File size: 11,537 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 | # -*- coding: utf-8 -*-
"""
Trainer Class
"""
from __future__ import annotations
import asyncio
import time
import traceback
from abc import ABC, abstractmethod
from typing import Dict, List, Tuple
import pandas as pd
import ray
from trinity.algorithm import SAMPLE_STRATEGY
from trinity.algorithm.sample_strategy.sample_strategy import SampleStrategy
from trinity.common.config import Config
from trinity.common.constants import RunningStatus, SyncMethod, SyncStyle
from trinity.common.experience import Experience
from trinity.manager.state_manager import StateManager
from trinity.manager.synchronizer import Synchronizer
from trinity.utils.log import get_logger
from trinity.utils.monitor import MONITOR
from trinity.utils.plugin_loader import load_plugins
from trinity.utils.timer import Timer
class Trainer:
"""Consume the experience and train the model."""
def __init__(self, config: Config) -> None:
self.config = config
self.logger = get_logger(config.trainer.name, in_ray_actor=True)
load_plugins()
self.synchronizer = Synchronizer.get_actor(config)
self.engine = get_trainer_wrapper(config)
self.state = StateManager(
path=config.checkpoint_job_dir, trainer_name=config.trainer.name, config=config
)
trainer_state = self.state.load_trainer()
self.monitor = MONITOR.get(config.monitor.monitor_type)(
project=config.project,
group=self.config.group,
name=config.name,
role=config.trainer.name,
config=config,
)
self._sample_exps_to_log = []
self.sample_strategy: SampleStrategy = SAMPLE_STRATEGY.get(
config.algorithm.sample_strategy
)(
buffer_config=config.buffer,
**config.algorithm.sample_strategy_args,
)
if "latest_exp_index" in trainer_state:
sample_strategy_state = {"current_index": trainer_state["latest_exp_index"]}
else:
sample_strategy_state = trainer_state.get("sample_strategy_state", {})
self.sample_strategy.load_state_dict(sample_strategy_state)
self.save_interval = config.trainer.save_interval
self.last_sync_step = 0
self.last_sync_time = None
self.total_steps = config.trainer.total_steps or float("inf")
self.save_hf_checkpoint = config.trainer.save_hf_checkpoint
async def prepare(self) -> None:
"""Prepare the trainer."""
await self.engine.prepare()
self.last_sync_step = self.train_step_num
await self.synchronizer.set_trainer_status.remote(RunningStatus.RUNNING)
async def train(self) -> str:
"""Train the model."""
while self.train_step_num < self.total_steps:
try:
metrics = {}
# sample may be blocked due to explorer does not generate enough data
self.logger.info(f"Sample data for step {self.train_step_num + 1} started.")
sample_task = asyncio.create_task(self._sample_data())
while not sample_task.done():
# sync weight to make sure the explorer can continue to explore and generate enough data
if await self.need_sync():
metrics.update(await self.sync_weight())
await asyncio.sleep(1)
exps, sample_metrics, repr_samples = await sample_task
metrics.update(sample_metrics)
self.logger.info(f"Sample data for step {self.train_step_num + 1} finished.")
metrics.update(await self.train_step(exps))
if await self.need_sync():
metrics.update(await self.sync_weight())
if self.need_save():
metrics.update(
self.save_checkpoint(save_as_hf=self.save_hf_checkpoint == "always")
)
if self.config.trainer.enable_preview:
self._log_experiences(repr_samples)
self.monitor.log(metrics, self.train_step_num)
except StopAsyncIteration:
self.logger.info("No more samples to train. Stopping training.")
break
except Exception:
self.logger.error(f"Error in Trainer:\n{traceback.format_exc()}")
break
self.save_checkpoint(block_until_saved=True, save_as_hf=self.save_hf_checkpoint != "never")
await self.synchronizer.set_trainer_status.remote(RunningStatus.STOPPED)
self.logger.info("--------------------\n> Trainer finished.\n--------------------")
return self.config.trainer.name
async def train_step(self, exps: List[Experience]) -> Dict:
"""Train one step.
Returns:
bool: Whether to continue training.
Dict: Metrics of the training step.
"""
self.logger.info(f"Training at step {self.train_step_num + 1} started.")
metrics = {}
with Timer(metrics, "time/train_step"):
train_metrics = await self.engine.train_step(exps)
self.logger.info(f"Training at step {self.train_step_num} finished.")
metrics.update(train_metrics)
return metrics
async def _sample_data(self) -> Tuple[List[Experience], Dict, List[Dict]]:
"""Sample a batch of experiences.
Returns:
List[Experience]: A batch of experiences.
Dict: Metrics of the sampling step.
List[Dict]: A list of representative samples for logging.
"""
batch, metrics, repr_samples = await self.sample_strategy.sample(self.train_step_num + 1)
metrics["sample/task_count"] = len(set(exp.eid.tid for exp in batch))
return batch, metrics, repr_samples
async def need_sync(self) -> bool:
"""Whether to sync the model weight."""
if self.config.synchronizer.sync_style == SyncStyle.FIXED:
return (
self.last_sync_step != self.train_step_num
and self.train_step_num % self.config.synchronizer.sync_interval == 0
)
else:
if self.config.synchronizer.sync_style == SyncStyle.DYNAMIC_BY_TRAINER:
delta = self.train_step_num - self.last_sync_step
if delta >= self.config.synchronizer.sync_interval:
await self.synchronizer.set_trainer_status.remote(RunningStatus.REQUIRE_SYNC)
explorer_status_counts = await self.synchronizer.get_explorer_status_counts.remote()
if self.config.synchronizer.sync_method == SyncMethod.NCCL:
return explorer_status_counts[RunningStatus.WAITING_SYNC] > 0
else: # memory & checkpoint
return (
self.last_sync_step != self.train_step_num
and explorer_status_counts[RunningStatus.REQUIRE_SYNC] > 0
)
def need_save(self) -> bool:
"""Whether to save the checkpoint."""
return self.save_interval > 0 and self.train_step_num % self.save_interval == 0
async def sync_weight(self) -> Dict:
"""Sync the model weight."""
self.logger.info(f"Trainer sync_weights at step {self.train_step_num} started.")
metrics = {}
if self.last_sync_time is not None:
metrics["time/trainer_sync_interval"] = time.time() - self.last_sync_time
with Timer(metrics, "time/sync_weight"):
if self.config.synchronizer.sync_method == SyncMethod.NCCL:
result = await self.synchronizer.ready_to_nccl_sync.remote(
"trainer", self.train_step_num
)
if result is None:
self.logger.error("Trainer sync_weights failed.")
else:
self.engine.sync_weight()
elif self.config.synchronizer.sync_method == SyncMethod.CHECKPOINT:
self.engine.save_state_dict()
elif self.config.synchronizer.sync_method == SyncMethod.MEMORY:
self.engine.upload_state_dict()
self.last_sync_step = self.train_step_num
self.last_sync_time = time.time()
await self.synchronizer.set_trainer_status.remote(RunningStatus.RUNNING)
self.logger.info(f"Trainer sync_weights at step {self.train_step_num} finished.")
return metrics
def _log_experiences(self, samples: List[Dict]) -> None:
self._sample_exps_to_log.extend(samples)
if self.train_step_num % self.config.synchronizer.sync_interval == 0:
self.monitor.log_table(
"rollout_examples", pd.DataFrame(self._sample_exps_to_log), self.train_step_num
)
self._sample_exps_to_log.clear()
def save_checkpoint(self, block_until_saved: bool = False, save_as_hf: bool = False) -> Dict:
metrics = {}
with Timer(metrics, "time/save_checkpoint"):
self.logger.info(f"Saving checkpoint at step {self.train_step_num}...")
self.engine.save_checkpoint(block_until_saved=block_until_saved, save_as_hf=save_as_hf)
self.state.save_trainer(
current_step=self.train_step_num,
sample_strategy_state=self.sample_strategy.state_dict(),
)
return metrics
async def shutdown(self) -> None:
self.monitor.close()
@property
def train_step_num(self) -> int:
"""Get the current training step number."""
return self.engine.train_step_num
async def is_alive(self) -> bool:
"""Check if the trainer is alive."""
return True
@classmethod
def get_actor(cls, config: Config):
"""Get a Ray actor for the trainer."""
return (
ray.remote(cls)
.options(name=config.trainer.name, namespace=config.ray_namespace)
.remote(config)
)
class TrainEngineWrapper(ABC):
"""A wrapper class to wrap various training engines."""
@abstractmethod
async def prepare(self) -> None:
"""Do some preparation before training started."""
@property
@abstractmethod
def train_step_num(self) -> int:
"""Get the current training step number."""
@abstractmethod
async def train_step(self, batch_exps: List[Experience]) -> Dict:
"""Training one step.
Args:
batch_exps (List[Experience]): A batch of experiences to train.
Returns:
Dict: Metrics of the training step.
"""
@abstractmethod
def save_checkpoint(self, block_until_saved: bool = False, save_as_hf: bool = False) -> None:
"""Save the checkpoint."""
@abstractmethod
def sync_weight(self) -> None:
"""Sync the model weight."""
@abstractmethod
def upload_state_dict(self) -> None:
"""Upload the state dict to Synchronizer."""
@abstractmethod
def save_state_dict(self) -> None:
"""Only save the model state dict for Synchronizer."""
def get_trainer_wrapper(config: Config) -> TrainEngineWrapper:
"""Get a trainer wrapper."""
if config.trainer.trainer_type == "verl":
from trinity.trainer.verl_trainer import VerlPPOTrainerWrapper
return VerlPPOTrainerWrapper(config)
elif config.trainer.trainer_type == "tinker":
from trinity.trainer.tinker_trainer import TinkerTrainerWrapper
return TinkerTrainerWrapper(config)
else:
raise NotImplementedError
|