mirror of
https://github.com/ROCm/ROCm.git
synced 2026-04-05 03:01:17 -04:00
5231d57c71
Prior to this PR, matmul on sm_89 (RTX 4070) (`test/unit/operators/test_matmul.py::test_op`) would result in test failure due to too strict atol/rtol. To avoid having to choose strictness ourselves, and to have better defaults based on dtype, use the non-deprecated torch testing util. See: https://github.com/pytorch/pytorch/issues/61844 Replace: https://github.com/openai/triton/pull/2242
41 lines
1.4 KiB
Python
41 lines
1.4 KiB
Python
import pytest
|
|
import torch
|
|
|
|
import triton
|
|
import triton.ops
|
|
|
|
|
|
@pytest.mark.parametrize("M, N, dtype, mode",
|
|
[
|
|
(M, N, dtype, mode) for M in [1024, 821]
|
|
for N in [512, 857, 1871, 2089, 8573, 31000]
|
|
for dtype in ['float16', 'float32']
|
|
for mode in ['forward', 'backward']
|
|
]
|
|
)
|
|
def test_op(M, N, dtype, mode):
|
|
capability = torch.cuda.get_device_capability()
|
|
if capability[0] < 8 and dtype == "bfloat16":
|
|
pytest.skip("Only test bfloat16 on devices with sm >= 80")
|
|
dtype = {'bfloat16': torch.bfloat16, 'float16': torch.float16, 'float32': torch.float32}[dtype]
|
|
# create inputs
|
|
x = torch.randn(M, N, dtype=dtype, device='cuda', requires_grad=True)
|
|
idx = 4 + torch.ones(M, dtype=torch.int64, device='cuda')
|
|
# forward pass
|
|
tt_y = triton.ops.cross_entropy(x, idx)
|
|
th_y = torch.nn.CrossEntropyLoss(reduction="none")(x, idx)
|
|
if mode == 'forward':
|
|
torch.testing.assert_close(th_y, tt_y)
|
|
# backward pass
|
|
elif mode == 'backward':
|
|
dy = torch.randn_like(tt_y)
|
|
# triton backward
|
|
tt_y.backward(dy)
|
|
tt_dx = x.grad.clone()
|
|
# torch backward
|
|
x.grad = None
|
|
th_y.backward(dy)
|
|
th_dx = x.grad.clone()
|
|
|
|
torch.testing.assert_close(th_dx, tt_dx)
|