Skip to contents

Quantize a Matrix

mlx_quantize.Rd

Quantizes a weight matrix to low-precision representation (typically 4-bit or 8-bit). This reduces memory usage and enables faster computation during inference.

Usage

mlx_quantize(
  w,
  group_size = 64L,
  bits = 4L,
  mode = "affine",
  device = mlx_default_device()
)

Arguments

w

An mlx array (the weight matrix to quantize)

group_size

The group size for quantization. Smaller groups provide better accuracy but slightly higher memory. Default: 64

bits

The number of bits for quantization (typically 4 or 8). Default: 4

mode

The quantization mode: "affine" (with scales and biases) or "mxfp4" (4-bit floating point with group_size=32). Default: "affine"

device

Execution target: supply "gpu", "cpu", or an mlx_stream created via mlx_new_stream(). Default: mlx_default_device().

Value

A list containing:

w_q

The quantized weight matrix (packed as uint32)

scales

The quantization scales for dequantization

biases

The quantization biases (NULL for symmetric mode)

Details

Quantization converts floating-point weights to low-precision integers, reducing memory by up to 8x for 4-bit quantization. The scales (and optionally biases) are stored to enable approximate reconstruction of the original values.

See also

mlx_dequantize(), mlx_quantized_matmul()

Examples

if (FALSE) { # \dontrun{
w <- mlx_random_normal(c(512, 256))
quant <- mlx_quantize(w, group_size = 64, bits = 4)
# Use quant$w_q, quant$scales, quant$biases with mlx_quantized_matmul()
} # }