Home Creators Posts Import Register Favorites Logout

Downloads

Content

NVFP4 Model Converter app quantizes FLUX models to NVFP4, which runs 100%+ faster than BF16, FP8 or GGUF; the FLUX SRPO NVFP4 result is only 6.32 GB.

Based on NVIDIA Model-Optimizer with major fixes; V1.2 adds Mixed NVFP4 quantization for FLUX. Compiling FLUX needs a 48 GB GPU, so SimplePod, RunPod and Massed Compute instructions are included besides Windows and Linux.

Patreon 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)

=======

Latest installer zip file : NVFP4_Quantizer_v2.zip

  • NVFP4 models are 100+% faster than BF16 or FP8 or GGUF models.

  • However compiling them is not trivial task at all

  • This app is based on https://github.com/NVIDIA/Model-Optimizer

  • But I had to do massive fixes and improvements to make it work just with FLUX

  • Currently only FLUX models tested and verified and I have compiled NVFP4 for FLUX SRPO model

  • You can download FLUX SRPO with our Model downloader app : https://www.patreon.com/posts/114517862

  • I have used max Quantization Algorithm since svdquant was so slow and I had to spent over 24 hours on RTX PRO 6000 to constant test and compile to make it work

  • Currently compiling FLUX model requiring 48 GB GPU and I recommend SimplePod RTX PRO 6000 if you want to try

  • You can of course use RunPod or Massed Compute

  • Massed Compute H200 would work amazing

  • I plan to add Qwen 2512 hopefully later and compile a better quality FLUX SRPO with svdquant

  • Generated FLUX SRPO NVFP4 is only 6.32 GB so it is amazing for low VRAM GPUs

  • More about NVFP4 shown in this tutorial : https://youtu.be/yOj9PYq3XYM

15 January 2026 V1.2

Windows Requirements

For RunPod and Massed Compute

  • Follow

    • RunPod_SimplePod_Install_NVFP4_Instructions.txt

    • Massed_Compute_Install_Instructions.txt

SIMPLEPOD CHEAPER AND FASTER THAN RUNPOD

  • Now we fully support SimplePod as well please use this link to register : https://simplepod.ai/ref?user=secourses

    • SimplePod is faster and cheaper than RunPod and works exactly same

      • E.g. RTX 5090 on RunPod is 0.89 USD per hour, on SimplePod it is 0.45$ per hour,

      • RTX PRO 6000 on RunPod is 1.84 USD per hour and on SimplePod it is 0.79 USD per hour

  • Please use this template on SimplePod : https://dash.simplepod.ai/account/explore/100/ref-secourses/

    • For permanent storage, generate it from Storage tab with any name and size you want and when selecting template with above link, click Edit and Use, select Persistence Volume and change mount point to /workspace

  • Up-to-date SimplePod tutorial starting from 21:51 : https://youtu.be/yOj9PYq3XYM?si=Z86wZZLBeYzWo1Qo&t=1311

  • As usual follow Massed_Compute_Instructions_READ.txt and Runpod_SimplePod_Instructions_Musubi_Trainer.txt to install and use and watch the tutorials

Related posts

All posts in this topic: AI Image Generation and Editing Apps · New to SECourses? Start Here: Best and Most Popular SECourses Posts

Files

Comments

Glen Carpenter

Just to confirm, NVFP4 only works on Blackwell, meaning won't work on 4090?

SECourses

it will work but it has no advantage to use. so use FP8 Scaled on 4090

Rizwan Sheikh

Any insight into whether Loras work in comfy for NVFP4?

SECourses

Yes flux and z turbo working so far I tested. And they are improving and adding more support

Taiga

how many people have '50' cards now? Compared to those with measly 4090s?

SECourses

true nvidia should still work on 4090 support

Roni Bliss

Yes but fp4 has Zero effect on any other cards other than 50x up, NVFP4 is only effective on Blackwell chips. Any others will see no benefit. It works on RTX Pro so many MANYYY thanks for bringing this tAWEOME GUIo the comminuty brother! .. Z-Image generates images before I have even clicked left mouse button! Its so fast it has "pre-generation" .. Its like "pre-crime" in minority report, it know the image i want ahead of time, AI is moving so fast!

Nick

I am trying to quantize SRPO with RTX 6000 Pro from simplepod. "2026-04-17 21:12:51 | INFO | __main__ | Starting Enhanced Diffusion Model Quantization 2026-04-17 21:12:51 | INFO | __main__ | ================================================================================ 2026-04-17 21:12:51 | INFO | __main__ | 🔍 EXPORT CONFIGURATION DEBUG 2026-04-17 21:12:51 | INFO | __main__ | args.quantized_torch_ckpt_save_path = /workspace/SwarmUI/Models/Stable-Diffusion/Ray_NVFP4_mixed_custom.pt 2026-04-17 21:12:51 | INFO | __main__ | Type: 2026-04-17 21:12:51 | INFO | __main__ | Truthy: True 2026-04-17 21:12:51 | INFO | __main__ | ================================================================================ 2026-04-17 21:12:51 | INFO | __main__ | ================================================================================ 2026-04-17 21:12:51 | INFO | __main__ | 🔍 CREATED EXPORT CONFIG 2026-04-17 21:12:51 | INFO | __main__ | export_config.quantized_torch_ckpt_path = /workspace/SwarmUI/Models/Stable-Diffusion/Ray_NVFP4_mixed_custom.pt 2026-04-17 21:12:51 | INFO | __main__ | Type: 2026-04-17 21:12:51 | INFO | __main__ | ================================================================================ 2026-04-17 21:12:51 | INFO | __main__ | Validating configurations... 2026-04-17 21:12:51 | INFO | __main__ | Creating pipeline for flux-dev 2026-04-17 21:12:51 | INFO | __main__ | Model path: /workspace/SwarmUI/Models/Stable-Diffusion/Ray.safetensors 2026-04-17 21:12:51 | INFO | __main__ | Data type: {'default': torch.bfloat16} 2026-04-17 21:12:51 | INFO | __main__ | Detected single .safetensors file: /workspace/SwarmUI/Models/Stable-Diffusion/Ray.safetensors 2026-04-17 21:12:51 | INFO | __main__ | Loading pipeline components from local paths... 2026-04-17 21:12:51 | INFO | __main__ | Loading FLUX pipeline from local components... 2026-04-17 21:12:51 | INFO | __main__ | Loading transformer from: /workspace/SwarmUI/Models/Stable-Diffusion/Ray.safetensors 2026-04-17 21:12:52 | INFO | __main__ | Transformer loaded successfully (dtype: torch.bfloat16) 2026-04-17 21:12:52 | INFO | __main__ | Loading FLUX VAE from: /workspace/SwarmUI/Models/VAE/Flux/ae.safetensors 2026-04-17 21:12:52 | INFO | __main__ | FLUX VAE loaded successfully (dtype: torch.bfloat16) 2026-04-17 21:12:52 | INFO | __main__ | Loading CLIP text encoder from: /workspace/SwarmUI/Models/clip/clip_l.safetensors 2026-04-17 21:12:55 | INFO | __main__ | Loading T5 text encoder from: /workspace/SwarmUI/Models/clip/t5xxl_enconly.safetensors You set `add_prefix_space`. The tokenizer needs to be converted from the slow tokenizers 2026-04-17 21:13:36 | INFO | __main__ | Successfully created FLUX pipeline from local components 2026-04-17 21:13:36 | INFO | __main__ | Pipeline created successfully 2026-04-17 21:13:36 | INFO | __main__ | Moving pipeline to CUDA 2026-04-17 21:13:40 | INFO | __main__ | Initializing calibration... 2026-04-17 21:13:40 | INFO | __main__ | Loading calibration prompts from {'name': 'Gustavosta/Stable-Diffusion-Prompts', 'split': 'train', 'column': 'Prompt'} 2026-04-17 21:13:45 | INFO | __main__ | Building quantization config for fp4_mixed_dev 2026-04-17 21:13:45 | INFO | __main__ | Checking for LoRA layers... 2026-04-17 21:13:45 | INFO | __main__ | Starting model quantization... Inserted 1797 quantizers 2026-04-17 21:13:45 | INFO | __main__ | Starting calibration with 256 batches Calibration: 0%| | 0/256 [00:00 77). Running this sequence through the model will result in indexing errors The following part of your input was truncated because CLIP can only handle sequences up to 77 tokens: ['incev , in style of lee souder , in plastic , dark atmosphere , tilt shift , depth of field ,', 'trending on art station <|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|>'] Calibration: 0%| | 0/256 [00:00= 0), ( AssertionError: detected negative values after abs, could be torch or cuda bug" do you know what this error is?