File size: 6,651 Bytes
c179c12 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 | """MLX runtime for Qwen3.6 ModelOpt hybrid FP8/NVFP4 checkpoints.
The converter stores ModelOpt FP8 weights losslessly in MLX's packed MXFP8
carrier with unit E8M0 block scales, then applies the original per-tensor
ModelOpt scale to the output. ModelOpt NVFP4 weights and E4M3 block scales
are likewise retained bit-for-bit; their FP32 tensor scale is applied after
the matrix multiplication.
Activations remain in the model dtype. This avoids adding a second lossy
activation requantization scheme while still using MLX's native quantized
weight kernels.
"""
from dataclasses import dataclass, field
from typing import Dict
import mlx.core as mx
import mlx.nn as nn
from mlx.utils import tree_flatten, tree_unflatten
from mlx_lm.models.base import BaseModelArgs
from mlx_lm.models.qwen3_5_moe import Model as BaseModel
from mlx_lm.models.switch_layers import SwitchLinear
@dataclass
class ModelArgs(BaseModelArgs):
model_type: str
text_config: dict
mlx_modelopt_quantization: Dict[str, str] = field(default_factory=dict)
class ScaledQuantizedLinear(nn.Module):
"""Weight-quantized dense linear with an additional tensor scale."""
def __init__(
self,
input_dims: int,
output_dims: int,
*,
group_size: int,
bits: int,
mode: str,
bias: bool = False,
):
super().__init__()
if input_dims % group_size:
raise ValueError(
f"input_dims={input_dims} is not divisible by group_size={group_size}"
)
if (input_dims * bits) % 32:
raise ValueError(
f"input_dims={input_dims}, bits={bits} cannot be packed into uint32"
)
self.group_size = group_size
self.bits = bits
self.mode = mode
self.weight = mx.zeros(
(output_dims, input_dims * bits // 32), dtype=mx.uint32
)
self.scales = mx.zeros(
(output_dims, input_dims // group_size), dtype=mx.uint8
)
self.global_scale = mx.ones((), dtype=mx.float32)
if bias:
self.bias = mx.zeros((output_dims,))
self.freeze()
@classmethod
def from_linear(cls, linear: nn.Module, kind: str):
output_dims, input_dims = linear.weight.shape
has_bias = linear.get("bias") is not None
if kind == "scaled_mxfp8":
params = dict(group_size=32, bits=8, mode="mxfp8")
elif kind == "scaled_nvfp4":
params = dict(group_size=16, bits=4, mode="nvfp4")
else:
raise ValueError(f"Unsupported dense quantization kind: {kind}")
return cls(input_dims, output_dims, bias=has_bias, **params)
def __call__(self, x):
y = mx.quantized_matmul(
x,
self["weight"],
self["scales"],
transpose=True,
group_size=self.group_size,
bits=self.bits,
mode=self.mode,
)
# Avoid promoting the residual stream to float32.
y = y * self["global_scale"].astype(y.dtype)
if "bias" in self:
y = y + self["bias"]
return y
class ScaledNVFP4SwitchLinear(nn.Module):
"""Expert linear using MLX gather_qmm and per-expert tensor scales."""
group_size = 16
bits = 4
mode = "nvfp4"
def __init__(
self,
input_dims: int,
output_dims: int,
num_experts: int,
*,
bias: bool = False,
):
super().__init__()
if input_dims % self.group_size:
raise ValueError(
f"input_dims={input_dims} is not divisible by {self.group_size}"
)
self.weight = mx.zeros(
(num_experts, output_dims, input_dims * self.bits // 32),
dtype=mx.uint32,
)
self.scales = mx.zeros(
(num_experts, output_dims, input_dims // self.group_size),
dtype=mx.uint8,
)
self.global_scales = mx.ones((num_experts,), dtype=mx.float32)
if bias:
self.bias = mx.zeros((num_experts, output_dims))
self.freeze()
@classmethod
def from_switch_linear(cls, linear: SwitchLinear):
num_experts, output_dims, input_dims = linear.weight.shape
has_bias = linear.get("bias") is not None
return cls(
input_dims,
output_dims,
num_experts,
bias=has_bias,
)
@property
def input_dims(self):
return self.scales.shape[2] * self.group_size
@property
def output_dims(self):
return self.weight.shape[1]
@property
def num_experts(self):
return self.weight.shape[0]
def __call__(self, x, indices, sorted_indices=False):
y = mx.gather_qmm(
x,
self["weight"],
self["scales"],
rhs_indices=indices,
transpose=True,
group_size=self.group_size,
bits=self.bits,
mode=self.mode,
sorted_indices=sorted_indices,
)
scale = self["global_scales"][indices].astype(y.dtype)[..., None, None]
y = y * scale
if "bias" in self:
y = y + mx.expand_dims(self["bias"][indices], -2)
return y
def _replace_quantized_modules(model: nn.Module, quantization: Dict[str, str]):
leaves = dict(
tree_flatten(model.leaf_modules(), is_leaf=lambda m: isinstance(m, nn.Module))
)
missing = sorted(set(quantization) - set(leaves))
if missing:
preview = "\n ".join(missing[:20])
raise ValueError(f"Quantized module paths are absent from the model:\n {preview}")
for path, kind in quantization.items():
module = leaves[path]
if kind in ("scaled_mxfp8", "scaled_nvfp4"):
if not isinstance(module, nn.Linear):
raise TypeError(f"{path} is {type(module).__name__}, expected Linear")
leaves[path] = ScaledQuantizedLinear.from_linear(module, kind)
elif kind == "scaled_nvfp4_switch":
if not isinstance(module, SwitchLinear):
raise TypeError(
f"{path} is {type(module).__name__}, expected SwitchLinear"
)
leaves[path] = ScaledNVFP4SwitchLinear.from_switch_linear(module)
else:
raise ValueError(f"Unknown quantization kind {kind!r} for {path}")
model.update_modules(tree_unflatten(list(leaves.items())))
class Model(BaseModel):
def __init__(self, args: ModelArgs):
super().__init__(args)
_replace_quantized_modules(self, args.mlx_modelopt_quantization)
|