def get_model(neox_args, inference=False, get_key_value=True):
"""Build the model."""
print_rank_0('building GPT2 model ...')
# Build model on cpu.
model = GPT2ModelPipe(neox_args=neox_args,
num_tokentypes=0,
parallel_output=True,
topology=mpu.get_topology(),
inference=inference,
get_key_value=get_key_value)
if not neox_args.is_pipe_parallel:
# Export PipeParallel model to nn.Sequential model to avoid the overhead of deepspeed's pipe parallel training
model = model.to_sequential()
else:
# This is a hack to give us a reference to get_batch_pipe from within training.py
# We need to call model.set_batch_fn after deepspeed.initialize
model._megatron_batch_fn = partial(get_batch_pipe, neox_args=neox_args)
if neox_args.deepspeed:
# DeepSpeed handles CUDA, FP16, and DDP components.
return model
else:
raise ValueError("Must be using deepspeed to run neox")
def model_provider():
"""Build the model."""
args = get_args()
print_rank_0('building GPT2 model ...')
if args.pipe_parallel_size == 0:
model = GPT2Model(num_tokentypes=0, parallel_output=True)
else:
model = GPT2ModelPipe(num_tokentypes=0,
parallel_output=True,
topology=mpu.get_topology())
# This is a hack to give us a reference to get_batch_pipe from within training.py
# We need to call model.set_batch_fn after deepspeed.initialize
model._megatron_batch_fn = get_batch_pipe
return model
def model_provider():
"""Build the model."""
args = get_args()
print_rank_0('building GPT2 model ...')
if args.pipe_parallel_size == 0:
model = GPT2Model(num_tokentypes=0, parallel_output=True)
else:
model = GPT2ModelPipe(num_tokentypes=0,
parallel_output=True,
topology=mpu.get_topology())
# This is a hack to give us a reference to get_batch_pipe from within training.py
# We need to call model.set_batch_fn after deepspeed.initialize
model._megatron_batch_fn = get_batch_pipe
## Wandb
use_wandb = get_wandb_api_key() is not None
set_use_wandb(use_wandb)
args_dict = vars(args)
if use_wandb:
# only display system stats from one worker per machine
wandb_settings = wandb.Settings() if is_local_main(
) else wandb.Settings(_disable_stats=True)
group_name = args_dict.get('wandb_group')
name = f'{socket.gethostname()}-{local_rank()}' if group_name else None
try:
wandb.init(project="neox",
group=group_name,
name=name,
save_code=False,
force=False,
entity=args_dict.get('wandb_team'),
settings=wandb_settings)
except UsageError as e:
set_use_wandb(False)
print(e)
print(
'Skipping wandb. Execute `wandb login` on local or main node machine to enable.'
)
if use_wandb:
wandb.config.update(args_dict)
return model
def get_model(neox_args, use_cache=False):
"""Build the model."""
print_rank_0("building GPT2 model ...")
# Build model on cpu.
model = GPT2ModelPipe(
neox_args=neox_args,
num_tokentypes=0,
parallel_output=True,
topology=mpu.get_topology(),
use_cache=use_cache,
)
### soft prompt tuning stuff ###
if neox_args.soft_prompt_tuning is not None and neox_args.soft_prompt_tuning.get(
"enabled", False):
soft_prompt = SoftEmbedding(
neox_args,
wte=getattr(model, "0").word_embeddings,
n_tokens=neox_args.soft_prompt_tuning.get("n_tokens", 10),
init_string=neox_args.soft_prompt_tuning.get("init_string", ""),
init_range=neox_args.soft_prompt_tuning.get("init_range", 0.5),
)
model.insert_layers(
layers=soft_prompt, idx=1
) # insert the soft prompt layer directly after the word embeddings
# freeze everything but the soft prompt
for name, param in model.named_parameters():
if not "soft_embedding" in name:
param.requires_grad = False
if not neox_args.is_pipe_parallel:
# Export PipeParallel model to nn.Sequential model to avoid the overhead of deepspeed's pipe parallel training
model = model.to_sequential()
if neox_args.deepspeed:
# DeepSpeed handles CUDA, FP16, and DDP components.
return model
else:
raise ValueError("Must be using deepspeed to run neox")
