def main():
"""
Launches translation (inference).
Inference is executed on a single GPU, implementation supports beam search
with length normalization and coverage penalty.
"""
args = parse_args()
device = utils.set_device(args.cuda, args.local_rank)
utils.init_distributed(args.cuda)
args.rank = utils.get_rank()
utils.setup_logging()
if args.env:
utils.log_env_info()
logging.info(f'Run arguments: {args}')
if not args.cuda and torch.cuda.is_available():
warnings.warn('cuda is available but not enabled')
if not args.cudnn:
torch.backends.cudnn.enabled = False
# load checkpoint and deserialize to CPU (to save GPU memory)
checkpoint = torch.load(args.model, map_location={'cuda:0': 'cpu'})
# build GNMT model
tokenizer = Tokenizer()
tokenizer.set_state(checkpoint['tokenizer'])
model_config = checkpoint['model_config']
model_config['batch_first'] = args.batch_first
model_config['vocab_size'] = tokenizer.vocab_size
model = GNMT(**model_config)
model.load_state_dict(checkpoint['state_dict'])
# construct the dataset
if args.input:
data = RawTextDataset(
raw_datafile=args.input,
tokenizer=tokenizer,
sort=args.sort,
)
elif args.input_text:
data = RawTextDataset(
raw_data=args.input_text,
tokenizer=tokenizer,
sort=args.sort,
)
latency_table = tables.LatencyTable(args.percentiles)
throughput_table = tables.ThroughputTable(args.percentiles)
accuracy_table = tables.AccuracyTable('BLEU')
dtype = {'fp32': torch.FloatTensor, 'fp16': torch.HalfTensor}
for (math, batch_size, beam_size) in product(args.math, args.batch_size,
args.beam_size):
logging.info(f'math: {math}, batch size: {batch_size}, '
f'beam size: {beam_size}')
model.type(dtype[math])
model = model.to(device)
model.eval()
# build the data loader
loader = data.get_loader(
batch_size=batch_size,
batch_first=args.batch_first,
pad=True,
repeat=args.repeat[batch_size],
num_workers=0,
)
# build the translator object
translator = Translator(
model=model,
tokenizer=tokenizer,
loader=loader,
beam_size=beam_size,
max_seq_len=args.max_seq_len,
len_norm_factor=args.len_norm_factor,
len_norm_const=args.len_norm_const,
cov_penalty_factor=args.cov_penalty_factor,
print_freq=args.print_freq,
)
# execute the inference
output, stats = translator.run(
calc_bleu=args.bleu,
eval_path=args.output,
summary=True,
warmup=args.warmup,
reference_path=args.reference,
)
# print translated outputs
if not args.output and args.rank == 0:
logging.info(f'Translated output:')
for out in output:
print(out)
key = (batch_size, beam_size)
latency_table.add(key, {math: stats['runtimes']})
throughput_table.add(key, {math: stats['throughputs']})
accuracy_table.add(key, {math: stats['bleu']})
if args.tables:
accuracy_table.write('Inference accuracy', args.math)
if 'fp16' in args.math and 'fp32' in args.math:
relative = 'fp32'
else:
relative = None
if 'fp32' in args.math:
throughput_table.write('Inference throughput', 'fp32')
if 'fp16' in args.math:
throughput_table.write('Inference throughput',
'fp16',
relative=relative)
if 'fp32' in args.math:
latency_table.write('Inference latency', 'fp32')
if 'fp16' in args.math:
latency_table.write('Inference latency',
'fp16',
relative=relative,
reverse_speedup=True)
passed = utils.benchmark(stats['bleu'], args.target_bleu,
stats['tokens_per_sec'], args.target_perf)
return passed
def main():
"""
Launches translation (inference).
Inference is executed on a single GPU, implementation supports beam search
with length normalization and coverage penalty.
"""
args = parse_args()
args.batch_first = False
if args.cuda:
torch.cuda.set_device(0)
if not args.cuda and torch.cuda.is_available():
warnings.warn('cuda is available but not enabled')
if args.math == 'fp16' and not args.cuda:
raise RuntimeError('fp16 requires cuda')
if not args.cudnn:
torch.backends.cudnn.enabled = False
num_stages = args.num_stages
# compute BLEU score for every epoch
print("Epoch\tBLEU score")
epoch = 0
while True:
# no more epochs to run, since desired file not available
if not os.path.isfile(
os.path.join(args.checkpoint_path,
f"checkpoint.0.pth.tar.epoch.{epoch}")):
break
module = importlib.import_module(args.module)
model = module.model(None)
num_modules = len(model)
key_to_module_mapping = OrderedDict()
all_stages_state_dict = OrderedDict()
module_id = 0
stage_id = 0
for stage_id in range(num_stages):
# load the checkpoint associated with a stage
full_checkpoint_path = os.path.join(
args.checkpoint_path,
f"checkpoint.{stage_id}.pth.tar.epoch.{epoch}")
checkpoint = torch.load(full_checkpoint_path,
map_location=torch.device('cpu'))
# iterate through all modules in stage_id's checkpoint
local_module_id = 0
# quit when checkpoints for all modules in full model are loaded
while module_id < num_modules:
# load checkpoint corresponding to different modules in our runtime
state_dict = checkpoint["state_dict"]
state_dict_key = "module%d" % local_module_id
if state_dict_key not in state_dict:
break
state_dict = checkpoint["state_dict"][state_dict_key]
# remove mask buffer
keys_to_delete = []
for key in state_dict:
if "mask" in key:
keys_to_delete.append(key)
for key in keys_to_delete:
del state_dict[key]
if checkpoint_from_distributed(state_dict):
state_dict = unwrap_distributed(state_dict)
# collect all state_dicts in a single OrderedDict
for key in state_dict:
all_stages_state_dict[(stage_id, local_module_id,
key)] = state_dict[key]
stage_module, _, _ = model[module_id]
for key in state_dict:
# key_to_module_mapping maps key (in state_dict) to the
# torch.nn.Module wrapping the parameter and the name
# of parameter (weight, bias, etc.)
key_to_module_mapping[(
stage_id, local_module_id,
key)] = get_submodule_and_parameter_name(
stage_module, key)
# load tokenizer state
tokenizer = Tokenizer()
tokenizer.set_state(checkpoint['tokenizer'])
vocab_size = tokenizer.vocab_size
local_module_id += 1
module_id += 1
epoch += 1
# build model, and load state dict
model_config = {
'vocab_size': vocab_size,
'batch_first': args.batch_first,
'hidden_size': 1024,
'num_layers': args.num_layers,
'dropout': 0.2,
'share_embedding': False
}
model = GNMT(**model_config)
model_state_dict = OrderedDict()
for real_key in model.state_dict():
(module, parameter_name) = get_submodule_and_parameter_name(
model, real_key)
# find key in all_stages_state_dict that corresponds to real_key in
# model's state_dict
for key in key_to_module_mapping:
(module2, parameter_name2) = key_to_module_mapping[key]
if parameter_name == parameter_name2 and str(module) == str(
module2):
break
if parameter_name == parameter_name2 and str(module) == str(
module2):
model_state_dict[real_key] = all_stages_state_dict[key]
del key_to_module_mapping[key]
del all_stages_state_dict[key]
# load state_dict into model, and perform inference
model.load_state_dict(model_state_dict)
if args.math == 'fp32':
dtype = torch.FloatTensor
if args.math == 'fp16':
dtype = torch.HalfTensor
model.type(dtype)
model = model.cuda()
model.eval()
# construct the dataset
test_data = TextDataset(src_fname=args.input,
tokenizer=tokenizer,
sort=False)
# build the data loader
test_loader = test_data.get_loader(world_size=1,
rank=0,
batch_size=args.batch_size,
batch_first=args.batch_first,
shuffle=False,
pad=True,
num_workers=0)
# build the translator object
translator = Translator(model=model,
tokenizer=tokenizer,
loader=test_loader,
beam_size=args.beam_size,
max_seq_len=args.max_seq_len,
len_norm_factor=args.len_norm_factor,
len_norm_const=args.len_norm_const,
cov_penalty_factor=args.cov_penalty_factor,
cuda=args.cuda,
print_freq=args.print_freq,
dataset_dir=args.dataset_dir)
# execute the inference
test_bleu, _ = translator.run(calc_bleu=args.bleu,
eval_path=args.output,
reference_path=args.reference,
summary=True)
print(f'{epoch}\t{test_bleu:.2f}')
def main():
"""
Launches translation (inference).
Inference is executed on a single GPU, implementation supports beam search
with length normalization and coverage penalty.
"""
args = parse_args()
if args.affinity != 'disabled':
nproc_per_node = torch.cuda.device_count()
affinity = gpu_affinity.set_affinity(args.local_rank, nproc_per_node,
args.affinity)
print(f'{args.local_rank}: thread affinity: {affinity}')
device = utils.set_device(args.cuda, args.local_rank)
utils.init_distributed(args.cuda)
args.rank = utils.get_rank()
os.makedirs(args.save_dir, exist_ok=True)
utils.setup_logging()
dllog_file = os.path.join(args.save_dir, args.dllog_file)
utils.setup_dllogger(enabled=True, filename=dllog_file)
if args.profile:
try:
pyprof.init(enable_function_stack=True)
except NameError:
warnings.warn('Called pyprof.init() but pyprof is not available')
if args.env:
utils.log_env_info()
logging.info(f'Run arguments: {args}')
dllogger.log(step='PARAMETER', data=vars(args))
if not args.cuda and torch.cuda.is_available():
warnings.warn('cuda is available but not enabled')
if not args.cudnn:
torch.backends.cudnn.enabled = False
# load checkpoint and deserialize to CPU (to save GPU memory)
if args.model:
checkpoint = torch.load(args.model, map_location={'cuda:0': 'cpu'})
# build GNMT model
tokenizer = Tokenizer()
tokenizer.set_state(checkpoint['tokenizer'])
model_config = checkpoint['model_config']
model_config['batch_first'] = args.batch_first
model_config['vocab_size'] = tokenizer.vocab_size
model = GNMT(**model_config)
model.load_state_dict(checkpoint['state_dict'])
elif args.synthetic:
model = GNMT(args.synthetic_vocab, batch_first=args.batch_first)
tokenizer = None
else:
raise RuntimeError(
'Specify model either with --synthetic or with --model flag')
# construct the dataset
if args.input:
data = RawTextDataset(
raw_datafile=args.input,
tokenizer=tokenizer,
sort=args.sort,
)
elif args.input_text:
data = RawTextDataset(
raw_data=args.input_text,
tokenizer=tokenizer,
sort=args.sort,
)
elif args.synthetic:
data = SyntheticDataset(args.synthetic_vocab, args.synthetic_len,
args.batch_size[0] * args.synthetic_batches)
latency_table = tables.LatencyTable(args.percentiles)
throughput_table = tables.ThroughputTable(args.percentiles)
accuracy_table = tables.AccuracyTable('BLEU')
dtype = {
'fp32': torch.FloatTensor,
'tf32': torch.FloatTensor,
'fp16': torch.HalfTensor
}
for (math, batch_size, beam_size) in product(args.math, args.batch_size,
args.beam_size):
logging.info(f'math: {math}, batch size: {batch_size}, '
f'beam size: {beam_size}')
model.type(dtype[math])
model = model.to(device)
model.eval()
# build the data loader
loader = data.get_loader(
batch_size=batch_size,
batch_first=args.batch_first,
pad=True,
repeat=args.repeat[batch_size],
num_workers=0,
)
# build the translator object
translator = Translator(
model=model,
tokenizer=tokenizer,
loader=loader,
beam_size=beam_size,
max_seq_len=args.max_seq_len,
len_norm_factor=args.len_norm_factor,
len_norm_const=args.len_norm_const,
cov_penalty_factor=args.cov_penalty_factor,
print_freq=args.print_freq,
)
# execute the inference
with torch.autograd.profiler.emit_nvtx(enabled=args.profile):
output, stats = translator.run(
calc_bleu=args.bleu,
eval_path=args.output,
summary=True,
warmup=args.warmup,
reference_path=args.reference,
)
# print translated outputs
if not args.synthetic and (not args.output and args.rank == 0):
logging.info(f'Translated output:')
for out in output:
print(out)
key = (batch_size, beam_size)
latency_table.add(key, {math: stats['runtimes']})
throughput_table.add(key, {math: stats['throughputs']})
accuracy_table.add(key, {math: stats['bleu']})
if args.tables:
accuracy_table.write('Inference accuracy', args.math)
if 'fp16' in args.math and 'fp32' in args.math:
relative = 'fp32'
elif 'fp16' in args.math and 'tf32' in args.math:
relative = 'tf32'
else:
relative = None
if 'fp32' in args.math:
throughput_table.write('Inference throughput', 'fp32')
if 'tf32' in args.math:
throughput_table.write('Inference throughput', 'tf32')
if 'fp16' in args.math:
throughput_table.write('Inference throughput',
'fp16',
relative=relative)
if 'fp32' in args.math:
latency_table.write('Inference latency', 'fp32')
if 'tf32' in args.math:
latency_table.write('Inference latency', 'tf32')
if 'fp16' in args.math:
latency_table.write('Inference latency',
'fp16',
relative=relative,
reverse_speedup=True)
avg_throughput = np.array(stats['throughputs']).mean()
avg_latency = np.array(stats['runtimes']).mean()
summary = {
'eval_throughput': avg_throughput,
'eval_bleu': stats['bleu'],
'eval_avg_latency': avg_latency,
}
for p in args.percentiles:
summary[f'eval_{p}%_latency'] = np.percentile(stats['runtimes'], p)
dllogger.log(step=tuple(), data=summary)
passed = utils.benchmark(stats['bleu'], args.target_bleu,
stats['tokens_per_sec'], args.target_perf)
return passed
def main():
"""
Launches translation (inference).
Inference is executed on a single GPU, implementation supports beam search
with length normalization and coverage penalty.
"""
args = parse_args()
# initialize distributed backend
distributed = args.world_size > 1
if distributed:
backend = 'nccl' if args.cuda else 'gloo'
dist.init_process_group(backend=backend,
rank=args.rank,
init_method=args.dist_url,
world_size=args.world_size)
setup_logging()
logging.info(f'Run arguments: {args}')
if args.cuda:
torch.cuda.set_device(args.rank)
if not args.cuda and torch.cuda.is_available():
warnings.warn('cuda is available but not enabled')
if args.math == 'fp16' and not args.cuda:
raise RuntimeError('fp16 requires cuda')
if not args.cudnn:
torch.backends.cudnn.enabled = False
# load checkpoint and deserialize to CPU (to save GPU memory)
checkpoint = torch.load(args.model, map_location={'cuda:0': 'cpu'})
# build GNMT model
tokenizer = Tokenizer()
tokenizer.set_state(checkpoint['tokenizer'])
vocab_size = tokenizer.vocab_size
model_config = dict(vocab_size=vocab_size,
math=checkpoint['config'].math,
**literal_eval(checkpoint['config'].model_config))
model_config['batch_first'] = args.batch_first
model = GNMT(**model_config)
state_dict = checkpoint['state_dict']
if checkpoint_from_distributed(state_dict):
state_dict = unwrap_distributed(state_dict)
model.load_state_dict(state_dict)
if args.math == 'fp32':
dtype = torch.FloatTensor
if args.math == 'fp16':
dtype = torch.HalfTensor
model.type(dtype)
if args.cuda:
model = model.cuda()
model.eval()
# construct the dataset
test_data = TextDataset(src_fname=args.input,
tokenizer=tokenizer,
sort=False)
# build the data loader
test_loader = test_data.get_loader(batch_size=args.batch_size,
batch_first=args.batch_first,
shuffle=False,
pad=True,
num_workers=0,
drop_last=False)
# build the translator object
translator = Translator(model=model,
tokenizer=tokenizer,
loader=test_loader,
beam_size=args.beam_size,
max_seq_len=args.max_seq_len,
len_norm_factor=args.len_norm_factor,
len_norm_const=args.len_norm_const,
cov_penalty_factor=args.cov_penalty_factor,
cuda=args.cuda,
print_freq=args.print_freq,
dataset_dir=args.dataset_dir)
# execute the inference
translator.run(calc_bleu=args.bleu,
eval_path=args.output,
reference_path=args.reference,
summary=True)
def main():
"""
Launches translation (inference).
Inference is executed on a single GPU, implementation supports beam search
with length normalization and coverage penalty.
"""
args = parse_args()
utils.set_device(args.cuda, args.local_rank)
utils.init_distributed(args.cuda)
setup_logging()
if args.env:
utils.log_env_info()
logging.info(f'Run arguments: {args}')
if not args.cuda and torch.cuda.is_available():
warnings.warn('cuda is available but not enabled')
if not args.cudnn:
torch.backends.cudnn.enabled = False
# load checkpoint and deserialize to CPU (to save GPU memory)
checkpoint = torch.load(args.model, map_location={'cuda:0': 'cpu'})
# build GNMT model
tokenizer = Tokenizer()
tokenizer.set_state(checkpoint['tokenizer'])
vocab_size = tokenizer.vocab_size
model_config = checkpoint['model_config']
model_config['batch_first'] = args.batch_first
model = GNMT(vocab_size=vocab_size, **model_config)
model.load_state_dict(checkpoint['state_dict'])
for (math, batch_size, beam_size) in product(args.math, args.batch_size,
args.beam_size):
logging.info(f'math: {math}, batch size: {batch_size}, '
f'beam size: {beam_size}')
if math == 'fp32':
dtype = torch.FloatTensor
if math == 'fp16':
dtype = torch.HalfTensor
model.type(dtype)
if args.cuda:
model = model.cuda()
model.eval()
# construct the dataset
test_data = TextDataset(src_fname=args.input,
tokenizer=tokenizer,
sort=args.sort)
# build the data loader
test_loader = test_data.get_loader(batch_size=batch_size,
batch_first=args.batch_first,
shuffle=False,
pad=True,
num_workers=0)
# build the translator object
translator = Translator(model=model,
tokenizer=tokenizer,
loader=test_loader,
beam_size=beam_size,
max_seq_len=args.max_seq_len,
len_norm_factor=args.len_norm_factor,
len_norm_const=args.len_norm_const,
cov_penalty_factor=args.cov_penalty_factor,
cuda=args.cuda,
print_freq=args.print_freq,
dataset_dir=args.dataset_dir)
# execute the inference
translator.run(calc_bleu=args.bleu,
eval_path=args.output,
reference_path=args.reference,
summary=True)
