What is the recommended GPU VRAM for training FireRed-Image-Edit-1.1?

[xu-peng-7]

I would like to ask about the GPU VRAM requirements for training or fine-tuning this model.

• What is the recommended GPU VRAM for training?

[Maycbj]

I would like to ask about the GPU VRAM requirements for training or fine-tuning this model.

• What is the recommended GPU VRAM for training?

• Full fine-tuning of Diffusion models typically requires approximately 60～80GB of VRAM.
• LoRA fine-tuning reduces the memory footprint to around 50GB.
• Enable gradient checkpointing to trade computation for significant memory savings.
• Convert all modules to BF16 precision for improved numerical stability and efficiency.
• Use bnb.optim.AdamW8bit to quantize optimizer states and drastically cut memory overhead.
• Implement DeepSpeed ZeRO-2 to partition optimizer states across GPUs and eliminate redundancy.

[xu-peng-7]

Thank you for your reply.

I think there may have been a misunderstanding. I am specifically referring to training FireRed-Image-Edit-1.1 using the officially released SFT training code.

In my setup, I am using 2 × RTX P6000 GPUs, with 96GB VRAM per GPU (192GB total VRAM). However, when I use the official SFT code and the default training settings, I still encounter CUDA out-of-memory errors, even with:

• batch size = 1
• image_sample_size = 512

The error is:
torch.OutOfMemoryError: CUDA out of memory

So my question is not about the general theoretical VRAM recommendation, but rather about the actual memory requirement of the released FireRed-Image-Edit-1.1 training pipeline in practice.

Could you please clarify the following:

1. What is the actual minimum GPU memory requirement to run the released FireRed-Image-Edit-1.1 SFT training code successfully?
2. Are the default SFT training settings intended for higher-memory GPUs or a specific multi-GPU training setup?
3. Is there an officially recommended memory-saving configuration for this model?
4. Are there any specific options that must be enabled, such as gradient checkpointing, FSDP, DeepSpeed ZeRO, 8-bit optimizer, or freezing some modules, in order to make training feasible?
5. If possible, could you provide a minimal verified training configuration that can run successfully?

Thank you.

[xu-peng-7]

During debugging, I also noticed that there may be some issues or inconsistencies in the current fine-tuning code. It is possible that I have not fully understood the intended setup, so at the moment I am still unable to run the whole pipeline successfully.

I would sincerely appreciate more specific guidance from your side, such as:

• the exact runnable configuration,
• which options need to be changed,
• or which parts of the code need to be fixed.

If needed, I am also very willing to communicate further in a more convenient way (for example, WeChat or another direct contact method), so that I can fully reproduce the training pipeline under your guidance.

If I can finally run it through successfully with your help, I would be very happy to give back to the community:

1. write a complete step-by-step tutorial,
2. and submit a PR to fix the related issues I encountered.

Thank you again for your time and support.