File size: 11,778 Bytes
9118991 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 | """Shared Kronecker transform and exact no-quant folding for LoopQ LQ3.
LoopQ Equations (2)-(3) require ``X @ P`` to be paired with
``W @ P^{-T}``. The paper additionally says that ``P`` uses a
FlatQuant-style Kronecker decomposition. This module implements that minimal
shared contract without choosing LoopQ's underspecified optimizer or SVD/direct
calibration parameterization.
"""
from __future__ import annotations
import torch
from torch import nn
from torch.nn.utils import parametrizations
class MaterializedKroneckerTransform:
# One-forward execution view reusing effective factors and inverses.
def __init__(self, transform: nn.Module) -> None:
self.feature_size = transform.feature_size
self.left = transform.left
self.right = transform.right
log_diagonal = transform.log_diagonal
self.diagonal = None if log_diagonal is None else log_diagonal.exp()
self.left_inverse_transpose = torch.linalg.inv(
self.left.to(torch.float64)
).T.to(dtype=self.left.dtype)
self.right_inverse_transpose = torch.linalg.inv(
self.right.to(torch.float64)
).T.to(dtype=self.right.dtype)
def __call__(self, activation: torch.Tensor) -> torch.Tensor:
if activation.ndim == 0 or activation.shape[-1] != self.feature_size:
raise ValueError(
f"activation last dimension must be {self.feature_size}, got {tuple(activation.shape)}"
)
value = activation
if self.diagonal is not None:
value = value * self.diagonal.to(value)
return SharedKroneckerTransform._apply_kronecker(
value, self.left.to(value), self.right.to(value)
)
def fold_weight(self, weight: torch.Tensor) -> torch.Tensor:
if weight.ndim != 2 or weight.shape[1] != self.feature_size:
raise ValueError(
f"weight must have shape (out_features, {self.feature_size}), got {tuple(weight.shape)}"
)
value = weight
if self.diagonal is not None:
value = value / self.diagonal.to(value)
return SharedKroneckerTransform._apply_kronecker(
value,
self.left_inverse_transpose.to(value),
self.right_inverse_transpose.to(value),
)
class SharedKroneckerTransform(nn.Module):
"""One loop-shared invertible transform ``diag(d) @ kron(L, R)``.
The explicit factors make architecture-selected factor dimensions part of
the artifact rather than silently guessing them from hidden size. Identity
initialization exactly preserves the original BF16 function.
"""
def __init__(
self,
left_size: int,
right_size: int,
*,
add_diagonal: bool = False,
dtype: torch.dtype = torch.float32,
) -> None:
super().__init__()
if left_size <= 0 or right_size <= 0:
raise ValueError("Kronecker factor sizes must be positive")
self.left_size = int(left_size)
self.right_size = int(right_size)
self.add_diagonal = bool(add_diagonal)
self.left = nn.Parameter(torch.eye(left_size, dtype=dtype))
self.right = nn.Parameter(torch.eye(right_size, dtype=dtype))
if add_diagonal:
self.log_diagonal = nn.Parameter(torch.zeros(self.feature_size, dtype=dtype))
else:
self.register_parameter("log_diagonal", None)
@property
def feature_size(self) -> int:
return self.left_size * self.right_size
def matrix(self) -> torch.Tensor:
"""Materialize ``P`` for correctness paths and export."""
kronecker = torch.kron(self.left, self.right)
if self.log_diagonal is None:
return kronecker
return self.log_diagonal.exp().diag() @ kronecker
def forward(self, activation: torch.Tensor) -> torch.Tensor:
if activation.ndim == 0 or activation.shape[-1] != self.feature_size:
raise ValueError(
f"activation last dimension must be {self.feature_size}, "
f"got {tuple(activation.shape)}"
)
value = activation
if self.log_diagonal is not None:
value = value * self.log_diagonal.exp().to(value)
return self._apply_kronecker(value, self.left.to(value), self.right.to(value))
@staticmethod
def _apply_kronecker(
value: torch.Tensor, left: torch.Tensor, right: torch.Tensor
) -> torch.Tensor:
"""Compute ``value @ kron(left, right)`` without materializing kron."""
shape = value.shape
reshaped = value.reshape(-1, left.shape[0], right.shape[0])
reshaped = torch.matmul(reshaped, right)
reshaped = torch.matmul(left.T, reshaped)
return reshaped.reshape(shape)
def materialize(self) -> MaterializedKroneckerTransform:
return MaterializedKroneckerTransform(self)
def inverse_transpose(self) -> torch.Tensor:
"""Return ``P^{-T}``, computed in FP64 for stable offline folding."""
matrix = self.matrix()
inverse_transpose = torch.linalg.inv(matrix.to(torch.float64)).T
return inverse_transpose.to(dtype=matrix.dtype)
def fold_weight(self, weight: torch.Tensor) -> torch.Tensor:
"""Return ``W @ P^{-T}`` without mutating the supplied shared weight."""
if weight.ndim != 2 or weight.shape[1] != self.feature_size:
raise ValueError(
f"weight must have shape (out_features, {self.feature_size}), "
f"got {tuple(weight.shape)}"
)
value = weight
if self.log_diagonal is not None:
value = value / self.log_diagonal.exp().to(value)
left_inverse_transpose = torch.linalg.inv(self.left.to(torch.float64)).T.to(value)
right_inverse_transpose = torch.linalg.inv(self.right.to(torch.float64)).T.to(value)
return self._apply_kronecker(value, left_inverse_transpose, right_inverse_transpose)
def fresh_copy(self) -> "SharedKroneckerTransform":
result = type(self)(
self.left_size, self.right_size, add_diagonal=self.add_diagonal,
dtype=self.left.dtype,
).to(device=self.left.device)
result.load_state_dict(self.state_dict())
return result
def export_state(self) -> dict[str, object]:
return {
"format": "loopq_shared_kronecker_transform",
"format_version": 1,
"left_size": self.left_size,
"right_size": self.right_size,
"add_diagonal": self.add_diagonal,
"state_dict": {key: value.detach().cpu() for key, value in self.state_dict().items()},
}
@classmethod
def from_export_state(cls, state: dict[str, object]) -> "SharedKroneckerTransform":
if state.get("format") != "loopq_shared_kronecker_transform" or state.get("format_version") != 1:
raise ValueError("unsupported shared-transform export format or version")
transform = cls(
int(state["left_size"]),
int(state["right_size"]),
add_diagonal=bool(state["add_diagonal"]),
dtype=state["state_dict"]["left"].dtype,
)
transform.load_state_dict(state["state_dict"])
return transform
def _random_orthogonal(size: int, *, dtype: torch.dtype) -> torch.Tensor:
"""Return a torch-RNG-controlled analogue of FlatQuant get_init_weight."""
value = torch.randn(size, size, dtype=dtype)
q, r = torch.linalg.qr(value)
signs = torch.where(torch.diagonal(r) < 0, -torch.ones((), dtype=dtype),
torch.ones((), dtype=dtype))
return q @ torch.diag(signs)
class FlatQuantSVDKroneckerTransform(nn.Module):
"""FlatQuant-default SVD parameterization of a Kronecker transform.
Each small Kronecker factor is ``U diag(s) V.T``. ``U`` and ``V`` use
PyTorch's Cayley orthogonal parameterization, matching the cited official
FlatQuant implementation. Export stores only the effective factors so
inference does not depend on the training parameterization.
"""
def __init__(self, left_size: int, right_size: int, *,
dtype: torch.dtype = torch.float32) -> None:
super().__init__()
if left_size <= 0 or right_size <= 0:
raise ValueError("Kronecker factor sizes must be positive")
self.left_size = int(left_size)
self.right_size = int(right_size)
self.add_diagonal = False
self.left_u = self._orthogonal_linear(left_size, dtype)
self.left_v = self._orthogonal_linear(left_size, dtype)
self.right_u = self._orthogonal_linear(right_size, dtype)
self.right_v = self._orthogonal_linear(right_size, dtype)
self.left_singular = nn.Parameter(torch.ones(left_size, dtype=dtype))
self.right_singular = nn.Parameter(torch.ones(right_size, dtype=dtype))
@staticmethod
def _orthogonal_linear(size: int, dtype: torch.dtype) -> nn.Module:
linear = nn.Linear(size, size, bias=False, dtype=dtype)
with torch.no_grad():
linear.weight.copy_(_random_orthogonal(size, dtype=dtype))
return parametrizations.orthogonal(
linear, orthogonal_map="cayley", use_trivialization=False
)
@property
def feature_size(self) -> int:
return self.left_size * self.right_size
@property
def left(self) -> torch.Tensor:
return self.left_u.weight @ torch.diag(self.left_singular) @ self.left_v.weight.T
@property
def right(self) -> torch.Tensor:
return self.right_u.weight @ torch.diag(self.right_singular) @ self.right_v.weight.T
@property
def log_diagonal(self):
return None
def matrix(self) -> torch.Tensor:
return torch.kron(self.left, self.right)
def materialize(self) -> MaterializedKroneckerTransform:
return MaterializedKroneckerTransform(self)
def forward(self, activation: torch.Tensor) -> torch.Tensor:
if activation.ndim == 0 or activation.shape[-1] != self.feature_size:
raise ValueError(
f"activation last dimension must be {self.feature_size}, got {tuple(activation.shape)}"
)
return SharedKroneckerTransform._apply_kronecker(
activation, self.left.to(activation), self.right.to(activation)
)
def fold_weight(self, weight: torch.Tensor) -> torch.Tensor:
if weight.ndim != 2 or weight.shape[1] != self.feature_size:
raise ValueError(
f"weight must have shape (out_features, {self.feature_size}), got {tuple(weight.shape)}"
)
left_inverse_transpose = torch.linalg.inv(self.left.to(torch.float64)).T.to(weight)
right_inverse_transpose = torch.linalg.inv(self.right.to(torch.float64)).T.to(weight)
return SharedKroneckerTransform._apply_kronecker(
weight, left_inverse_transpose, right_inverse_transpose
)
def fresh_copy(self) -> "FlatQuantSVDKroneckerTransform":
result = type(self)(self.left_size, self.right_size).to(
device=self.left_singular.device, dtype=self.left_singular.dtype
)
result.load_state_dict(self.state_dict())
return result
def export_state(self) -> dict[str, object]:
return {
"format": "loopq_shared_kronecker_transform",
"format_version": 1,
"left_size": self.left_size,
"right_size": self.right_size,
"add_diagonal": False,
"parameterization": "flatquant_svd_cayley",
"state_dict": {
"left": self.left.detach().cpu(),
"right": self.right.detach().cpu(),
},
}
|