Skip to content

Fix linear and matmul accuracy tests #66

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
factnn merged 1 commit into from
Aug 12, 2026
Merged

Fix linear and matmul accuracy tests #66

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
67 changes: 59 additions & 8 deletions src/kernelgenbench/accuracy/test_ops_with_benchmark.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9797,9 +9797,38 @@ def test_accuracy_le_scalar(shape, dtype):


# ========== linear ==========
@label("baddbmm")
@label("linear")
@label("matmul")
@parametrize("M, N, K", MNK_SHAPES)
@parametrize("dtype", FLOAT_DTYPES)
@parametrize("has_bias", [True, False])
def test_accuracy_linear(M, N, K, dtype, has_bias):
inp = torch.randn((M, K), dtype=dtype, device=device)
weight = torch.randn((N, K), dtype=dtype, device=device)
bias = torch.randn((N,), dtype=dtype, device=device) if has_bias else None
ref_inp = to_reference(inp, True)
ref_weight = to_reference(weight, True)
ref_bias = to_reference(bias, True) if bias is not None else None

ref_out = torch.ops.aten.linear(ref_inp, ref_weight, ref_bias)
with kernelgenbench.use_ops(REGISTERED_OPS):
res_out = torch.ops.aten.linear(inp, weight, bias)
kernelgenbench_assert_close(res_out, ref_out, dtype, reduce_dim=K)

from sandbox.utils.accuracy_utils import CustomBenchmarkResult
quantiles = [0.5, 0.2, 0.8]
ms_torch, _, _ = get_triton_testing().do_bench(
lambda: torch.ops.aten.linear(ref_inp, ref_weight, ref_bias), rep=100, quantiles=quantiles
)
with kernelgenbench.use_ops(REGISTERED_OPS):
ms_triton, _, _ = get_triton_testing().do_bench(
lambda: torch.ops.aten.linear(inp, weight, bias), rep=100, quantiles=quantiles
)
return CustomBenchmarkResult(
ref_time=ms_torch, res_time=ms_triton, speedup=ms_torch / ms_triton
)


@label("baddbmm")
@parametrize("M, N, K", MNK_SHAPES)
@parametrize("scalar", SCALARS)
@parametrize("dtype", FLOAT_DTYPES)
Expand Down Expand Up @@ -9839,8 +9868,6 @@ def test_accuracy_baddbmm(M, N, K, scalar, dtype):

# ========== linear ==========
@label("baddbmm_backward")
@label("linear")
@label("matmul")
@parametrize("M, N, K", MNK_SHAPES)
@parametrize("scalar", SCALARS)
@parametrize("dtype", FLOAT_DTYPES)
Expand Down Expand Up @@ -9955,10 +9982,36 @@ def test_accuracy_masked_fill_(shape, dtype, threshold, value):


# ========== matmul ==========
@label("baddbmm")
@label("linear")
@label("matmul")
@parametrize("M, N, K", MNK_SHAPES)
@parametrize("dtype", FLOAT_DTYPES)
def test_accuracy_matmul(M, N, K, dtype):
inp = torch.randn((M, K), dtype=dtype, device=device)
other = torch.randn((K, N), dtype=dtype, device=device)
ref_inp = to_reference(inp, True)
ref_other = to_reference(other, True)

ref_out = torch.ops.aten.matmul(ref_inp, ref_other)
with kernelgenbench.use_ops(REGISTERED_OPS):
res_out = torch.ops.aten.matmul(inp, other)
kernelgenbench_assert_close(res_out, ref_out, dtype, reduce_dim=K)

from sandbox.utils.accuracy_utils import CustomBenchmarkResult
quantiles = [0.5, 0.2, 0.8]
ms_torch, _, _ = get_triton_testing().do_bench(
lambda: torch.ops.aten.matmul(ref_inp, ref_other), rep=100, quantiles=quantiles
)
with kernelgenbench.use_ops(REGISTERED_OPS):
ms_triton, _, _ = get_triton_testing().do_bench(
lambda: torch.ops.aten.matmul(inp, other), rep=100, quantiles=quantiles
)
return CustomBenchmarkResult(
ref_time=ms_torch, res_time=ms_triton, speedup=ms_torch / ms_triton
)


@label("baddbmm")
@parametrize("M, N, K", MNK_SHAPES)
@parametrize("scalar", SCALARS)
@parametrize("dtype", FLOAT_DTYPES)
def test_accuracy_baddbmm(M, N, K, scalar, dtype):
Expand Down Expand Up @@ -9997,8 +10050,6 @@ def test_accuracy_baddbmm(M, N, K, scalar, dtype):

# ========== matmul ==========
@label("baddbmm_backward")
@label("linear")
@label("matmul")
@parametrize("M, N, K", MNK_SHAPES)
@parametrize("scalar", SCALARS)
@parametrize("dtype", FLOAT_DTYPES)
Expand Down
Loading