Is your feature request related to a problem? Please describe.
The standard FLUX DreamBooth LoRA example supports FP16 and BF16, but does not provide an explicit, documented FP8 training path. Users with FP8-capable NVIDIA GPUs would benefit from an example showing how to use FP8 computation while preserving training quality and checkpoint compatibility.
Describe the solution you'd like.
Add an optional Transformer Engine FP8 training example or guide for FLUX, using Accelerate where practical.
The example should document supported hardware, FP8 scaling recipes, and compatible training modes. Validation should compare against BF16 for training throughput, peak GPU memory, generated-image quality, and checkpoint save/reload compatibility. Full fine-tuning and LoRA should be evaluated separately.
Describe alternatives you've considered.
A Diffusers example would make this workflow easier to discover, reproduce, and validate within the existing training ecosystem.
Additional context.
Relevant references:
Is your feature request related to a problem? Please describe.
The standard FLUX DreamBooth LoRA example supports FP16 and BF16, but does not provide an explicit, documented FP8 training path. Users with FP8-capable NVIDIA GPUs would benefit from an example showing how to use FP8 computation while preserving training quality and checkpoint compatibility.
Describe the solution you'd like.
Add an optional Transformer Engine FP8 training example or guide for FLUX, using Accelerate where practical.
The example should document supported hardware, FP8 scaling recipes, and compatible training modes. Validation should compare against BF16 for training throughput, peak GPU memory, generated-image quality, and checkpoint save/reload compatibility. Full fine-tuning and LoRA should be evaluated separately.
Describe alternatives you've considered.
A Diffusers example would make this workflow easier to discover, reproduce, and validate within the existing training ecosystem.
Additional context.
Relevant references: