def prepare_model_and_optimizer(args, device):
# Prepare model
config = modeling.BertConfig.from_json_file(args.config_file)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
modeling.ACT2FN["bias_gelu"] = modeling.bias_gelu_training
model = modeling.BertForPreTraining(config)
if args.disable_weight_tying:
import torch.nn as nn
print ("WARNING!!!!!!! Disabling weight tying for this run")
print ("BEFORE ", model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight)
model.cls.predictions.decoder.weight = torch.nn.Parameter(model.cls.predictions.decoder.weight.clone().detach())
print ("AFTER ", model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight)
assert (model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight) == False
checkpoint = None
if not args.resume_from_checkpoint:
global_step = 0
else:
if args.resume_step == -1 and not args.init_checkpoint:
model_names = [f for f in os.listdir(args.output_dir) if f.endswith(".pt")]
args.resume_step = max([int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names])
global_step = args.resume_step if not args.init_checkpoint else 0
if not args.init_checkpoint:
checkpoint = torch.load(os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)), map_location="cpu")
else:
checkpoint = torch.load(args.init_checkpoint, map_location="cpu")
model.load_state_dict(checkpoint['model'], strict=False)
if args.phase2 and not args.init_checkpoint:
global_step -= args.phase1_end_step
if is_main_process():
print("resume step from ", args.resume_step)
model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta', 'LayerNorm']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate)
lr_scheduler = PolyWarmUpScheduler(optimizer,
warmup=args.warmup_proportion,
total_steps=args.max_steps,
degree=1)
if args.fp16:
if args.loss_scale == 0:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale="dynamic", cast_model_outputs=torch.float16)
else:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale=args.loss_scale, cast_model_outputs=torch.float16)
amp._amp_state.loss_scalers[0]._loss_scale = args.init_loss_scale
model.checkpoint_activations(args.checkpoint_activations)
if args.resume_from_checkpoint:
if args.phase2 or args.init_checkpoint:
keys = list(checkpoint['optimizer']['state'].keys())
#Override hyperparameters from previous checkpoint
for key in keys:
checkpoint['optimizer']['state'][key]['step'] = global_step
for iter, item in enumerate(checkpoint['optimizer']['param_groups']):
checkpoint['optimizer']['param_groups'][iter]['step'] = global_step
checkpoint['optimizer']['param_groups'][iter]['t_total'] = args.max_steps
checkpoint['optimizer']['param_groups'][iter]['warmup'] = args.warmup_proportion
checkpoint['optimizer']['param_groups'][iter]['lr'] = args.learning_rate
optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False)
# Restore AMP master parameters
if args.fp16:
optimizer._lazy_init_maybe_master_weights()
optimizer._amp_stash.lazy_init_called = True
optimizer.load_state_dict(checkpoint['optimizer'])
for param, saved_param in zip(amp.master_params(optimizer), checkpoint['master params']):
param.data.copy_(saved_param.data)
if args.local_rank != -1:
if not args.allreduce_post_accumulation:
model = DDP(model, message_size=250000000, gradient_predivide_factor=get_world_size())
else:
flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0,) )
elif args.n_gpu > 1:
model = torch.nn.DataParallel(model)
criterion = BertPretrainingCriterion(config.vocab_size)
if args.disable_weight_tying:
# Sanity Check that new param is in optimizer
print ("SANITY CHECK OPTIMIZER: ", id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']])
assert id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']]
return model, optimizer, lr_scheduler, checkpoint, global_step, criterion
def prepare_model_and_optimizer(args, device):
# Prepare model
config = BertConfig.from_json_file(args.config_file)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
model = BertForPreTraining(config)
checkpoint = None
if not args.resume_from_checkpoint:
global_step = 0
else:
if args.resume_step == -1 and not args.init_checkpoint:
model_names = [
f for f in os.listdir(args.output_dir) if f.endswith(".pt")
]
args.resume_step = max([
int(x.split('.pt')[0].split('_')[1].strip())
for x in model_names
])
global_step = args.resume_step if not args.init_checkpoint else 0
if not args.init_checkpoint:
checkpoint = torch.load(os.path.join(
args.output_dir, "ckpt_{}.pt".format(global_step)),
map_location="cpu")
else:
checkpoint = torch.load(args.init_checkpoint, map_location="cpu")
model.load_state_dict(checkpoint['model'], strict=False)
if args.phase2:
global_step -= args.phase1_end_step
if is_main_process():
print("resume step from ", args.resume_step)
model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta', 'LayerNorm']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = FusedLAMB(optimizer_grouped_parameters, lr=args.learning_rate)
lr_scheduler = PolyWarmUpScheduler(optimizer,
warmup=args.warmup_proportion,
total_steps=args.max_steps)
if args.fp16:
if args.loss_scale == 0:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
loss_scale="dynamic")
else:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
loss_scale=args.loss_scale)
amp._amp_state.loss_scalers[0]._loss_scale = 2**20
if args.resume_from_checkpoint:
if args.phase2 or args.init_checkpoint:
keys = list(checkpoint['optimizer']['state'].keys())
#Override hyperparameters from previous checkpoint
for key in keys:
checkpoint['optimizer']['state'][key]['step'] = global_step
for iter, item in enumerate(
checkpoint['optimizer']['param_groups']):
checkpoint['optimizer']['param_groups'][iter][
'step'] = global_step
checkpoint['optimizer']['param_groups'][iter][
't_total'] = args.max_steps
checkpoint['optimizer']['param_groups'][iter][
'warmup'] = args.warmup_proportion
checkpoint['optimizer']['param_groups'][iter][
'lr'] = args.learning_rate
optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False)
# Restore AMP master parameters
if args.fp16:
optimizer._lazy_init_maybe_master_weights()
optimizer._amp_stash.lazy_init_called = True
optimizer.load_state_dict(checkpoint['optimizer'])
for param, saved_param in zip(amp.master_params(optimizer),
checkpoint['master params']):
param.data.copy_(saved_param.data)
if args.local_rank != -1:
if not args.allreduce_post_accumulation:
model = DDP(
model,
message_size=250000000,
gradient_predivide_factor=torch.distributed.get_world_size())
else:
flat_dist_call([param.data for param in model.parameters()],
torch.distributed.broadcast, (0, ))
elif args.n_gpu > 1:
model = torch.nn.DataParallel(model)
return model, optimizer, lr_scheduler, checkpoint, global_step
def prepare_model_and_optimizer(args, device):
# Prepare model
config = modeling.BertConfig.from_json_file(args.config_file)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
if args.use_sequential > 0:
config.use_sequential = True
else:
config.use_sequential = False
modeling.ACT2FN["bias_gelu"] = modeling.bias_gelu_training
model = modeling.BertForPreTraining(config)
model.checkpoint_activations(args.checkpoint_activations)
if args.smp > 0:
# SMP: Use the DistributedModel container to provide the model
# to be partitioned across different ranks. For the rest of the script,
# the returned DistributedModel object should be used in place of
# the model provided for DistributedModel class instantiation.
model = smp.DistributedModel(model)
checkpoint = None
if not args.resume_from_checkpoint:
global_step = 0
else:
if not args.init_checkpoint:
if not args.s3_checkpoint_uri:
raise ValueError(
"Need to set s3_checkpoint_uri, if init_checkpoint not set"
)
if smp.local_rank() == 0:
sync_s3_checkpoints_to_local(args.output_dir,
args.s3_checkpoint_uri)
smp.barrier()
if args.resume_step == -1 and not args.init_checkpoint:
model_names = [
f for f in os.listdir(args.output_dir) if ".pt" in f
]
args.resume_step = max([
int(x.split(".pt")[0].split("_")[1].strip())
for x in model_names
])
global_step = args.resume_step if not args.init_checkpoint else 0
# SMP: Load a model that was saved with smp.save
if not args.init_checkpoint:
checkpoint = smp.load(
os.path.join(args.output_dir,
"ckpt_{}.pt".format(global_step)),
partial=args.partial_checkpoint,
)
else:
checkpoint = smp.load(args.init_checkpoint)
model.load_state_dict(checkpoint["model"], strict=False)
if args.phase2 and not args.init_checkpoint:
global_step -= args.phase1_end_step
if is_main_process():
print("resume step from ", args.resume_step)
model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ["bias", "gamma", "beta", "LayerNorm"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
"weight_decay":
0.01,
},
{
"params":
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
"weight_decay":
0.0,
},
]
optimizer = FusedLAMB(optimizer_grouped_parameters, lr=args.learning_rate)
if args.smp > 0:
# SMP: Use Distributed Optimizer which allows the loading of optimizer state for a distributed model
# Also provides APIs to obtain local optimizer state for the current mp_rank.
optimizer = smp.DistributedOptimizer(optimizer)
lr_scheduler = PolyWarmUpScheduler(optimizer,
warmup=args.warmup_proportion,
total_steps=args.max_steps)
if args.fp16:
if args.loss_scale == 0:
model, optimizer = amp.initialize(
model,
optimizer,
opt_level="O2",
loss_scale="dynamic",
cast_model_outputs=torch.float16,
)
else:
model, optimizer = amp.initialize(
model,
optimizer,
opt_level="O2",
loss_scale=args.loss_scale,
cast_model_outputs=torch.float16,
)
amp._amp_state.loss_scalers[0]._loss_scale = args.init_loss_scale
if args.resume_from_checkpoint:
if args.phase2 or args.init_checkpoint:
keys = list(checkpoint["optimizer"]["state"].keys())
# Override hyperparameters from previous checkpoint
for key in keys:
checkpoint["optimizer"]["state"][key]["step"] = global_step
for iter, item in enumerate(
checkpoint["optimizer"]["param_groups"]):
checkpoint["optimizer"]["param_groups"][iter][
"step"] = global_step
checkpoint["optimizer"]["param_groups"][iter][
"t_total"] = args.max_steps
checkpoint["optimizer"]["param_groups"][iter][
"warmup"] = args.warmup_proportion
checkpoint["optimizer"]["param_groups"][iter][
"lr"] = args.learning_rate
optimizer.load_state_dict(checkpoint["optimizer"]) # , strict=False)
# Restore AMP master parameters
if args.fp16:
optimizer._lazy_init_maybe_master_weights()
optimizer._amp_stash.lazy_init_called = True
optimizer.load_state_dict(checkpoint["optimizer"])
for param, saved_param in zip(amp.master_params(optimizer),
checkpoint["master params"]):
param.data.copy_(saved_param.data)
# if args.local_rank != -1:
# if not args.allreduce_post_accumulation:
# model = DDP(model, message_size=250000000, gradient_predivide_factor=get_world_size())
# else:
# flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0,) )
# elif args.n_gpu > 1:
# model = torch.nn.DataParallel(model)
criterion = BertPretrainingCriterion(config.vocab_size)
return model, optimizer, lr_scheduler, checkpoint, global_step, criterion
def prepare_model_and_optimizer(args, device):
# Prepare model
config = modeling.BertConfig.from_json_file(args.config_file)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
modeling.ACT2FN["bias_gelu"] = modeling.bias_gelu_training
model = modeling.BertForPreTraining(config)
checkpoint = None
if not args.resume_from_checkpoint:
global_step = 0
else:
if args.resume_step == -1 and not args.init_checkpoint:
model_names = [
f for f in os.listdir(args.output_dir) if f.endswith(".pt")
]
args.resume_step = max([
int(x.split('.pt')[0].split('_')[1].strip())
for x in model_names
])
global_step = args.resume_step if not args.init_checkpoint else 0
if not args.init_checkpoint:
checkpoint = torch.load(os.path.join(
args.output_dir, "ckpt_{}.pt".format(global_step)),
map_location="cpu")
else:
checkpoint = torch.load(args.init_checkpoint, map_location="cpu")
model.load_state_dict(checkpoint['model'], strict=False)
if args.phase2 and not args.init_checkpoint:
global_step -= args.phase1_end_step
if is_main_process():
print("resume step from ", args.resume_step)
model.to(device)
# BERT modeling uses weight sharing between word embedding and prediction decoder.
# So make sure the storage is pointing properly even after model is moved to device.
if args.use_habana:
model.cls.predictions.decoder.weight = model.bert.embeddings.word_embeddings.weight
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta', 'LayerNorm']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.use_habana:
if args.use_fused_lamb:
try:
from hb_custom import FusedLamb
except ImportError:
raise ImportError("Please install hbopt.")
optimizer = FusedLamb(optimizer_grouped_parameters,
lr=args.learning_rate)
else:
optimizer = NVLAMB(optimizer_grouped_parameters,
lr=args.learning_rate)
else:
if torch.cuda.is_available():
optimizer = FusedLAMB(optimizer_grouped_parameters,
lr=args.learning_rate)
else:
optimizer = NVLAMB(optimizer_grouped_parameters,
lr=args.learning_rate)
lr_scheduler = PolyWarmUpScheduler(optimizer,
warmup=args.warmup_proportion,
total_steps=args.max_steps)
if args.fp16:
if args.loss_scale == 0:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
loss_scale="dynamic",
cast_model_outputs=torch.float16)
else:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
loss_scale=args.loss_scale,
cast_model_outputs=torch.float16)
amp._amp_state.loss_scalers[0]._loss_scale = args.init_loss_scale
model.checkpoint_activations(args.checkpoint_activations)
if args.resume_from_checkpoint:
if args.phase2 or args.init_checkpoint:
keys = list(checkpoint['optimizer']['state'].keys())
#Override hyperparameters from previous checkpoint
for key in keys:
checkpoint['optimizer']['state'][key]['step'] = global_step
for iter, item in enumerate(
checkpoint['optimizer']['param_groups']):
checkpoint['optimizer']['param_groups'][iter][
'step'] = global_step
checkpoint['optimizer']['param_groups'][iter][
't_total'] = args.max_steps
checkpoint['optimizer']['param_groups'][iter][
'warmup'] = args.warmup_proportion
checkpoint['optimizer']['param_groups'][iter][
'lr'] = args.learning_rate
optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False)
# Restore AMP master parameters
if args.fp16:
optimizer._lazy_init_maybe_master_weights()
optimizer._amp_stash.lazy_init_called = True
optimizer.load_state_dict(checkpoint['optimizer'])
for param, saved_param in zip(amp.master_params(optimizer),
checkpoint['master params']):
param.data.copy_(saved_param.data)
if args.local_rank != -1:
if not args.allreduce_post_accumulation:
if not args.use_jit_trace:
if args.use_habana:
model = DDP(model)
else:
model = DDP(model,
message_size=250000000,
gradient_predivide_factor=get_world_size())
else:
flat_dist_call([param.data for param in model.parameters()],
torch.distributed.broadcast, (0, ))
elif args.n_pu > 1:
model = torch.nn.DataParallel(model)
criterion = BertPretrainingCriterion(config.vocab_size)
return model, optimizer, lr_scheduler, checkpoint, global_step, criterion
def prepare_model_and_optimizer(args, device, sequence_output_is_dense):
# Prepare model
config = modeling.BertConfig.from_json_file(args.config_file)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
model = modeling.BertForPreTraining(config, sequence_output_is_dense=sequence_output_is_dense)
checkpoint = None
if not args.resume_from_checkpoint:
global_step = 0
else:
if args.resume_step == -1 and not args.init_checkpoint:
model_names = [f for f in os.listdir(args.output_dir) if f.endswith(".pt")]
args.resume_step = max([int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names])
global_step = args.resume_step if not args.init_checkpoint else 0
if not args.init_checkpoint:
checkpoint = torch.load(os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)), map_location=device)
else:
checkpoint = torch.load(args.init_checkpoint, map_location=device)
model.load_state_dict(checkpoint['model'], strict=False)
if args.phase2 and not args.init_checkpoint:
global_step -= args.phase1_end_step
if is_main_process():
print("resume step from ", args.resume_step)
model.to(device)
# If allreduce_post_accumulation_fp16 is not set, Native AMP Autocast is
# used along with FP32 gradient accumulation and all-reduce
if args.fp16 and args.allreduce_post_accumulation_fp16:
model.half()
if not args.disable_jit_fusions :
model = torch.jit.script(model)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta', 'LayerNorm']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}]
optimizer = FusedLAMBAMP(optimizer_grouped_parameters,
lr=args.learning_rate)
lr_scheduler = PolyWarmUpScheduler(optimizer,
warmup=args.warmup_proportion,
total_steps=args.max_steps,
base_lr=args.learning_rate,
device=device)
grad_scaler = torch.cuda.amp.GradScaler(init_scale=args.init_loss_scale, enabled=args.fp16)
model.checkpoint_activations(args.checkpoint_activations)
if args.resume_from_checkpoint:
# For phase2, need to reset the learning rate and step count in the checkpoint
if args.phase2 or args.init_checkpoint :
for group in checkpoint['optimizer']['param_groups'] :
group['step'].zero_()
group['lr'].fill_(args.learning_rate)
else :
if 'grad_scaler' in checkpoint and not args.phase2:
grad_scaler.load_state_dict(checkpoint['grad_scaler'])
optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False)
if args.local_rank != -1:
# Cuda Graphs requires that DDP is captured on a side stream
# It is important to synchronize the streams after the DDP initialization
# so anything after sees properly initialized model weights across GPUs
side_stream = torch.cuda.Stream()
with torch.cuda.stream(side_stream) :
model = DDP(model, device_ids=[args.local_rank], output_device=args.local_rank, bucket_cap_mb=torch.cuda.get_device_properties(device).total_memory, gradient_as_bucket_view=True)
torch.cuda.current_stream().wait_stream(side_stream)
from torch.distributed.algorithms.ddp_comm_hooks.default_hooks import allreduce_hook
def scale_by_grad_accum_steps_wrapper(hook: Callable[[Any, dist.GradBucket], torch.futures.Future[torch.Tensor]]) -> Callable[[Any, dist.GradBucket], torch.futures.Future[torch.Tensor]]:
def scale_by_grad_accum_steps_wrapper_hook(
hook_state, bucket: dist.GradBucket
) -> torch.futures.Future[torch.Tensor]:
bucket.set_buffer(bucket.buffer().div_(args.gradient_accumulation_steps))
fut = hook(hook_state, bucket)
return fut
return scale_by_grad_accum_steps_wrapper_hook
# With gradient accumulation, the DDP comm hook divides the gradients by the number
# gradient accumulation steps
if args.gradient_accumulation_steps > 1:
model.register_comm_hook(None, scale_by_grad_accum_steps_wrapper(allreduce_hook))
optimizer.setup_fp32_params()
criterion = BertPretrainingCriterion(config.vocab_size, sequence_output_is_dense=sequence_output_is_dense)
if args.resume_from_checkpoint and args.init_checkpoint:
start_epoch = checkpoint['epoch']
else:
start_epoch = 0
return model, optimizer, grad_scaler, lr_scheduler, checkpoint, global_step, criterion, start_epoch