J3ff/IQ.Pilot
1
0
Fork 0
forked from IQ.Lvbs/IQ.Pilot
Code Pull Requests Activity

IQ.Pilot Prebuilt Release @ ab07000

Browse Source
This commit is contained in:
IQ.Lvbs history cleanup
2026-08-22 23:42:42 -05:00
commit 9f9c9a70cc
3729 changed files with 778697 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
tinygrad_repo/test/external/external_benchmark_bert_matmuls.py vendored Normal file
View File

@@ -0,0 +1,18 @@
from tinygrad import Tensor, dtypes
dtypes.default_float = dtypes.float16
from tinygrad.dtype import to_dtype
from tinygrad.helpers import getenv
if __name__ == "__main__":
# matmuls in bert layers
BS = getenv("BS", 96//6)
acc_dtype = to_dtype(getenv("ACC_DTYPE", "half"))
tensors = [
(Tensor.empty(BS, 512, 1024), Tensor.empty(1024, 1024).T), # linear to get qkv
(Tensor.empty(BS, 512, 16, 64).permute(0,2,1,3), Tensor.empty(BS, 512, 16, 64).permute(0,2,3,1)), # q@k
(Tensor.empty(BS, 16, 512, 512), Tensor.empty(BS, 512, 16, 64).permute(0,2,1,3)), # qk@v
]
for t0, t1 in tensors:
print(f"{t0.shape=}, {t0.uop.st.is_expanded()=}, {t1.shape=}, {t1.uop.st.is_expanded()=}")
for _ in range(5):
t0.dot(t1, dtype=acc_dtype).realize()