Skip to content

Enable 5-bit support on MLX #20811

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
metascroy merged 1 commit into from
Jul 9, 2026
Merged

Enable 5-bit support on MLX #20811

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
6 changes: 3 additions & 3 deletions backends/mlx/builder/op_helpers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -669,10 +669,10 @@ def parse_dequant_node(
return None

# MLX supports 2,3,4,5,6,8-bit affine quantization. to_mlx_qparams packs
# 2/4/8 via fast paths and other widths (e.g. 6) via a general contiguous
# bit-packer, so enable 6 here too.
# 2/4/8 via fast paths and other widths (e.g. 5, 6) via a general
# contiguous bit-packer, so enable 5 and 6 here too.
bits = (qmax - qmin + 1).bit_length() - 1
if bits not in [2, 4, 6, 8]:
if bits not in [2, 4, 5, 6, 8]:
return None

@Gasoonjia Gasoonjia Jul 9, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

wondering what will be happend if return None? rasie error?

return qdata, scale, zero_point, group_size, bits, out_dtype, quantized_dim

Expand Down
12 changes: 6 additions & 6 deletions examples/models/gemma4_31b/quant/quantize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,8 +7,8 @@
"""Quantize weights to torchao tensor subclasses.

``quantize_weight`` quantizes a single tensor given a ``QuantConfig``,
returning an ``Int4Tensor`` (4-bit) or ``IntxUnpackedToInt8Tensor`` (6- or
8-bit).
returning an ``Int4Tensor`` (4-bit) or ``IntxUnpackedToInt8Tensor`` (5-, 6-,
or 8-bit).

``quantize_model`` walks a model's parameters, applies a ``QuantRecipe``,
and returns a single state dict containing both quantized subclass tensors
Expand Down Expand Up @@ -152,7 +152,7 @@ def _to_intx_tensor(
weight: torch.Tensor,
config: QuantConfig,
) -> torch.Tensor:
"""Quantize to 6- or 8-bit and wrap in IntxUnpackedToInt8Tensor.
"""Quantize to 5-, 6-, or 8-bit and wrap in IntxUnpackedToInt8Tensor.

Quantizes in float32 for numerical precision, then constructs the
subclass directly. We avoid ``from_hp`` because it quantizes in the
Expand Down Expand Up @@ -229,10 +229,10 @@ def _to_intx_tensor(
def quantize_weight(weight: torch.Tensor, config: QuantConfig) -> torch.Tensor:
"""Quantize ``weight`` to a torchao tensor subclass.

Returns ``Int4Tensor`` for 4-bit or ``IntxUnpackedToInt8Tensor`` for 6-
and 8-bit.
Returns ``Int4Tensor`` for 4-bit or ``IntxUnpackedToInt8Tensor`` for 5-,
6-, and 8-bit.
"""
if config.bits in (6, 8):
if config.bits in (5, 6, 8):
return _to_intx_tensor(weight, config)

if config.bits != 4:
Expand Down
Loading