Skip to content

Add D256 to GQA-8 two-pass vector attention #4477

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

Closed
wyanzhao wants to merge 2 commits into from
+28 −1
Closed

Add D256 to GQA-8 two-pass vector attention #4477

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
7b03ec2
Add D256 to GQA-8 two-pass vector attention
wyanzhao Sep 8, 2026
288906a
Merge upstream/main and retain D256 GQA and D512 tests
wyanzhao Sep 8, 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
1 change: 1 addition & 0 deletions mlx/backend/metal/kernels/scaled_dot_product_attention.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,7 @@ using namespace metal;
instantiate_sdpa_vector(type, 512, 512) \
instantiate_sdpa_vector_gqa(type, 64, 64, 8, 8) \
instantiate_sdpa_vector_gqa(type, 128, 128, 8, 4) \
instantiate_sdpa_vector_gqa(type, 256, 256, 8, 2) \
instantiate_sdpa_vector_gqa(type, 128, 128, 12, 4) \
instantiate_sdpa_vector_gqa(type, 128, 128, 16, 2) \
instantiate_sdpa_vector_aggregation(type, 64) \
Expand Down
3 changes: 2 additions & 1 deletion mlx/backend/metal/scaled_dot_product_attention.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -538,7 +538,8 @@ void sdpa_vector_2pass(
kname += "sdpa_vector_2pass_1";
int gqa_factor = q.shape(1) / k.shape(1);
bool gqa_dims =
(gqa_factor == 8 && (q.shape(-1) == 64 || q.shape(-1) == 128)) ||
(gqa_factor == 8 &&
(q.shape(-1) == 64 || q.shape(-1) == 128 || q.shape(-1) == 256)) ||
((gqa_factor == 12 || gqa_factor == 16) && q.shape(-1) == 128);
if (!mask && !sinks && q.shape(2) == 1 &&
q.shape(1) == gqa_factor * k.shape(1) && q.shape(-1) == v.shape(-1) &&
Expand Down
25 changes: 25 additions & 0 deletions python/tests/test_fast_sdpa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -410,6 +410,31 @@ def test_sdpa_vector_gqa_long(self):
out = mx.fast.scaled_dot_product_attention(q, k, v, scale=scale)
self.assertTrue(mx.allclose(ref, out, atol=1e-4, rtol=1e-4))

@unittest.skipUnless(mx.metal.is_available(), "Metal is not available")
def test_sdpa_vector_gqa_d256(self):

@RohanGautam RohanGautam Sep 12, 2026 •
edited
Loading

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Same concern i had in #4476 , this mainly tests numerics and not necessarily the changes you introduced. For your changes, good benchmarking should be sufficient. But let me know if you disagree!

if mx.default_device() != mx.gpu:
self.skipTest("requires GPU")
mx.random.seed(0)
for dtype, atol in [
(mx.float32, 1e-4),
(mx.float16, 1e-3),
(mx.bfloat16, 5e-3),
]:
for B, length in [(1, 8192), (1, 8201), (2, 8192), (2, 32768)]:
with self.subTest(dtype=dtype, B=B, length=length):
q = mx.random.normal((B, 16, 1, 256)).astype(dtype)
k = mx.random.normal((B, 2, length + 32, 256)).astype(dtype)
v = mx.random.normal((B, 2, length + 32, 256)).astype(dtype)
k, v = k[:, :, :length], v[:, :, :length]
ref = mlx_primitives_sdpa(
q.astype(mx.float32),
mx.repeat(k.astype(mx.float32), 8, axis=1),
mx.repeat(v.astype(mx.float32), 8, axis=1),
256**-0.5,
)
out = mx.fast.scaled_dot_product_attention(q, k, v, scale=256**-0.5)
self.assertTrue(mx.allclose(ref, out, atol=atol, rtol=1e-3))

@unittest.skipIf(not mx.metal.is_available(), "Metal kernel path only")
def test_sdpa_vector_head_dim_512(self):
if mx.default_device() != mx.gpu:
Expand Down
Loading