Skip to content
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

Fix quantization of all 0s #1028

Merged
merged 2 commits into from
Apr 24, 2024
Merged

Fix quantization of all 0s #1028

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
a319ea8
Fix quantization of all 0s
angeloskath Apr 23, 2024
bfc186a
Fix the fix
angeloskath Apr 24, 2024
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
5 changes: 4 additions & 1 deletion mlx/ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3274,7 +3274,10 @@ std::tuple<array, array, array> quantize(
reshape(w, {w.shape(0), w.shape(1) / group_size, group_size}, s);
array w_max = max(packed_w, /* axis= */ -1, /* keepdims= */ true, s);
array w_min = min(packed_w, /* axis= */ -1, /* keepdims= */ true, s);
array delta = divide(subtract(w_max, w_min, s), array(n_bins, w.dtype()), s);
array delta = maximum(
divide(subtract(w_max, w_min, s), array(n_bins, w.dtype()), s),
array(1e-7, w.dtype()),
s);
array scales = squeeze(delta, -1, s);
array biases = squeeze(w_min, -1, s);

Expand Down
8 changes: 8 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 @@ -18,6 +18,14 @@ def test_quantize_dequantize(self):
eps = 1e-6
self.assertTrue((errors <= (scales[..., None] + eps)).all())

# test quantize/dequantize 0s
a = mx.zeros((256, 512))
for gs in [32, 64, 128]:
for b in [2, 4, 8]:
w_q, scales, biases = mx.quantize(a, gs, b)
a_hat = mx.dequantize(w_q, scales, biases, gs, b)
self.assertTrue(mx.all(a_hat == 0))

def test_qmm(self):
key = mx.random.key(0)
k1, k2 = mx.random.split(key)
Expand Down