def train(
rank: int,
args: argparse.Namespace,
backend: str = "gloo",
optim_type: OptimType = OptimType.vanilla,
check_regression: bool = True,
):
logging.basicConfig(
level=logging.INFO if not args.debug else logging.DEBUG)
use_multi_tensor = args.multi_tensor_optim and hasattr(
torch.optim, "_multi_tensor")
OPTIM = torch.optim._multi_tensor.RMSprop if use_multi_tensor else torch.optim.RMSprop # type: ignore # attr is checked but mypy misses that
logging.info("Multi tensor optimizer: {}".format(use_multi_tensor))
# DDP
dist_init(rank=rank, world_size=args.world_size, backend=backend)
# Setup
if not args.cpu:
torch.cuda.set_device(rank)
torch.cuda.manual_seed(0)
torch.manual_seed(0) # also sets the cuda seed
np.random.seed(0)
if backend == "nccl":
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
device = torch.device("cpu") if args.cpu else torch.device(rank)
model, dataloader, loss_fn = get_problem(rank, args.world_size,
args.batch_size, device,
args.model)
# Shard the optimizer
optimizer: Optional[torch.optim.Optimizer] = None
model = cast(nn.Module, model)
scaler = (TorchGradScaler() if args.optim_type == OptimType.vanilla else
ShardedGradScaler()) if args.amp else None
if optim_type == OptimType.oss_sharded_ddp:
optimizer = OSS(params=model.parameters(),
optim=OPTIM,
lr=1e-4,
momentum=0.9)
# Single node run typically, no need for reduce buckets
model = ShardedDDP(model, optimizer, reduce_buffer_size=0)
else:
device_ids = None if args.cpu else [rank]
model = DDP(model, device_ids=device_ids,
find_unused_parameters=False) # type: ignore
optimizer = (OSS(
params=model.parameters(), optim=OPTIM, lr=1e-4, momentum=0.9)
if optim_type == OptimType.oss_ddp else OPTIM(
model.parameters(), lr=1e-4, momentum=0.9))
optimizer = cast(torch.optim.Optimizer, optimizer)
# Reset the memory use counter
if not args.cpu:
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats(rank)
torch.cuda.synchronize(rank)
# Standard training loop
training_start = time.monotonic()
model.train()
measurements = []
final_loss: Optional[float] = -1.0
need_profiling = args.profile
for epoch in range(args.epochs):
n_items = 0
epoch_runtime = 0.0
for batch in dataloader:
if not args.cpu:
torch.cuda.synchronize(rank)
batch_start = time.monotonic()
def closure(data=batch, grad_scaler=None):
model.zero_grad()
if args.debug and rank == 0 and next(
model.parameters()).grad is not None:
logging.debug("\nbefore: param {} -- grad {}".format(
next(model.parameters()).norm().item(),
next(model.parameters()).grad.norm().item()))
if grad_scaler is not None:
# Automatically computes the FW pass in half precision
with torch.cuda.amp.autocast():
outputs = model(data["inputs"])
loss = loss_fn(outputs, data["label"])
# Accumulates scaled gradients.
grad_scaler.scale(loss).backward()
else:
outputs = model(data["inputs"])
loss = loss_fn(outputs, data["label"])
loss.backward()
if args.debug and rank == 0 and next(
model.parameters()).grad is not None:
logging.debug("after BW: param {} -- grad {}".format(
next(model.parameters()).norm().item(),
next(model.parameters()).grad.norm().item()))
return loss
def run_closure(closure, scaler, optimizer):
if scaler is not None:
final_loss = closure(
grad_scaler=scaler
) # AMP scaler.step does not support closures
scaler.step(optimizer)
scaler.update()
return final_loss
else:
return optimizer.step(closure)
if need_profiling and not args.cpu:
logging.info("Profiling the run")
with profiler.profile(
use_cuda=True, record_shapes=True,
profile_memory=True) as prof: # type: ignore
with profiler.record_function("batch"):
final_loss = run_closure(closure, scaler, optimizer)
prof.export_chrome_trace(
f"{optim_type}_trace_rank_{rank}.json")
need_profiling = False # only profile once
else:
final_loss = run_closure(closure, scaler, optimizer)
if args.debug and rank == 0:
logging.debug("buffer: {}".format(
next(model.buffers()).norm().item()))
logging.debug("after update: param {} -- grad {}".format(
next(model.parameters()).norm().item(),
next(model.parameters()).grad.norm().item()))
n_items += args.batch_size
if not args.cpu:
# make sure that the cuda kernels are finished before taking a timestamp
torch.cuda.synchronize(rank)
batch_end = time.monotonic()
epoch_runtime += batch_end - batch_start
if optim_type == OptimType.oss_ddp or optim_type == OptimType.oss_sharded_ddp:
# Check the checkpointing in the case of the OSS optimizer
# Memory usage could spill over from there
optimizer = cast(OSS, optimizer)
optimizer.consolidate_state_dict()
if dist.get_rank() == 0:
_ = optimizer.state_dict()
logging.info("... State dict collected")
measurements.append(n_items / epoch_runtime)
if dist.get_rank() == 0:
logging.info(
f"Epoch {epoch} - processed {measurements[-1]:.2f} img per sec. Loss {final_loss:.3f}"
)
training_stop = time.monotonic()
img_per_sec = n_items / (training_stop - training_start) * args.epochs
logging.info(
f"[{dist.get_rank()}] : Training done. {img_per_sec:.2f} img per sec inc. checkpoint"
)
validate_benchmark(measurements, final_loss, args, check_regression)
dist.destroy_process_group() # type: ignore
def train(
rank: int,
args: argparse.Namespace,
backend: str = "gloo",
optim_type: OptimType = OptimType.vanilla,
check_regression: bool = True,
):
logging.basicConfig(
level=logging.INFO if not args.debug else logging.DEBUG)
# DDP
dist_init(rank=rank, world_size=args.world_size, backend=backend)
# Setup
if not args.cpu:
torch.cuda.set_device(rank)
torch.cuda.manual_seed(0)
torch.manual_seed(0) # also sets the cuda seed
np.random.seed(0)
if backend == "nccl":
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
device = torch.device("cpu") if args.cpu else torch.device(rank)
model, dataloader, loss_fn = get_problem(rank, args.world_size,
args.batch_size, device,
args.torchvision_model)
# Shard the optimizer
optimizer: Optional[torch.optim.Optimizer] = None
model = cast(nn.Module, model)
scaler = (TorchGradScaler() if args.optim_type == OptimType.vanilla else
ShardedGradScaler()) if args.amp else None
if optim_type == OptimType.oss_sharded_ddp:
optimizer = OSS(params=model.parameters(),
optim=OPTIM,
lr=1e-4,
momentum=0.9)
model = ShardedDDP(model, optimizer)
else:
device_ids = None if args.cpu else [rank]
model = DDP(model, device_ids=device_ids,
find_unused_parameters=False) # type: ignore
optimizer = (OSS(
params=model.parameters(), optim=OPTIM, lr=1e-4, momentum=0.9)
if optim_type == OptimType.oss_ddp else OPTIM(
model.parameters(), lr=1e-4, momentum=0.9))
optimizer = cast(torch.optim.Optimizer, optimizer)
# Reset the memory use counter
if not args.cpu:
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats(rank)
torch.cuda.synchronize(rank)
# Standard training loop
training_start = time.monotonic()
model.train()
measurements = []
final_loss: Optional[float] = -1.0
need_profiling = args.profile
for epoch in range(args.epochs):
n_items = 0
epoch_runtime = 0.0
for batch in dataloader:
if not args.cpu:
torch.cuda.synchronize(rank)
batch__start = time.monotonic()
def closure(data=batch, grad_scaler=None):
model.zero_grad()
if args.debug and rank == 0 and next(
model.parameters()).grad is not None:
logging.debug("\nbefore: param {} -- grad {}".format(
next(model.parameters()).norm().item(),
next(model.parameters()).grad.norm().item()))
if grad_scaler is not None:
# Automatically computes the FW pass in half precision
with torch.cuda.amp.autocast():
outputs = model(data["inputs"])
loss = loss_fn(outputs, data["label"])
# Accumulates scaled gradients.
grad_scaler.scale(loss).backward()
else:
outputs = model(data["inputs"])
loss = loss_fn(outputs, data["label"])
loss.backward()
if args.debug and rank == 0 and next(
model.parameters()).grad is not None:
logging.debug("after BW: param {} -- grad {}".format(
next(model.parameters()).norm().item(),
next(model.parameters()).grad.norm().item()))
return loss
if need_profiling and not args.cpu:
logging.info("Profiling the run")
with profiler.profile(
use_cuda=True, record_shapes=True,
profile_memory=True) as prof: # type: ignore
with profiler.record_function("batch"):
if scaler is not None:
final_loss = closure(
grad_scaler=scaler
) # AMP scaler.step does not support closures
scaler.step(optimizer)
scaler.update()
else:
final_loss = optimizer.step(closure)
prof.export_chrome_trace(
f"{optim_type}_trace_rank_{rank}.json")
need_profiling = False # only profile once
else:
if scaler is not None:
final_loss = closure(
grad_scaler=scaler
) # AMP scaler.step does not support closures
scaler.step(optimizer)
scaler.update()
else:
final_loss = optimizer.step(closure)
if args.debug and rank == 0:
logging.debug("buffer: {}".format(
next(model.buffers()).norm().item()))
logging.debug("after update: param {} -- grad {}".format(
next(model.parameters()).norm().item(),
next(model.parameters()).grad.norm().item()))
n_items += args.batch_size
if not args.cpu:
# make sure that the cuda kernels are finished before taking a timestamp
torch.cuda.synchronize(rank)
batch_end = time.monotonic()
epoch_runtime += batch_end - batch__start
if optim_type == OptimType.oss_ddp or optim_type == OptimType.oss_sharded_ddp:
# Check the checkpointing in the case of the OSS optimizer
# Memory usage could spill over from there
optimizer = cast(OSS, optimizer)
optimizer.consolidate_state_dict()
if dist.get_rank() == 0:
_ = optimizer.state_dict()
logging.info("... State dict collected")
measurements.append(n_items / epoch_runtime)
if dist.get_rank() == 0:
logging.info(
f"Epoch {epoch} - processed {measurements[-1]:.2f} img per sec. Loss {final_loss:.3f}"
)
max_memory = -1.0
if not args.cpu:
torch.cuda.synchronize(rank)
max_memory = torch.cuda.max_memory_allocated(rank) / 2**20
logging.info(f"[{dist.get_rank()}] : Peak memory {max_memory:.1f}MiB")
training_stop = time.monotonic()
img_per_sec = n_items / (training_stop - training_start) * args.epochs
logging.info(
f"[{dist.get_rank()}] : Training done. {img_per_sec:.2f} img per sec inc. checkpoint"
)
# Compute the median and median of absolute differences img per second
measurements.sort()
median = measurements[len(measurements) // 2]
abs_diff = list(map(lambda x: abs(x - median), measurements))
abs_diff.sort()
mad = abs_diff[len(measurements) // 2] if args.epochs > 2 else -1
logging.info(
f"[{dist.get_rank()}] : Median speed: {median:.2f} +/- {mad:.2f}")
if check_regression and dist.get_rank() == 0:
assert (median +
3.0 * mad) > args.reference_speed, "Speed regression detected"
assert max_memory < 1.05 * args.reference_memory, "Memory use regression detected"
assert abs(cast(float, final_loss) -
args.reference_loss) < 1e-3, "Loss regression detected"
logging.info("[Regression Test] VALID")
dist.destroy_process_group() # type: ignore
