Upload LoRa_Merge_Script.ipynb
Browse files- LoRa_Merge_Script.ipynb +224 -26
LoRa_Merge_Script.ipynb
CHANGED
@@ -36,10 +36,10 @@
|
|
36 |
],
|
37 |
"metadata": {
|
38 |
"id": "CBVTifA_ZwdC",
|
39 |
-
"outputId": "8ce58389-8263-4016-8ebe-f61708ffef95",
|
40 |
"colab": {
|
41 |
"base_uri": "https://localhost:8080/"
|
42 |
-
}
|
|
|
43 |
},
|
44 |
"execution_count": 1,
|
45 |
"outputs": [
|
@@ -55,27 +55,225 @@
|
|
55 |
{
|
56 |
"cell_type": "code",
|
57 |
"source": [
|
|
|
|
|
58 |
"from safetensors.torch import load_file, save_file\n",
|
59 |
-
"
|
60 |
-
"
|
|
|
|
|
|
|
|
|
|
|
61 |
"\n",
|
62 |
-
"#
|
63 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
"\n",
|
65 |
-
"
|
66 |
-
"
|
67 |
-
"
|
68 |
-
"
|
69 |
-
"
|
70 |
-
"
|
71 |
-
"
|
72 |
-
" #print(puff[f'{key}'])\n",
|
73 |
-
" continue\n",
|
74 |
-
" puff[f'{key}'] = 2*_puff[f'{key}'].to(device=device , dtype = torch.float16)\n",
|
75 |
"\n",
|
76 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
"\n",
|
78 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
],
|
80 |
"metadata": {
|
81 |
"id": "SKYzFxehkfG8"
|
@@ -180,10 +378,10 @@
|
|
180 |
"metadata": {
|
181 |
"id": "1oxeJYHRqxQC",
|
182 |
"collapsed": true,
|
183 |
-
"outputId": "12e3a407-f9d1-403e-949b-31330be59577",
|
184 |
"colab": {
|
185 |
"base_uri": "https://localhost:8080/"
|
186 |
-
}
|
|
|
187 |
},
|
188 |
"execution_count": 12,
|
189 |
"outputs": [
|
@@ -1203,11 +1401,11 @@
|
|
1203 |
"#print(torch.matmul(tgt[f'{up}'],tgt[f'{down}']).shape)\n"
|
1204 |
],
|
1205 |
"metadata": {
|
1206 |
-
"id": "GoDfgENYaWD7",
|
1207 |
-
"outputId": "9336ae1a-6244-4e76-f291-82cda4482831",
|
1208 |
"colab": {
|
1209 |
"base_uri": "https://localhost:8080/"
|
1210 |
-
}
|
|
|
|
|
1211 |
},
|
1212 |
"execution_count": 17,
|
1213 |
"outputs": [
|
@@ -1230,11 +1428,11 @@
|
|
1230 |
" print(oily[f'{key}'].shape)"
|
1231 |
],
|
1232 |
"metadata": {
|
1233 |
-
"id": "xQhVLouEfmGE",
|
1234 |
-
"outputId": "662176b3-480d-48eb-f5db-97ec71b5e970",
|
1235 |
"colab": {
|
1236 |
"base_uri": "https://localhost:8080/"
|
1237 |
-
}
|
|
|
|
|
1238 |
},
|
1239 |
"execution_count": 18,
|
1240 |
"outputs": [
|
|
|
36 |
],
|
37 |
"metadata": {
|
38 |
"id": "CBVTifA_ZwdC",
|
|
|
39 |
"colab": {
|
40 |
"base_uri": "https://localhost:8080/"
|
41 |
+
},
|
42 |
+
"outputId": "8ce58389-8263-4016-8ebe-f61708ffef95"
|
43 |
},
|
44 |
"execution_count": 1,
|
45 |
"outputs": [
|
|
|
55 |
{
|
56 |
"cell_type": "code",
|
57 |
"source": [
|
58 |
+
"\n",
|
59 |
+
"import torch\n",
|
60 |
"from safetensors.torch import load_file, save_file\n",
|
61 |
+
"import torch.nn as nn\n",
|
62 |
+
"from torch import linalg as LA\n",
|
63 |
+
"import os\n",
|
64 |
+
"import math\n",
|
65 |
+
"import random\n",
|
66 |
+
"import numpy as np\n",
|
67 |
+
"device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
|
68 |
"\n",
|
69 |
+
"# For pcnt = 30 , 'filter_and_save' will keep all top 30 % values\n",
|
70 |
+
"#, and the lowest (negative) 30% values for each layer delta_W in this lora\n",
|
71 |
+
"# Then save the new filtered lora as a .safetensor file\n",
|
72 |
+
"def filter_and_save(_lora , savefile_name, new_rank , new_alpha, resolution):\n",
|
73 |
+
" lora = {}\n",
|
74 |
+
" count = 0\n",
|
75 |
+
" for key in _lora:count = count + 1\n",
|
76 |
+
" NUM_ITEMS = count\n",
|
77 |
+
" count = 0\n",
|
78 |
+
" thresh = resolution*0.000001 # 1e-6\n",
|
79 |
+
" #-------#\n",
|
80 |
+
" for key in _lora:\n",
|
81 |
+
" if f'{key}'.find('alpha') > -1:\n",
|
82 |
+
" lora[f'{key}'] = torch.tensor(new_alpha).to(device = device , dtype = torch.float32)\n",
|
83 |
+
" count = count + 1\n",
|
84 |
+
" print(f'{count} / {NUM_ITEMS}')\n",
|
85 |
+
" continue\n",
|
86 |
+
" #------#\n",
|
87 |
+
" if not f'{key}'.find('lora_down') > -1: continue\n",
|
88 |
+
" up = f'{key}'.replace('lora_down' , 'lora_up')\n",
|
89 |
+
" down = f'{key}'\n",
|
90 |
+
" #-------#\n",
|
91 |
+
" delta_W = torch.matmul(_lora[up],_lora[down]).to(device = device , dtype=torch.float32)\n",
|
92 |
+
" #---#\n",
|
93 |
+
" N = delta_W.numel()\n",
|
94 |
+
" y = delta_W.flatten().to(device = device , dtype=torch.float32)\n",
|
95 |
+
" values,indices = torch.sort(y, descending = False) # smallest -> largest elements\n",
|
96 |
+
" y = torch.zeros(y.shape).to(device = device , dtype=torch.float32)\n",
|
97 |
+
" y[indices[values>thresh]] = 1\n",
|
98 |
+
" y[indices[values<-thresh]] = 1\n",
|
99 |
+
" y = y.unflatten(0,delta_W.shape).to(device = device , dtype=torch.float32)\n",
|
100 |
+
" delta_W = torch.mul(delta_W,y).to(device = device , dtype=torch.float32)\n",
|
101 |
+
" #------#\n",
|
102 |
+
" tmp={}\n",
|
103 |
+
" tmp['u'], tmp['s'], tmp['Vh'] = torch.svd(delta_W)\n",
|
104 |
+
" tmp['u'] = tmp['u'][:,: new_rank]\n",
|
105 |
+
" tmp['s'] = tmp['s'][: new_rank]\n",
|
106 |
+
" #-------#\n",
|
107 |
+
" tmp['u'] = torch.round(torch.matmul(tmp['u'], torch.diag(tmp['s'])),decimals=6)\n",
|
108 |
+
" tmp['Vh'] = torch.round(tmp['Vh'].t()[: new_rank,:],decimals=6)\n",
|
109 |
+
" #-------#\n",
|
110 |
+
" for key in tmp:tmp[f'{key}'] = tmp[f'{key}'].contiguous()\n",
|
111 |
+
" lora[up] = tmp['u'].to(device = device , dtype=torch.float32)\n",
|
112 |
+
" lora[down] = tmp['Vh'].to(device = device , dtype=torch.float32)\n",
|
113 |
+
" #-------#\n",
|
114 |
+
" count = count + 2\n",
|
115 |
+
" print(f'{count} / {NUM_ITEMS}')\n",
|
116 |
+
" #-------#\n",
|
117 |
+
" print(f'done!')\n",
|
118 |
+
" print(f'casting params to fp16....')\n",
|
119 |
+
" for key in _lora: lora[f'{key}'] = lora[f'{key}'].to(device = device , dtype=torch.float16)\n",
|
120 |
+
" #-------#\n",
|
121 |
+
" print(f'done!')\n",
|
122 |
+
" print(f'saving {savefile_name}...')\n",
|
123 |
+
" save_file(lora , f'{savefile_name}')\n",
|
124 |
+
"#--------#\n",
|
125 |
"\n",
|
126 |
+
"def count_zeros(_lora, resolution):\n",
|
127 |
+
" count = 0\n",
|
128 |
+
" for key in _lora:count = count + 1\n",
|
129 |
+
" NUM_ITEMS = count\n",
|
130 |
+
" count = 0\n",
|
131 |
+
" #-----#\n",
|
132 |
+
" thresh = resolution*0.000001 # 1e-6\n",
|
|
|
|
|
|
|
133 |
"\n",
|
134 |
+
" print(f'at resolution = {resolution}e-6 :')\n",
|
135 |
+
" for key in _lora:\n",
|
136 |
+
" if f'{key}'.find('alpha') > -1:\n",
|
137 |
+
" count = count + 1\n",
|
138 |
+
" continue\n",
|
139 |
+
" #------#\n",
|
140 |
+
" if not f'{key}'.find('lora_down') > -1: continue\n",
|
141 |
+
" up = f'{key}'.replace('lora_down' , 'lora_up')\n",
|
142 |
+
" down = f'{key}'\n",
|
143 |
+
" #-------#\n",
|
144 |
+
" delta_W = torch.matmul(_lora[up],_lora[down]).to(device = device , dtype=torch.float32)\n",
|
145 |
+
" N = delta_W.numel()\n",
|
146 |
+
" y = delta_W.flatten().to(device = device , dtype=torch.float32)\n",
|
147 |
+
" values,indices = torch.sort(y, descending = False) # smallest -> largest elements\n",
|
148 |
+
" y = torch.ones(y.shape).to(device = device , dtype=torch.float32)\n",
|
149 |
+
" y[indices[values>thresh]] = 0\n",
|
150 |
+
" neg_pcnt = round((100*torch.sum(y) / N).item(),2)\n",
|
151 |
+
" y[indices[values<-thresh]] = 0\n",
|
152 |
+
" count = count + 2\n",
|
153 |
+
" pcnt = round((100*torch.sum(y) / N).item(),2)\n",
|
154 |
+
" neg_pcnt = round(neg_pcnt - pcnt,2) # remove zero % from neg_pcnt\n",
|
155 |
+
" pos_pcnt = round(100- pcnt - neg_pcnt,2)\n",
|
156 |
+
" print(f'at {count} / {NUM_ITEMS} : {pcnt} % zeros ,{pos_pcnt} % pos. , {neg_pcnt} % neg ')\n",
|
157 |
+
" #------#\n",
|
158 |
+
"#-----#\n",
|
159 |
+
"\n",
|
160 |
+
"def merge_and_save(_lora1 , _lora2 , _lora3, savefile_name, new_rank , new_alpha, resolution):\n",
|
161 |
+
" lora = {}\n",
|
162 |
+
" count = 0\n",
|
163 |
+
" for key in _lora1:count = count + 1\n",
|
164 |
+
" NUM_ITEMS = count\n",
|
165 |
+
" count = 0\n",
|
166 |
+
" thresh = resolution*0.000001 # 1e-6\n",
|
167 |
+
"\n",
|
168 |
+
" #-------#\n",
|
169 |
+
" for key in _lora1:\n",
|
170 |
+
" if f'{key}'.find('alpha') > -1:\n",
|
171 |
+
" lora[f'{key}'] = torch.tensor(new_alpha).to(device = device , dtype = torch.float32)\n",
|
172 |
+
" count = count + 1\n",
|
173 |
+
" print(f'{count} / {NUM_ITEMS}')\n",
|
174 |
+
" continue\n",
|
175 |
+
" #------#\n",
|
176 |
+
" #if count<462:\n",
|
177 |
+
" # count = count + 2\n",
|
178 |
+
" # continue\n",
|
179 |
+
" if not f'{key}'.find('lora_down') > -1: continue\n",
|
180 |
+
" up = f'{key}'.replace('lora_down' , 'lora_up')\n",
|
181 |
+
" down = f'{key}'\n",
|
182 |
+
" #-------#\n",
|
183 |
+
"\n",
|
184 |
+
" # Setup\n",
|
185 |
+
" delta_W = torch.matmul(_lora1[up]*0,_lora1[down]*0).to(device = device, dtype=torch.float32)\n",
|
186 |
+
" tgt_shape = delta_W.shape\n",
|
187 |
+
" N = delta_W.numel()\n",
|
188 |
+
" delta_W = torch.zeros(N).to(device = device , dtype=torch.float32)\n",
|
189 |
+
" #-----#\n",
|
190 |
+
"\n",
|
191 |
+
" #Positives\n",
|
192 |
+
" Y = torch.zeros(3,N).to(device = device , dtype=torch.float32)\n",
|
193 |
+
" Y[0] = torch.matmul(_lora1[up],_lora1[down]).flatten().to(device = device , dtype=torch.float32)\n",
|
194 |
+
" Y[1] = torch.matmul(_lora2[up],_lora2[down]).flatten().to(device = device , dtype=torch.float32)\n",
|
195 |
+
" Y[2] = torch.matmul(_lora2[up],_lora2[down]).flatten().to(device = device , dtype=torch.float32)\n",
|
196 |
+
" Y[torch.abs(Y)<thresh] = 0 #Trim\n",
|
197 |
+
" Y = torch.round(Y,decimals=6).t()\n",
|
198 |
+
" num_positives = torch.sum(Y>0,dim=1) + 0.1\n",
|
199 |
+
" elect = torch.sum(Y<0,dim=1) + 0.1\n",
|
200 |
+
" elect = (num_positives>=elect)\n",
|
201 |
+
" Y[Y<0] = 0\n",
|
202 |
+
" Y = torch.sum(Y, dim=1).to(device = device , dtype=torch.float32)\n",
|
203 |
+
" delta_W[elect] = torch.round((Y[elect]/num_positives[elect]),decimals=6).to(device = device , dtype=torch.float32)\n",
|
204 |
+
" #-----#\n",
|
205 |
+
"\n",
|
206 |
+
" #Negatives\n",
|
207 |
+
" Y = torch.zeros(3,N).to(device = device , dtype=torch.float32)\n",
|
208 |
+
" Y[0] = torch.matmul(_lora1[up],_lora1[down]).flatten().to(device = device , dtype=torch.float32)\n",
|
209 |
+
" Y[1] = torch.matmul(_lora2[up],_lora2[down]).flatten().to(device = device , dtype=torch.float32)\n",
|
210 |
+
" Y[2] = torch.matmul(_lora2[up],_lora2[down]).flatten().to(device = device , dtype=torch.float32)\n",
|
211 |
+
" Y[torch.abs(Y)<thresh] = 0 #Trim\n",
|
212 |
+
" Y = torch.round(Y,decimals=6).t()\n",
|
213 |
+
" num_negatives = torch.sum(Y<0,dim=1) + 0.1 # <-- to prevent divide by 0\n",
|
214 |
+
" elect = torch.sum(Y>0,dim=1) + 0.1\n",
|
215 |
+
" elect = (elect<num_negatives)\n",
|
216 |
+
" Y[Y>0] = 0\n",
|
217 |
+
" Y = torch.sum(Y, dim=1).to(device = device , dtype=torch.float32)\n",
|
218 |
+
" delta_W[elect] = torch.round(Y[elect]/num_positives[elect],decimals=6).to(device = device , dtype=torch.float32)\n",
|
219 |
+
" #----#\n",
|
220 |
+
"\n",
|
221 |
+
" # Free up memory prior to SVD\n",
|
222 |
+
" delta_W = delta_W.unflatten(0,tgt_shape).to(device = device , dtype=torch.float32)\n",
|
223 |
+
" delta_W = delta_W.clone().detach()\n",
|
224 |
+
" Y = {}\n",
|
225 |
+
" num_positives = {}\n",
|
226 |
+
" num_negatives = {}\n",
|
227 |
+
" elect = {}\n",
|
228 |
+
" #-----#\n",
|
229 |
+
"\n",
|
230 |
+
" # Run SVD (Single Value Decomposition)\n",
|
231 |
+
" #to get the new lora_up and lora_down for delta_W\n",
|
232 |
+
" tmp={}\n",
|
233 |
+
" tmp['u'], tmp['s'], tmp['Vh'] = torch.svd(delta_W)\n",
|
234 |
+
" tmp['u'] = tmp['u'][:,: new_rank]\n",
|
235 |
+
" tmp['s'] = tmp['s'][: new_rank]\n",
|
236 |
+
" tmp['u'] = torch.matmul(tmp['u'], torch.diag(tmp['s']))\n",
|
237 |
+
" tmp['Vh'] = tmp['Vh'].t()[: new_rank,:]\n",
|
238 |
+
" for key in tmp:tmp[f'{key}'] = tmp[f'{key}'].contiguous()\n",
|
239 |
+
" lora[up] = torch.round(tmp['u'],decimals=6).to(device = device , dtype=torch.float32)\n",
|
240 |
+
" lora[down] = torch.round(tmp['Vh'],decimals=6).to(device = device , dtype=torch.float32)\n",
|
241 |
+
" #-------#\n",
|
242 |
"\n",
|
243 |
+
" count = count + 2\n",
|
244 |
+
" print(f'{count} / {NUM_ITEMS}')\n",
|
245 |
+
" #----#\n",
|
246 |
+
" #--------#\n",
|
247 |
+
" print(f'done!')\n",
|
248 |
+
" print(f'casting params to fp16....')\n",
|
249 |
+
" for key in lora: lora[f'{key}'] = lora[f'{key}'].to(device = device , dtype=torch.float16)\n",
|
250 |
+
" #-------#\n",
|
251 |
+
" print(f'done!')\n",
|
252 |
+
" print(f'saving {savefile_name}...')\n",
|
253 |
+
" save_file(lora , f'{savefile_name}')\n",
|
254 |
+
"#------#\n",
|
255 |
+
"\n",
|
256 |
+
"new_rank = 32\n",
|
257 |
+
"new_alpha = math.floor(new_rank/2)\n",
|
258 |
+
"resolution = 200\n",
|
259 |
+
"name = 'yeero_euro_puff'\n",
|
260 |
+
"yeero = load_file('/kaggle/input/flux-loras/yeero_100_r32_16alpha.safetensors')\n",
|
261 |
+
"euro = load_file('/kaggle/input/flux-loras/euro_100_r32_16alpha.safetensors')\n",
|
262 |
+
"puff = load_file('/kaggle/input/flux-loras/puff_200_r32_16alpha.safetensors')\n",
|
263 |
+
"savefile_name = f'{name}_{resolution}_r{new_rank}_a{new_alpha}.safetensors'\n",
|
264 |
+
"\n",
|
265 |
+
"#tgt = load_file(f'/kaggle/input/flux-loras/{name}_{resolution}_r32_16alpha.safetensors')\n",
|
266 |
+
"for key in yeero:\n",
|
267 |
+
" yeero[f'{key}'] = yeero[f'{key}'].to(device = device , dtype = torch.float32)\n",
|
268 |
+
" euro[f'{key}'] = euro[f'{key}'].to(device = device , dtype = torch.float32)\n",
|
269 |
+
" puff[f'{key}'] = puff[f'{key}'].to(device = device , dtype = torch.float32)\n",
|
270 |
+
"#-----#\n",
|
271 |
+
"print(f'for {name}.safetensors at scale = (rank/alpha) = 0.5')\n",
|
272 |
+
"merge_and_save(yeero , euro , puff, savefile_name, new_rank , new_alpha, resolution)\n",
|
273 |
+
"\n",
|
274 |
+
"\n",
|
275 |
+
"#Yeero + Scale + Puff\n",
|
276 |
+
"#filter_and_save(tgt , f'{name}_{resolution}_r{new_rank}_{new_alpha}alpha.safetensors' , new_rank , new_alpha, resolution)\n"
|
277 |
],
|
278 |
"metadata": {
|
279 |
"id": "SKYzFxehkfG8"
|
|
|
378 |
"metadata": {
|
379 |
"id": "1oxeJYHRqxQC",
|
380 |
"collapsed": true,
|
|
|
381 |
"colab": {
|
382 |
"base_uri": "https://localhost:8080/"
|
383 |
+
},
|
384 |
+
"outputId": "12e3a407-f9d1-403e-949b-31330be59577"
|
385 |
},
|
386 |
"execution_count": 12,
|
387 |
"outputs": [
|
|
|
1401 |
"#print(torch.matmul(tgt[f'{up}'],tgt[f'{down}']).shape)\n"
|
1402 |
],
|
1403 |
"metadata": {
|
|
|
|
|
1404 |
"colab": {
|
1405 |
"base_uri": "https://localhost:8080/"
|
1406 |
+
},
|
1407 |
+
"id": "GoDfgENYaWD7",
|
1408 |
+
"outputId": "9336ae1a-6244-4e76-f291-82cda4482831"
|
1409 |
},
|
1410 |
"execution_count": 17,
|
1411 |
"outputs": [
|
|
|
1428 |
" print(oily[f'{key}'].shape)"
|
1429 |
],
|
1430 |
"metadata": {
|
|
|
|
|
1431 |
"colab": {
|
1432 |
"base_uri": "https://localhost:8080/"
|
1433 |
+
},
|
1434 |
+
"id": "xQhVLouEfmGE",
|
1435 |
+
"outputId": "662176b3-480d-48eb-f5db-97ec71b5e970"
|
1436 |
},
|
1437 |
"execution_count": 18,
|
1438 |
"outputs": [
|