>>1058 Show the error log and output of "pip list" inside sd_scripts venv

>>1063 I ended up getting one trainer to work and trained a few and it works fine my only complaint now with it is I can't figure out how to queue up multiple things for training assuming it can, have to re-investigate ai toolkit again maybe. Only error thing though on new one was the 10061 windows. Can't find that pip list either, dug out my old complaint list on what the sequence was I think >lora_easy_training_scripts\backend\main.py", line 2, in <module> >from starlette.applications import Starlette >ModuleNotFoundError: No Module named 'starlette' Can get into the gui regardless, load up settings and hit training >Failed to establish a new connection: [WinError 10061] No connection could be made because the target machine actively refused it Went and installed starlette, still didn't change the message. This is all of fresh clone I did nothing to it since it had cuda 13 already. new version now on this roll back is it's missing "uvicorn" same thing though, lets me into gui, runs the training, fails on 10061 connection error.

>>1064 Easy scripts are using primitive 8000 default port to connect front and back, I don't think you can change that, so I suppose it's a derrian flaw for choosing popular port. If you are running something else at the same time, for example sillytavern with default port it will have a conflict, so check the port. You should put "pip list" inside cmd, just open it inside backend\sd_scripts, write venv\scripts\activate and then pip list

>>1065 anyo 4.11.0 certifi 2025.11.12 filelock 3.19.1 fsspec 2025.9.0 h11 0.16.0 httpcore 1.0.9 httpx 0.28.1 idna 3.11 itsdangerous 2.2.0 Jinja2 3.1.6 MarkupSafe 2.1.5 mpmath 1.3.0 networkx 3.5 numpy 2.3.3 pillow 11.3.0 pip 25.3 python-mulipart 0.0.20 PyYAML 6.0.3 setuptools 70.2.0 sniffio 1.3.1 starlette 0.50.0 sympy 1.14.0 torch 2.9 torchaudio 2.9.0 torchvision 0.24.0 typing_extension 4.15.0

>>1066 If this is your venv inside sd_scripts folder, then it's completely wrong, there should be alot more of packages and torch with torchvision should also have +cu128 or simillar appendix, for example: torch 2.7.0+cu128. Your torch is without cuda support. Idk just reinstall the thing ig

>>1067 That is a fresh install, that's what its' pulling. >git clone https://github.com/derrian-distro/LoRA_Easy_Training_Scripts >cd LoRA_Easy_Training_Scripts >install.bat New folder raw, naked install that's what it gives me on drop.

>>1068 >>1067 So now that I kinda have an idea about that backend venv I didn't know existed the >could not find install.py >python.exe installer.py local returned non-zero Makes more sense, though if I run python in there it says "requirements already satisfied" on running any of the pip setups for 12.8 or higher from torch's website.

>>1069 Further investigative journalism assuming it's not just something completely fucked up on my end, if I install the commit at Sept 15th, it installs everything fine but says no xformers and won't run without it for some reason. >Install xformers doesn't work Whatever, if I install anything past that commit it full bricks the backend install and gives me the same errors I had earlier

>>1070 Got run around for an hour at this point by google AI trouble shooting to end up at this stupid loop of >ImportError: No xformers But if I do a pip check on xformers it's in the venv, I also remade the venv in 3.10 since I have system 3.12 which I think was causing part of the issue. Either way got to the point where google ai said good luck and peaced out on me

>>1072 Basically google has come to the conclusion it's a circular firing squad where I can't get xformers to work cause I need to use torch 2.9.0 + cu130, and if I disable xformers in the training window to not use it throws me a cuda error no kernel image is available. Jesus christ kill me I just want to train some fucking loras

Here is some contributions to balance the bitching.

>>1074 >Here is some contributions to balance the bitching At this point, having a proper technical discussion in an AI thread is its own reward. Thanks for the pic tho.

>>1075 I don't even feel good about it because it was a net gain of zero besides rip 3 hours. I got One Trainer to work for vpred training cause honestly I'd rather use ai toolkit but I couldn't figure out how to enable vpred, though if we get that noob update to eps sometime soon it won't matter or wait on the vpred of that. But going back the one trainer thing it's hard to tell what the fuck it's doing in the scope, have to train some wildly different artists to see the learning cause the ones I was running weren't insanely different but it was having a hard time getting details it was just getting a vague genericness to all the sets.

>>1072 >3.10 since I have system 3.12 It's probably better to just use 3.10 for everything old like xl, if you don't know what you are doing >xformers It was always painfull on windows, not every pytorch version has xformers version builded for it on pypi, so either you building it yourself, or finding one that was builded for specific version of pytorch. You can also change that to --sdpa anyway You can run through all the versions here https://pypi.org/project/xformers/#history with command like pip install xformers==0.0.34.dev1100 --index-url https://download.pytorch.org/whl/cu130 to try install any compatible with your pytorch and cuda version, but it's better to find already builded wheel, like panchito did here https://huggingface.co/Panchovix/xformers-windows-blackwell2.0-nightly/tree/main but it's for 3.12 2.7.0 12.8 probably, and will not be compatible with your venv Overall 5000 is a disaster for anyone not familiar with all this bullshit, when devs are just tired of updating their stuff, you just need to know how to do it yourself. But looks like derrian updated easy scripts https://github.com/derrian-distro/LoRA_Easy_Training_scripts_Backend/commit/162e56de18e4d4ab96b066380606d657a68c66fa so I don't know why you have problems installing it

>>1078 At first I was kinda bummed down that not only most anons did not find this place but, that some even did either not return or, as this anon points out: >>1028 >I only barely remembered where this one was Some simply forgot this thread (or the way to it) altogether. But boy is the attack on the 4chan's threads is still very much a thing. Bunker Thread rides on.

>>955 That one looks way more better, too much even probably

>>1081 Woah, yeah that's way better, it's way too clean compared to >>1043 but the lines are very cool

>>1082 Well I said before even starting that it's gonna perform worse as a dedicated lora

>>1083 I wonder why are you doing 1280x1856 on local?

>>1083 So you end up doing another iteration of your finetune? rad

>>1084 I was probably testing something that I don't remember and liked that res more, than 1248x1824 or 2x, haven't even changed that since >>1085 No, it's just a fat lora, I just added betabeet there too

>>1086 >I was probably testing something that I don't remember and liked that res more, than 1248x1824 or 2x, haven't even changed that since Interesting because it has different aspect ratio from the 832x1216 res, so you either have to crop or slightly stretch the pic

>>1086 >No, it's just a fat lora, Oh I see, it's another version of your "regular" lora then? or a new approach? will it be able to do the same styles as the icbinf-extract?

>>1087 It's very insignificant, can't even tell if it was cropped or stretched >>1088 It's another version, yeah, not all styles transferred there https://files.catbox.moe/f0wn49.txt alot was cut

>>1089 Forgot to change it back, well it really looks just a little bigger

>>1089 Looking good, hope it will be flexible enough

>>1091 I know that it's ends up with forgetting some of the other artists that was working on noob with the latest versions, but trying to train less doesn't output good style results

>>1077 Yeah I dont' get it either I saw that one, but it fails installing the backend completely on anything recent and those pre that commit it bricks somewhat, I found one that did a full install so I grabbed it the pushed it to head with the full install and it still doesn't work. Believe me there's no reason it shouldn't work but it's not working and it's frustating beyond reason, only thing at this point is it's some forced python shit with 3.12 which I installed for some reason I can't even remember why now, I'm not sure I'm even using it. I can uninstall it so the system doesn't try to force it and pray that's all I need to do but even forcing a 3.10 venv into the folder and building on that with the install gave failure. Yeah 50x is a disaster for sure if you're not familiar with this shit at all it's been rough. Doesn't look like it's getting better either the way recent trends are showing.

>>1077 Booted up reforge again to even check it to be sure and it's loading 3.10, so I dunno what the fuck I have 3.12 for maybe it was a stroke when I was installing python libraries to get this pre-worked. I will say though the gemini AI help was almost useful cause yeah it actually pulled the proper set ups / commits I needed for specific pytorch / xformers anyway with what the shit it thought was the problem, even a different numpy when that threw me an error. Now still again a failure since we're at where we started on the day, so guess next step is to uninstall 3.12, restart PC and see if it just wants to play fresh install venv. If that doesn't work how do you force it to sdpa cause I think part of it too is like I think I mentioned is 50x doesn't have / use xformers from what I saw? If you wanna humor me even more honestly git bash a fresh install of that and just let it run cause I'm curious if it's just a broken to launch installer and I got no way to test that since I have the single pc.

>>1094 You should also install cuda toolkit https://developer.nvidia.com/cuda-downloads I have 12.8 installed a while ago and can confirm that picrel versions works with dummy venv, installed them with pip install torch==2.7.0 torchvision xformers --index-url https://download.pytorch.org/whl/cu128 Don't want to reinstall for 13.0, but it's should be the same process --sdpa is training arg, not installation, you can use either that one attention optimization or xformers

>>1094 Here, but 5000 series should have no problems with xformers either, once everything setup properly

>>1095 Don't know what that does for me but just tried another install with it and failed at the installer. I hadn't uninstalled 3.12 yet since it might be hijacking the process but I dunno shit is so annoying. And yeah when I toggle sdpa it's in the box but when I do that it throws some random missing depency like numpy or what have you that somehow got missed in installer. I guess at this point it's a literal "you" problem on my end cause no fixes have worked

>>1098 Now I might be smooth braining a bunch here but like the before / after on this shit didn't have these install310,311,312.sh files in here which are an enigma since original purely was the installer

>>1099 Also tangentially related I realized just now that xformers doesn't actually work on reforge currently with 50x, but people claim it still speeds up gens so was gonna force it and it won't force. So I was trying to install sage-attention and getting my ass handed again because I'm not a coder. And my issue with >>1095 was literally just the base installer ripped from the github run, just has a failure at the backend portion that even when I seemed to install all the missing pieces it's not working which assuming you did install fresh is just a fuck me I guess situation that I have no idea what to do about

>>1096 And are you running on win11 by chance just making sure to figure out how fucked my shit is