM_TYPE.get(config.algorithm.algorithm_type)
self.engine.global_steps += 1
self.logger.info(f"Training at step {self.engine.global_steps} started.")
await asyncio.sleep(0.1)
time.sleep(intervals[self.engine.global_steps - 1])
metrics = {"actor/step": self.engine.global_steps}
self.logger.info(f"Training at step {self.engine.global_steps} finished.")
return metrics
Trainer.train_step = new_train_step
Trainer._sample_data = new_sample_data
def explorer_monkey_patch(config: Config, max_steps: int, intervals: List[int]):
async def new_explore_step(self):
if self.explore_step_num == max_steps:
await self.save_checkpoint(sync_weight=False)
self.explore_step_num += 1
return self.explore_step_num <= max_steps
def wrapper(old_save_checkpoint):
async def new_save_checkpoint(self, sync_weight: bool = False):
await asyncio.sleep(intervals.pop(0))
await old_save_checkpoint(self, sync_weight)
return new_save_checkpoint
async def new_finish_explore_step(self, step: int, model_version: int) -> None:
metric = {"rollout/model_version": model_version}
self.monitor.log(metric, step=step)
Explorer.explore_step = new_explore_step
Explorer.save_checkpoint = wrapper(Explorer.save_checkpoint)
Explorer._finish_explore_step = new_finish_explore_step
def run_trainer(config: Config, max_steps: int, intervals: List[int]) -> None:
ray.init(ignore_reinit_error=True, namespace=config.ray_namespace)
trainer_monkey_patch(config, max_steps, intervals)
train(config)
ray.shutdown()
def run_explorer(config: Config, max_steps: int, intervals: List[int]) -> None:
ray.init(ignore_reinit_error=True, namespace=config.ray_namespace)
explorer_monkey_patch(config, max_steps, intervals)
explore(config)
ray.shutdown()
def run_both(
config: Config, max_steps: int, trainer_intervals: List[int], explorer_intervals: List[int]
) -> None:
ray.init(ignore_reinit_error=True, namespace=config.ray_namespace)
trainer_monkey_patch(config, max_steps, trainer_intervals)
explorer_monkey_patch(config, max_steps, explorer_intervals)
both(config)
ray.shutdown()
class BaseTestSynchronizer(unittest.TestCase):
def setUp(self):
if multiprocessing.get_start_method(allow_none=True) != "spawn":
multiprocessing.set_start_method("spawn", force=True)
self.process_list = []
def tearDown(self):
ray.shutdown(_exiting_interpreter=True)
if os.path.exists(CHECKPOINT_ROOT_DIR):
shutil.rmtree(CHECKPOINT_ROOT_DIR, ignore_errors=True)
for process in self.process_list:
if process.is_alive():
process.terminate()
process.join(timeout=10)
if process.is_alive():
process.kill()
process.join()
class TestSynchronizerExit(BaseTestSynchronizer):
def test_synchronizer(self):
config = get_template_config()
config.project = "unittest"
config.name = f"test_synchronizer_{datetime.now().strftime('%Y%m%d%H%M%S')}"
config.checkpoint_root_dir = get_checkpoint_path()
config.buffer.total_epochs = 1
config.buffer.batch_size = 4
config.cluster.gpu_per_node = 2
config.cluster.node_num = 1
config.model.model_path = get_model_path()
config.buffer.explorer_input.taskset = get_unittest_dataset_config("countdown")
config.buffer.trainer_input.experience_buffer = ExperienceBufferConfig(
name="exp_buffer",
storage_type=StorageType.QUEUE.value,
)
config.synchronizer.sync_method = SyncMethod.CHECKPOINT
config.synchronizer.sync_style = SyncStyle.DYNAMIC_BY_EXPLORER
config.synchronizer.sync_interval = 2
config.trainer.save_interval = 100
config.monitor.monitor_type = "tensorboard"
trainer_config = deepcopy(config)
trainer_config.mode = "train"
trainer_config.buffer.train_batch_size = 4
trainer_config.check_and_update()
explorer1_config = deepcopy(config)
explorer1_config.mode = "explore"
explorer1_config.explorer.name = "explorer1"
explorer1_config.explorer.rollout_model.engine_num = 1
explorer1_config.explorer.rollout_model.tensor_parallel_size = 1
explorer1_config.buffer.explorer_output = ExperienceBufferConfig(
name="exp_buffer",
storage_type=StorageType.QUEUE.value,
)
explorer1_config.check_and_update()
trainer_process = multiprocessing.Process(
target=run_trainer, args=(trainer_config, 8, [2, 1, 2, 1, 2, 1, 2, 1])
)
trainer_process.start()
self.process_list.append(trainer_process)
ray.init(ignore_reinit_error=True)
while True:
try:
synchronizer = ray.get_actor("synchronizer", namespace=trainer_config.ray_namespace)
break
except ValueError:
print("waiting for trainer to start.")
time.sleep(5)
explorer_process_1 = multiprocessing.Process(
target=run_explorer,
args=(explorer1_config, 8, [0, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5]),
)
explorer_process_1.start()
self.process_list.append(explorer_process_1)
self.assertEqual(
synchronizer, ray.get_actor("synchronizer", namespace=trainer_config.ray_namespace)
)
for _ in range(12): # Wait for up to 60 seconds
try:
explorer1 = ray.get_actor("explorer1", namespace=trainer_config.ray_namespace)
ray.get(explorer1.is_alive.remote())
break
except ValueError:
print("waiting for explorer1 to start.")
time.sleep(5)
trainer_process.join(timeout=200)
self.assertEqual(
synchronizer, ray.get_actor("synchronizer", namespace=trainer_config.ray_namespace)
)
explorer_process_1.join(timeout=200)
time.sleep(6)
with self.assertRaises(ValueError):
ray.get_actor("synchronizer", namespace=trainer_config.ray_namespace)
@parameterized_class(
(
"sync_method",
"sync_style",
"max_steps",
"trainer_intervals",
"explorer1_intervals",
"explorer2_intervals",
),
[
(
SyncMethod.CHECKPOINT,
SyncStyle.FIXED,
8,
[2, 1, 2, 1, 2, 1, 2, 1],
[0, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5],
[0, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5],
),
(
SyncMethod.CHECKPOINT,
SyncStyle.DYNAMIC_BY_EXPLORER,
8,
[2, 1, 2, 1, 2, 1, 2, 1],
[0, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5],
[0, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5],
),
(
SyncMethod.MEMORY,
SyncStyle.FIXED,
8,
[2, 1, 2, 1, 2, 1, 2, 1],
[0, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5],
[0, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5],
),
(
SyncMethod.MEMORY,
SyncStyle.DYNAMIC_BY_EXPLORER,
8,
[2, 1, 2, 1, 2, 1, 2, 1],
[0, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5, 2.5],
[0, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5],
),
],
)
class TestStateDictBasedSynchronizer(BaseTestSynchronizer):
def test_synchronizer(self):
config = get_template_config()
config.project = "unittest"
config.name = f"test_synchronizer_{datetime.now().strftime('%Y%m%d%H%M%S')}"
config.checkpoint_root_dir = get_checkpoint_path()
config.buffer.total_epochs = 1
config.buffer.batch_size = 4
config.cluster.gpu_per_node = 2
config.cluster.node_num = 1
config.model.model_path = get_model_path()
config.buffer.explorer_input.taskset = get_unittest_dataset_config("countdown")
config.buffer.trainer_input.experience_buffer = ExperienceBufferConfig(
name="exp_buffer",
storage_type=StorageType.QUEUE.value,
)
config.synchronizer.sync_method = self.sync_method
config.synchronizer.sync_style = self.sync_style
config.synchronizer.sync_interval = 2
config.trainer.save_interval = 100
config.trainer.total_steps = self.max_steps
config.monitor.monitor_type = "tensorboard"
trainer_config = deepcopy(config)
trainer_config.mode = "train"
trainer_config.buffer.train_batch_size = 4
trainer_config.check_and_update()
explorer1_config = deepcopy(config)
explorer1_config.mode = "explore"
explorer1_config.explorer.name = "explorer1"
explorer1_config.explorer.rollout_model.engine_num = 1
explorer1_config.explorer.rollout_model.tensor_parallel_size = 1
explorer1_config.buffer.explorer_output = ExperienceBufferConfig(
name="exp_buffer",
storage_type=StorageType.QUEUE.value,
)
explorer2_config = deepcopy(explorer1_config)
explorer2_config.explorer.name = "explorer2"
explorer1_config.check_and_update()
explorer2_config.check_and_update()
trainer_process = multiprocessing.Process(
target=run_trainer, args=(trainer_config, self.max_steps, self.trainer_intervals)
)
trainer_process.start()
self.process_list.append(trainer_process)
ray.init(ignore_reinit_error=True)
while True:
try:
ray.get_actor("queue-exp_buffer", namespace=trainer_config.ray_namespace)
break
except ValueError:
print("waiting for trainer to start.")
time.sleep(5)
explorer_process_1 = multiprocessing.Process(
target=run_explorer,
args=(explorer1_config, self.max_steps, self.explorer1_intervals),
)
explorer_process_1.start()
self.process_list.append(explorer_process_1)
explorer_process_2 = multiprocessing.Process(
target=run_explorer, args=(explorer2_config, self.max_steps, self.explorer2_intervals)
)
explorer_process_2.start()
self.process_list.append(explorer_process_2)
explorer_process_1.join(timeout=200)
explorer_process_2.join(timeout=200)
trainer_process.join(timeout=200)
# check the tensorboard
parser = TensorBoardParser(
os.path.join(trainer_config.monitor.cache_dir, "tensorboard", "trainer")
)
actor_metrics = parser.metric_list("actor")
self.assertEqual(parser.metric_max_step(actor_metrics[0]), 8)
parser = TensorBoardParser(
os.path.join(explorer1_config.monitor.cache_dir, "tensorboard", "explorer1")
)
rollout_metrics = parser.metric_list("rollout")
self.assertEqual(parser.metric_max_step(rollout_metrics[0]), 8)
parser = TensorBoardParser(
os.path.join(explorer2_config.monitor.cache_dir, "tensorboard", "explorer2")
)
rollout_metrics = parser.metric_list("rollout")
self.assertEqual(parser.metric_max_step(rollout_metrics[0]), 8)
@parameterized_class(
("sync_style", "max_steps", "trainer_intervals", "explorer_intervals"),
[
(
SyncStyle.FIXED,
8,
[2, 1, 2, 1, 2, 1, 2, 1],
[0, 2.5, 2.5, 2.5, 2.5, 0],
),
(
SyncStyle.DYNAMIC_BY_EXPLORER,
8,
[2, 1, 2, 1, 2, 1, 2, 1],
[0, 0.5, 0.5, 0.5, 0.5, 0],
),
],
)
class TestNCCLBasedSynchronizer(BaseTestSynchronizer):
def test_synchronizer(self):
config = get_template_config()
config.project = "unittest"
config.name = f"test_synchronizer_{datetime.now().strftime('%Y%m%d%H%M%S')}"
config.checkpoint_root_dir = get_checkpoint_path()
config.buffer.total_epochs = 1
config.buffer.batch_size = 4
config.trainer.total_steps = self.max_steps
config.model.model_path = get_model_path()
config.buffer.explorer_input.taskset = get_unittest_dataset_config("countdown")
config.buffer.trainer_input.experience_buffer = ExperienceBufferConfig(
name="exp_buffer",
storage_type=StorageType.QUEUE.value,
)
config.synchronizer.sync_method = SyncMethod.NCCL
config.synchronizer.sync_style = self.sync_style
config.synchronizer.sync_interval = 2
config.trainer.save_interval = 100
config.monitor.monitor_type = "tensorboard"
config.mode = "both"
config.check_and_update()
# TODO: test more interval cases
both_process = multiprocessing.Process(
target=run_both,
args=(config, self.max_steps, self.trainer_intervals, self.explorer_intervals),
)
both_process.start()
self.process_list.append(both_process)
both_process.join(timeout=200)
# check the tensorboard
parser = TensorBoardParser(os.path.join(config.monitor.cache_dir, "tensorboard", "trainer"))
actor_metrics = parser.metric_list("actor")
self.assertEqual(parser.metric_max_step(actor_metrics[0]), 8)
parser = TensorBoardParser(
os.path.join(config.monitor.cache_dir, "tensorboard", "explorer")
)
rollout_metrics = parser.metric_list("rollout")
self.assertEqual(parser.metric_max_step(rollout_metrics[0]), 8)
|