File size: 20,689 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 | """A centralized synchronizer for coordinating explorer and trainer."""
import asyncio
import os
import shutil
from collections import defaultdict
from typing import Dict, List, Optional, Tuple, Union
import ray
from trinity.common.config import Config
from trinity.common.constants import RunningStatus, SyncMethod
from trinity.common.models.utils import (
get_checkpoint_dir_with_step_num,
load_state_dict,
)
from trinity.utils.log import get_logger
class Synchronizer:
"""
A central component to manage synchronization of models and states between
the trainer and one or more explorers in a distributed training setup.
Attributes:
trainer_status: Current status of the trainer (e.g., running, waiting).
explorer_status_counts: Dictionary tracking the number of explorers in each status.
_ready_condition: Async condition variable for signaling state changes.
model_state_dict: The latest model weights.
model_version: Version number of the current model.
checkpoint_shard_counter: Tracks how many shards are received from trainer for a specific train step.
"""
def __init__(self, config: Config, module_ref: ray.actor.ActorHandle):
self.logger = get_logger("synchronizer", in_ray_actor=True)
self.config = config
self.enable_lora = config.explorer.rollout_model.enable_lora
self.trainer_status = RunningStatus.STOPPED
self.explorer_status_counts: Dict[RunningStatus, int] = defaultdict(lambda: 0)
self._ready_condition = asyncio.Condition()
self.model_state_dict = None
self.model_version = 0
self.named_model_state_dict: Dict[str, Union[dict, None, str, Tuple[str, str]]] = {}
self.named_model_version: Dict[str, int] = {}
self.checkpoint_shard_counter = defaultdict(lambda: 0)
self.ref_count = 0
self._modules = {module_ref}
self._modules_lock = asyncio.Lock()
asyncio.create_task(self._check_modules())
if (
self.config.mode != "bench"
and self.config.synchronizer.sync_method == SyncMethod.CHECKPOINT
):
asyncio.create_task(self._find_latest_state_dict())
async def add_module(self, module_ref: ray.actor.ActorHandle) -> None:
"""Adds a module to be tracked by the synchronizer.
Args:
module_ref: The Ray actor handle of the module to track.
"""
async with self._modules_lock:
if module_ref not in self._modules:
self._modules.add(module_ref)
async def _check_modules(self) -> None:
while len(self._modules) > 0:
alive_modules = set()
async with self._modules_lock:
for module in self._modules:
try:
is_alive_ref = module.is_alive.remote()
await asyncio.wait_for(is_alive_ref, timeout=5.0)
alive_modules.add(module)
except ray.exceptions.RayActorError:
pass
except asyncio.TimeoutError:
ray.cancel(is_alive_ref)
alive_modules.add(module)
self._modules = alive_modules
await asyncio.sleep(1)
self.logger.info("Synchronizer stopped.")
try:
ray.actor.exit_actor()
except Exception:
pass
async def _find_latest_state_dict(self) -> None:
if self.config.trainer.trainer_type == "verl":
await self._find_verl_latest_state_dict()
elif self.config.trainer.trainer_type == "tinker":
await self._find_tinker_latest_state_dict()
else:
self.logger.warning(
"Synchronizer does not support this trainer type. Please use `verl` or `tinker`."
)
async def _find_verl_latest_state_dict(self) -> None:
default_local_dir = self.config.checkpoint_job_dir
local_latest_state_dict_iteration = os.path.join(
default_local_dir, "latest_state_dict_iteration.txt"
)
while True:
if os.path.exists(local_latest_state_dict_iteration):
current_model_version = self.model_version
try:
with open(local_latest_state_dict_iteration, "r") as f:
latest_model_version = int(f.read().strip())
except (IOError, ValueError) as e:
self.logger.warning(f"Failed to read or parse state dict iteration file: {e}")
continue
if latest_model_version > current_model_version:
self.logger.info(
f"Synchronizer has found a new model state dict at step {latest_model_version}."
)
model_state_dict = (
load_state_dict(
os.path.join(
default_local_dir, f"global_step_{latest_model_version}", "actor"
),
self.config.trainer,
)
if not self.enable_lora
else {}
)
self.logger.info(
f"Synchronizer has loaded model state dict from checkpoint {latest_model_version}."
)
await self.set_model_state_dict(model_state_dict, latest_model_version)
# remove the previous checkpoints to save disk space
await self._remove_previous_state_dict(current_model_version)
await asyncio.sleep(1)
async def _remove_previous_state_dict(self, previous_model_version: int) -> None:
previous_state_dict_dir = os.path.join(
self.config.checkpoint_job_dir, f"global_step_{previous_model_version}"
)
if os.path.exists(previous_state_dict_dir):
# check if it's a full checkpoint, only remove checkpoints for sync
if not os.path.exists(os.path.join(previous_state_dict_dir, ".full_checkpoint")):
self.logger.info(
f"Removing previous checkpoint for sync at step {previous_model_version}."
)
shutil.rmtree(previous_state_dict_dir, ignore_errors=True)
async def _find_tinker_latest_state_dict(self) -> None:
default_local_dir = self.config.checkpoint_job_dir
local_latest_state_dict_iteration = os.path.join(
default_local_dir, "latest_state_dict_iteration.txt"
)
while True:
if os.path.exists(local_latest_state_dict_iteration):
try:
with open(local_latest_state_dict_iteration, "r") as f:
latest_model_version = int(f.read().strip())
except (IOError, ValueError) as e:
self.logger.warning(f"Failed to read or parse state dict iteration file: {e}")
continue
if latest_model_version > self.model_version:
self.logger.info(
f"Synchronizer has found a new remote tinker sampler path at step {latest_model_version}."
)
remote_path_file = os.path.join(
default_local_dir,
f"global_step_{latest_model_version}",
"remote_sampler_path.txt",
)
with open(remote_path_file, "r") as f:
remote_sampler_path = f.read().strip()
await self.set_model_state_dict(remote_sampler_path, latest_model_version)
await asyncio.sleep(1)
async def set_trainer_status(self, status: RunningStatus):
"""Update the status of the trainer."""
async with self._ready_condition:
self.trainer_status = status
if status == RunningStatus.STOPPED:
self._ready_condition.notify_all()
def get_trainer_status(self) -> RunningStatus:
"""Get the current status of the trainer."""
return self.trainer_status
async def set_explorer_status(
self, status: RunningStatus, old_status: Optional[RunningStatus] = None
):
"""
Update the status count for an explorer.
Args:
status: New status of the explorer.
old_status: Previous status if changing from one to another.
"""
if old_status is not None:
assert (
old_status in self.explorer_status_counts
), f"Invalid explorer status {old_status}"
assert old_status != status, f"Invalid status change from {old_status} to {status}"
self.explorer_status_counts[old_status] -= 1
assert (
self.explorer_status_counts[old_status] >= 0
), f"Invalid status count {old_status} (new status {status})"
if status not in self.explorer_status_counts:
self.explorer_status_counts[status] = 0
self.explorer_status_counts[status] += 1
def get_explorer_status_counts(self) -> Dict[RunningStatus, int]:
"""Return the current status counts for all explorers."""
return self.explorer_status_counts
async def set_model_state_dict_with_step_num(
self, step_num: Optional[int] = None, world_size: Optional[int] = None
) -> int:
"""
Load and set the model state dictionary from a checkpoint at a specific step.
Args:
step_num: Training step number corresponding to the checkpoint.
world_size: Number of shards expected for this checkpoint.
Returns:
The updated model version (step number).
"""
if world_size is not None: # Used when trainer updates the model
assert step_num is not None
assert self.checkpoint_shard_counter[step_num] < world_size, "World size mismatch!"
self.checkpoint_shard_counter[step_num] += 1
self.logger.info(
f"Synchronizer has received {self.checkpoint_shard_counter[step_num]} out of {world_size} shards from the checkpoint {step_num}."
)
if self.checkpoint_shard_counter[step_num] < world_size:
return step_num
checkpoint_dir, checkpoint_step_num = get_checkpoint_dir_with_step_num(
checkpoint_root_path=self.config.checkpoint_job_dir,
trainer_type=self.config.trainer.trainer_type,
step_num=step_num,
)
if checkpoint_step_num != self.model_version:
model_state_dict = (
load_state_dict(
os.path.join(checkpoint_dir, "actor"),
self.config.trainer,
)
if not self.enable_lora
else {}
)
# lora weights are stored in 'lora_adapter' subfolder and cannot be loaded directly
await self.set_model_state_dict(model_state_dict, checkpoint_step_num)
return checkpoint_step_num
async def set_model_state_dict(
self, model_state_dict: Union[dict, None, str, Tuple[str, str]], trainer_step: int
):
"""
Set the new model state and update the version.
Args:
model_state_dict: The PyTorch model state dictionary.
trainer_step: Step number associated with this model version.
"""
async with self._ready_condition:
self.model_state_dict = model_state_dict
self.model_version = trainer_step
self.logger.info(f"Set model state dict version to {trainer_step}.")
self._ready_condition.notify_all()
async def set_named_model_state_dict(
self,
name: str,
model_state_dict: Union[dict, None, str, Tuple[str, str]],
trainer_step: int,
):
"""Set a named model state dict and its version (e.g. teacher EMA)."""
async with self._ready_condition:
self.named_model_state_dict[name] = model_state_dict
self.named_model_version[name] = trainer_step
self.logger.info(f"Set named model state dict `{name}` version to {trainer_step}.")
self._ready_condition.notify_all()
def get_model_state_dict(self, source: str = "student"):
"""Return model state dict and version for a given source."""
if source == "student":
return self.model_state_dict, self.model_version
return self.named_model_state_dict.get(source), self.named_model_version.get(source, -1)
async def get_state_dict_meta(self, source: str = "student"):
"""
Return metadata about the model state (names, data types, shapes).
Returns:
List of tuples: (name, dtype, shape).
"""
state_dict = self.model_state_dict if source == "student" else self.named_model_state_dict.get(source)
if state_dict is None:
return None
if isinstance(state_dict, tuple):
async with self._ready_condition:
await self._ready_condition.wait_for(
lambda: not isinstance(
self.model_state_dict
if source == "student"
else self.named_model_state_dict.get(source),
tuple,
)
)
state_dict = (
self.model_state_dict if source == "student" else self.named_model_state_dict.get(source)
)
update_weight_args_list = []
for name, param in state_dict.items():
update_weight_args_list.append((name, str(param.dtype), tuple(param.shape)))
return update_weight_args_list
async def setup_weight_sync_group(
self, master_address: str, master_port: int, state_dict_meta: List = None
):
"""
Notify the explorer actor to setup weight sync group.
This is used to initialize NCCL-based synchronization for distributed training.
Args:
master_address: IP address of the master node.
master_port: Port used for synchronization.
state_dict_meta: Metadata of the model parameters.
"""
explorer = ray.get_actor(self.config.explorer.name, namespace=self.config.ray_namespace)
await explorer.setup_weight_sync_group.remote(master_address, master_port, state_dict_meta)
async def wait_new_model_state_dict(self, current_version: int, no_wait: bool = False) -> int:
"""
Wait until a new model state is available.
Args:
current_version: Current model version known to one explorer.
Returns:
The new model version after it has been updated.
"""
async with self._ready_condition:
assert (
self.model_version >= current_version
), f"The model version in Synchronizer ({self.model_version}) should be no smaller than that in Explorer ({current_version})!"
if self.model_version == current_version:
if not no_wait and self.trainer_status != RunningStatus.STOPPED:
# TODO: explorer need support no wait
# TODO: handle timeout
await asyncio.wait_for(
self._ready_condition.wait(),
timeout=self.config.synchronizer.sync_timeout,
)
if self.model_version > current_version:
await self.set_explorer_status(
RunningStatus.RUNNING, old_status=RunningStatus.REQUIRE_SYNC
)
return self.model_version
async def get_latest_model_version(self) -> int:
"""
Get the latest model version available in the synchronizer.
Returns:
The current model version.
"""
async with self._ready_condition:
return self.model_version
async def ready_to_nccl_sync(self, module: str, trainer_step: int) -> Union[int, None]:
"""
Prepare for NCCL-based synchronization between modules.
Only supports one explorer currently.
Args:
module: Either 'trainer' or 'explorer'.
trainer_step: Step number from the trainer.
Returns:
The model version if both sides are ready; otherwise None.
"""
assert (
sum(self.explorer_status_counts.values()) == 1
), "NCCL sync is only supported for one explorer."
async def sync_failed():
if module == "explorer":
another_module = "Trainer"
await self.set_explorer_status(
RunningStatus.REQUIRE_SYNC, old_status=RunningStatus.WAITING_SYNC
)
else:
another_module = "Explorer"
self.trainer_status = RunningStatus.REQUIRE_SYNC
self.logger.error(f"{another_module} is not ready for model weight sync.")
return None
non_stop_cnt = sum(
value
for key, value in self.explorer_status_counts.items()
if key != RunningStatus.STOPPED
)
if non_stop_cnt == 0:
return await sync_failed()
async with self._ready_condition:
try:
if module == "trainer":
self.model_version = trainer_step
self.trainer_status = RunningStatus.WAITING_SYNC
self._ready_condition.notify_all()
if self.explorer_status_counts[RunningStatus.WAITING_SYNC] != 1:
await asyncio.wait_for(
self._ready_condition.wait_for(
lambda: self.explorer_status_counts[RunningStatus.WAITING_SYNC]
+ self.explorer_status_counts[RunningStatus.STOPPED]
== 1,
),
timeout=self.config.synchronizer.sync_timeout,
)
if self.explorer_status_counts[RunningStatus.STOPPED] == 1:
return await sync_failed()
await self.set_explorer_status(
RunningStatus.RUNNING,
old_status=RunningStatus.WAITING_SYNC,
)
elif module == "explorer":
await self.set_explorer_status(
RunningStatus.WAITING_SYNC, old_status=RunningStatus.REQUIRE_SYNC
)
self._ready_condition.notify_all()
if self.trainer_status != RunningStatus.WAITING_SYNC:
await asyncio.wait_for(
self._ready_condition.wait_for(
lambda: self.trainer_status
in {RunningStatus.WAITING_SYNC, RunningStatus.STOPPED},
),
timeout=self.config.synchronizer.sync_timeout,
)
if self.trainer_status == RunningStatus.STOPPED:
return await sync_failed()
self.trainer_status = RunningStatus.RUNNING
return self.model_version
except asyncio.TimeoutError:
return await sync_failed()
@classmethod
def get_actor(cls, config: Optional[Config] = None, namespace: Optional[str] = None):
"""
Get or create a remote Ray actor for the Synchronizer.
Args:
config: Optional configuration to use for creating the actor.
namespace: Optional Ray namespace for the actor.
Returns:
A reference to the Synchronizer actor.
"""
if config is not None:
module_ref = ray.get_runtime_context().current_actor
synchronizer = (
ray.remote(cls)
.options(
name="synchronizer",
namespace=config.ray_namespace,
get_if_exists=True,
lifetime="detached",
)
.remote(config, module_ref=module_ref)
)
synchronizer.add_module.remote(module_ref)
return synchronizer
return ray.get_actor("synchronizer", namespace=namespace)
|