Add support for the latest GPTQ models with group-size #530

oobabooga · 2023-03-24T03:52:20Z

Example: https://huggingface.co/ozcur/alpaca-native-4bit

Usage:

python server.py --model alpaca-native-4bit --wbits 4 --groupsize 128

sgsdxzy · 2023-03-24T05:04:40Z

I think maybe you can make groupsize a parameter that defaults to 128, not a hard-coded one. That can also support -1 to load old 4bit models.

Ph0rk0z · 2023-03-24T05:12:09Z

I tried loading old 4 bit models with -1 and it would error. Thought everything has to be re-quantized for it to work with the new GPTQ.

oobabooga · 2023-03-24T05:25:49Z

This need for re-quantization is a killer when combined with the fact that LLaMA is not a publicly available model that we can just re-quantize and upload to Hugging Face freely.

Ph0rk0z · 2023-03-24T05:48:28Z

The alpaca native can just be quantized the old way. I didn't see anyone delete any of the PT files from hugging face, have they been doing that?

Is the new GPTQ any faster?

RandomInternetPreson · 2023-03-24T07:05:49Z

https://huggingface.co/ozcur/alpaca-native-4bit

This guy has quantized the alpaca-native model from chavinlo.

Ph0rk0z · 2023-03-24T11:48:38Z

I wish they made a 13b.. 7b just runs fast on everything.

USBhost · 2023-03-24T19:27:57Z

GPTQ 4bit does not load if it was made with act-order. I am currently testing true-sequential if that's okay.
Edit: true-sequential loads so only act-order is broken.

USBhost · 2023-03-24T20:27:49Z

modules/GPTQ_loader.py


    if not pt_path:
        print(f"Could not find {pt_model}, exiting...")
        exit()

    # qwopqwop200's offload
    if shared.args.gptq_pre_layer:
-        model = load_quant(str(path_to_model), str(pt_path), shared.args.gptq_bits, shared.args.gptq_pre_layer)
+        model = load_quant(str(path_to_model), str(pt_path), shared.args.gptq_bits, 128, shared.args.gptq_pre_layer)
    else:


Could these 128 break act-order? GPTQ states "Currently, groupsize and act-order do not work together and you must choose one of them." So I would imagen using 128 when you are not supposed to will cause issues.

This function is imported from here: https://github.com/qwopqwop200/GPTQ-for-LLaMa/blob/main/llama_inference.py#L26

Maybe it hasn't been updated yet to work with act-order?

I changed both 128 to -1 and I was able to load act-order. So I guess just make groupsize a parameter as @sgsdxzy said. If act-order becomes the default just have groupsize default to -1 as it already does on GPTQ.

That's good to hear! I will make it a parameter.

oobabooga · 2023-03-24T22:38:16Z

I have added the --gptq-group-size parameter with the value set to -1 by default.

Should this be a paramater? If it has to match the group size used during quantization, it would be better to store this number in a config.json that is distributed with the model. But we can leave it like this for now.

USBhost · 2023-03-24T22:43:54Z

We could follow the naming of the GPTQ example https://github.com/qwopqwop200/GPTQ-for-LLaMa/blob/main/README.md?plain=1#L119 ? Having to use a file just for one parameter seems wasteful.

oobabooga · 2023-03-24T23:13:19Z

Sure, I have renamed all parameters and added deprecation warnings to the old parameter names.

TwitchPid · 2023-03-24T23:26:10Z

Error with command python server.py --model alpaca-native-4bit --wbits 4 --model_type llama --groupsize 128:

Loading alpaca-native-4bit...
Traceback (most recent call last):
File "/run/media/user/disk/text-generation-webui/server.py", line 234, in
shared.model, shared.tokenizer = load_model(shared.model_name)
File "/run/media/user/disk/text-generation-webui/modules/models.py", line 101, in load_model
model = load_quantized(model_name)
File "/run/media/user/disk/text-generation-webui/modules/GPTQ_loader.py", line 64, in load_quantized
model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits)
TypeError: load_quant() missing 1 required positional argument: 'groupsize'

I'm assuming it has to do with --pre_layer

oobabooga · 2023-03-24T23:30:42Z

You are right, I messed up. Can you see if it works now?

TwitchPid · 2023-03-24T23:37:13Z

The model loads fine, I think GPTQ-for-LLaMA is the problem: https://pastebin.com/irGyfc6L

TwitchPid · 2023-03-24T23:49:20Z

Update: works! Used git reset --hard 5cdfad2a15dffebad5d1c24b443bc5dc291d8372 on GPTQ-for-LLaMA (branch that alpaca-native-4bit used)

oobabooga · 2023-03-30T00:05:40Z

@USBhost does that require the (currently) slow Pytorch version of GPTQ-for-LLaMa at inference time?

USBhost · 2023-03-30T00:08:25Z

@USBhost does that require the (currently) slow Pytorch version of GPTQ-for-LLaMa at inference time?

Yeah... But if you live in 128+ context it's faster. Like character cards on cai-chat

USBhost · 2023-03-30T00:25:00Z

Kind of funny you call it slow, here in 65b lands our delay was 50+ seconds

borriodelrio · 2023-03-30T01:36:31Z

Hello, a few questions if someone has the time, and I apologize if this isn't the place for this. I've gone over the new text-generation-webui install documentation. Am I correct in my understanding that this is only for Windows/Linux machines at this time due to the need of Cuda (Nvidia)?

The only model compatible with text-generation-webui on Apple silicon seems to be Chavinlo's 7b Alpaca-Native (too slow for practical use on 16GB RAM).

For Apple Silicon there appear to be two options:

[CPU] Llama.cpp/Alpaca.cpp with GGML
[GPU] LLaMA_MPS (https://github.com/jankais3r/LLaMA_MPS), - this project seems as though it could make use of Apple's Neural Engine (ANE) transformers

I've seen little interest and documentation regarding getting Alpaca or LLAMA Apple Silicon. If anyone has additional information, I'd be interested

mjlbach · 2023-03-30T07:46:57Z

Has anyone documented the RAM (not VRAM) costs for loading these? I don't see anything in the wiki, but it currently crashes for me when I try to load MetaIX_Alpaca_30B python server.py --wbits 4 --model MetaIX_Alpaca-30B-Int4 on a 3090 system with only 16gb of RAM

neuhaus · 2023-03-30T08:42:05Z

@mjlbach i suggest you add some swapspace (perhaps 32GB) and then run the command with "/usr/bin/time -v" in front of it to determine the peak memory use.

Ph0rk0z · 2023-03-30T14:25:26Z

@USBhost does that require the (currently) slow Pytorch version of GPTQ-for-LLaMa at inference time?

Yeah... But if you live in 128+ context it's faster. Like character cards on cai-chat

Can I use act order + true sequential in the cuda kernel implementation without group size for both conversion and inference?

I have a lot of models to re-do because I was using GPT-J/Neox and opts. It was doing ok before these new additions and the models weren't exactly "dumb". My card is old so the difference for me will be much larger, probably to the point where chatting becomes a huge frustration.

USBhost · 2023-03-30T14:50:27Z

@USBhost does that require the (currently) slow Pytorch version of GPTQ-for-LLaMa at inference time?

Yeah... But if you live in 128+ context it's faster. Like character cards on cai-chat

Can I use act order + true sequential in the cuda kernel implementation without group size for both conversion and inference?

I have a lot of models to re-do because I was using GPT-J/Neox and opts. It was doing ok before these new additions and the models weren't exactly "dumb". My card is old so the difference for me will be much larger, probably to the point where chatting becomes a huge frustration.

Yes the only issue with the cuda kernel you can not use act-order and group size together.

Highlyhotgames · 2023-03-31T10:39:30Z

Groupsize 128 version. TLDR this torrent 13b up is slightly better than the baseline one but also uses slightly more VRAM.

magnet:?xt=urn:btih:88f7d9d2460ffcaf78b21e83012de00939eacb65&dn=LLaMA-HF-4bit-128g&tr=http%3a%2f%2fbt2.archive.org%3a6969%2fannounce&tr=http%3a%2f%2fbt1.archive.org%3a6969%2fannounce

Or unzip for the torrent file. LLaMA-HF-4bit-128g.zip
for f in *; do sha256sum $f/*.safetensors; done
d3073ef1a2c0b441f95a5d4f8a5aa3b82884eef45d8997270619cb29bcc994b8  llama-13b-4bit-128g/llama-13b-4bit-128g.safetensors
8b7d75d562938823c4503b956cb4b8af6ac0a5afbce2278566cc787da0f8f682  llama-30b-4bit-128g/llama-30b-4bit-128g.safetensors
f1418091e3307611fb0a213e50a0f52c80841b9c4bcba67abc1f6c64c357c850  llama-65b-4bit-128g/llama-65b-4bit-128g.safetensors
ed8ec9c9f0ebb83210157ad0e3c5148760a4e9fd2acfb02cf00f8f2054d2743b  llama-7b-4bit-128g/llama-7b-4bit-128g.safetensors
Edit: Below is the original post with more information.

If Your on the the main branch to try out this pull do the following:
git revert dcfd866402dfbbc849bd4441fd1de9448de18c75
git pull origin --no-ff pull/530/head
Here's my second torrent using groupsize 128 + true-sequentia. Do to groupsize it will be slightly bigger than normal. However It is better but how much idk... I'm just a noob with a old server lol. Lower is better: 7B is included for completeness. So only use this torrent for 13b up.
**7B**
wikitext2 6.259988784790039
wikitext2 6.237235069274902 (groupsize 128)
ptb-new 10.817036628723145
ptb-new 11.199039459228516 (groupsize 128)
c4-new 7.802077293395996
c4-new 8.000247955322266 (groupsize 128)

**13B**
wikitext2 5.341851711273193
wikitext2 5.242600440979004 (groupsize 128)
ptb-new 9.474738121032715
ptb-new 9.225408554077148 (groupsize 128)
c4-new 7.071592330932617
c4-new 6.912217617034912 (groupsize 128)

**30B**
wikitext2 4.45449686050415
wikitext2 4.230341911315918 (groupsize 128)
ptb-new 8.377615928649902
ptb-new 8.243087768554688 (groupsize 128)
c4-new 6.390762805938721
c4-new 6.231330394744873 (groupsize 128)

**65B**
wikitext2 3.8416879177093506
wikitext2 3.658999443054199 (groupsize 128)
ptb-new 7.881875991821289
ptb-new 7.780252456665039 (groupsize 128)
c4-new 5.998412609100342
c4-new 5.896479606628418 (groupsize 128)
magnet:?xt=urn:btih:88f7d9d2460ffcaf78b21e83012de00939eacb65&dn=LLaMA-HF-4bit-128g&tr=http%3a%2f%2fbt2.archive.org%3a6969%2fannounce&tr=http%3a%2f%2fbt1.archive.org%3a6969%2fannounce

Or unzip for the torrent file. LLaMA-HF-4bit-128g.zip

To run this one you need to use groupsize 128: python server.py --listen --wbits 4 --cai-chat --groupsize 128 --model llama-7b-4bit-128g

After 3 consecutive days trying to find a way of running server.py on WSL I finally managed to get it working, so I'm developing a simple straightforward tutorial to share with the community, yet I want to include the download from this models (which are the ones that work flawlessly on this specific context) but downloading the torrent through the WSL terminal is a pain as idk how to specifically download only the required model instead of downloading the complete torrent. So my question is if this files are available elsewhere (like HF for example) or if someone knows how can I specifically download only one of the folders (models) from the torrent (through WSL terminal) I'll greatly appreciate it because the tutorial is lacking that last step to be completed :) Keep up the amazing work guys

wywywywy · 2023-03-31T10:43:34Z

downloading the torrent through the WSL terminal is a pain as idk how to specifically download only the required model instead of downloading the complete torrent. So my question is if this files are available elsewhere (like HF for example) or if someone knows how can I specifically download only one of the folders (models) from the torrent (through WSL terminal) I'll greatly appreciate it because the tutorial is lacking that last step to be completed :) Keep up the amazing work guys

In Win11, you can run Linux GUI apps through WSL. Or you can run a torrent client with a web UI inside of WSL. That way it's much easier to manage the torrents and select only the files you need.

Highlyhotgames · 2023-03-31T11:16:50Z

downloading the torrent through the WSL terminal is a pain as idk how to specifically download only the required model instead of downloading the complete torrent. So my question is if this files are available elsewhere (like HF for example) or if someone knows how can I specifically download only one of the folders (models) from the torrent (through WSL terminal) I'll greatly appreciate it because the tutorial is lacking that last step to be completed :) Keep up the amazing work guys

In Win11, you can run Linux GUI apps through WSL. Or you can run a torrent client with a web UI inside of WSL. That way it's much easier to manage the torrents and select only the files you need.

Thanks! I'm using Win10 latest version so as it uses WSL2 (just like Win11) it will run Linux GUI apps too. Yet, I want this tutorial to be very straightforward so a simple command line to retrieve only the needed folder from the torrent would be great! I was researching on ways to make this though.. but through command line I can only find info on how to download the torrent on its entirety, so if someone knows how to do this through linux terminal please let me know. Oh, and sorry if this appears to be out of context, but I thought asking here where the torrent link is so it would save time for someone.

USBhost · 2023-04-05T05:09:59Z

My last update for visibility: #530 (comment)
I have updated Stock LLaMA to include the latest tokenizer fixes. See my edit for info. Anyways I will be uploading them to HF tomorrow.

Edit: https://huggingface.co/Neko-Institute-of-Science

Highlyhotgames · 2023-04-07T18:24:26Z

downloading the torrent through the WSL terminal is a pain as idk how to specifically download only the required model instead of downloading the complete torrent. So my question is if this files are available elsewhere (like HF for example) or if someone knows how can I specifically download only one of the folders (models) from the torrent (through WSL terminal) I'll greatly appreciate it because the tutorial is lacking that last step to be completed :) Keep up the amazing work guys

In Win11, you can run Linux GUI apps through WSL. Or you can run a torrent client with a web UI inside of WSL. That way it's much easier to manage the torrents and select only the files you need.

Thanks! I'm using Win10 latest version so as it uses WSL2 (just like Win11) it will run Linux GUI apps too. Yet, I want this tutorial to be very straightforward so a simple command line to retrieve only the needed folder from the torrent would be great! I was researching on ways to make this though.. but through command line I can only find info on how to download the torrent on its entirety, so if someone knows how to do this through linux terminal please let me know. Oh, and sorry if this appears to be out of context, but I thought asking here where the torrent link is so it would save time for someone.

Guys I finally made it. So now you can use the installation script I created to install this on WSL in a very easy way. The link is:
https://github.com/Highlyhotgames/fast_txtgen_7B

I'm very new to all of this github world, so comments - negative or positive - will be of great help! I want to include a part saying like "Thanks to oobabooga for the text-generation-webui, USBhost for the 4-bit quantized models and qwopqwop200 for the GPTQ-for-Llama" can anyone help me on how I should properly post this thanks message?

And I made a video tutorial too, but I don't even know if it's ok to share it here, I hope so! --> https://youtu.be/RcHIOVtYB7g

thistleknot · 2023-04-16T01:39:58Z

Loading alpaca-native-4bit...
Loading model ...
╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮
│ /mnt/distvol/text-generation-webui/server.py:308 in │
│ │
│ 305 │ │ i = int(input()) - 1 │
│ 306 │ │ print() │
│ 307 │ shared.model_name = available_models[i] │
│ ❱ 308 shared.model, shared.tokenizer = load_model(shared.model_name) │
│ 309 if shared.args.lora: │
│ 310 │ add_lora_to_model(shared.args.lora) │
│ 311 │
│ │
│ /mnt/distvol/text-generation-webui/modules/models.py:102 in load_model │
│ │
│ 99 │ elif shared.args.wbits > 0: │
│ 100 │ │ from modules.GPTQ_loader import load_quantized │
│ 101 │ │ │
│ ❱ 102 │ │ model = load_quantized(model_name) │
│ 103 │ │
│ 104 │ # llamacpp model │
│ 105 │ elif shared.is_llamacpp: │
│ │
│ /mnt/distvol/text-generation-webui/modules/GPTQ_loader.py:135 in load_quantized │
│ │
│ 132 │ │ model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.a │
│ 133 │ else: │
│ 134 │ │ threshold = False if model_type == 'gptj' else 128 │
│ ❱ 135 │ │ model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.a │
│ 136 │ │ │
│ 137 │ │ # accelerate offload (doesn't work properly) │
│ 138 │ │ if shared.args.gpu_memory: │
│ │
│ /mnt/distvol/text-generation-webui/modules/GPTQ_loader.py:63 in _load_quant │
│ │
│ 60 │ │ from safetensors.torch import load_file as safe_load │
│ 61 │ │ model.load_state_dict(safe_load(checkpoint), strict=False) │
│ 62 │ else: │
│ ❱ 63 │ │ model.load_state_dict(torch.load(checkpoint), strict=False) │
│ 64 │ model.seqlen = 2048 │
│ 65 │ print('Done.') │
│ 66 │
│ │
│ /mnt/distvol/python_user/python_user/gpt/lib/python3.9/site-packages/torch/serialization.py:795 │
│ in load │
│ │
│ 792 │ │ │ │ return _legacy_load(opened_file, map_location, _weights_only_unpickler, │
│ 793 │ │ │ except RuntimeError as e: │
│ 794 │ │ │ │ raise pickle.UnpicklingError(UNSAFE_MESSAGE + str(e)) from None │
│ ❱ 795 │ │ return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args │
│ 796 │
│ 797 │
│ 798 # Register pickling support for layout instances such as │
│ │
│ /mnt/distvol/python_user/python_user/gpt/lib/python3.9/site-packages/torch/serialization.py:1002 │
│ in _legacy_load │
│ │
│ 999 │ │ │ f"Received object of type "{type(f)}". Please update to Python 3.8.2 or ne │
│ 1000 │ │ │ "functionality.") │
│ 1001 │ │
│ ❱ 1002 │ magic_number = pickle_module.load(f, **pickle_load_args) │
│ 1003 │ if magic_number != MAGIC_NUMBER: │
│ 1004 │ │ raise RuntimeError("Invalid magic number; corrupt file?") │
│ 1005 │ protocol_version = pickle_module.load(f, **pickle_load_args) │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
UnpicklingError: invalid load key, 'v'.

thistleknot · 2023-04-16T03:02:09Z

fixed (lfs newbie)

	yum install git-lfs
	cd into cloned dir
	git lfs install
	git lfs pull
	git lfs smudge

**Warning: old 4-bit weights will not work anymore!** See here how to get up to date weights: https://github.com/oobabooga/text-generation-webui/wiki/LLaMA-model#step-2-get-the-pre-converted-weights

Bortus-AI · 2023-04-22T04:06:23Z

downloading the torrent through the WSL terminal is a pain as idk how to specifically download only the required model instead of downloading the complete torrent. So my question is if this files are available elsewhere (like HF for example) or if someone knows how can I specifically download only one of the folders (models) from the torrent (through WSL terminal) I'll greatly appreciate it because the tutorial is lacking that last step to be completed :) Keep up the amazing work guys

In Win11, you can run Linux GUI apps through WSL. Or you can run a torrent client with a web UI inside of WSL. That way it's much easier to manage the torrents and select only the files you need.

Thanks! I'm using Win10 latest version so as it uses WSL2 (just like Win11) it will run Linux GUI apps too. Yet, I want this tutorial to be very straightforward so a simple command line to retrieve only the needed folder from the torrent would be great! I was researching on ways to make this though.. but through command line I can only find info on how to download the torrent on its entirety, so if someone knows how to do this through linux terminal please let me know. Oh, and sorry if this appears to be out of context, but I thought asking here where the torrent link is so it would save time for someone.

Guys I finally made it. So now you can use the installation script I created to install this on WSL in a very easy way. The link is: https://github.com/Highlyhotgames/fast_txtgen_7B

I'm very new to all of this github world, so comments - negative or positive - will be of great help! I want to include a part saying like "Thanks to oobabooga for the text-generation-webui, USBhost for the 4-bit quantized models and qwopqwop200 for the GPTQ-for-Llama" can anyone help me on how I should properly post this thanks message?

And I made a video tutorial too, but I don't even know if it's ok to share it here, I hope so! --> https://youtu.be/RcHIOVtYB7g

That worked perfectly thank you!

gmm005 · 2023-05-02T13:42:15Z

What is meant by these do we need to update these lines of code after downloading . i have listened to your tutorial but could not successfully run it

oobabooga added 5 commits March 24, 2023 00:51

Update GPTQ_loader.py

ba6b5b5

Update download-model.py

edd5a7a

Fix offloading

f4b35ee

Update server.py

5fa482a

Add a default prompt for alpaca models

c7598a0

oobabooga mentioned this pull request Mar 24, 2023

8-bit precision not working on Windows #20

Closed

USBhost mentioned this pull request Mar 24, 2023

Move to updated GPTQ with new PTB and C4 eval #541

Closed

USBhost reviewed Mar 24, 2023

View reviewed changes

oobabooga added 2 commits March 24, 2023 19:32

Update shared.py

aabf072

Update GPTQ_loader.py

04eb089

oobabooga added 4 commits March 24, 2023 19:53

Update shared.py

40e0cab

Update GPTQ_loader.py

b865a41

Update models.py

4b9d45b

Update shared.py

9771017

Update GPTQ_loader.py

5dd9208

oobabooga mentioned this pull request Mar 25, 2023

Allow loading of .safetensors through GPTQ-for-LLaMa #529

Merged

EyeDeck mentioned this pull request Mar 30, 2023

make_quant() got an unexpected keyword argument 'faster' #667

Closed

1 task

VldmrB mentioned this pull request Mar 31, 2023

Strange generation speed #391

Closed

Roerib mentioned this pull request Mar 31, 2023

4bit model loads but doesn't output. UnboundLocalError: local variable 'reply' referenced before assignment #681

Closed

1 task

loeken mentioned this pull request Apr 1, 2023

creating a layer with Docker/docker-compose #633

Merged

strangePause mentioned this pull request Apr 1, 2023

Error while loading 4-bit model #708

Closed

1 task

AndreyRGW mentioned this pull request Apr 3, 2023

Help me... #745

Closed

oobabooga deleted the gptq-group-size branch April 4, 2023 15:52

oobabooga referenced this pull request Apr 7, 2023

Bump transformers (16-bit llama must be reconverted/redownloaded)

113f94b

darth-veitcher mentioned this pull request Apr 9, 2023

any advice for getting this running with gpt-x-alpaca models? 1b5d/llm-api#1

Open

Panchovix mentioned this pull request Apr 12, 2023

Multi GPU 4-bit doesn't work on either ooba GPTQ or qwopqwop200 GPTQ #1112

Closed

1 task

oobabooga mentioned this pull request May 4, 2023

Implement support for AutoGPTQ for loading GPTQ quantized models. #1668

Merged

oobabooga mentioned this pull request May 23, 2023

Improve CPU offload AutoGPTQ/AutoGPTQ#100

Merged

Add support for the latest GPTQ models with group-size #530

Add support for the latest GPTQ models with group-size #530

Conversation

oobabooga commented Mar 24, 2023 • edited Loading

sgsdxzy commented Mar 24, 2023

Ph0rk0z commented Mar 24, 2023

oobabooga commented Mar 24, 2023 • edited Loading

Ph0rk0z commented Mar 24, 2023

RandomInternetPreson commented Mar 24, 2023

Ph0rk0z commented Mar 24, 2023

USBhost commented Mar 24, 2023 • edited Loading

USBhost Mar 24, 2023

Choose a reason for hiding this comment

oobabooga Mar 24, 2023 • edited Loading

Choose a reason for hiding this comment

USBhost Mar 24, 2023 • edited Loading

Choose a reason for hiding this comment

oobabooga Mar 24, 2023

Choose a reason for hiding this comment

oobabooga commented Mar 24, 2023

USBhost commented Mar 24, 2023 • edited Loading

oobabooga commented Mar 24, 2023

TwitchPid commented Mar 24, 2023

oobabooga commented Mar 24, 2023

TwitchPid commented Mar 24, 2023

TwitchPid commented Mar 24, 2023

oobabooga commented Mar 30, 2023

USBhost commented Mar 30, 2023

USBhost commented Mar 30, 2023

borriodelrio commented Mar 30, 2023

mjlbach commented Mar 30, 2023 • edited Loading

neuhaus commented Mar 30, 2023

Ph0rk0z commented Mar 30, 2023

USBhost commented Mar 30, 2023

Highlyhotgames commented Mar 31, 2023

Edit: Below is the original post with more information.

wywywywy commented Mar 31, 2023

Highlyhotgames commented Mar 31, 2023

USBhost commented Apr 5, 2023 • edited Loading

Highlyhotgames commented Apr 7, 2023

thistleknot commented Apr 16, 2023

thistleknot commented Apr 16, 2023 • edited Loading

Bortus-AI commented Apr 22, 2023

gmm005 commented May 2, 2023

oobabooga commented Mar 24, 2023 •

edited

Loading

oobabooga commented Mar 24, 2023 •

edited

Loading

USBhost commented Mar 24, 2023 •

edited

Loading

oobabooga Mar 24, 2023 •

edited

Loading

USBhost Mar 24, 2023 •

edited

Loading

USBhost commented Mar 24, 2023 •

edited

Loading

mjlbach commented Mar 30, 2023 •

edited

Loading

USBhost commented Apr 5, 2023 •

edited

Loading

thistleknot commented Apr 16, 2023 •

edited

Loading