forked from EngineX-Cambricon/enginex-mlu370-vllm
47 lines
2.0 KiB
Python
47 lines
2.0 KiB
Python
import torch
|
|
import torch_mlu
|
|
import torch_mlu_ops as tmo
|
|
from common import benchmark_forward, save_to_csv
|
|
import argparse
|
|
from tabulate import tabulate
|
|
import os
|
|
|
|
e2e_time_param_dict_list = [{"input_shape": [100, 100, 100], "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"input_shape": [100, 100], "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"input_shape": [50, 50, 50], "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"input_shape": [1, 100, 1000], "input_dtype": [torch.float16, torch.bfloat16]}
|
|
]
|
|
def main():
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument('--repeat_times', type=int, default=10, help='repeat times for testing')
|
|
parser.add_argument('--csv', action='store_true', help='write the report data to csv')
|
|
parser.add_argument('-o', type=str, help='specify the output folder name under --csv mode')
|
|
|
|
args = parser.parse_args()
|
|
device = 'mlu'
|
|
titles = ["input_shape", "input_dtype", "hardware_time(us)", "e2e_latency(us)"]
|
|
contents = []
|
|
for params_dict in e2e_time_param_dict_list:
|
|
input_shape = params_dict["input_shape"]
|
|
input_dtype_list = params_dict["input_dtype"]
|
|
for dtype in input_dtype_list:
|
|
if dtype == torch.bfloat16 and not torch_mlu.mlu.is_bf16_supported():
|
|
continue
|
|
input = torch.randn(input_shape).to(device).to(dtype)
|
|
hardware_time, e2e_time = benchmark_forward(tmo.preload,
|
|
input,
|
|
input.element_size() * input.numel(),
|
|
repeats=args.repeat_times)
|
|
content = [f"{input_shape}", f"{dtype}", f"{hardware_time}", f"{e2e_time}"]
|
|
contents.append(content)
|
|
table = [titles] + contents
|
|
print(tabulate(table, headers="firstrow", tablefmt="grid"))
|
|
|
|
if args.csv:
|
|
current_file_path = __file__
|
|
_, file_name = os.path.split(current_file_path)
|
|
save_to_csv(table, args.o, file_name)
|
|
|
|
if __name__=="__main__":
|
|
main()
|