Skip to content

Fix quantizing sliced arrays #4381

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

Merged
zcbenz merged 1 commit into from
Aug 24, 2026
+17 −5
Merged

Fix quantizing sliced arrays #4381

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
10 changes: 5 additions & 5 deletions mlx/backend/metal/quantized.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -249,28 +249,28 @@ void quantize_impl(
auto w = ensure_row_contiguous(w_pre, d, s);
if (dequantize) {
auto scales = ensure_row_contiguous(inputs[1], d, s);
compute_encoder.set_input_array(w, 0);
compute_encoder.set_input_array(scales, 1);
if (has_biases) {
auto biases = ensure_row_contiguous(inputs[2], d, s);
compute_encoder.set_input_array(biases, 2);
} else if (has_global_scale) {
compute_encoder.set_input_array(inputs[2], 2);
}
compute_encoder.set_input_array(w, 0);
compute_encoder.set_input_array(scales, 1);
compute_encoder.set_output_array(out, 3);
} else {
auto& scales = outputs[1];
scales.set_data(allocator::malloc(scales.nbytes()));
compute_encoder.set_input_array(w, 0);
compute_encoder.set_output_array(out, 1);
compute_encoder.set_output_array(scales, 2);
if (has_biases) {
auto& biases = outputs[2];
biases.set_data(allocator::malloc(biases.nbytes()));
compute_encoder.set_output_array(biases, 3);
} else if (has_global_scale) {
compute_encoder.set_input_array(inputs[1], 3);
}
compute_encoder.set_input_array(w, 0);
compute_encoder.set_output_array(out, 1);
compute_encoder.set_output_array(scales, 2);
}

auto type_string = dequantize ? get_type_string(out.dtype())
Expand Down
12 changes: 12 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 @@ -39,6 +39,18 @@ def test_quantize_dequantize(self):
a_hat = mx.dequantize(w_q, scales, biases, gs, b)
self.assertTrue(mx.all(a_hat == 0))

# slices
if mx.default_device() == mx.gpu:
w = mx.random.normal(shape=(2, 256, 32))
quant = {"group_size": 32, "bits": 4}
wq, scales, biases = mx.quantize(w, **quant)
wq_s = wq[:, :16, :]
scales_s = scales[:, :16, :]
biases_s = biases[:, :16, :]
dq_cpu = mx.dequantize(wq_s, scales_s, biases_s, **quant, stream=mx.cpu)
dq_gpu = mx.dequantize(wq_s, scales_s, biases_s, **quant, stream=mx.gpu)
self.assertTrue(mx.abs(dq_cpu - dq_gpu).max().item() < 1e-6)

def test_mxfp4_quantize_dequantize(self):
lut = mx.array(
[
Expand Down
Loading