def main(argv):
rank, world_size, gpu = dist.init_distributed_mode()
top_mlp = create_top_mlp().to("cuda")
print(top_mlp)
optimizer = torch.optim.SGD(top_mlp.parameters(), lr=1.)
if FLAGS.fp16:
top_mlp, optimizer = amp.initialize(top_mlp,
optimizer,
opt_level="O1",
loss_scale=1)
if world_size > 1:
top_mlp = parallel.DistributedDataParallel(top_mlp)
model_without_ddp = top_mlp.module
dummy_bottom_mlp_output = torch.rand(FLAGS.batch_size,
EMBED_DIM,
device="cuda")
dummy_embedding_output = torch.rand(FLAGS.batch_size,
26 * EMBED_DIM,
device="cuda")
dummy_target = torch.ones(FLAGS.batch_size, device="cuda")
if FLAGS.fp16:
dummy_bottom_mlp_output = dummy_bottom_mlp_output.to(torch.half)
dummy_embedding_output = dummy_embedding_output.to(torch.half)
# warm up GPU
for _ in range(100):
interaction_out = dot_interaction(dummy_bottom_mlp_output,
[dummy_embedding_output],
FLAGS.batch_size)
output = top_mlp(interaction_out)
start_time = utils.timer_start()
for _ in range(FLAGS.num_iters):
interaction_out = dot_interaction(dummy_bottom_mlp_output,
[dummy_embedding_output],
FLAGS.batch_size)
output = top_mlp(interaction_out).squeeze()
dummy_loss = output.mean()
optimizer.zero_grad()
if FLAGS.fp16:
with amp.scale_loss(dummy_loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
dummy_loss.backward()
optimizer.step()
stop_time = utils.timer_stop()
elapsed_time = (stop_time - start_time) / FLAGS.num_iters * 1e3
print(F"Average step time: {elapsed_time:.4f} ms.")
def main(argv):
if FLAGS.seed is not None:
torch.manual_seed(FLAGS.seed)
np.random.seed(FLAGS.seed)
# Initialize distributed mode
use_gpu = "cpu" not in FLAGS.device.lower()
rank, world_size, gpu = dist.init_distributed_mode(backend=FLAGS.backend,
use_gpu=use_gpu)
if world_size == 1:
raise NotImplementedError(
"This file is only for distributed training.")
mlperf_logger.mlperf_submission_log('dlrm')
mlperf_logger.log_event(key=mlperf_logger.constants.SEED, value=FLAGS.seed)
mlperf_logger.log_event(key=mlperf_logger.constants.GLOBAL_BATCH_SIZE,
value=FLAGS.batch_size)
# Only print cmd args on rank 0
if rank == 0:
print("Command line flags:")
pprint(FLAGS.flag_values_dict())
# Check arguments sanity
if FLAGS.batch_size % world_size != 0:
raise ValueError(
F"Batch size {FLAGS.batch_size} is not divisible by world_size {world_size}."
)
if FLAGS.test_batch_size % world_size != 0:
raise ValueError(
F"Test batch size {FLAGS.test_batch_size} is not divisible by world_size {world_size}."
)
# Load config file, create sub config for each rank
with open(FLAGS.model_config, "r") as f:
config = json.loads(f.read())
wolrd_categorical_feature_sizes = np.asarray(
config.pop('categorical_feature_sizes'))
device_mapping = dist_model.get_criteo_device_mapping(world_size)
vectors_per_gpu = device_mapping['vectors_per_gpu']
# Get sizes of embeddings each GPU is gonna create
categorical_feature_sizes = wolrd_categorical_feature_sizes[
device_mapping['embedding'][rank]].tolist()
bottom_mlp_sizes = config.pop('bottom_mlp_sizes')
if rank != device_mapping['bottom_mlp']:
bottom_mlp_sizes = None
model = dist_model.DistDlrm(
categorical_feature_sizes=categorical_feature_sizes,
bottom_mlp_sizes=bottom_mlp_sizes,
world_num_categorical_features=len(wolrd_categorical_feature_sizes),
**config,
device=FLAGS.device,
use_embedding_ext=FLAGS.use_embedding_ext)
print(model)
dist.setup_distributed_print(rank == 0)
# DDP introduces a gradient average through allreduce(mean), which doesn't apply to bottom model.
# Compensate it with further scaling lr
scaled_lr = FLAGS.lr / FLAGS.loss_scale if FLAGS.fp16 else FLAGS.lr
scaled_lrs = [scaled_lr / world_size, scaled_lr]
embedding_optimizer = torch.optim.SGD([
{
'params': model.bottom_model.joint_embedding.parameters(),
'lr': scaled_lrs[0]
},
])
mlp_optimizer = apex_optim.FusedSGD([{
'params':
model.bottom_model.bottom_mlp.parameters(),
'lr':
scaled_lrs[0]
}, {
'params': model.top_model.parameters(),
'lr': scaled_lrs[1]
}])
if FLAGS.fp16:
(model.top_model,
model.bottom_model.bottom_mlp), mlp_optimizer = amp.initialize(
[model.top_model, model.bottom_model.bottom_mlp],
mlp_optimizer,
opt_level="O2",
loss_scale=1,
cast_model_outputs=torch.float16)
if use_gpu:
model.top_model = parallel.DistributedDataParallel(model.top_model)
else: # Use other backend for CPU
model.top_model = torch.nn.parallel.DistributedDataParallel(
model.top_model)
loss_fn = torch.nn.BCEWithLogitsLoss(reduction="mean")
# Too many arguments to pass for distributed training. Use plain train code here instead of
# defining a train function
# Print per 16384 * 2000 samples by default
default_print_freq = 16384 * 2000 // FLAGS.batch_size
print_freq = default_print_freq if FLAGS.print_freq is None else FLAGS.print_freq
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter(
'loss', utils.SmoothedValue(window_size=1, fmt='{avg:.4f}'))
metric_logger.add_meter(
'step_time', utils.SmoothedValue(window_size=1, fmt='{avg:.4f} ms'))
metric_logger.add_meter(
'lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
# Accumulating loss on GPU to avoid memcpyD2H every step
moving_loss = torch.zeros(1, device=FLAGS.device)
moving_loss_stream = torch.cuda.Stream()
local_embedding_device_mapping = torch.tensor(
device_mapping['embedding'][rank],
device=FLAGS.device,
dtype=torch.long)
# LR is logged twice for now because of a compliance checker bug
mlperf_logger.log_event(key=mlperf_logger.constants.OPT_BASE_LR,
value=FLAGS.lr)
mlperf_logger.log_event(key=mlperf_logger.constants.OPT_LR_WARMUP_STEPS,
value=FLAGS.warmup_steps)
# use logging keys from the official HP table and not from the logging library
mlperf_logger.log_event(key='sgd_opt_base_learning_rate', value=FLAGS.lr)
mlperf_logger.log_event(key='lr_decay_start_steps',
value=FLAGS.decay_start_step)
mlperf_logger.log_event(key='sgd_opt_learning_rate_decay_steps',
value=FLAGS.decay_steps)
mlperf_logger.log_event(key='sgd_opt_learning_rate_decay_poly_power',
value=FLAGS.decay_power)
lr_scheduler = utils.LearningRateScheduler(
optimizers=[mlp_optimizer, embedding_optimizer],
base_lrs=[scaled_lrs, [scaled_lrs[0]]],
warmup_steps=FLAGS.warmup_steps,
warmup_factor=FLAGS.warmup_factor,
decay_start_step=FLAGS.decay_start_step,
decay_steps=FLAGS.decay_steps,
decay_power=FLAGS.decay_power,
end_lr_factor=FLAGS.decay_end_lr / FLAGS.lr)
data_stream = torch.cuda.Stream()
eval_data_cache = [] if FLAGS.cache_eval_data else None
start_time = time()
stop_time = time()
print("Creating data loaders")
dist_dataset_args = {
"numerical_features": rank == 0,
"categorical_features": device_mapping['embedding'][rank]
}
mlperf_logger.barrier()
mlperf_logger.log_end(key=mlperf_logger.constants.INIT_STOP)
mlperf_logger.barrier()
mlperf_logger.log_start(key=mlperf_logger.constants.RUN_START)
mlperf_logger.barrier()
data_loader_train, data_loader_test = dataset.get_data_loader(
FLAGS.dataset,
FLAGS.batch_size,
FLAGS.test_batch_size,
FLAGS.device,
dataset_type=FLAGS.dataset_type,
shuffle=FLAGS.shuffle,
**dist_dataset_args)
steps_per_epoch = len(data_loader_train)
# Default 20 tests per epoch
test_freq = FLAGS.test_freq if FLAGS.test_freq is not None else steps_per_epoch // 20
for epoch in range(FLAGS.epochs):
epoch_start_time = time()
mlperf_logger.barrier()
mlperf_logger.log_start(key=mlperf_logger.constants.BLOCK_START,
metadata={
mlperf_logger.constants.FIRST_EPOCH_NUM:
epoch + 1,
mlperf_logger.constants.EPOCH_COUNT:
1
})
mlperf_logger.barrier()
mlperf_logger.log_start(
key=mlperf_logger.constants.EPOCH_START,
metadata={mlperf_logger.constants.EPOCH_NUM: epoch + 1})
if FLAGS.profile_steps is not None:
torch.cuda.profiler.start()
for step, (numerical_features, categorical_features,
click) in enumerate(
dataset.prefetcher(iter(data_loader_train),
data_stream)):
torch.cuda.current_stream().wait_stream(data_stream)
global_step = steps_per_epoch * epoch + step
lr_scheduler.step()
# Slice out categorical features if not using the "dist" dataset
if FLAGS.dataset_type != "dist":
categorical_features = categorical_features[:,
local_embedding_device_mapping]
if FLAGS.fp16 and categorical_features is not None:
numerical_features = numerical_features.to(torch.float16)
last_batch_size = None
if click.shape[0] != FLAGS.batch_size: # last batch
last_batch_size = click.shape[0]
logging.debug("Pad the last batch of size %d to %d",
last_batch_size, FLAGS.batch_size)
padding_size = FLAGS.batch_size - last_batch_size
padding_numiercal = torch.empty(
padding_size,
numerical_features.shape[1],
device=numerical_features.device,
dtype=numerical_features.dtype)
numerical_features = torch.cat(
(numerical_features, padding_numiercal), dim=0)
if categorical_features is not None:
padding_categorical = torch.ones(
padding_size,
categorical_features.shape[1],
device=categorical_features.device,
dtype=categorical_features.dtype)
categorical_features = torch.cat(
(categorical_features, padding_categorical), dim=0)
padding_click = torch.empty(padding_size,
device=click.device,
dtype=click.dtype)
click = torch.cat((click, padding_click))
bottom_out = model.bottom_model(numerical_features,
categorical_features)
batch_size_per_gpu = FLAGS.batch_size // world_size
from_bottom = dist_model.bottom_to_top(bottom_out,
batch_size_per_gpu,
config['embedding_dim'],
vectors_per_gpu)
if last_batch_size is not None:
partial_rank = math.ceil(last_batch_size / batch_size_per_gpu)
if rank == partial_rank:
top_out = model.top_model(
from_bottom[:last_batch_size %
batch_size_per_gpu]).squeeze().float()
loss = loss_fn(
top_out, click[rank * batch_size_per_gpu:(rank + 1) *
batch_size_per_gpu][:last_batch_size %
batch_size_per_gpu])
elif rank < partial_rank:
loss = loss_fn(
model.top_model(from_bottom).squeeze().float(),
click[rank * batch_size_per_gpu:(rank + 1) *
batch_size_per_gpu])
else:
# Back propgate nothing for padded samples
loss = 0. * model.top_model(
from_bottom).squeeze().float().mean()
else:
loss = loss_fn(
model.top_model(from_bottom).squeeze().float(),
click[rank * batch_size_per_gpu:(rank + 1) *
batch_size_per_gpu])
# We don't need to accumulate gradient. Set grad to None is faster than optimizer.zero_grad()
for param_group in itertools.chain(
embedding_optimizer.param_groups,
mlp_optimizer.param_groups):
for param in param_group['params']:
param.grad = None
if FLAGS.fp16:
loss *= FLAGS.loss_scale
with amp.scale_loss(loss, mlp_optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
mlp_optimizer.step()
embedding_optimizer.step()
moving_loss_stream.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(moving_loss_stream):
moving_loss += loss
if step == 0:
print(F"Started epoch {epoch}...")
elif step % print_freq == 0:
torch.cuda.synchronize()
# Averaging cross a print_freq period to reduce the error.
# An accurate timing needs synchronize which would slow things down.
metric_logger.update(step_time=(time() - stop_time) * 1000 /
print_freq,
loss=moving_loss.item() / print_freq /
(FLAGS.loss_scale if FLAGS.fp16 else 1),
lr=mlp_optimizer.param_groups[1]["lr"] *
(FLAGS.loss_scale if FLAGS.fp16 else 1))
stop_time = time()
eta_str = datetime.timedelta(
seconds=int(metric_logger.step_time.avg / 1000 *
(steps_per_epoch - step)))
metric_logger.print(
header=
F"Epoch:[{epoch}/{FLAGS.epochs}] [{step}/{steps_per_epoch}] eta: {eta_str}"
)
moving_loss = 0.
with torch.cuda.stream(moving_loss_stream):
moving_loss = 0.
if global_step % test_freq == 0 and global_step > 0 and global_step / steps_per_epoch >= FLAGS.test_after:
mlperf_epoch_index = global_step / steps_per_epoch + 1
mlperf_logger.barrier()
mlperf_logger.log_start(key=mlperf_logger.constants.EVAL_START,
metadata={
mlperf_logger.constants.EPOCH_NUM:
mlperf_epoch_index
})
auc = dist_evaluate(model, data_loader_test, eval_data_cache)
mlperf_logger.log_event(
key=mlperf_logger.constants.EVAL_ACCURACY,
value=float(auc),
metadata={
mlperf_logger.constants.EPOCH_NUM: mlperf_epoch_index
})
print(F"Epoch {epoch} step {step}. auc {auc:.6f}")
stop_time = time()
mlperf_logger.barrier()
mlperf_logger.log_end(key=mlperf_logger.constants.EVAL_STOP,
metadata={
mlperf_logger.constants.EPOCH_NUM:
mlperf_epoch_index
})
if auc > FLAGS.auc_threshold:
mlperf_logger.barrier()
mlperf_logger.log_end(key=mlperf_logger.constants.RUN_STOP,
metadata={
mlperf_logger.constants.STATUS:
mlperf_logger.constants.SUCCESS
})
mlperf_logger.barrier()
mlperf_logger.log_end(
key=mlperf_logger.constants.EPOCH_STOP,
metadata={
mlperf_logger.constants.EPOCH_NUM: epoch + 1
})
mlperf_logger.barrier()
mlperf_logger.log_end(
key=mlperf_logger.constants.BLOCK_STOP,
metadata={
mlperf_logger.constants.FIRST_EPOCH_NUM: epoch + 1
})
run_time_s = int(stop_time - start_time)
print(
F"Hit target accuracy AUC {FLAGS.auc_threshold} at epoch "
F"{global_step/steps_per_epoch:.2f} in {run_time_s}s. "
F"Average speed {global_step * FLAGS.batch_size / run_time_s:.1f} records/s."
)
return
if FLAGS.profile_steps is not None and global_step == FLAGS.profile_steps:
torch.cuda.profiler.stop()
logging.warning("Profile run, stopped at step %d.",
global_step)
return
mlperf_logger.barrier()
mlperf_logger.log_end(
key=mlperf_logger.constants.EPOCH_STOP,
metadata={mlperf_logger.constants.EPOCH_NUM: epoch + 1})
mlperf_logger.barrier()
mlperf_logger.log_end(
key=mlperf_logger.constants.BLOCK_STOP,
metadata={mlperf_logger.constants.FIRST_EPOCH_NUM: epoch + 1})
epoch_stop_time = time()
epoch_time_s = epoch_stop_time - epoch_start_time
print(
F"Finished epoch {epoch} in {datetime.timedelta(seconds=int(epoch_time_s))}. "
F"Average speed {steps_per_epoch * FLAGS.batch_size / epoch_time_s:.1f} records/s."
)
mlperf_logger.barrier()
mlperf_logger.log_end(key=mlperf_logger.constants.RUN_STOP,
metadata={
mlperf_logger.constants.STATUS:
mlperf_logger.constants.ABORTED
})
def main(argv):
torch.manual_seed(FLAGS.seed)
utils.init_logging(log_path=FLAGS.log_path)
use_gpu = "cpu" not in FLAGS.base_device.lower()
rank, world_size, gpu = dist.init_distributed_mode(backend=FLAGS.backend,
use_gpu=use_gpu)
device = FLAGS.base_device
if is_main_process():
dllogger.log(data=FLAGS.flag_values_dict(), step='PARAMETER')
print("Command line flags:")
pprint(FLAGS.flag_values_dict())
print("Creating data loaders")
FLAGS.set_default("test_batch_size",
FLAGS.test_batch_size // world_size * world_size)
categorical_feature_sizes = get_categorical_feature_sizes(FLAGS)
world_categorical_feature_sizes = np.asarray(categorical_feature_sizes)
device_mapping = get_device_mapping(categorical_feature_sizes,
num_gpus=world_size)
batch_sizes_per_gpu = get_gpu_batch_sizes(FLAGS.batch_size,
num_gpus=world_size)
batch_indices = tuple(np.cumsum([0] + list(batch_sizes_per_gpu)))
# sizes of embeddings for each GPU
categorical_feature_sizes = world_categorical_feature_sizes[
device_mapping['embedding'][rank]].tolist()
bottom_mlp_sizes = FLAGS.bottom_mlp_sizes if rank == device_mapping[
'bottom_mlp'] else None
data_loader_train, data_loader_test = get_data_loaders(
FLAGS, device_mapping=device_mapping)
model = DistributedDlrm(
vectors_per_gpu=device_mapping['vectors_per_gpu'],
embedding_device_mapping=device_mapping['embedding'],
embedding_type=FLAGS.embedding_type,
embedding_dim=FLAGS.embedding_dim,
world_num_categorical_features=len(world_categorical_feature_sizes),
categorical_feature_sizes=categorical_feature_sizes,
num_numerical_features=FLAGS.num_numerical_features,
hash_indices=FLAGS.hash_indices,
bottom_mlp_sizes=bottom_mlp_sizes,
top_mlp_sizes=FLAGS.top_mlp_sizes,
interaction_op=FLAGS.interaction_op,
fp16=FLAGS.amp,
use_cpp_mlp=FLAGS.optimized_mlp,
bottom_features_ordered=FLAGS.bottom_features_ordered,
device=device)
print(model)
print(device_mapping)
print(f"Batch sizes per gpu: {batch_sizes_per_gpu}")
dist.setup_distributed_print(is_main_process())
# DDP introduces a gradient average through allreduce(mean), which doesn't apply to bottom model.
# Compensate it with further scaling lr
scaled_lr = FLAGS.lr / FLAGS.loss_scale if FLAGS.amp else FLAGS.lr
if FLAGS.Adam_embedding_optimizer:
embedding_model_parallel_lr = scaled_lr
else:
embedding_model_parallel_lr = scaled_lr / world_size
if FLAGS.Adam_MLP_optimizer:
MLP_model_parallel_lr = scaled_lr
else:
MLP_model_parallel_lr = scaled_lr / world_size
data_parallel_lr = scaled_lr
if is_main_process():
mlp_params = [{
'params': list(model.top_model.parameters()),
'lr': data_parallel_lr
}, {
'params': list(model.bottom_model.mlp.parameters()),
'lr': MLP_model_parallel_lr
}]
mlp_lrs = [data_parallel_lr, MLP_model_parallel_lr]
else:
mlp_params = [{
'params': list(model.top_model.parameters()),
'lr': data_parallel_lr
}]
mlp_lrs = [data_parallel_lr]
if FLAGS.Adam_MLP_optimizer:
mlp_optimizer = apex_optim.FusedAdam(mlp_params)
else:
mlp_optimizer = apex_optim.FusedSGD(mlp_params)
embedding_params = [{
'params':
list(model.bottom_model.embeddings.parameters()),
'lr':
embedding_model_parallel_lr
}]
embedding_lrs = [embedding_model_parallel_lr]
if FLAGS.Adam_embedding_optimizer:
embedding_optimizer = torch.optim.SparseAdam(embedding_params)
else:
embedding_optimizer = torch.optim.SGD(embedding_params)
checkpoint_writer = make_distributed_checkpoint_writer(
device_mapping=device_mapping,
rank=rank,
is_main_process=is_main_process(),
config=FLAGS.flag_values_dict())
checkpoint_loader = make_distributed_checkpoint_loader(
device_mapping=device_mapping, rank=rank)
if FLAGS.load_checkpoint_path:
checkpoint_loader.load_checkpoint(model, FLAGS.load_checkpoint_path)
model.to(device)
if FLAGS.amp:
(model.top_model,
model.bottom_model.mlp), mlp_optimizer = amp.initialize(
[model.top_model, model.bottom_model.mlp],
mlp_optimizer,
opt_level="O2",
loss_scale=1)
if use_gpu:
model.top_model = parallel.DistributedDataParallel(model.top_model)
else: # Use other backend for CPU
model.top_model = torch.nn.parallel.DistributedDataParallel(
model.top_model)
if FLAGS.mode == 'test':
auc = dist_evaluate(model, data_loader_test)
results = {'auc': auc}
dllogger.log(data=results, step=tuple())
if auc is not None:
print(f"Finished testing. Test auc {auc:.4f}")
return
if FLAGS.save_checkpoint_path and not FLAGS.bottom_features_ordered and is_main_process(
):
logging.warning(
"Saving checkpoint without --bottom_features_ordered flag will result in "
"a device-order dependent model. Consider using --bottom_features_ordered "
"if you plan to load the checkpoint in different device configurations."
)
loss_fn = torch.nn.BCEWithLogitsLoss(reduction="mean")
# Print per 16384 * 2000 samples by default
default_print_freq = 16384 * 2000 // FLAGS.batch_size
print_freq = default_print_freq if FLAGS.print_freq is None else FLAGS.print_freq
steps_per_epoch = len(data_loader_train)
test_freq = FLAGS.test_freq if FLAGS.test_freq is not None else steps_per_epoch - 1
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter(
'loss', utils.SmoothedValue(window_size=1, fmt='{avg:.4f}'))
metric_logger.add_meter(
'step_time', utils.SmoothedValue(window_size=1, fmt='{avg:.6f}'))
metric_logger.add_meter(
'lr', utils.SmoothedValue(window_size=1, fmt='{value:.4f}'))
# Accumulating loss on GPU to avoid memcpyD2H every step
moving_loss = torch.zeros(1, device=device)
moving_loss_stream = torch.cuda.Stream()
lr_scheduler = utils.LearningRateScheduler(
optimizers=[mlp_optimizer, embedding_optimizer],
base_lrs=[mlp_lrs, embedding_lrs],
warmup_steps=FLAGS.warmup_steps,
warmup_factor=FLAGS.warmup_factor,
decay_start_step=FLAGS.decay_start_step,
decay_steps=FLAGS.decay_steps,
decay_power=FLAGS.decay_power,
end_lr_factor=FLAGS.decay_end_lr / FLAGS.lr)
data_stream = torch.cuda.Stream()
timer = utils.StepTimer()
best_auc = 0
best_epoch = 0
start_time = time()
stop_time = time()
for epoch in range(FLAGS.epochs):
epoch_start_time = time()
batch_iter = prefetcher(iter(data_loader_train), data_stream)
for step in range(len(data_loader_train)):
timer.click()
numerical_features, categorical_features, click = next(batch_iter)
torch.cuda.synchronize()
global_step = steps_per_epoch * epoch + step
if FLAGS.max_steps and global_step > FLAGS.max_steps:
print(
f"Reached max global steps of {FLAGS.max_steps}. Stopping."
)
break
lr_scheduler.step()
if click.shape[0] != FLAGS.batch_size: # last batch
logging.error("The last batch with size %s is not supported",
click.shape[0])
else:
output = model(numerical_features, categorical_features,
batch_sizes_per_gpu).squeeze()
loss = loss_fn(
output, click[batch_indices[rank]:batch_indices[rank + 1]])
if FLAGS.Adam_embedding_optimizer or FLAGS.Adam_MLP_optimizer:
model.zero_grad()
else:
# We don't need to accumulate gradient. Set grad to None is faster than optimizer.zero_grad()
for param_group in itertools.chain(
embedding_optimizer.param_groups,
mlp_optimizer.param_groups):
for param in param_group['params']:
param.grad = None
if FLAGS.amp:
loss *= FLAGS.loss_scale
with amp.scale_loss(loss, mlp_optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
if FLAGS.Adam_MLP_optimizer:
scale_MLP_gradients(mlp_optimizer, world_size)
mlp_optimizer.step()
if FLAGS.Adam_embedding_optimizer:
scale_embeddings_gradients(embedding_optimizer, world_size)
embedding_optimizer.step()
moving_loss_stream.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(moving_loss_stream):
moving_loss += loss
if timer.measured is None:
# first iteration, no step time etc. to print
continue
if step == 0:
print(f"Started epoch {epoch}...")
elif step % print_freq == 0:
torch.cuda.current_stream().wait_stream(moving_loss_stream)
# Averaging across a print_freq period to reduce the error.
# An accurate timing needs synchronize which would slow things down.
if global_step < FLAGS.benchmark_warmup_steps:
metric_logger.update(
loss=moving_loss.item() / print_freq /
(FLAGS.loss_scale if FLAGS.amp else 1),
lr=mlp_optimizer.param_groups[0]["lr"] *
(FLAGS.loss_scale if FLAGS.amp else 1))
else:
metric_logger.update(
step_time=timer.measured,
loss=moving_loss.item() / print_freq /
(FLAGS.loss_scale if FLAGS.amp else 1),
lr=mlp_optimizer.param_groups[0]["lr"] *
(FLAGS.loss_scale if FLAGS.amp else 1))
stop_time = time()
eta_str = datetime.timedelta(
seconds=int(metric_logger.step_time.global_avg *
(steps_per_epoch - step)))
metric_logger.print(
header=
f"Epoch:[{epoch}/{FLAGS.epochs}] [{step}/{steps_per_epoch}] eta: {eta_str}"
)
with torch.cuda.stream(moving_loss_stream):
moving_loss = 0.
if global_step % test_freq == 0 and global_step > 0 and global_step / steps_per_epoch >= FLAGS.test_after:
auc = dist_evaluate(model, data_loader_test)
if auc is None:
continue
print(f"Epoch {epoch} step {step}. auc {auc:.6f}")
stop_time = time()
if auc > best_auc:
best_auc = auc
best_epoch = epoch + ((step + 1) / steps_per_epoch)
if FLAGS.auc_threshold and auc >= FLAGS.auc_threshold:
run_time_s = int(stop_time - start_time)
print(
f"Hit target accuracy AUC {FLAGS.auc_threshold} at epoch "
f"{global_step / steps_per_epoch:.2f} in {run_time_s}s. "
f"Average speed {global_step * FLAGS.batch_size / run_time_s:.1f} records/s."
)
sys.exit()
epoch_stop_time = time()
epoch_time_s = epoch_stop_time - epoch_start_time
print(
f"Finished epoch {epoch} in {datetime.timedelta(seconds=int(epoch_time_s))}. "
f"Average speed {steps_per_epoch * FLAGS.batch_size / epoch_time_s:.1f} records/s."
)
avg_throughput = FLAGS.batch_size / metric_logger.step_time.avg
if FLAGS.save_checkpoint_path:
checkpoint_writer.save_checkpoint(model, FLAGS.save_checkpoint_path,
epoch, step)
results = {
'best_auc': best_auc,
'best_epoch': best_epoch,
'average_train_throughput': avg_throughput
}
dllogger.log(data=results, step=tuple())
def main(argv):
torch.manual_seed(FLAGS.seed)
utils.init_logging(log_path=FLAGS.log_path)
use_gpu = "cpu" not in FLAGS.base_device.lower()
rank, world_size, gpu = dist.init_distributed_mode(backend=FLAGS.backend, use_gpu=use_gpu)
device = FLAGS.base_device
feature_spec = load_feature_spec(FLAGS)
cat_feature_count = len(get_embedding_sizes(feature_spec, None))
validate_flags(cat_feature_count)
if is_main_process():
dllogger.log(data=FLAGS.flag_values_dict(), step='PARAMETER')
FLAGS.set_default("test_batch_size", FLAGS.test_batch_size // world_size * world_size)
feature_spec = load_feature_spec(FLAGS)
world_embedding_sizes = get_embedding_sizes(feature_spec, max_table_size=FLAGS.max_table_size)
world_categorical_feature_sizes = np.asarray(world_embedding_sizes)
device_mapping = get_device_mapping(world_embedding_sizes, num_gpus=world_size)
batch_sizes_per_gpu = get_gpu_batch_sizes(FLAGS.batch_size, num_gpus=world_size)
batch_indices = tuple(np.cumsum([0] + list(batch_sizes_per_gpu))) # todo what does this do
# Embedding sizes for each GPU
categorical_feature_sizes = world_categorical_feature_sizes[device_mapping['embedding'][rank]].tolist()
num_numerical_features = feature_spec.get_number_of_numerical_features()
bottom_mlp_sizes = FLAGS.bottom_mlp_sizes if rank == device_mapping['bottom_mlp'] else None
data_loader_train, data_loader_test = get_data_loaders(FLAGS, device_mapping=device_mapping,
feature_spec=feature_spec)
model = DistributedDlrm(
vectors_per_gpu=device_mapping['vectors_per_gpu'],
embedding_device_mapping=device_mapping['embedding'],
embedding_type=FLAGS.embedding_type,
embedding_dim=FLAGS.embedding_dim,
world_num_categorical_features=len(world_categorical_feature_sizes),
categorical_feature_sizes=categorical_feature_sizes,
num_numerical_features=num_numerical_features,
hash_indices=FLAGS.hash_indices,
bottom_mlp_sizes=bottom_mlp_sizes,
top_mlp_sizes=FLAGS.top_mlp_sizes,
interaction_op=FLAGS.interaction_op,
fp16=FLAGS.amp,
use_cpp_mlp=FLAGS.optimized_mlp,
bottom_features_ordered=FLAGS.bottom_features_ordered,
device=device
)
dist.setup_distributed_print(is_main_process())
# DDP introduces a gradient average through allreduce(mean), which doesn't apply to bottom model.
# Compensate it with further scaling lr
if FLAGS.Adam_embedding_optimizer:
embedding_model_parallel_lr = FLAGS.lr
else:
embedding_model_parallel_lr = FLAGS.lr / world_size
if FLAGS.Adam_MLP_optimizer:
MLP_model_parallel_lr = FLAGS.lr
else:
MLP_model_parallel_lr = FLAGS.lr / world_size
data_parallel_lr = FLAGS.lr
if is_main_process():
mlp_params = [
{'params': list(model.top_model.parameters()), 'lr': data_parallel_lr},
{'params': list(model.bottom_model.mlp.parameters()), 'lr': MLP_model_parallel_lr}
]
mlp_lrs = [data_parallel_lr, MLP_model_parallel_lr]
else:
mlp_params = [
{'params': list(model.top_model.parameters()), 'lr': data_parallel_lr}
]
mlp_lrs = [data_parallel_lr]
if FLAGS.Adam_MLP_optimizer:
mlp_optimizer = apex_optim.FusedAdam(mlp_params)
else:
mlp_optimizer = apex_optim.FusedSGD(mlp_params)
embedding_params = [{
'params': list(model.bottom_model.embeddings.parameters()),
'lr': embedding_model_parallel_lr
}]
embedding_lrs = [embedding_model_parallel_lr]
if FLAGS.Adam_embedding_optimizer:
embedding_optimizer = torch.optim.SparseAdam(embedding_params)
else:
embedding_optimizer = torch.optim.SGD(embedding_params)
checkpoint_writer = make_distributed_checkpoint_writer(
device_mapping=device_mapping,
rank=rank,
is_main_process=is_main_process(),
config=FLAGS.flag_values_dict()
)
checkpoint_loader = make_distributed_checkpoint_loader(device_mapping=device_mapping, rank=rank)
if FLAGS.load_checkpoint_path:
checkpoint_loader.load_checkpoint(model, FLAGS.load_checkpoint_path)
model.to(device)
scaler = torch.cuda.amp.GradScaler(enabled=FLAGS.amp, growth_interval=int(1e9))
def parallelize(model):
if world_size <= 1:
return model
if use_gpu:
model.top_model = parallel.DistributedDataParallel(model.top_model)
else: # Use other backend for CPU
model.top_model = torch.nn.parallel.DistributedDataParallel(model.top_model)
return model
if FLAGS.mode == 'test':
model = parallelize(model)
auc = dist_evaluate(model, data_loader_test)
results = {'auc': auc}
dllogger.log(data=results, step=tuple())
if auc is not None:
print(f"Finished testing. Test auc {auc:.4f}")
return
elif FLAGS.mode == 'inference_benchmark':
if world_size > 1:
raise ValueError('Inference benchmark only supports singleGPU mode.')
results = {}
if FLAGS.amp:
# can use pure FP16 for inference
model = model.half()
for batch_size in FLAGS.inference_benchmark_batch_sizes:
FLAGS.test_batch_size = batch_size
_, data_loader_test = get_data_loaders(FLAGS, device_mapping=device_mapping, feature_spec=feature_spec)
latencies = inference_benchmark(model=model, data_loader=data_loader_test,
num_batches=FLAGS.inference_benchmark_steps,
cuda_graphs=FLAGS.cuda_graphs)
# drop the first 10 as a warmup
latencies = latencies[10:]
mean_latency = np.mean(latencies)
mean_inference_throughput = batch_size / mean_latency
subresult = {f'mean_inference_latency_batch_{batch_size}': mean_latency,
f'mean_inference_throughput_batch_{batch_size}': mean_inference_throughput}
results.update(subresult)
dllogger.log(data=results, step=tuple())
return
if FLAGS.save_checkpoint_path and not FLAGS.bottom_features_ordered and is_main_process():
logging.warning("Saving checkpoint without --bottom_features_ordered flag will result in "
"a device-order dependent model. Consider using --bottom_features_ordered "
"if you plan to load the checkpoint in different device configurations.")
loss_fn = torch.nn.BCEWithLogitsLoss(reduction="mean")
# Print per 16384 * 2000 samples by default
default_print_freq = 16384 * 2000 // FLAGS.batch_size
print_freq = default_print_freq if FLAGS.print_freq is None else FLAGS.print_freq
# last one will be dropped in the training loop
steps_per_epoch = len(data_loader_train) - 1
test_freq = FLAGS.test_freq if FLAGS.test_freq is not None else steps_per_epoch - 2
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('loss', utils.SmoothedValue(window_size=1, fmt='{avg:.8f}'))
metric_logger.add_meter('step_time', utils.SmoothedValue(window_size=1, fmt='{avg:.6f}'))
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.4f}'))
# Accumulating loss on GPU to avoid memcpyD2H every step
moving_loss = torch.zeros(1, device=device)
lr_scheduler = utils.LearningRateScheduler(optimizers=[mlp_optimizer, embedding_optimizer],
base_lrs=[mlp_lrs, embedding_lrs],
warmup_steps=FLAGS.warmup_steps,
warmup_factor=FLAGS.warmup_factor,
decay_start_step=FLAGS.decay_start_step,
decay_steps=FLAGS.decay_steps,
decay_power=FLAGS.decay_power,
end_lr_factor=FLAGS.decay_end_lr / FLAGS.lr)
def zero_grad(model):
if FLAGS.Adam_embedding_optimizer or FLAGS.Adam_MLP_optimizer:
model.zero_grad()
else:
# We don't need to accumulate gradient. Set grad to None is faster than optimizer.zero_grad()
for param_group in itertools.chain(embedding_optimizer.param_groups, mlp_optimizer.param_groups):
for param in param_group['params']:
param.grad = None
def forward_backward(model, *args):
numerical_features, categorical_features, click = args
with torch.cuda.amp.autocast(enabled=FLAGS.amp):
output = model(numerical_features, categorical_features, batch_sizes_per_gpu).squeeze()
loss = loss_fn(output, click[batch_indices[rank]: batch_indices[rank + 1]])
scaler.scale(loss).backward()
return loss
def weight_update():
if not FLAGS.freeze_mlps:
if FLAGS.Adam_MLP_optimizer:
scale_MLP_gradients(mlp_optimizer, world_size)
scaler.step(mlp_optimizer)
if not FLAGS.freeze_embeddings:
if FLAGS.Adam_embedding_optimizer:
scale_embeddings_gradients(embedding_optimizer, world_size)
scaler.unscale_(embedding_optimizer)
embedding_optimizer.step()
scaler.update()
trainer = CudaGraphWrapper(model, forward_backward, parallelize, zero_grad,
cuda_graphs=FLAGS.cuda_graphs)
data_stream = torch.cuda.Stream()
timer = utils.StepTimer()
best_auc = 0
best_epoch = 0
start_time = time()
for epoch in range(FLAGS.epochs):
epoch_start_time = time()
batch_iter = prefetcher(iter(data_loader_train), data_stream)
for step in range(len(data_loader_train)):
timer.click()
numerical_features, categorical_features, click = next(batch_iter)
torch.cuda.synchronize()
global_step = steps_per_epoch * epoch + step
if FLAGS.max_steps and global_step > FLAGS.max_steps:
print(f"Reached max global steps of {FLAGS.max_steps}. Stopping.")
break
# One of the batches will be smaller because the dataset size
# isn't necessarily a multiple of the batch size. #TODO isn't dropping here a change of behavior
if click.shape[0] != FLAGS.batch_size:
continue
lr_scheduler.step()
loss = trainer.train_step(numerical_features, categorical_features, click)
# need to wait for the gradients before the weight update
torch.cuda.current_stream().wait_stream(trainer.stream)
weight_update()
moving_loss += loss
if timer.measured is None:
# first iteration, no step time etc. to print
continue
if step == 0:
print(f"Started epoch {epoch}...")
elif step % print_freq == 0:
# Averaging across a print_freq period to reduce the error.
# An accurate timing needs synchronize which would slow things down.
if global_step < FLAGS.benchmark_warmup_steps:
metric_logger.update(
loss=moving_loss.item() / print_freq,
lr=mlp_optimizer.param_groups[0]["lr"])
else:
metric_logger.update(
step_time=timer.measured,
loss=moving_loss.item() / print_freq,
lr=mlp_optimizer.param_groups[0]["lr"])
eta_str = datetime.timedelta(seconds=int(metric_logger.step_time.global_avg * (steps_per_epoch - step)))
metric_logger.print(header=f"Epoch:[{epoch}/{FLAGS.epochs}] [{step}/{steps_per_epoch}] eta: {eta_str}")
moving_loss = 0.
if global_step % test_freq == 0 and global_step > 0 and global_step / steps_per_epoch >= FLAGS.test_after:
auc = dist_evaluate(trainer.model, data_loader_test)
if auc is None:
continue
print(f"Epoch {epoch} step {step}. auc {auc:.6f}")
stop_time = time()
if auc > best_auc:
best_auc = auc
best_epoch = epoch + ((step + 1) / steps_per_epoch)
if FLAGS.auc_threshold and auc >= FLAGS.auc_threshold:
run_time_s = int(stop_time - start_time)
print(f"Hit target accuracy AUC {FLAGS.auc_threshold} at epoch "
f"{global_step / steps_per_epoch:.2f} in {run_time_s}s. ")
sys.exit()
epoch_stop_time = time()
epoch_time_s = epoch_stop_time - epoch_start_time
print(f"Finished epoch {epoch} in {datetime.timedelta(seconds=int(epoch_time_s))}. ")
avg_throughput = FLAGS.batch_size / metric_logger.step_time.avg
if FLAGS.save_checkpoint_path:
checkpoint_writer.save_checkpoint(model, FLAGS.save_checkpoint_path, epoch, step)
results = {'best_auc': best_auc,
'best_epoch': best_epoch,
'average_train_throughput': avg_throughput}
if is_main_process():
dllogger.log(data=results, step=tuple())