mirror of
https://github.com/ml-explore/mlx.git
synced 2025-09-10 21:37:50 +08:00
Fused Affine Quantize/Dequantize ops (#1282)
* Add fast affine dequantize * add full quantize kernel * fused kernel with scale/bias computation * fix docstring * fix no jit error * fix test * test fix * reduce fast api to only affine_quantize
This commit is contained in:
@@ -2,6 +2,7 @@
|
||||
|
||||
#include <nanobind/nanobind.h>
|
||||
#include <nanobind/stl/optional.h>
|
||||
#include <nanobind/stl/tuple.h>
|
||||
#include <nanobind/stl/variant.h>
|
||||
|
||||
#include "mlx/fast.h"
|
||||
@@ -138,4 +139,47 @@ void init_fast(nb::module_& parent_module) {
|
||||
Returns:
|
||||
array: The output array.
|
||||
)pbdoc");
|
||||
|
||||
m.def(
|
||||
"affine_quantize",
|
||||
nb::overload_cast<
|
||||
const array&,
|
||||
const array&,
|
||||
const array&,
|
||||
int,
|
||||
int,
|
||||
StreamOrDevice>(&fast::affine_quantize),
|
||||
"w"_a,
|
||||
"scales"_a,
|
||||
"biases"_a,
|
||||
"group_size"_a = 64,
|
||||
"bits"_a = 4,
|
||||
nb::kw_only(),
|
||||
"stream"_a = nb::none(),
|
||||
nb::sig(
|
||||
"def affine_quantize(w: array, /, scales: array, biases: array, group_size: int = 64, bits: int = 4, *, stream: Union[None, Stream, Device] = None) -> array"),
|
||||
R"pbdoc(
|
||||
Quantize the matrix ``w`` using the provided ``scales`` and
|
||||
``biases`` and the ``group_size`` and ``bits`` configuration.
|
||||
|
||||
Formally, given the notation in :func:`quantize`, we compute
|
||||
:math:`w_i` from :math:`\hat{w_i}` and corresponding :math:`s` and
|
||||
:math:`\beta` as follows
|
||||
|
||||
.. math::
|
||||
|
||||
w_i = s (\hat{w_i} + \beta)
|
||||
|
||||
Args:
|
||||
w (array): Matrix to be quantize
|
||||
scales (array): The scales to use per ``group_size`` elements of ``w``
|
||||
biases (array): The biases to use per ``group_size`` elements of ``w``
|
||||
group_size (int, optional): The size of the group in ``w`` that shares a
|
||||
scale and bias. (default: ``64``)
|
||||
bits (int, optional): The number of bits occupied by each element in
|
||||
``w``. (default: ``4``)
|
||||
|
||||
Returns:
|
||||
array: The quantized version of ``w``
|
||||
)pbdoc");
|
||||
}
|
||||
|
Reference in New Issue
Block a user