Vous avez un workflow ComfyUI ? Obtenez une documentation lisible directement dans votre navigateur.Rendre un workflow lisible.Essayer maintenant ↗
The workflow should run on the DisTorch2MultiGPUv2 nodes to control VRAM (RAM) allocation, preventing VRAM overflow and the resulting swapping. Unfortunately, these nodes are currently broken due to the latest...
Profil d'exécution
Description de la source
The workflow should run on the DisTorch2MultiGPUv2 nodes to control VRAM (RAM) allocation, preventing VRAM overflow and the resulting swapping. Unfortunately, these nodes are currently broken due to the latest ComfyUI updates. As a fallback, the older MultiGPUv1 nodes are used (with up to 10–15% lower inference speed compared to DisTorch2).
GitHub: pollockjj/ComfyUI-MultiGPU
If you have a working SageAttention setup on your system, the workflow with fp8.safetensors should run without issues, depending on the available VRAM (RAM).
With GGUF :
The acceleration provided by SageAttention is significantly lower than with FP8 models and may sometimes not take effect at all, since GGUF formats are primarily designed for CPU-optimized inference and do not fully leverage the GPU kernels of SageAttention.
Commentaire généré par l’IA
Explication générée par l’IA à partir des détails de la source et de la configuration. Les suggestions sont clairement signalées.
Simplified t2i Workflow for Flux2D. The workflow should run on the DisTorch2MultiGPUv2 nodes to control VRAM (RAM) allocation, preventing VRAM overflow and the resulting swapping. Unfortunately, these nodes are currently broken due to the latest...
The current graph contains 1 retained relationship for this resource.
Current retained source (civitai_workflow) is published by GegenDenTag.
flux2-dev-Q8_0.gguf: unresolved.
Mistral-Small-3.2-24B-Instruct-2506-UD-Q8_K_XL.gguf: unresolved.
comfyui-multigpu: resolved.
flux2-vae.safetensors: ambiguous.
Sweet Tea records presentation quality as ready with no listed presentation blockers.
Confirm flux2-dev-Q8_0.gguf before use because its retained dependency state is unresolved.
Suggestion · non vérifié
Confirm Mistral-Small-3.2-24B-Instruct-2506-UD-Q8_K_XL.gguf before use because its retained dependency state is unresolved.
Suggestion · non vérifié
This commentary explains retained evidence; it does not independently assert that the resource is recommended or runtime-tested.
Suggestion · non vérifié
Faut-il modifier cela ?
Connectez-vous pour envoyer une demande de modification.
Sources
1 sourceExtraits de sources
2 extraitsSource context: 208 downloads · Type Workflows · Base model Flux.2 D
-> Disable --use-sage-attention
-> Use --fast (standard PyTorch optimization)
-> Rely on the internal optimizations of the GGUF Nodes (backend)
For RTX30xx, RTX40xx (and RTX50xx) systems with less than 24 GB VRAM , please refer to the table “Quick Reference: FLUX.2 + Mistral-3-Small GGUF” on my HF/Modelpage.
GegenDenTag/comfyUI-Flux2D-t2i-workflow · Hugging Face
You will also find notes there regarding memory management, run_nvidia_gpu.bat, SageAttention installation guide , and some CMD Console Outputs (see performance as follows)
Performace
Test Setup rtx3090 24GB VRAM + 32GB RAM:
Flux2 fp8mixed.safetensors Nvidia (35.5GB), Mistral Text Encoder fp8.safetensors (18GB)
Flux2 Q8_0 (35GB), Mistral Text Encoder Q8_K (29GB)
flux2-vae.safetensors (336MB)
Guidance: 4 | Steps: 20 (Production: Guidance: 2-2.5, 30-40 Steps)
Based on: ~80 runs with different resolutions
First Run Initial loading of required layers into VRAM|RAM, first inference; Further inferences are then significantly faster as memory management is already initialized. Exact timings, loaded partially etc. see Console-Output / Screenshots.
FP8 Format
First Run
832×1216px: ~380-400s
Subsequent Runs:
832×1216px: 75-80s (~3.70-3.90s/it)
1080×1920px: 135-150s (~6.75-7.50s/it)
1440×2160px: 225-240s (~11.00-11.50s/it)
GGUF Format (expectedly higher runtimes)
First Run
832×1216px: ~420-440s
Subsequent Runs:
832×1216px: 105-120s (~5.30-5.50s/it)
1440×2160px: 250-260s (~12.00-12.75s/it)
Addendum: No metadata is embedded in the sample images. I am running the workflow in Automation n8n-upscayl_1440px_ultrasharp-4x and have not yet managed to prevent the upscaler from overwriting the metadata.
Estimation des besoins VRAM
Estimation indisponible
160 MB sur 1 de 3 fichiers de modèle. Total des fichiers du modèle + 25 % de surcharge de chargement + 2 Go de tampon d'exécution, arrondi à l'unité supérieure.
Exigences
Exigences 4flux2-dev-Q8_0.gguf
Non résoluCheckpoint · Unknown
VAE · 160 MB · SAFETENSORS · Hugging Face · darkmaniac7/tokforge-bonsai-image-4b-gguf
Mistral-Small-3.2-24B-Instruct-2506-UD-Q8_K_XL.gguf
Non résoluText encoder · Unknown
Pack de nœud · Registry
Source context: 199 downloads · Type Workflows · Base model Flux.2 D
