adam_test1.py

'''Copyright The Microsoft DeepSpeed Team'''

import torch
from deepspeed.ops.adam import DeepSpeedCPUAdam
import time
from deepspeed.accelerator import get_accelerator

device = 'cpu'
model_size = 1 * 1024**3
param = torch.nn.Parameter(torch.ones(model_size, device=device))
param_fp16 = torch.nn.Parameter(
    torch.ones(model_size,
               dtype=torch.half,
               device=get_accelerator().device_name(0)))

optimizer = DeepSpeedCPUAdam([param])
#torch.set_num_threads(128)
param.grad = torch.ones(model_size, device=device)
avg = 0
for i in range(100):
    start = time.time()
    optimizer.step(fp16_param_groups=[param_fp16])
    stop = time.time()
    avg += (stop - start)
    param.grad = torch.ones(model_size, device=device) * 2
print("Elapsed Time is ", avg / 100)