Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio MLX convert: pack 3 / 5 / 6-bit codes from Fortran-order or strided arrays (#17082)

Daniel Han750d728b91a6
diff --git a/studio/backend/core/inference/mlx_diffusion_convert.py b/studio/backend/core/inference/mlx_diffusion_convert.py
index c73047d62a3..317938671b7 100644
--- a/studio/backend/core/inference/mlx_diffusion_convert.py
+++ b/studio/backend/core/inference/mlx_diffusion_convert.py
@@ -139,7 +139,9 @@ def _pack(q: np.ndarray, bits: int) -> np.ndarray:
# 8 elements = `bits` bytes: the low bytes of a little-endian uint64.
for start in range(0, rows, _ROW_CHUNK):
lanes = q[start : start + _ROW_CHUNK].reshape(-1, cols // 8, 8)
- acc = lanes[..., 0].astype("<u8")
+ acc = lanes[..., 0].astype(
+ "<u8", order = "C"
+ ) # C order: the byte view below needs a contiguous last axis
for k in range(1, 8):
acc |= lanes[..., k].astype("<u8") << np.uint64(bits * k)
stream[start : start + _ROW_CHUNK] = (
diff --git a/studio/backend/tests/test_mlx_diffusion_convert.py b/studio/backend/tests/test_mlx_diffusion_convert.py
index 0a0a15a17ee..36f9f01a838 100644
--- a/studio/backend/tests/test_mlx_diffusion_convert.py
+++ b/studio/backend/tests/test_mlx_diffusion_convert.py
@@ -187,6 +187,14 @@ def test_pack_is_byte_identical_to_the_reference(bits, dtype):
assert conv._pack(edge, bits).tobytes() == _pack_reference(edge, bits).tobytes()
[email protected]("bits", [2, 3, 4, 5, 6, 8])
+def test_pack_takes_fortran_order_and_strided_inputs(bits):
+ q = _rng(bits).integers(0, 1 << bits, size = (300, 64), dtype = np.uint8)
+ ref = _pack_reference(q, bits).tobytes()
+ assert conv._pack(np.asfortranarray(q), bits).tobytes() == ref
+ assert conv._pack(np.ascontiguousarray(q.T).T, bits).tobytes() == ref
+
+
@pytest.mark.parametrize("bits", [4, 8])
def test_pack_does_not_widen_to_uint32_per_element(bits):
import tracemalloc