Tinghong-Ye22's picture
Add files using upload-large-folder tool
6eb47b8 verified
Raw History Blame Contribute Delete
23.4 kB
from __future__ import annotations
from dotenv import load_dotenv
assert load_dotenv(), "Failed to load .env file. Make sure it exists and is properly formatted."
import os
import re
import torch
import gc
from tqdm import tqdm
import shutil
import json
import datetime
import sys
import itertools
import threading
import tyro
import importlib
from typing import TYPE_CHECKING
import time
if TYPE_CHECKING:
from psi.config.config import LaunchConfig
from psi.utils import batch_str_to_tensor, seed_everything, initialize_overwatch, nice, flatten
from psi.trainers import Trainer
overwatch = initialize_overwatch(__name__)
from accelerate import Accelerator, DeepSpeedPlugin
from accelerate.data_loader import (
DataLoaderStateMixin as AcceleratorDataLoaderStateMixin,
)
from accelerate.utils import ProjectConfiguration
import random, numpy as np
MAX_TRAINING_EPOCHS = 1_000_000
def _start_wait_progress(desc: str, *, position: int = 0) -> tuple[threading.Event, threading.Thread | None]:
done = threading.Event()
if not overwatch.is_rank_zero():
return done, None
def _run() -> None:
bar = tqdm(
total=None,
desc=desc,
unit="s",
position=position,
leave=True,
dynamic_ncols=True,
bar_format="{desc}: {n}s elapsed [{elapsed}]",
)
try:
while not done.wait(1.0):
bar.update(1)
finally:
bar.close()
thread = threading.Thread(target=_run, name=f"wait-progress-{desc}", daemon=True)
thread.start()
return done, thread
def _auto_tag_run(run_name: str):
"""commit the code and tag the run with the run name"""
import subprocess
subprocess.run(["git", "add", "."], check=False)
subprocess.run(["git", "commit", "-m", f"Auto-commit for {run_name}"], check=False)
subprocess.run(["git", "tag", run_name], check=False)
# subprocess.run(["git", "push", "--tags"], check=True)
def _initialize_accelerator(trainer: Trainer) -> Accelerator:
os.makedirs(trainer.project_dir, exist_ok=True)
logging_dir = os.path.join(trainer.project_dir, trainer.cfg.log.logging_dir)
accelerator_project_config = ProjectConfiguration(
project_dir=trainer.project_dir, logging_dir=logging_dir
)
# setup FSDP training strategy if needed
fsdp_plugin = None
if trainer.cfg.train.data_parallel == "fsdp":
fsdp_plugin = trainer.get_fsdp_plugin()
deepspeed_plugin = None
if trainer.cfg.train.data_parallel == "deepspeed":
# SONGLIN: use accelerate launch config instead
ds_config_path = trainer.cfg.train.deepspeed_config
deepspeed_plugin = DeepSpeedPlugin(zero_stage=3, hf_ds_config=ds_config_path)
accelerator = Accelerator(
gradient_accumulation_steps=trainer.cfg.train.gradient_accumulation_steps,
mixed_precision=trainer.cfg.train.mixed_precision,
# downcast_bf16=True,
log_with=trainer.cfg.log.report_to,
project_config=accelerator_project_config,
fsdp_plugin=fsdp_plugin,
deepspeed_plugin=deepspeed_plugin
)
if trainer.cfg.train.data_parallel == "deepspeed":
# When using DeepSpeed, accelerate.prepare() requires you
# to pass at least one of training or evaluation dataloaders
# with batch_size attribute returning an integer value or alternatively
# set an integer value in train_micro_batch_size_per_gpu
# in the deepspeed config file or assign integer value to
accelerator.deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] = trainer.device_train_batch_size # type: ignore
# Initialize the trackers unless in eval or debug mode
if overwatch.is_rank_zero() and not trainer.cfg.eval:
if trainer.cfg.auto_tag_run:
_auto_tag_run(trainer.run_name)
if trainer.cfg.log.report_to == "wandb":
tracker_config = (
trainer.cfg.model_dump()
) # cfg.model_dump() # dict(vars(cfg.task))
# Add environment variables to tracker config for wandb logging
if "_ENV_INFO_JSON" in os.environ:
env_info = json.loads(os.environ["_ENV_INFO_JSON"])
tracker_config["environment_variables"] = env_info
wandb_config = dict(
trainer.cfg.wandb
) # cfg.wandb.model_dump() #vars(copy.deepcopy(cfg.wandb))
wandb_config["name"] = trainer.run_name
wandb_config["dir"] = os.path.abspath(
os.path.expanduser(trainer.project_dir)
)
wandb_config["group"] = trainer.cfg.train.name
project = wandb_config.pop("project", "default")
# wandb resume with old hisotry
if trainer.cfg.train.resume_from_checkpoint is not None and trainer.cfg.wandb.resume != "never":
run_id = wandb_config.get("id") or getattr(trainer.cfg.wandb, "id", None)
if not run_id :
run_dir = os.path.join(*trainer.cfg.train.resume_from_checkpoint.split("/")[:3])
if os.path.exists(os.path.join(run_dir, "run_config.json")):
with open(os.path.join(run_dir, "run_config.json")) as f:
run_id = (json.load(f).get("wandb") or {}).get("id")
overwatch.info(f"resume wandb with run id: {run_id}")
wandb_config["id"] = run_id
accelerator.init_trackers(project, tracker_config, {"wandb": wandb_config})
for tracker in accelerator.trackers:
if tracker.name == "wandb":
trainer.cfg.wandb.id = tracker.run.id
trainer.cfg.wandb.name = tracker.run.name
trainer.cfg.wandb.entity = tracker.run.entity
trainer.cfg.wandb.group = tracker.run.group
# Log the configuration
os.makedirs(trainer.project_dir, exist_ok=True)
# draccus.dump(cfg, open(f'{project_dir}/run_config.yaml','w'))
with open(f"{trainer.project_dir}/run_config.json", "w") as f:
f.write(trainer.cfg.model_dump_json(indent=4))
overwatch.info(f"Saved configuration to {trainer.project_dir}")
# Write down the sys.argv to argv.txt
argv_path = os.path.join(trainer.project_dir, "argv.txt")
with open(argv_path, "w") as f:
i = 0
while i < len(sys.argv):
arg = sys.argv[i]
if arg.startswith('--'):
# Start of a key, check for list values
line = [arg]
j = i + 1
# Collect all following args that do not start with '--'
while j < len(sys.argv) and not sys.argv[j].startswith('--'):
line.append(sys.argv[j])
j += 1
f.write(' '.join(line) + '\n')
i = j
else:
f.write(arg + '\n')
i += 1
# write system envs to envs.txt
if "_ENV_INFO_JSON" in os.environ:
env_info = json.loads(os.environ["_ENV_INFO_JSON"])
envs_path = os.path.join(trainer.project_dir, "envs.txt")
with open(envs_path, "w") as f:
for k, v in env_info.items():
f.write(f"{k}={v}\n")
# copy dataset statistics file if exists
if hasattr(trainer.cfg.data.transform.field, "stat_path"):
from psi.utils import resolve_path
stat_path = resolve_path(trainer.cfg.data.transform.field.stat_path) # type: ignore
if os.path.exists(stat_path):
dst_stat_path = os.path.join(trainer.project_dir, "dataset_statistics.json")
shutil.copy2(stat_path, dst_stat_path)
trainer.accelerator = accelerator
return accelerator
def train(config: LaunchConfig):
torch.cuda.set_device(device_id := overwatch.local_rank())
torch.cuda.empty_cache()
startup_bar = tqdm(
total=5,
desc="Startup",
disable=not overwatch.is_rank_zero(),
position=0,
leave=False,
)
if config.seed:
overwatch.info(f"Seed everything with {config.seed}")
seed_everything(config.seed)
trainer = Trainer.instantiate(config, device_id) # PosttrainTrainer
overwatch.info("Initialize models ... ")
trainer.init_models()
startup_bar.set_postfix_str("models ready")
startup_bar.update(1)
accelerator = _initialize_accelerator(trainer) # 分布式训练框架,先不看了
startup_bar.set_postfix_str("accelerator ready")
startup_bar.update(1)
overwatch.info(f"Training configurations:")
overwatch.info(f"training task: '{config.train.name}'", ctx_level=1)
overwatch.info(f"run name: {trainer.run_name}", ctx_level=1)
overwatch.info(f"seed: {config.seed}", ctx_level=1)
overwatch.info(f"mixed precision: {trainer.dtype}", ctx_level=1)
overwatch.info(f"warmup steps: {trainer.num_warmup_steps}", ctx_level=1)
overwatch.info(f"validation steps: {config.train.validation_steps}", ctx_level=1)
overwatch.info(f"checkpoint steps: {config.train.checkpointing_steps}", ctx_level=1)
overwatch.info(f"max gradient norm: {config.train.max_grad_norm}", ctx_level=1)
overwatch.info("Build datasets ...")
train_dataset, val_dataset = trainer.create_datasets()
startup_bar.set_postfix_str("datasets ready")
startup_bar.update(1)
overwatch.info(f"Num training samples:", ctx_level=1)
overwatch.info(f"Training dataset size: {trainer.len_train_dataset:,}", ctx_level=2) # type: ignore
if val_dataset:
overwatch.info(f"Val dataset size: {trainer.len_val_dataset:,}", ctx_level=2) # type: ignore
overwatch.info("Build dataloaders ...")
trainer.create_dataloaders(train_dataset, val_dataset)
startup_bar.set_postfix_str("dataloaders ready")
startup_bar.update(1)
overwatch.info("Initialize optimizers and schedulers...")
trainer.create_optimizer_and_scheduler(trainer.max_training_steps)
startup_bar.set_postfix_str("optim ready")
startup_bar.update(1)
startup_bar.close()
# fmt: off
overwatch.info("***** Running training *****")
overwatch.info(f"Num training examples = {trainer.len_train_dataset}", ctx_level=1)
overwatch.info(f"Max training Epochs = {trainer.max_training_epochs}", ctx_level=1)
overwatch.info(f"Total optimization steps = {trainer.max_training_steps}", ctx_level=2)
overwatch.info(f"Num steps Per Epoch = {trainer.num_steps_per_epoch}", ctx_level=2)
overwatch.info(f"Effective training epochs = {trainer.max_training_steps/trainer.num_steps_per_epoch*trainer.world_size:.2f}", ctx_level=2)
overwatch.info(f"Global train batch size (w. parallel, distributed & accumulation) = {trainer.global_train_batch_size}", ctx_level=1)
overwatch.info(f"Device train batch size = {trainer.device_train_batch_size}", ctx_level=2)
overwatch.info(f"Gradient Accumulation steps = {trainer.gradient_accumulation_steps}", ctx_level=2)
overwatch.info(f"Num processes (GPUs) = {trainer.world_size}", ctx_level=2)
# fmt: on
overwatch.info("Prepare accelerator / DeepSpeed ...")
prepare_start = time.time()
trainer.prepare(accelerator)
overwatch.info(f"Accelerator prepare took {time.time() - prepare_start:.1f}s")
overwatch.info("Resume checkpoint state ...")
resume_start = time.time()
global_step = initial_global_step = trainer.resume_from_checkpoint()[0]
overwatch.info(f"Resume state took {time.time() - resume_start:.1f}s")
epoch_start = global_step // trainer.num_steps_per_epoch
eval_prewarm_enabled = os.environ.get(
"PSI_OMNI_OSS_EVAL_PREWARM", "0"
).strip().lower() in {"1", "true", "yes", "on"}
prepare_evaluation_cache = getattr(trainer, "prepare_evaluation_cache", None)
if (
eval_prewarm_enabled
and config.train.validation_steps > 0
and callable(prepare_evaluation_cache)
):
eval_prewarm_start = time.time()
overwatch.info("Prewarm fixed validation batches before training ...")
cache_ready = prepare_evaluation_cache(int(config.train.validation_steps))
overwatch.info(
f"Validation batch prewarm ready={int(bool(cache_ready))} "
f"in {time.time() - eval_prewarm_start:.1f}s"
)
overwatch.info(f"Accelerator runs in: {trainer.project_dir}")
trainer.set_train()
is_max_train_steps_reached = (initial_global_step >= trainer.max_training_steps)
# skip_resumed_steps = False
skip = 0
progress_bar = None
first_batch_prefetched = False
for epoch in range(epoch_start, MAX_TRAINING_EPOCHS):
trainer.next_epoch(epoch)
accelerator.wait_for_everyone()
if not first_batch_prefetched:
overwatch.info("Fetch first training batch ...")
first_batch_wait_done, first_batch_wait_thread = _start_wait_progress(
"First batch wait",
position=0,
)
first_batch_start = time.time()
train_iter = iter(trainer.train_dataloader)
try:
first_batch = next(train_iter)
finally:
first_batch_wait_done.set()
if first_batch_wait_thread is not None:
first_batch_wait_thread.join(timeout=2.0)
overwatch.info(f"First batch fetched in {time.time() - first_batch_start:.1f}s")
epoch_batches = itertools.chain([first_batch], train_iter)
first_batch_prefetched = True
else:
epoch_batches = trainer.train_dataloader
if progress_bar is None:
progress_bar = tqdm(
range(0, trainer.max_training_steps),
initial=initial_global_step,
desc="Traing steps",
disable=not overwatch.is_rank_zero(),
position=0,
)
for local_step, batch in enumerate(epoch_batches):
# breakpoint()
if (
config.train.skip_resumed_steps
and skip < initial_global_step % trainer.num_steps_per_epoch
):
# skip to inital global step
skip += 1
continue
sync_gradients, losses = trainer.step(batch_str_to_tensor(batch), global_step, local_step)
if sync_gradients:
# Start rank-local CPU validation workers shortly before an
# eval boundary. Their prefetched batches overlap the final
# training steps; no CUDA work is issued by this path.
eval_prefetch_lead = max(
int(os.environ.get("PSI_OMNI_OSS_EVAL_PREFETCH_LEAD_STEPS", "0")),
0,
)
prefetch_evaluation = getattr(trainer, "prefetch_evaluation", None)
if (
eval_prefetch_lead > 0
and callable(prefetch_evaluation)
and config.train.validation_steps > 0
):
validation_interval = int(config.train.validation_steps)
regular_distance = (-global_step) % validation_interval
if regular_distance == 0:
regular_distance = validation_interval
final_distance = trainer.max_training_steps - 1 - global_step
targets = [(regular_distance, global_step + regular_distance)]
if final_distance > 0:
targets.append((final_distance, trainer.max_training_steps - 1))
distance, target_step = min(targets)
if distance <= eval_prefetch_lead:
prefetch_evaluation(target_step)
# log metrics
trainer.log(flatten({**losses, "epoch": epoch}, parent_key="train"))
# save checkpoints
if (global_step + 1) % config.train.checkpointing_steps == 0 or global_step == trainer.max_training_steps - 1:
accelerator.wait_for_everyone() # ensures fsdp checkpointing without deadlock
save_path = trainer.save_checkpoint(global_step+1)
accelerator.wait_for_everyone()
if overwatch.is_rank_zero():
tqdm.write(f"Saved state to {save_path}")
# validation
if config.train.validation_steps > 0 and global_step > 0 and (
global_step % config.train.validation_steps == 0 or
global_step == trainer.max_training_steps - 1
) :
gc.collect()
torch.cuda.empty_cache()
trainer.set_eval()
with torch.inference_mode(), torch.autocast(
device_type="cuda", dtype=trainer.dtype
):
eval_losses = trainer.evaluate()
if eval_losses is not None: # FIXME
trainer.log(flatten(eval_losses, parent_key="eval"))
# <modify> eval loss更低时额外保存best ckpt,路径固定为checkpoints/best_ckpt。
best_save_path = trainer.save_best_checkpoint(eval_losses, global_step + 1)
if best_save_path is not None and overwatch.is_rank_zero():
tqdm.write(f"Saved best eval checkpoint to {best_save_path}")
trainer.set_train()
""" NOTE:
This is a workaround for iterating over val_dataloader in the middle of training_dataloader.
When val dataloader dost not end explicitly, eg., the val dataloader is wrapped in a tqdm progress bar,
and the bar is closed before the end of the dataloader, it will cause an inconsistent accelerate gradient state
"""
val_dataloaders = getattr(trainer, "val_dataloader", None)
used_cpu_prefetch = bool(
getattr(trainer, "_last_eval_used_cpu_prefetch", False)
)
if used_cpu_prefetch:
# The raw rank-sharded CPU loaders never call
# DataLoaderShard.begin(), so there is no Accelerate
# validation state to end.
val_dataloaders_to_end = []
trainer._last_eval_used_cpu_prefetch = False
elif isinstance(val_dataloaders, dict):
val_dataloaders_to_end = list(val_dataloaders.values())
elif val_dataloaders is not None:
val_dataloaders_to_end = [val_dataloaders]
else:
val_dataloaders_to_end = []
for val_dataloader in val_dataloaders_to_end:
if isinstance(val_dataloader, AcceleratorDataLoaderStateMixin):
# ``end`` only resets Accelerate's gradient-state
# bookkeeping; it does not tear down persistent
# PyTorch workers or their worker-local OSS cache.
if (
accelerator.gradient_state.active_dataloader
!= trainer.train_dataloader
):
try:
val_dataloader.end()
overwatch.warning_once(
"Warning only Once: "
"Val dataloader does not end after evaluation. I killed it for you."
)
except ValueError:
overwatch.warning_once("Even I can not save you.")
progress_bar.update()
progress_bar.set_postfix(dict(loss=losses["loss"], lr=nice(trainer.lr)))
global_step += 1
if global_step >= trainer.max_training_steps:
if overwatch.is_rank_zero():
tqdm.write("Training has reached maximum steps.")
is_max_train_steps_reached = True # set to break outer loop
break
if is_max_train_steps_reached:
break
accelerator.wait_for_everyone()
trainer.finalize()
overwatch.info("Happy Ending!")
accelerator.end_training()
if torch.distributed.is_initialized():
torch.distributed.destroy_process_group()
def _is_slurm_job_process() -> bool:
return "SLURM_JOB_ID" in os.environ and not os.isatty(sys.stdout.fileno())
if __name__ == "__main__":
if _is_slurm_job_process():
overwatch.info("SLURM job detected. Setting up distributed environment variables...")
os.environ["LOCAL_WORLD_SIZE"] = str(
int(os.environ["WORLD_SIZE"]) // int(os.environ["SLURM_JOB_NUM_NODES"])
)
# Print and log important environment variables
env_vars_to_track = [
"OMP_NUM_THREADS", "HF_HOME", "TORCH_HOME", "HF_TOKEN", "HF_LEROBOT_HOME",
"DATA_HOME", "UV_CACHE_DIR", "WANDB_API_KEY",
"PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION", "CUDA_VISIBLE_DEVICES",
"WORLD_SIZE", "LOCAL_WORLD_SIZE", "RANK", "LOCAL_RANK", "MASTER_ADDR", "MASTER_PORT"
]
env_info = {}
overwatch.info("=== Environment Variables ===")
for var in env_vars_to_track:
value = os.environ.get(var, "Not Set")
# Mask sensitive tokens for logging
if "TOKEN" in var or "KEY" in var:
display_value = f"{value[:3]}...{value[-4:]}" if value != "Not Set" and len(value) > 7 else value
else:
display_value = value
env_info[var] = display_value
overwatch.info(f"{var}: {display_value}")
# Store env info for wandb logging later
os.environ["_ENV_INFO_JSON"] = json.dumps(env_info)
start = datetime.datetime.now()
overwatch.info("Parsing configuration...")
try:
# By convention, the first argument after the script name is the config module name,
# eg., {trainer}_{data}_{model}_config, which corresponds to psi.config.train.pretrain_egodex_qwen3vl_config
module = importlib.import_module(f"psi.config.train.{sys.argv[1]}")
DynamicLaunchConfigClass = getattr(module, "DynamicLaunchConfig")
config = tyro.cli(DynamicLaunchConfigClass, config=(tyro.conf.ConsolidateSubcommandArgs,), args=sys.argv[2:])
# breakpoint()
except Exception as e:
overwatch.error(f"Failed to import config module 'psi.config.train.{sys.argv[1]}'")
raise e
end = datetime.datetime.now()
overwatch.info(f"Config parsing took {(end - start).total_seconds():.2f}s")
train(config) # type: ignore