24 lines
636 B
Python
24 lines
636 B
Python
|
import torch
|
||
|
|
|
||
|
|
import triton
|
||
|
|
import triton.language as tl
|
||
|
|
|
||
|
|
# trigger the torch.device implicitly to ensure cuda context initialization
|
||
|
|
torch.zeros([10], device=torch.device('cuda'))
|
||
|
|
|
||
|
|
|
||
|
|
def test_empty_kernel_cubin_compile():
|
||
|
|
@triton.jit
|
||
|
|
def kernel(X, stride_xm, stride_xn, BLOCK: tl.constexpr):
|
||
|
|
pass
|
||
|
|
|
||
|
|
device = torch.cuda.current_device()
|
||
|
|
cubin = triton.compile(kernel,
|
||
|
|
"*fp32,i32,i32",
|
||
|
|
device=device,
|
||
|
|
constants={"BLOCK": 256},
|
||
|
|
output="cubin")
|
||
|
|
|
||
|
|
print('cubin size:', len(cubin))
|
||
|
|
assert len(cubin) > 0
|