Skip to content

Fix CUDA RMSNorm small-row dispatch #3792

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
dhiltgen wants to merge 2 commits into
base: main
Choose a base branch
from
+15 −13
Open

Fix CUDA RMSNorm small-row dispatch #3792

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e486ac4
Fix CUDA RMSNorm small-row dispatch
dhiltgen Jul 3, 2026
98323e8
Simplify tests
zcbenz Jul 3, 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
4 changes: 3 additions & 1 deletion mlx/backend/cuda/rms_norm.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -298,7 +298,7 @@ void dispatch_group_dim(int axis_size, F&& f) {
} else if (axis_size <= n_per_thread * 32 * 2) {
f(std::integral_constant<int, 32>{},
std::integral_constant<int, 2>(),
std::integral_constant<int, 2>());
std::integral_constant<int, 1>());
} else if (axis_size <= n_per_thread * 32 * 4) {
f(std::integral_constant<int, 32>{},
std::integral_constant<int, 4>(),
Expand Down Expand Up @@ -369,6 +369,7 @@ void RMSNorm::eval_gpu(
dispatch_group_dim<N_READS>(
axis_size, [&](auto group_dim, auto n_groups, auto groups_per_block) {
constexpr int block_dim = n_groups() * group_dim();
static_assert(block_dim <= 32 || groups_per_block() == 1);
auto kernel =
cu::rms_norm_small<DataType, block_dim, group_dim(), N_READS>;
auto n_blocks =
Expand Down Expand Up @@ -479,6 +480,7 @@ void RMSNormVJP::eval_gpu(
axis_size,
[&](auto group_dim, auto n_groups, auto groups_per_block) {
constexpr int block_dim = group_dim() * n_groups();
static_assert(block_dim <= 32 || groups_per_block() == 1);
auto kernel = cu::rms_norm_vjp_small<
DataType,
has_w_constant.value,
Expand Down
24 changes: 12 additions & 12 deletions python/tests/test_fast.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -428,7 +428,7 @@ def test_rms_norm(self):

dtypes = [mx.float32, mx.float16, mx.bfloat16]
epss = [1e-3, 1e-5]
dimss = [31, 32, 33]
dimss = [31, 32, 33, 256, 512]
defaults = (mx.float32, 1e-5, 32)

for dtype in dtypes:
Expand Down Expand Up @@ -483,23 +483,23 @@ def test_rms_norm(self):
mx.fast.rms_norm(x, mx.ones((4,)), 1e-5)

def test_rms_norm_grad(self):
D = 32
eps = 1e-5
f1 = lambda x, w, y: (rms_norm(x, w, eps) * y).sum()
f2 = lambda x, w, y: (mx.fast.rms_norm(x, w, eps) * y).sum()
f3 = lambda x, y: (rms_norm(x, mx.ones((x.shape[-1],)), eps) * y).sum()
f4 = lambda x, y: (mx.fast.rms_norm(x, None, eps) * y).sum()

x = mx.random.uniform(shape=(8, 100, D))
w = mx.random.uniform(shape=(D,))
y = mx.random.uniform(shape=(8, 100, D))
gx1, gw1 = mx.grad(f1, argnums=(0, 1))(x, w, y)
gx2, gw2 = mx.grad(f2, argnums=(0, 1))(x, w, y)
self.assertLess(mx.abs(gx1 - gx2).max(), 1e-5)
self.assertLess(mx.abs(gw1 - gw2).max() / mx.abs(gw1).mean(), 1e-5)
gx1 = mx.grad(f3, argnums=(0,))(x, y)
gx2 = mx.grad(f4, argnums=(0,))(x, y)
self.assertLess(mx.abs(gx1 - gx2).max(), 1e-5)
for D in [32, 256]:
x = mx.random.uniform(shape=(8, 100, D))
w = mx.random.uniform(shape=(D,))
y = mx.random.uniform(shape=(8, 100, D))
gx1, gw1 = mx.grad(f1, argnums=(0, 1))(x, w, y)
gx2, gw2 = mx.grad(f2, argnums=(0, 1))(x, w, y)
self.assertLess(mx.abs(gx1 - gx2).max(), 1e-5)
self.assertLess(mx.abs(gw1 - gw2).max() / mx.abs(gw1).mean(), 1e-5)
gx1 = mx.grad(f3, argnums=(0,))(x, y)
gx2 = mx.grad(f4, argnums=(0,))(x, y)
self.assertLess(mx.abs(gx1 - gx2).max(), 1e-5)

D = 8192
x = mx.random.uniform(shape=(2, 2, D))
Expand Down
Loading