Skip to content

[CI/Build] Updated rmsnorm test to improve OOT device coverage #36246

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

Open
romitjain wants to merge 7 commits into
base: main
Choose a base branch
from
+16 −7
Open

[CI/Build] Updated rmsnorm test to improve OOT device coverage #36246

Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
320f9fe
Updated rmsnorm test
romitjain Mar 6, 2026
09ac3b2
Fixed unavail var
romitjain Mar 6, 2026
a655e21
Updated device
romitjain Mar 6, 2026
cf853d1
Fixed precommit checks
romitjain Mar 6, 2026
1e11e1f
Merge branch 'main' into romit/rmsnorm-test
romitjain Mar 13, 2026
a3b591a
Fixed lint
romitjain Mar 13, 2026
e6b5fa0
Merge branch 'main' into romit/rmsnorm-test
romitjain Mar 17, 2026
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
23 changes: 16 additions & 7 deletions tests/kernels/core/test_layernorm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,17 +14,19 @@
HIDDEN_SIZES = [8, 768, 769, 5120, 5125, 8192] # Arbitrary values for testing
ADD_RESIDUAL = [False, True]
SEEDS = [0]
CUDA_DEVICES = [
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
]
DEVICES = ["cpu"]
if torch.cuda.is_available():
DEVICES = [
f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2)
]


@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
@pytest.mark.parametrize("add_residual", ADD_RESIDUAL)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
@pytest.mark.parametrize("device", CUDA_DEVICES)
@pytest.mark.parametrize("device", DEVICES)
@pytest.mark.parametrize("strided_input", [False, True])
@torch.inference_mode()
def test_rms_norm(
Expand All @@ -46,12 +48,19 @@ def test_rms_norm(
x = torch.randn(num_tokens, last_dim, dtype=dtype)
x = x[..., :hidden_size]
assert x.is_contiguous() != strided_input
x *= scale
x = x * scale if add_residual else x
residual = torch.randn_like(x) * scale if add_residual else None

# NOTE(woosuk): The reference implementation should be executed first
# because the custom kernel is in-place.
ref_out = layer.forward_native(x, residual)
ref_out = layer.forward_static(
Copy link
Copy Markdown
Collaborator

@jikunshang jikunshang Mar 11, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why can't keep forward_native here?

Copy link
Copy Markdown
Author

@romitjain romitjain Mar 11, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@jikunshang We can, this was just an opiniated approach.
forward_native is directly calling the forward_static.

Just from the nomenclature I infered:

  • forward_native can be native to OOT platform
  • forward_static can remain the golden reference implementation - not to be inherited or over riden

x,
variance_epsilon=layer.variance_epsilon,
orig_dtype=x.dtype,
weight=layer.weight,
hidden_size=layer.hidden_size,
residual=residual,
)
out = layer(x, residual)
# NOTE(woosuk): LayerNorm operators (including RMS) typically have larger
# numerical errors than other operators because they involve reductions.
Expand Down Expand Up @@ -79,7 +88,7 @@ def test_rms_norm(
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("quant_scale", [0.01, 1.0, 10.0])
@pytest.mark.parametrize("seed", SEEDS)
@pytest.mark.parametrize("device", CUDA_DEVICES)
@pytest.mark.parametrize("device", DEVICES)
@pytest.mark.parametrize("strided_input", [False, True])
def test_fused_rms_norm_quant(
num_tokens: int,
Expand Down
Loading