Furkan Gözükara
patreon
FLUX Models 1-Click Auto Downloaders for SwarmUI for Windows, RunPod and Massed Compute
🕑 Added 2024-08-14 21:11:00 +0000 UTCPatreon exclusive posts index to find our scripts easily, Patreon scripts updates history to see which updates arrived to which scripts and amazing Patreon special generative scripts list that you can use in any of your task.
Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord
Please also Star, Watch and Fork our Stable Diffusion & Generative AI GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)
=======
Automatic Black Forest Labs FLUX models 1-Click downloaders for Windows, RunPod, Massed Compute
The downloader scripts have resume capability, skip if fully downloaded capability and will auto retry 100 times if an error occurs and continue from wherever left automatically
15 August 2024 Update
Downloader scripts file updated to v6 : FLUX_models_Auto_Downloaders_v7.zip
Newest optimized and quantized models added to downloads like below
Now you can overwrite previously automatically downloaded T5 FP8 text encoder
Comprehensive model and precision tests have been made and published at below post
2 August 2024 Update
aws FP8 model weights downloader added - If you are under 30 GB GPU use them no difference
FP16 T5 XXL clip downloader added - 24 GB can use this
By default SwarmUI downloads FP8 version of T5 text encoder
More comparison tests zip file added
How To Use
Download and extract zip file in the attachments into SwarmUI parent folder like below
Full Windows, Massed Compute and RunPod tutorials here : https://www.patreon.com/posts/106135985
For Kaggle we already have updated notebook here : https://www.patreon.com/posts/106650931
I have compared FP8 vs FP16 mode and Turbo model in below imgsli in details
Running at FP16 requires 28 GB VRAM but produces better quality
FP8 vs FP16 vs Turbo vs FP16 T5 model detailed comparison (16 different tests) : https://imgsli.com/MjgzODYy/7/6
Turbo model and dev model has same VRAM usage and per step speed
Only difference is number of necessary steps needed
When running FP16 model at FP8 it gives identical results as FP8 version - half size
How to run the model at FP16 shown below
Comments
Furkan Gözükara
this tutorial fully explains sorry for late reply : https://youtu.be/bupRePUOA18
Dylan Normandin
i don't get what to do after downloading installer into SwarUI folder. I am on massed compute.
Furkan Gözükara
use most up to date swarmui. comfyui is harder : https://youtu.be/bupRePUOA18
Furkan Gözükara
yep that should work best
Steve
after getting it up ran into different issue: Invalid operation: ComfyUI execution error: Error(s) in loading state_dict for Flux: size mismatch for img_in.weight: copying a param with shape torch.Size([98304, 1]) from checkpoint, the shape in current model is torch.Size([3072, 64]). size mismatch for time_in.in_layer.weight: copying a param with shape torch.Size([393216, 1]) from checkpoint, the shape in current model is torch.Size([3072, 256]). size mismatch for time_in.out_layer.weight: copying a param with shape torch.Size([4718592, 1]) from checkpoint, the shape in current model is torch.Size([3072, 3072]). size mismatch for vector_in.in_layer.weight: copying a param with shape torch.Size([1179648, 1]) from checkpoint, the shape in current model is torch.Size([3072, 768]). size mismatch for vector_in.out_layer.weight: copying a param with shape torch.Size([4718592, 1]) from checkpoint, the shape in current model is torch.Size([3072, 3072]). size mismatch for guidance_in.in_layer.weight: copying a param with shape torch.Size([393216, 1]) from checkpoint, the shape in current model is torch.Size([3072, 256]). size mismatch for guidance_in.out_layer.weight: copying a param with shape torch.Size([4718592, 1]) from checkpoint, the shape in current model is torch.Size([3072, 3072]). size mismatch for txt_in.weight: copying a param with shape torch.Size([6291456, 1]) from checkpoint, the shape in current model is torch.Size([3072, 4096]). size mismatch for double_blocks.0.img_mod.lin.weight: copying a param with shape torch.Size([28311552, 1]) from checkpoint, the shape in current model is torch.Size([18432, 3072]). size mismatch for double_blocks.0.img_attn.qkv.weight: copying a param with shape torch.Size([14155776, 1]) from
Steve
never mind found it in your benchmarks file: "If you have a 12 GB VRAM or below GPU use flux1-dev-bnb-nf4-v2 - 30 steps"
Steve
any suggestion for which option for 12gb? (4070)
Furkan Gözükara
hello it is true. you need to delete pre-downloaded T5 and redownload. when you first time generate an image swarmui auto downloads fp8 version of T5. I am hopefully going to update downloader and fix this issue and add new versions as well like n4f and merged of turbo and dev version.
N V
Hi. Everything is very well documented. Thank you. I did get this error when trying to use the downloader for the extra encoder. Here is the error: Attempting to download t5xxl_fp16.safetensors from comfyanonymous/flux_text_encoders... t5xxl_fp16.safetensors: 100%|█████████████████████████████████████████████████████| 9.79G/9.79G [08:44<00:00, 18.7MB/s] Error occurred on attempt 1 for t5xxl_fp16.safetensors from comfyanonymous/flux_text_encoders: [WinError 183] Cannot create a file when that file already exists: 'E:\\AI 2024\\SD 2024\\Swarm\\SwarmUI\\Models\\clip\\t5xxl_fp16.safetensors' -> 'E:\\AI 2024\\SD 2024\\Swarm\\SwarmUI\\Models\\clip\\t5xxl_enconly.safetensors' Retrying in 1 seconds... Error occurred on attempt 2 for t5xxl_fp16.safetensors from comfyanonymous/flux_text_encoders: [WinError 183] Cannot create a file when that file already exists: 'E:\\AI 2024\\SD 2024\\Swarm\\SwarmUI\\Models\\clip\\t5xxl_fp16.safetensors' -> 'E:\\AI 2024\\SD 2024\\Swarm\\SwarmUI\\Models\\clip\\t5xxl_enconly.safetensors' Retrying in 2 seconds...
Furkan Gözükara
use scripts inside Massed_Compute_Downloaders folder
Furkan Gözükara
you dont need to use massed compute instructions
Furkan Gözükara
massed compute installer doesnt check it. it directly downloads into accurate folder.
James Kokonas
When I rename the folder to get the downloader to run, it errors at: "ModuleNotFoundError: No module named 'dataclasses'" and then quits. Any ideas?
James Kokonas
I'm noticing another error which claims I'm not running the autodownloader in the right folder, because they've renamed the base folder from SwarmUI -> StableSwarmUI
Furkan Gözükara
hello sorry for delay. they renamed files. so i updated zip files. please redownload FLUX_models_Auto_Downloaders_v5.zip
FliX
Hi. Unfortunatly your script FLUX Auto download Models doesn't work. I got this error Entry Not Found for url: https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/ae.sft. Retrying in 1 seconds... Error occurred on attempt 2 for ae.sft from black-forest-labs/FLUX.1-schnell: 404 Client Error. (Request ID: Root=1-66b22ae6-301ae63614e9dfc309f16ef1;e729ca78-800f-48e0-a645-0d31df3842a7)
Furkan Gözükara
you are welcome
Tobe2d
Thank you!
Furkan Gözükara
nope it won't this is completely different architecture. but i will look for it if possible
Furkan Gözükara
added to attachments please refresh the page
Tobe2d
Where can I find SwarmUI 1-Click installer?
Guillaume
Would it be possible to have Flux 4steps fp8 compatible with Supir? It does load in and I tried to make it work by making a few modification to the files, but no luck so far.
Furkan Gözükara
I am looking forward to it but not possible yet
楠 陈
How to fine-tune FLUX
Furkan Gözükara
hello updated with new zip file and info
Furkan Gözükara
I am about to update with more accurate info and new files few minutes please
Jeff Penny
The downloader doesn't let you select option 4 for FP8, you can only type 1, 2 or 3. I just manually ran python FLUX_1_dev_FP8_Download.py and that seems to work.
