Skip to content

Fix int16 overflow in the gather_qmm sorted row bound #4010

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
erwinzhang7 wants to merge 1 commit into from
Closed

Fix int16 overflow in the gather_qmm sorted row bound #4010

Show file tree
Hide file tree
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
8 changes: 4 additions & 4 deletions mlx/backend/metal/kernels/quantized_nax.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1529,10 +1529,10 @@ template <
const short tm = SM * (simd_group_id / WN);
const short tn = SN * (simd_group_id % WN);

const short sgp_sm =
align_M ? SM : min(SM, short(max(0, (M - (y_row + tm)))));
const short sgp_sn =
align_N ? SN : min(SN, short(max(0, (N - (y_col + tn)))));
// Take the min in int and narrow afterwards. Narrowing first overflows for
// M or N above the int16 range, which wraps the bound negative.
const short sgp_sm = align_M ? SM : min(int(SM), max(0, (M - (y_row + tm))));
const short sgp_sn = align_N ? SN : min(int(SN), max(0, (N - (y_col + tn))));

const bool is_unaligned_sm = align_M ? false : (sgp_sm != SM);
const bool is_unaligned_bn = align_N ? false : (tgp_bn != BN);
Expand Down
30 changes: 30 additions & 0 deletions python/tests/test_quantized.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1403,6 +1403,36 @@ def scatter_unsort(x, inv_order, shape=None):
self.assertTrue(mx.allclose(y1, y3, atol=tol))
self.assertTrue(mx.allclose(y1, y4, atol=tol))

def test_gather_qmm_sorted_large_m(self):
# The sorted path bounded the per-simdgroup row count by narrowing to
# short before taking the min, which wraps negative once the row count
# passes the int16 range. The guard only runs when M % BM != 0, so both
# a large M and an unaligned M are needed to reach it. In a MoE this is
# tokens * experts_per_token, which passes 32767 at long context.
E, N, K = 2, 128, 256
key = mx.random.key(0)
k1, k2, k3 = mx.random.split(key, 3)
on_gpu = mx.default_device() == mx.gpu
dtype = mx.float16 if on_gpu else mx.float32

for M in (32704, 32800, 40001):
with self.subTest(M=M):
indices = mx.sort(
(mx.random.uniform(shape=(M,), key=k1) * E).astype(mx.uint32)
)
x = (mx.random.normal((M, 1, K), key=k2) / K**0.5).astype(dtype)
w = (mx.random.normal((E, N, K), key=k3) / K**0.5).astype(dtype)
wq = mx.quantize(w, group_size=64, bits=8)

kwargs = dict(
group_size=64, bits=8, transpose=True, rhs_indices=indices
)
y_sorted = mx.gather_qmm(x, *wq, sorted_indices=True, **kwargs)
y_unsorted = mx.gather_qmm(x, *wq, sorted_indices=False, **kwargs)

tol = 1e-3 if on_gpu else 1.5e-5
self.assertLess((y_sorted - y_unsorted).abs().max(), tol)

def test_gather_qmm_grad(self):
def gather_qmm_ref(x, w, s, b, lhs, rhs, trans, sort):
if lhs is not None:
Expand Down