forked from EngineX-Cambricon/enginex-mlu370-vllm
70 lines
2.9 KiB
Python
70 lines
2.9 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 = [{"m": 1024, "k": 4096, "n": 14336, "has_c": True, "has_bias": True,
|
|
"act_mode": "none", "output_dtype": [torch.float16, torch.bfloat16]},
|
|
{"m": 1024, "k": 5120, "n": 13824, "has_c": False, "has_bias": True,
|
|
"act_mode": "silu", "output_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 = ["m", "k", "n", "has_c", "has_bias", "act_mode", "output_dtype", "hardware_time(us)", "e2e_latency(us)"]
|
|
contents = []
|
|
for params_dict in e2e_time_param_dict_list:
|
|
m = params_dict["m"]
|
|
k = params_dict["k"]
|
|
n = params_dict["n"]
|
|
has_c = params_dict["has_c"]
|
|
has_bias = params_dict["has_bias"]
|
|
act_mode = params_dict["act_mode"]
|
|
output_dtype_list = params_dict["output_dtype"]
|
|
for dtype in output_dtype_list:
|
|
if dtype == torch.bfloat16 and not torch_mlu.mlu.is_bf16_supported():
|
|
continue
|
|
a = torch.randn(m, k).to(device).to(torch.int8)
|
|
b = torch.randn(n, k).to(device).to(torch.int8)
|
|
a_scale = torch.randn(m).to(device)
|
|
b_scale = torch.randn(n).to(device)
|
|
c = None
|
|
if has_c:
|
|
c = torch.randn(m, n).to(device).to(dtype)
|
|
bias = None
|
|
if has_bias:
|
|
bias = torch.randn(n).to(device).to(dtype)
|
|
hardware_time, e2e_time = benchmark_forward(tmo.smooth_quant_matmul,
|
|
a,
|
|
a_scale,
|
|
b,
|
|
b_scale,
|
|
dtype,
|
|
bias,
|
|
c,
|
|
act_mode,
|
|
1.0,
|
|
1.0,
|
|
repeats=args.repeat_times)
|
|
content = [f"{m}", f"{k}", f"{n}", f"{has_c}", f"{has_bias}", f"{act_mode}", 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()
|