def inference_benchmark_nongraphed(model, data_loader, num_batches=100):
model.eval()
base_device = FLAGS.base_device
latencies = []
y_true = []
y_score = []
with torch.no_grad():
for step, (numerical_features, categorical_features, click) in enumerate(data_loader):
if step > num_batches:
break
step_start_time = time()
numerical_features = numerical_features.to(base_device)
if FLAGS.amp:
numerical_features = numerical_features.half()
categorical_features = categorical_features.to(device=base_device, dtype=torch.int64)
inference_result = model(numerical_features, categorical_features).squeeze()
torch.cuda.synchronize()
step_time = time() - step_start_time
if step >= FLAGS.benchmark_warmup_steps:
latencies.append(step_time)
y_true.append(click)
y_score.append(inference_result.reshape([-1]).clone())
y_true = torch.cat(y_true)
y_score = torch.sigmoid(torch.cat(y_score)).float()
auc = utils.roc_auc_score(y_true, y_score)
print('auc: ', auc)
return latencies
def evaluate(model, loss_fn, data_loader):
"""Test dlrm model
Args:
model (dlrm):
loss_fn (torch.nn.Module): Loss function
data_loader (torch.utils.data.DataLoader):
"""
model.eval()
print_freq = FLAGS.print_freq
prefetching_enabled = is_data_prefetching_enabled()
steps_per_epoch = len(data_loader)
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}'))
if prefetching_enabled:
data_stream = torch.cuda.Stream()
with torch.no_grad():
y_true = []
y_score = []
timer = utils.StepTimer()
timer.click()
input_pipeline = iter(data_loader)
if prefetching_enabled:
input_pipeline = prefetcher(input_pipeline, data_stream)
for step, (numerical_features, categorical_features,
click) in enumerate(input_pipeline):
if FLAGS.amp:
numerical_features = numerical_features.half()
if prefetching_enabled:
torch.cuda.synchronize()
output = model(numerical_features, categorical_features).squeeze()
loss = loss_fn(output, click)
y_true.append(click)
y_score.append(output)
loss_value = loss.item()
timer.click()
if timer.measured is not None:
metric_logger.update(loss=loss_value, step_time=timer.measured)
if step % print_freq == 0 and step > 0:
metric_logger.print(
header=f"Test: [{step}/{steps_per_epoch}]")
y_true = torch.cat(y_true)
y_score = torch.cat(y_score)
before_auc_timestamp = time()
auc = utils.roc_auc_score(y_true=y_true, y_score=y_score)
print(f'AUC computation took: {time() - before_auc_timestamp:.2f} [s]')
model.train()
return metric_logger.loss.global_avg, auc, metric_logger.step_time.avg
def dist_evaluate(model, data_loader):
"""Test distributed DLRM model
Args:
model (DistDLRM):
data_loader (torch.utils.data.DataLoader):
"""
model.eval()
device = FLAGS.base_device
world_size = dist.get_world_size()
batch_sizes_per_gpu = [
FLAGS.test_batch_size // world_size for _ in range(world_size)
]
test_batch_size = sum(batch_sizes_per_gpu)
if FLAGS.test_batch_size != test_batch_size:
print(f"Rounded test_batch_size to {test_batch_size}")
print(f"Batch sizes per GPU {batch_sizes_per_gpu}")
# Test bach size could be big, make sure it prints
default_print_freq = max(524288 * 100 // test_batch_size, 1)
print_freq = default_print_freq if FLAGS.print_freq is None else FLAGS.print_freq
steps_per_epoch = len(data_loader)
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter(
'step_time', utils.SmoothedValue(window_size=1, fmt='{avg:.4f}'))
with torch.no_grad():
timer = utils.StepTimer()
# ROC can be computed per batch and then compute AUC globally, but I don't have the code.
# So pack all the outputs and labels together to compute AUC. y_true and y_score naming follows sklearn
y_true = []
y_score = []
data_stream = torch.cuda.Stream()
batch_iter = prefetcher(iter(data_loader), data_stream)
timer.click()
for step in range(len(data_loader)):
numerical_features, categorical_features, click = next(batch_iter)
torch.cuda.synchronize()
last_batch_size = None
if click.shape[0] != test_batch_size: # last batch
last_batch_size = click.shape[0]
logging.warning("Pad the last test batch of size %d to %d",
last_batch_size, test_batch_size)
padding_size = test_batch_size - last_batch_size
if numerical_features is not None:
padding_numerical = torch.empty(
padding_size,
numerical_features.shape[1],
device=numerical_features.device,
dtype=numerical_features.dtype)
numerical_features = torch.cat(
(numerical_features, padding_numerical), 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)
output = model(numerical_features, categorical_features,
batch_sizes_per_gpu).squeeze()
output_receive_buffer = torch.empty(test_batch_size, device=device)
torch.distributed.all_gather(
list(output_receive_buffer.split(batch_sizes_per_gpu)), output)
if last_batch_size is not None:
output_receive_buffer = output_receive_buffer[:last_batch_size]
if FLAGS.auc_device == "CPU":
click = click.cpu()
output_receive_buffer = output_receive_buffer.cpu()
y_true.append(click)
y_score.append(output_receive_buffer)
timer.click()
if timer.measured is not None:
metric_logger.update(step_time=timer.measured)
if step % print_freq == 0 and step > 0:
metric_logger.print(
header=f"Test: [{step}/{steps_per_epoch}]")
if is_main_process():
auc = utils.roc_auc_score(
torch.cat(y_true), torch.sigmoid(torch.cat(y_score).float()))
else:
auc = None
torch.distributed.barrier()
model.train()
return auc
def inference_benchmark_graphed(model, data_loader, num_batches=100):
model.eval()
base_device = FLAGS.base_device
latencies = []
data_iter = iter(data_loader)
numerical, categorical, _ = next(data_iter)
# Warmup before capture
s = torch.cuda.Stream()
static_numerical = numerical.to(base_device)
static_categorical = categorical.to(device=base_device, dtype=torch.int64)
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
for i in range(10):
if FLAGS.amp:
numerical = static_numerical.half()
else:
numerical = static_numerical
inference_result = model(numerical, static_categorical).squeeze()
torch.cuda.synchronize()
# Graph capture
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph):
if FLAGS.amp:
numerical = static_numerical.half()
else:
numerical = static_numerical
inference_result = model(numerical, static_categorical).squeeze()
torch.cuda.synchronize()
# Inference
y_true = []
y_score = []
with torch.no_grad():
for step, (numerical_features, categorical_features, click) in enumerate(data_loader):
if step > num_batches:
break
torch.cuda.synchronize()
step_start_time = time()
numerical_features = numerical_features.to(base_device)
categorical_features = categorical_features.to(device=base_device, dtype=torch.int64)
static_categorical.copy_(categorical_features)
static_numerical.copy_(numerical_features)
graph.replay()
torch.cuda.synchronize()
step_time = time() - step_start_time
if step >= FLAGS.benchmark_warmup_steps:
latencies.append(step_time)
y_true.append(click)
y_score.append(inference_result.reshape([-1]).clone())
y_true = torch.cat(y_true)
y_score = torch.sigmoid(torch.cat(y_score)).float()
auc = utils.roc_auc_score(y_true, y_score)
print('auc: ', auc)
return latencies
