forked from EngineX-Cambricon/enginex-mlu370-vllm
75 lines
4.1 KiB
Python
75 lines
4.1 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 = [{"batch": 1, "seq_len": 1024, "input_size": 1600, "hidden_size": 1600,
|
|
"has_residual": True, "has_bias": True, "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"batch": 1, "seq_len": 1024, "input_size": 2048, "hidden_size": 2048,
|
|
"has_residual": True, "has_bias": True, "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"batch": 1, "seq_len": 1024, "input_size": 4096, "hidden_size": 4096,
|
|
"has_residual": True, "has_bias": True, "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"batch": 1, "seq_len": 1024, "input_size": 6144, "hidden_size": 6144,
|
|
"has_residual": True, "has_bias": True, "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"batch": 1, "seq_len": 1024, "input_size": 6656, "hidden_size": 6656,
|
|
"has_residual": True, "has_bias": True, "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"batch": 1, "seq_len": 1024, "input_size": 8192, "hidden_size": 8192,
|
|
"has_residual": True, "has_bias": True, "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"batch": 1, "seq_len": 1024, "input_size": 12288, "hidden_size": 12288,
|
|
"has_residual": True, "has_bias": True, "input_dtype": [torch.float16, torch.bfloat16]},
|
|
{"batch": 1, "seq_len": 1024, "input_size": 14336, "hidden_size": 14336,
|
|
"has_residual": True, "has_bias": True, "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 = ["batch", "seq_len", "input_size", "hidden_size", "has_residual", "has_bias", "input_dtype", "hardware_time(us)", "e2e_latency(us)"]
|
|
contents = []
|
|
for params_dict in e2e_time_param_dict_list:
|
|
batch = params_dict["batch"]
|
|
seq_len = params_dict["seq_len"]
|
|
input_size = params_dict["input_size"]
|
|
hidden_size = params_dict["hidden_size"]
|
|
has_residual = params_dict["has_residual"]
|
|
has_bias = params_dict["has_bias"]
|
|
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
|
|
x = torch.randn(batch, seq_len, hidden_size).to(dtype).to(device)
|
|
weight = torch.randn(hidden_size, input_size).to(dtype).to(device)
|
|
residual, bias = None, None
|
|
if has_residual:
|
|
residual = torch.randn(batch, seq_len, hidden_size).to(dtype).to(device)
|
|
if has_bias:
|
|
bias = torch.randn(hidden_size).to(dtype).to(device)
|
|
hardware_time, e2e_time = benchmark_forward(tmo.attention_project,
|
|
x,
|
|
weight,
|
|
bias,
|
|
residual,
|
|
1.0,
|
|
1.0,
|
|
repeats=args.repeat_times)
|
|
content = [f"{batch}", f"{seq_len}", f"{input_size}", f"{hidden_size}", f"{has_residual}", f"{has_bias}", 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()
|