import sys
import atheris
import torch
import torch_npu
import random
with atheris.instrument_imports():
from apex.optimizers import NpuFusedBertAdam
@atheris.instrument_func
def Test_NpuFusedBertAdam(input_bytes):
input_list = [True, False]
Schedules = ['warmup_cosine','warmup_constant','warmup_linear','warmup_poly']
check_list_input = input_bytes.decode('utf-8', 'ignore').strip().split(',')
if not check_list_input or len(check_list_input) != 10:
return False
try:
for i in range(10):
check_list_input[i] = float(check_list_input[i])
except Exception as e:
return False
lr = random.random() * check_list_input[1]
warmup = random.random() * check_list_input[2]
t_total = random.random() * check_list_input[3]
b1 = random.random() * check_list_input[4]
b2 = random.random() * check_list_input[5]
e = check_list_input[6]
weight_decay = check_list_input[7]
max_grad_norm = check_list_input[8]
number_of_params = int(check_list_input[9])
schedule = Schedules[random.randint(0, 3)]
try:
params = []
for i in range(number_of_params):
input_tensor_size = [random.randint(0, check_list_input[0]) for _ in range(random.randint(0, 5))]
if input_list[random.randint(0, 1)]:
input_tensor = torch.randn(input_tensor_size).float().npu()
else:
input_tensor = torch.randn(input_tensor_size).half().npu()
params.append(input_tensor)
for i, p in enumerate(params):
if i < len(params) - 1:
p.requires_grad = True
p.grad = p.clone().detach() / 100
NpuFusedBertAdam(params, lr, warmup, t_total, schedule, b1, b2, e, weight_decay, max_grad_norm)
except Exception as e:
print(e)
return True
if __name__ == "__main__":
atheris.Setup(sys.argv, Test_NpuFusedBertAdam)
atheris.Fuzz()