Lokr support
This commit is contained in:
parent
58c3df32f3
commit
46466f09d0
@ -1,4 +1,3 @@
|
||||
import lyco_helpers
|
||||
import network
|
||||
import network_lyco
|
||||
|
||||
|
65
extensions-builtin/Lora/network_lokr.py
Normal file
65
extensions-builtin/Lora/network_lokr.py
Normal file
@ -0,0 +1,65 @@
|
||||
import torch
|
||||
|
||||
import lyco_helpers
|
||||
import network
|
||||
import network_lyco
|
||||
|
||||
|
||||
class ModuleTypeLokr(network.ModuleType):
|
||||
def create_module(self, net: network.Network, weights: network.NetworkWeights):
|
||||
has_1 = "lokr_w1" in weights.w or ("lokr_w1a" in weights.w and "lokr_w1b" in weights.w)
|
||||
has_2 = "lokr_w2" in weights.w or ("lokr_w2a" in weights.w and "lokr_w2b" in weights.w)
|
||||
if has_1 and has_2:
|
||||
return NetworkModuleLokr(net, weights)
|
||||
|
||||
return None
|
||||
|
||||
|
||||
def make_kron(orig_shape, w1, w2):
|
||||
if len(w2.shape) == 4:
|
||||
w1 = w1.unsqueeze(2).unsqueeze(2)
|
||||
w2 = w2.contiguous()
|
||||
return torch.kron(w1, w2).reshape(orig_shape)
|
||||
|
||||
|
||||
class NetworkModuleLokr(network_lyco.NetworkModuleLyco):
|
||||
def __init__(self, net: network.Network, weights: network.NetworkWeights):
|
||||
super().__init__(net, weights)
|
||||
|
||||
self.w1 = weights.w.get("lokr_w1")
|
||||
self.w1a = weights.w.get("lokr_w1_a")
|
||||
self.w1b = weights.w.get("lokr_w1_b")
|
||||
self.dim = self.w1b.shape[0] if self.w1b else self.dim
|
||||
self.w2 = weights.w.get("lokr_w2")
|
||||
self.w2a = weights.w.get("lokr_w2_a")
|
||||
self.w2b = weights.w.get("lokr_w2_b")
|
||||
self.dim = self.w2b.shape[0] if self.w2b else self.dim
|
||||
self.t2 = weights.w.get("lokr_t2")
|
||||
|
||||
def calc_updown(self, orig_weight):
|
||||
if self.w1 is not None:
|
||||
w1 = self.w1.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
else:
|
||||
w1a = self.w1a.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
w1b = self.w1b.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
w1 = w1a @ w1b
|
||||
|
||||
if self.w2 is not None:
|
||||
w2 = self.w2.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
elif self.t2 is None:
|
||||
w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
w2 = w2a @ w2b
|
||||
else:
|
||||
t2 = self.t2.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
w2a = self.w2a.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
w2b = self.w2b.to(orig_weight.device, dtype=orig_weight.dtype)
|
||||
w2 = lyco_helpers.make_weight_cp(t2, w2a, w2b)
|
||||
|
||||
output_shape = [w1.size(0) * w2.size(0), w1.size(1) * w2.size(1)]
|
||||
if len(orig_weight.shape) == 4:
|
||||
output_shape = orig_weight.shape
|
||||
|
||||
updown = make_kron(output_shape, w1, w2)
|
||||
|
||||
return self.finalize_updown(updown, orig_weight, output_shape)
|
@ -5,6 +5,7 @@ import network
|
||||
import network_lora
|
||||
import network_hada
|
||||
import network_ia3
|
||||
import network_lokr
|
||||
|
||||
import torch
|
||||
from typing import Union
|
||||
@ -15,6 +16,7 @@ module_types = [
|
||||
network_lora.ModuleTypeLora(),
|
||||
network_hada.ModuleTypeHada(),
|
||||
network_ia3.ModuleTypeIa3(),
|
||||
network_lokr.ModuleTypeLokr(),
|
||||
]
|
||||
|
||||
|
||||
|
Loading…
Reference in New Issue
Block a user