def set_amp_args(self):
"""
Two automatic mixed precision implementations are available: Apex's and PyTorch's.
- If Apex's AMP is enabled, amp_args is a dictionary containing arguments
to be passed to amp.initialize. Set to None to disable amp.
To enable mixed precision training, pass amp_args={"opt_level": "O1"} here.
See https://nvidia.github.io/apex/amp.html for more info.
- If Pytorch's AMP is enabled, no arguments are needed.
"""
if self.config.MODEL.AMP_PARAMS.USE_AMP:
assert (
self.device.type == "cuda"
), "Mixed precision is only available on CUDA devices for now"
# This will rightly fail if the setting is not correct
self.amp_type = AmpType[self.config.MODEL.AMP_PARAMS.AMP_TYPE.upper()]
# Check Apex availability
if self.amp_type == AmpType.APEX:
if not is_apex_available():
raise RuntimeError(
"Apex is not available. Can't use mixed precision"
)
# "amp_args" are actually Apex Amp args
self.amp_args = self.config.MODEL.AMP_PARAMS.AMP_ARGS
logging.info(f"Setting AMP: using apex, args {self.amp_args}")
elif self.amp_type == AmpType.PYTORCH:
# if the optimizer is sharded or FSDP data parallel is used, then the GradScaler
# needs to be shard-aware.
if (
self.config["TRAINER"]["TASK_NAME"] == "self_supervision_fsdp_task"
or self.config["OPTIMIZER"]["name"] == "zero"
):
assert is_fairscale_sharded_available(), (
"To use ZeRO with PyTorch AMP, ShardedGradScaler() "
"from fairscale is needed. Please upgrade fairscale"
)
from fairscale.optim.grad_scaler import ShardedGradScaler
self.amp_grad_scaler = ShardedGradScaler()
logging.info("Setting AMP: using sharded grad scaler")
else:
self.amp_grad_scaler = TorchGradScaler()
logging.info("Setting AMP: using pytorch grad scaler")
logging.info(f"Setting AMP: {self.amp_type} - args: {self.amp_args}")
else:
self.amp_args, self.amp_type = None, None
logging.info("Not using Automatic Mixed Precision")
def _init_pytorch_grad_scaler(self):
if self.config["OPTIMIZER"]["name"] == "zero":
assert is_fairscale_sharded_available(), (
"To use ZeRO with PyTorch AMP, ShardedGradScaler() "
"from fairscale is needed. Please upgrade fairscale")
from fairscale.optim.grad_scaler import ShardedGradScaler
self.amp_grad_scaler = ShardedGradScaler()
logging.info("Setting AMP: using sharded grad scaler")
else:
self.amp_grad_scaler = TorchGradScaler()
logging.info("Setting AMP: using pytorch grad scaler")
def set_amp_args(self, amp_args: Optional[Dict[str, Any]]):
"""Disable / enable apex.amp and set the automatic mixed precision parameters.
apex.amp can be utilized for mixed / half precision training.
Args:
amp_args: Dictionary containing arguments to be passed to
amp.initialize. Set to None to disable amp. To enable mixed
precision training, pass amp_args={"opt_level": "O1"} here.
See https://nvidia.github.io/apex/amp.html for more info.
Raises:
RuntimeError: If opt_level is not None and apex is not installed.
Warning: apex needs to be installed to utilize this feature.
"""
self.amp_args = amp_args
if amp_args is None:
logging.info("AMP disabled")
else:
# Check that the requested AMP type is known
try:
self.amp_type = AmpType[self.amp_args["amp_type"].upper()]
except KeyError:
logging.info("AMP type not specified, defaulting to Apex")
self.amp_type = AmpType.APEX
# Check for CUDA availability, required for both Apex and Pytorch AMP
if not torch.cuda.is_available():
raise RuntimeError(
"AMP is required but CUDA is not supported, cannot enable AMP"
)
# Check for Apex availability
if self.amp_type == AmpType.APEX and not apex_available:
raise RuntimeError(
"Apex AMP is required but Apex is not installed, cannot enable AMP"
)
# Set Torch AMP grad scaler, used to prevent gradient underflow
elif self.amp_type == AmpType.PYTORCH:
self.amp_grad_scaler = TorchGradScaler()
logging.info(f"AMP enabled with args {amp_args}")
return self
def set_amp_args(self):
"""
Two automatic mixed precision implementations are available: Apex's and PyTorch's.
- If Apex's AMP is enabled, amp_args is a dictionary containing arguments
to be passed to amp.initialize. Set to None to disable amp.
To enable mixed precision training, pass amp_args={"opt_level": "O1"} here.
See https://nvidia.github.io/apex/amp.html for more info.
- If Pytorch's AMP is enabled, no arguments are needed.
"""
if self.config.MODEL.AMP_PARAMS.USE_AMP:
assert (
self.device.type == "cuda"
), "Mixed precision is only available on CUDA devices for now"
# This will rightly fail if the setting is not correct
self.amp_type = AmpType[self.config.MODEL.AMP_PARAMS.AMP_TYPE.upper()]
# Check Apex availability
if self.amp_type == AmpType.APEX:
if not is_apex_available():
raise RuntimeError(
"Apex is not available. Can't use mixed precision"
)
# "amp_args" are actually Apex Amp args
self.amp_args = self.config.MODEL.AMP_PARAMS.AMP_ARGS
elif self.amp_type == AmpType.PYTORCH:
# If the optimizer is sharded, then the GradScaler needs to be shard-aware
self.amp_grad_scaler = (
ShardedGradScaler()
if self.config["OPTIMIZER"]["name"] == "zero"
else TorchGradScaler()
)
logging.info(f"Setting AMP: {self.amp_type} - args: {self.amp_args}")
else:
self.amp_args, self.amp_type = None, None
logging.info("Not using Automatic Mixed Precision")
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 run_ddp_parity(
rank,
world_size,
backend,
temp_file_name,
reduce_buffer_size,
grad_accumulation,
change_train_graph,
fp16_reduction,
clip_grad_norm,
amp,
manual_reduction,
multiple_fw,
):
dist.init_process_group(init_method="file://" + temp_file_name,
backend=backend,
rank=rank,
world_size=world_size)
device = torch.device("cuda")
torch.cuda.set_device(rank)
torch.manual_seed(rank)
np.random.seed(rank)
NUMBER_BATCHS = 5
# Test all combinations: AMP, Accumulate, Change train graph, reduce buckets
print(
f"{rank}: Checking configuration: accumulate {grad_accumulation}" +
f" - change train graph {change_train_graph}" + f" - amp {amp}" +
f" - manual reduction {manual_reduction}" +
f" - buffers {reduce_buffer_size}" + f" - multiple FW {multiple_fw}",
flush=True,
)
# The API should be the exact same in between the sharded and non-sharded variants, generic closure
def closure(model,
scaler,
input_tensor,
should_accumulate,
_manual_reduction=False):
accumulate_steps = 3 if should_accumulate else 1
model.zero_grad()
def step():
if scaler is not None:
with torch.cuda.amp.autocast():
loss = model(input_tensor).abs().sum()
scaler.scale(loss).backward()
else:
loss = model(input_tensor).abs().sum()
loss.backward()
with model.no_sync() if should_accumulate else suppress():
for _ in range(accumulate_steps - 1):
step()
if not _manual_reduction:
step()
else:
with model.no_sync():
step()
model.reduce()
# Any model works. Add one different buffer per rank
model = _get_mlp_emb(multiple_fw)
model.register_buffer("test_buffer", torch.ones((1)) * rank)
model.to(device)
# Make sure that the model starts with non-trainable, so that we check for the buckets to be
# properly reassigned when/if this changes
next(model.parameters()).requires_grad = False
sharded_optimizer = OSS(params=model.parameters(),
optim=torch.optim.SGD,
lr=1e-4,
momentum=0.99)
sharded_ddp_model = ShardedDataParallel(
module=model,
sharded_optimizer=sharded_optimizer,
broadcast_buffers=True,
reduce_buffer_size=reduce_buffer_size,
reduce_fp16=fp16_reduction,
)
ddp_model_single = copy.deepcopy(model)
ddp_optimizer = torch.optim.SGD(ddp_model_single.parameters(),
lr=1e-4,
momentum=0.99)
ddp_model = DDP(ddp_model_single,
device_ids=[rank],
broadcast_buffers=True,
find_unused_parameters=True)
if fp16_reduction:
from dist.algorithms.ddp_com_hooks.default_hooks import fp16_compress_hook
ddp_model.register_comm_hook(state=None,
hook=fp16_compress_hook) # type: ignore
ddp_scaler = TorchGradScaler() if amp else None
sharded_scaler = ShardedGradScaler() if amp else None
# The model should be synchronized in between the ranks at construction time, check that
check_same_model_params(sharded_ddp_model, ddp_model)
# Typical training loop, check that we get the exact same results as DDP
for i in range(NUMBER_BATCHS):
input_tensor = _get_random_inputs(device)
def ddp_closure(input_tensor=input_tensor):
return closure(ddp_model, ddp_scaler, input_tensor,
grad_accumulation)
def sharded_closure(input_tensor=input_tensor):
return closure(
sharded_ddp_model,
sharded_scaler,
input_tensor,
grad_accumulation,
_manual_reduction=manual_reduction,
)
# Step/scale both
for _scaler, _closure, _optimizer in (
(ddp_scaler, ddp_closure, ddp_optimizer),
(sharded_scaler, sharded_closure, sharded_optimizer),
):
if _scaler is not None:
_ = _closure(input_tensor)
_scaler.step(_optimizer)
_scaler.update()
else:
_optimizer.step(_closure())
check_same_model_params(sharded_ddp_model, ddp_model,
f"Rank: {rank} - Step {i} broke")
# Check that the two grad norm are equivalent
# NOTE: The grads can occasionally be NaNs, the scaler will skip the step in that case
# This is not ShardedDDP specific. If the grads are not NaN for DDP then they should also
# be valid for ShardedDDP
# NOTE: DDP does not handle parameters trainability being changed after the fact, see
# https://github.com/pytorch/pytorch/blob/5781aec74ef00284e0262817a649278c2e8072bf/torch/nn/parallel/distributed.py#L471
if clip_grad_norm and not change_train_graph:
if torch_version() >= (1, 9, 0):
total_norm = torch.nn.utils.clip_grad_norm_(
ddp_model.parameters(),
0.3,
norm_type=2.0,
error_if_nonfinite=False) # type: ignore
else:
total_norm = torch.nn.utils.clip_grad_norm_(
ddp_model.parameters(), 0.3, norm_type=2.0) # type: ignore
if not torch.isnan(total_norm):
oss_total_norm = sharded_optimizer.clip_grad_norm(
0.3, norm_type=2.0)
allclose = torch.allclose(oss_total_norm,
total_norm,
atol=1e-2 if amp else 1e-8)
if not allclose:
# Debug helper if this unit test does not pass, compare the gradients in between DDP and ShardedDDP
for idx, (p_ddp, p_sdp) in enumerate(
zip(ddp_model.parameters(),
sharded_ddp_model.parameters())):
if p_ddp.grad is not None:
if p_sdp.grad is not None:
print(rank,
idx,
torch.norm(p_ddp.grad),
torch.norm(p_sdp.grad),
flush=True)
else:
print(rank,
idx,
torch.norm(p_ddp.grad),
"not owned",
flush=True)
assert (
allclose
), f"torch and fairscale should return the same grad norm\n {oss_total_norm} vs {total_norm}"
else:
print(rank, "NaN grad norm in DDP", flush=True)
# Flip the trainability of the first parameter back and forth
if i == 0 and change_train_graph:
next(sharded_ddp_model.parameters()).requires_grad = not next(
sharded_ddp_model.parameters()).requires_grad
next(ddp_model.parameters()).requires_grad = not next(
ddp_model.parameters()).requires_grad
check_same_model_params(
sharded_ddp_model, ddp_model,
f"Rank: {rank} - Trainability refresh {i} broke")
dist.destroy_process_group()
def check_parity(manual_reduction: bool):
# The API should be the exact same in between the sharded and non-sharded variants, generic closure
def closure(model,
scaler,
input_tensor,
should_accumulate,
_manual_reduction=False):
accumulate_steps = 3 if should_accumulate else 1
model.zero_grad()
def step():
if scaler is not None:
with torch.cuda.amp.autocast():
loss = model(input_tensor).abs().sum()
scaler.scale(loss).backward()
else:
loss = model(input_tensor).abs().sum()
loss.backward()
with model.no_sync() if should_accumulate else suppress():
for _ in range(accumulate_steps - 1):
step()
if not _manual_reduction:
step()
else:
with model.no_sync():
step()
model.reduce()
# Any model works. Add one different buffer per rank
model = _get_mlp()
model.register_buffer("test_buffer", torch.ones((1)) * rank)
model.to(device)
# Make sure that the model starts with non-trainable, so that we check for the buckets to be
# properly reassigned when/if this changes
next(model.parameters()).requires_grad = False
sharded_optimizer = OSS(params=model.parameters(),
optim=torch.optim.SGD,
lr=1e-4,
momentum=0.99)
sharded_ddp_model = ShardedDataParallel(
module=model,
sharded_optimizer=sharded_optimizer,
broadcast_buffers=True,
reduce_buffer_size=reduce_buffer_size,
reduce_fp16=fp16_reduction,
)
ddp_model_single = copy.deepcopy(model)
ddp_optimizer = torch.optim.SGD(ddp_model_single.parameters(),
lr=1e-4,
momentum=0.99)
ddp_model = DDP(ddp_model_single,
device_ids=[rank],
broadcast_buffers=True,
find_unused_parameters=True)
if fp16_reduction:
from dist.algorithms.ddp_com_hooks.default_hooks import fp16_compress_hook
ddp_model.register_comm_hook(
state=None, hook=fp16_compress_hook) # type: ignore
ddp_scaler = TorchGradScaler() if amp else None
sharded_scaler = ShardedGradScaler() if amp else None
# The model should be synchronized in between the ranks at construction time, check that
check_same_model_params(sharded_ddp_model, ddp_model)
# Typical training loop, check that we get the exact same results as DDP
for i in range(NUMBER_BATCHS):
input_tensor = torch.rand((BATCH_SIZE, 2)).to(device)
def ddp_closure(input_tensor=input_tensor):
return closure(ddp_model, ddp_scaler, input_tensor,
grad_accumulation)
def sharded_closure(input_tensor=input_tensor):
return closure(
sharded_ddp_model,
sharded_scaler,
input_tensor,
grad_accumulation,
_manual_reduction=manual_reduction,
)
# Step/scale both
for _scaler, _closure, _optimizer in (
(ddp_scaler, ddp_closure, ddp_optimizer),
(sharded_scaler, sharded_closure, sharded_optimizer),
):
if _scaler is not None:
_ = _closure(input_tensor)
_scaler.step(_optimizer)
_scaler.update()
check_same_model_params(sharded_ddp_model, ddp_model,
f"Rank: {rank} - Step {i} broke")
# Check that the two grad norm are equivalent
# NOTE: The grads can occasionally be NaNs, the scaler will skip the step in that case
# This is not ShardedDDP specific. If the grads are not NaN for DDP then they should also
# be valid for ShardedDDP
if clip_grad_norm:
total_norm = torch.nn.utils.clip_grad_norm_(
ddp_model.parameters(), 0.3, norm_type=2.0) # type: ignore
if not torch.isnan(total_norm):
oss_total_norm = sharded_optimizer.clip_grad_norm(
0.3, norm_type=2.0)
assert torch.allclose(
oss_total_norm, total_norm, atol=1e-2 if amp else 1e-8
), f"torch and fairscale should return the same grad norm\n {oss_total_norm} vs {total_norm}"
else:
print(rank, "NaN grad norm in DDP", flush=True)
# Flip the trainability of the first parameter back and forth
if i == 0 and change_train_graph:
next(sharded_ddp_model.parameters()).requires_grad = not next(
sharded_ddp_model.parameters()).requires_grad
next(ddp_model.parameters()).requires_grad = not next(
ddp_model.parameters()).requires_grad
check_same_model_params(
sharded_ddp_model, ddp_model,
f"Rank: {rank} - Trainability refresh {i} broke")
def check_parity(amp: bool, accumulate: bool, change_train_graph: bool):
# The API should be the exact same in between the sharded and non-sharded variants, generic closure
def closure(model, scaler, input_tensor, should_accumulate):
accumulate_steps = 3 if should_accumulate else 1
model.zero_grad()
def step():
if scaler is not None:
with torch.cuda.amp.autocast():
loss = model(input_tensor).abs().sum()
scaler.scale(loss).backward()
else:
loss = model(input_tensor).abs().sum()
loss.backward()
with model.no_sync() if should_accumulate else suppress():
for _ in range(accumulate_steps - 1):
step()
step()
# Any model works. Add one different buffer per rank
model = Sequential(Linear(INPUTS, 3), Linear(3, 3), Linear(3, 3), Linear(3, 3), Linear(3, 3), Linear(3, 3))
model.register_buffer("test_buffer", torch.ones((1)) * rank)
model.to(device)
# Make sure that the model starts with non-trainable, so that we check for the buckets to be
# properly reassigned when/if this changes
next(model.parameters()).requires_grad = False
sharded_optimizer = OSS(params=model.parameters(), optim=torch.optim.SGD, lr=1e-5, momentum=0.99)
sharded_ddp_model = ShardedDataParallel(
module=model, sharded_optimizer=sharded_optimizer, broadcast_buffers=True
)
ddp_model_single = copy.deepcopy(model)
ddp_optimizer = torch.optim.SGD(ddp_model_single.parameters(), lr=1e-5, momentum=0.99)
ddp_model = DDP(ddp_model_single, device_ids=[rank], broadcast_buffers=True, find_unused_parameters=True)
ddp_scaler = TorchGradScaler() if amp else None
sharded_ddp_scaler = ShardedGradScaler() if amp else None
# The model should be synchronized in between the ranks at construction time, check that
check_same_model_params(sharded_ddp_model, ddp_model)
# Typical training loop, check that we get the exact same results as DDP
for i in range(NUMBER_BATCHS):
input_tensor = torch.rand((BATCH_SIZE, INPUTS)).to(device)
def closure_ddp(input_tensor=input_tensor):
return closure(ddp_model, ddp_scaler, input_tensor, accumulate)
def closure_sharded(input_tensor=input_tensor):
return closure(sharded_ddp_model, sharded_ddp_scaler, input_tensor, accumulate)
# Step/scale both
if ddp_scaler is not None:
_ = closure_ddp(input_tensor)
ddp_scaler.step(ddp_optimizer)
ddp_scaler.update()
else:
ddp_optimizer.step(closure=closure_ddp)
if sharded_ddp_scaler is not None:
_ = closure_sharded(input_tensor)
sharded_ddp_scaler.step(sharded_optimizer)
sharded_ddp_scaler.update()
else:
sharded_optimizer.step(closure=closure_sharded)
check_same_model_params(sharded_ddp_model, ddp_model, f"Rank: {rank} - Step {i} broke")
# Flip the trainability of the first parameter back and forth
if i == 0 and change_train_graph:
next(sharded_ddp_model.parameters()).requires_grad = not next(
sharded_ddp_model.parameters()
).requires_grad
next(ddp_model.parameters()).requires_grad = not next(ddp_model.parameters()).requires_grad
check_same_model_params(sharded_ddp_model, ddp_model, f"Rank: {rank} - Trainability refresh {i} broke")
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
def check_parity(amp: bool):
# Any model works. Add one different buffer per rank
model = Sequential(Linear(2, 3), Linear(3, 3), Linear(3, 3), Linear(3, 3), Linear(3, 3), Linear(3, 3))
model.register_buffer("test_buffer", torch.ones((1)) * rank)
model.to(device)
sharded_optimizer = OSS(params=model.parameters(), optim=torch.optim.SGD, lr=1e-3, momentum=0.99)
sharded_ddp_model = ShardedDataParallel(
module=model, sharded_optimizer=sharded_optimizer, broadcast_buffers=True
)
ddp_model_single = copy.deepcopy(model)
ddp_optimizer = torch.optim.SGD(ddp_model_single.parameters(), lr=1e-3, momentum=0.99)
ddp_model = DDP(ddp_model_single, device_ids=[rank], broadcast_buffers=True)
ddp_scaler = TorchGradScaler() if amp else None
sharded_ddp_scaler = ShardedGradScaler() if amp else None
def check_same_model_params():
for pg, ddp_pg in zip(sharded_optimizer.param_groups, ddp_optimizer.param_groups):
for p, ddp_p in zip(pg["params"], ddp_pg["params"]):
assert torch.allclose(
p, ddp_p, atol=1e-3
), f"Model parameters differ in between DDP and ShardedDDP {p} {ddp_p}"
for b, ddp_b in zip(sharded_ddp_model.buffers(), ddp_model.buffers()):
assert torch.allclose(
b, ddp_b, atol=1e-3
), f"Model buffers differ in between DDP and ShardedDDP. AMP {amp}"
# The model should be synchronized in between the ranks at construction time, check that
check_same_model_params()
# The models should stay the same in between the ranks
for i in range(10):
input_tensor = torch.rand((64, 2)).to(device)
def closure_ddp(input_tensor=input_tensor):
ddp_optimizer.zero_grad()
if ddp_scaler is not None:
with torch.cuda.amp.autocast():
ddp_loss = ddp_model(input_tensor).abs().sum()
ddp_scaler.scale(ddp_loss).backward()
else:
ddp_loss = ddp_model(input_tensor).abs().sum()
ddp_loss.backward()
return ddp_loss
def closure_sharded(input_tensor=input_tensor):
sharded_optimizer.zero_grad()
if sharded_ddp_scaler is not None:
with torch.cuda.amp.autocast():
sharded_loss = sharded_ddp_model(input_tensor).abs().sum()
sharded_ddp_scaler.scale(sharded_loss).backward()
else:
sharded_loss = sharded_ddp_model(input_tensor).abs().sum()
sharded_loss.backward()
return sharded_loss
# Step/scale both
if ddp_scaler is not None:
_ = closure_ddp(input_tensor)
ddp_scaler.step(ddp_optimizer)
ddp_scaler.update()
else:
ddp_optimizer.step(closure=closure_ddp)
if sharded_ddp_scaler is not None:
_ = closure_sharded(input_tensor)
sharded_ddp_scaler.step(sharded_optimizer)
sharded_ddp_scaler.update()
else:
sharded_optimizer.step(closure=closure_sharded)
check_same_model_params()