mirror of
https://github.com/dragonpilot/dragonpilot.git
synced 2026-07-23 13:52:05 +08:00
Vehicle Researcher
161005a0c1
date: 2025-11-19T19:10:11 master commit: a29fdbd02407d41ecbcc69d151bb4837bfba3cbc
18 lines
542 B
Python
18 lines
542 B
Python
from tinygrad import Tensor, dtypes, GlobalCounters
|
|
dtypes.default_float = dtypes.float16
|
|
from tinygrad.dtype import to_dtype
|
|
from tinygrad.helpers import getenv
|
|
from test.test_softmax_fusion import single_kernel_softmax
|
|
|
|
if __name__ == "__main__":
|
|
# softmax in bert layers
|
|
BS = getenv("BS", 96//6)
|
|
acc_dtype = to_dtype(getenv("ACC_DTYPE", "half"))
|
|
t = Tensor.empty(BS, 16, 512, 512)
|
|
t.softmax(-1, dtype="half").realize()
|
|
|
|
# test single kernel softmax
|
|
GlobalCounters.reset()
|
|
single_kernel_softmax(t, -1, acc_dtype).realize()
|
|
|