Tried to allocate 20.00 MiB (GPU 0; 14.76 GiB total capacity; 13.90 GiB already allocated; 14.75 MiB free; 14.14 GiB reserved in total by PyTorch about thechosenone

[JahnKhan]

I got the same message. Could you fix this?

[GraemeHarris]

@paratechnical @JahnKhan I've had some success using the following https://huggingface.co/docs/diffusers/optimization/memory#memoryefficient-attention, the pipeline is instantiated in the load_trained_pipeline function, where you should be able to try reduce the memory usage as per the hugging face article.

Because I was still low on VRAM I went to the pipe.enable_sequential_cpu_offload() option, which is much slower but working :). I haven't tried model offloading yet, but might be something to try to keep some speed.

[paratechnical]

@GraemeHarris

if model_path is not None:
        # TODO: long warning for lora
        pipe = DiffusionPipeline.from_pretrained(model_path, torch_dtype=torch.float16)
        if load_lora:
            pipe.load_lora_weights(lora_path)
    else:
        pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")
    pipe.to("cuda")
    pipe.enable_xformers_memory_efficient_attention()
    pipe.enable_sequential_cpu_offload()

tried it like this and I have the same problem

What kind of GPU configuration are you using?