ai-toolkit/toolkit/models/lokr.py

434 lines
17 KiB
Python

# based heavily on https://github.com/KohakuBlueleaf/LyCORIS/blob/eb460098187f752a5d66406d3affade6f0a07ece/lycoris/modules/lokr.py
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from toolkit.network_mixins import ToolkitModuleMixin
from typing import TYPE_CHECKING, Union, List
from optimum.quanto import QBytesTensor, QTensor
from torchao.dtypes import AffineQuantizedTensor
if TYPE_CHECKING:
from toolkit.lora_special import LoRASpecialNetwork
def factorization(dimension: int, factor: int = -1) -> tuple[int, int]:
'''
return a tuple of two value of input dimension decomposed by the number closest to factor
second value is higher or equal than first value.
In LoRA with Kroneckor Product, first value is a value for weight scale.
secon value is a value for weight.
Becuase of non-commutative property, A(kron)B != B(kron)A. Meaning of two matrices is slightly different.
examples)
factor
-1 2 4 8 16 ...
127 -> 127, 1 127 -> 127, 1 127 -> 127, 1 127 -> 127, 1 127 -> 127, 1
128 -> 16, 8 128 -> 64, 2 128 -> 32, 4 128 -> 16, 8 128 -> 16, 8
250 -> 125, 2 250 -> 125, 2 250 -> 125, 2 250 -> 125, 2 250 -> 125, 2
360 -> 45, 8 360 -> 180, 2 360 -> 90, 4 360 -> 45, 8 360 -> 45, 8
512 -> 32, 16 512 -> 256, 2 512 -> 128, 4 512 -> 64, 8 512 -> 32, 16
1024 -> 32, 32 1024 -> 512, 2 1024 -> 256, 4 1024 -> 128, 8 1024 -> 64, 16
'''
if factor > 0 and (dimension % factor) == 0:
m = factor
n = dimension // factor
return m, n
if factor == -1:
factor = dimension
m, n = 1, dimension
length = m + n
while m < n:
new_m = m + 1
while dimension % new_m != 0:
new_m += 1
new_n = dimension // new_m
if new_m + new_n > length or new_m > factor:
break
else:
m, n = new_m, new_n
if m > n:
n, m = m, n
return m, n
def make_weight_cp(t, wa, wb):
rebuild2 = torch.einsum('i j k l, i p, j r -> p r k l',
t, wa, wb) # [c, d, k1, k2]
return rebuild2
def make_kron(w1, w2, scale):
if len(w2.shape) == 4:
w1 = w1.unsqueeze(2).unsqueeze(2)
w2 = w2.contiguous()
rebuild = torch.kron(w1, w2)
return rebuild*scale
class LokrModule(ToolkitModuleMixin, nn.Module):
def __init__(
self,
lora_name,
org_module: nn.Module,
multiplier=1.0,
lora_dim=4,
alpha=1,
dropout=0.,
rank_dropout=0.,
module_dropout=0.,
use_cp=False,
decompose_both=False,
network: 'LoRASpecialNetwork' = None,
factor: int = -1, # factorization factor
**kwargs,
):
""" if alpha == 0 or None, alpha is rank (no scaling). """
ToolkitModuleMixin.__init__(self, network=network)
torch.nn.Module.__init__(self)
factor = int(factor)
self.lora_name = lora_name
self.lora_dim = lora_dim
self.cp = False
self.use_w1 = False
self.use_w2 = False
self.can_merge_in = True
# only the plain Linear path (no cp/Conv2d) gets the factorized forward;
# Conv2d and the cp branch keep using get_weight()/make_kron() below.
self._fast_linear = False
# avoid the weight property on quantized OstrisLinear: it dequantizes the
# whole weight just to answer .shape
if getattr(org_module, "is_ostris_quantized", False):
self.shape = torch.Size((org_module.out_features, org_module.in_features))
else:
self.shape = org_module.weight.shape
if org_module.__class__.__name__ == 'Conv2d':
in_dim = org_module.in_channels
k_size = org_module.kernel_size
out_dim = org_module.out_channels
in_m, in_n = factorization(in_dim, factor)
out_l, out_k = factorization(out_dim, factor)
# ((a, b), (c, d), *k_size)
shape = ((out_l, out_k), (in_m, in_n), *k_size)
self.cp = use_cp and k_size != (1, 1)
if decompose_both and lora_dim < max(shape[0][0], shape[1][0])/2:
self.lokr_w1_a = nn.Parameter(
torch.empty(shape[0][0], lora_dim))
self.lokr_w1_b = nn.Parameter(
torch.empty(lora_dim, shape[1][0]))
else:
self.use_w1 = True
self.lokr_w1 = nn.Parameter(torch.empty(
shape[0][0], shape[1][0])) # a*c, 1-mode
if lora_dim >= max(shape[0][1], shape[1][1])/2:
self.use_w2 = True
self.lokr_w2 = nn.Parameter(torch.empty(
shape[0][1], shape[1][1], *k_size))
elif self.cp:
self.lokr_t2 = nn.Parameter(torch.empty(
lora_dim, lora_dim, shape[2], shape[3]))
self.lokr_w2_a = nn.Parameter(
torch.empty(lora_dim, shape[0][1])) # b, 1-mode
self.lokr_w2_b = nn.Parameter(
torch.empty(lora_dim, shape[1][1])) # d, 2-mode
else: # Conv2d not cp
# bigger part. weight and LoRA. [b, dim] x [dim, d*k1*k2]
self.lokr_w2_a = nn.Parameter(
torch.empty(shape[0][1], lora_dim))
self.lokr_w2_b = nn.Parameter(torch.empty(
lora_dim, shape[1][1]*shape[2]*shape[3]))
# w1 (kron) (w2_a x w2_b) = (a, b)(kron)((c, dim)x(dim, d*k1*k2)) = (a, b)(kron)(c, d*k1*k2) = (ac, bd*k1*k2)
self.op = F.conv2d
self.extra_args = {
"stride": org_module.stride,
"padding": org_module.padding,
"dilation": org_module.dilation,
"groups": org_module.groups
}
else: # Linear
in_dim = org_module.in_features
out_dim = org_module.out_features
in_m, in_n = factorization(in_dim, factor)
out_l, out_k = factorization(out_dim, factor)
# ((a, b), (c, d)), out_dim = a*c, in_dim = b*d
shape = ((out_l, out_k), (in_m, in_n))
# stash factor pair for the factorized (no full-kron) forward path
self._in_m, self._in_n = in_m, in_n
self._out_l, self._out_k = out_l, out_k
self._fast_linear = True
# smaller part. weight scale
if decompose_both and lora_dim < max(shape[0][0], shape[1][0])/2:
self.lokr_w1_a = nn.Parameter(
torch.empty(shape[0][0], lora_dim))
self.lokr_w1_b = nn.Parameter(
torch.empty(lora_dim, shape[1][0]))
else:
self.use_w1 = True
self.lokr_w1 = nn.Parameter(torch.empty(
shape[0][0], shape[1][0])) # a*c, 1-mode
if lora_dim < max(shape[0][1], shape[1][1])/2:
# bigger part. weight and LoRA. [b, dim] x [dim, d]
self.lokr_w2_a = nn.Parameter(
torch.empty(shape[0][1], lora_dim))
self.lokr_w2_b = nn.Parameter(
torch.empty(lora_dim, shape[1][1]))
# w1 (kron) (w2_a x w2_b) = (a, b)(kron)((c, dim)x(dim, d)) = (a, b)(kron)(c, d) = (ac, bd)
else:
self.use_w2 = True
self.lokr_w2 = nn.Parameter(
torch.empty(shape[0][1], shape[1][1]))
self.op = F.linear
self.extra_args = {}
self.dropout = dropout
if dropout:
print("[WARN]LoKr haven't implemented normal dropout yet.")
self.rank_dropout = rank_dropout
self.module_dropout = module_dropout
if isinstance(alpha, torch.Tensor):
alpha = float(alpha.detach().float().item())
alpha = lora_dim if alpha is None or alpha == 0 else alpha
if self.use_w2 and self.use_w1:
# use scale = 1
alpha = lora_dim
self._set_runtime_scale(float(alpha) / self.lora_dim)
self.register_buffer('alpha', torch.tensor(alpha)) # treat as constant
if self.use_w2:
torch.nn.init.constant_(self.lokr_w2, 0)
else:
if self.cp:
torch.nn.init.kaiming_uniform_(self.lokr_t2, a=math.sqrt(5))
torch.nn.init.kaiming_uniform_(self.lokr_w2_a, a=math.sqrt(5))
torch.nn.init.constant_(self.lokr_w2_b, 0)
if self.use_w1:
torch.nn.init.kaiming_uniform_(self.lokr_w1, a=math.sqrt(5))
else:
torch.nn.init.kaiming_uniform_(self.lokr_w1_a, a=math.sqrt(5))
torch.nn.init.kaiming_uniform_(self.lokr_w1_b, a=math.sqrt(5))
self.multiplier = multiplier
self.org_module = [org_module]
weight = make_kron(
self.lokr_w1 if self.use_w1 else self.lokr_w1_a@self.lokr_w1_b,
(self.lokr_w2 if self.use_w2
else make_weight_cp(self.lokr_t2, self.lokr_w2_a, self.lokr_w2_b) if self.cp
else self.lokr_w2_a@self.lokr_w2_b),
self.multiplier * self.scale
)
assert torch.sum(torch.isnan(weight)) == 0, "weight is nan"
# Same as locon.py
def apply_to(self):
self.org_forward = self.org_module[0].forward
self.org_module[0].forward = self.forward
def get_weight(self, orig_weight=None):
weight = make_kron(
self.lokr_w1 if self.use_w1 else self.lokr_w1_a@self.lokr_w1_b,
(self.lokr_w2 if self.use_w2
else make_weight_cp(self.lokr_t2, self.lokr_w2_a, self.lokr_w2_b) if self.cp
else self.lokr_w2_a@self.lokr_w2_b),
self._runtime_scale
)
if orig_weight is not None:
weight = weight.reshape(orig_weight.shape)
if self.training and self.rank_dropout:
drop = torch.rand(weight.size(0)) < self.rank_dropout
weight *= drop.view(-1, [1] *
len(weight.shape[1:])).to(weight.device)
return weight
@torch.no_grad()
def merge_in(self, merge_weight=1.0):
if not self.can_merge_in:
return
# extract weight from org_module
org_sd = self.org_module[0].state_dict()
# todo find a way to merge in weights when doing quanto quantized model
if 'weight._data' in org_sd:
# quanto quantized weight
return
weight_key = "weight"
from toolkit.util.quantize import is_quantized_tensor
org_weight = self.org_module[0].weight
is_ao_quantized = is_quantized_tensor(org_weight)
orig_dtype = org_weight.dtype
# dequantize torchao weights so the delta can be merged in full precision
weight = (org_weight.dequantize() if is_ao_quantized else org_weight).float()
scale = self.scale
# handle trainable scaler method locon does
if hasattr(self, 'scalar'):
scale = scale * self.scalar
lokr_weight = self.get_weight(weight)
merged_weight = (
weight
+ (lokr_weight * merge_weight).to(weight.device, dtype=weight.dtype)
)
# write the merged weight back, re-quantizing if the original was torchao quantized so the
# model stays quantized across continuous merge/reset cycles
if is_ao_quantized:
from toolkit.util.quantize import get_torchao_config, requantize_module_weight
requantize_module_weight(
self.org_module[0], merged_weight, orig_dtype, get_torchao_config(self._get_base_qtype())
)
else:
org_sd[weight_key] = merged_weight.to(orig_dtype)
self.org_module[0].load_state_dict(org_sd)
def get_orig_weight(self, device):
weight = self.org_module[0].weight
if weight.device != device:
weight = weight.to(device)
if isinstance(weight, QTensor) or isinstance(weight, QBytesTensor):
return weight.dequantize().data.detach()
elif isinstance(weight, AffineQuantizedTensor):
return weight.dequantize().data.detach()
else:
return weight.data.detach()
def get_orig_bias(self, device):
if hasattr(self.org_module[0], 'bias') and self.org_module[0].bias is not None:
bias = self.org_module[0].bias
if bias.device != device:
bias = bias.to(device)
if isinstance(bias, QTensor) or isinstance(bias, QBytesTensor):
return bias.dequantize().data.detach()
elif isinstance(bias, AffineQuantizedTensor):
return bias.dequantize().data.detach()
else:
return self.org_module[0].bias.data.detach()
return None
def _get_delta_factors(self):
"""(A, w2_or_w2a, w2_b_or_None) without ever combining the two kron
factors into a full-size matrix."""
A = self.lokr_w1 if self.use_w1 else self.lokr_w1_a @ self.lokr_w1_b # (out_l, in_m), always small
if self.use_w2:
return A, self.lokr_w2, None
else:
return A, self.lokr_w2_a, self.lokr_w2_b
def _call_forward_fast_linear(self, x):
"""Factorized LoKr delta for the plain Linear case (quantized or not).
Never materializes a (out_dim, in_dim) tensor: uses the kron
mixed-product identity plus the existing low-rank factoring of w2, so
both compute and the backward-pass gradient are O(rank) instead of
O(out_dim * in_dim)."""
if isinstance(x, QTensor) or isinstance(x, QBytesTensor):
x = x.dequantize()
orig_dtype = x.dtype
if getattr(self.org_module[0], "is_ostris_quantized", False):
# org_forward here does raw CUDA stream/event orchestration and
# in-place buffer swapping (manager_modules.py _mm_forward) to
# move the quantized weight onto the device -- that can't be
# traced by inductor, so break the graph at exactly this call.
base_out = torch._dynamo.disable(self.org_forward)(x)
else:
base_out = self.org_forward(x)
A, w2a, w2b = self._get_delta_factors()
# match the base path's compute dtype (usually bf16) rather than
# upcasting the (much larger) activation tensor to the fp32 master
# params -- keeps every intermediate here at the same footprint the
# old full-kron GEMM had, instead of doubling it.
compute_dtype = base_out.dtype
x_ = x.to(compute_dtype) if x.dtype != compute_dtype else x
A = A.to(compute_dtype)
w2a = w2a.to(compute_dtype)
if w2b is not None:
w2b = w2b.to(compute_dtype)
X = x_.unflatten(-1, (self._in_m, self._in_n))
if w2b is None:
# use_w2: w2a is the full (out_k, in_n) factor -- still far smaller
# than the full kron product, so a single einsum here is fine.
tmp = torch.einsum('...qs,os->...qo', X, w2a) # (..., in_m, out_k)
else:
# low-rank w2 = w2a @ w2b, rank << out_k, in_n: fold the rank
# factor through first so we never touch an (out_k, in_n) tensor.
tmp = torch.einsum('...qs,rs->...qr', X, w2b) # (..., in_m, rank)
tmp = torch.einsum('...qr,or->...qo', tmp, w2a) # (..., in_m, out_k)
# scale folded into A (not applied to the reduction output) to avoid
# an inductor lowering bug under torch.compile
delta = torch.einsum('...qo,pq->...po', tmp, A * self._runtime_scale) # (..., out_l, out_k)
delta = delta.flatten(-2, -1)
if self.training and self.rank_dropout:
# equivalent to the old row-mask on the full weight: zeroing a
# weight ROW is the same as zeroing that output channel post-matmul.
drop = torch.rand(delta.size(-1), device=delta.device) < self.rank_dropout
delta = delta * drop.to(delta.dtype)
multiplier = torch.mean(self.network_ref().torch_multiplier).to(compute_dtype)
delta = delta * multiplier
return (base_out + delta).to(orig_dtype)
def _call_forward(self, x):
if self._fast_linear:
return self._call_forward_fast_linear(x)
# legacy path: Conv2d / cp branch. Still materializes the full kron
# product -- TODO: extend the factorized path to cover these.
if isinstance(x, QTensor) or isinstance(x, QBytesTensor):
x = x.dequantize()
orig_dtype = x.dtype
orig_weight = self.get_orig_weight(x.device)
lokr_weight = self.get_weight(orig_weight).to(dtype=orig_weight.dtype)
multiplier = self.network_ref().torch_multiplier
if x.dtype != orig_weight.dtype:
x = x.to(dtype=orig_weight.dtype)
# we do not currently support split batch multipliers for lokr. Just do a mean
multiplier = torch.mean(multiplier)
weight = (
orig_weight
+ lokr_weight * multiplier
)
bias = self.get_orig_bias(x.device)
if bias is not None:
bias = bias.to(weight.device, dtype=weight.dtype)
output = self.op(
x,
weight.view(self.shape),
bias,
**self.extra_args
)
return output.to(orig_dtype)