Jonathan Rahn

yes, it's currently not in the right place (should go into the `__init__.py` you mention), but i created a variable for that

this runs without exceptions now: ```python from diffusers import FlaxStableDiffusionPipeline, StableDiffusionPipeline from diffusers import AutoencoderKL, UNet2DConditionModel import jax.numpy as jnp pipe, params = FlaxStableDiffusionPipeline.from_pretrained('CompVis/stable-diffusion-v1-4', revision='bf16', dtype=jnp.bfloat16) pipe.save_pretrained('output-flax', params) pipe2 =...

this is my updated test scenario: ```python from diffusers import FlaxStableDiffusionPipeline, StableDiffusionPipeline from diffusers import AutoencoderKL, UNet2DConditionModel import jax.numpy as jnp import jax from flax.jax_utils import replicate from flax.training.common_utils import...

Thank you!

Thanks, I'll give it a try

Ok, so I have a lokal version, that works for this: ```python from diffusers import FlaxStableDiffusionPipeline, StableDiffusionPipeline from diffusers import AutoencoderKL, UNet2DConditionModel import jax.numpy as jnp pipe, params = FlaxStableDiffusionPipeline.from_pretrained('CompVis/stable-diffusion-v1-4',...

I've taken a quick look at this e.g. [here](https://stackoverflow.com/questions/19849077/error-loading-dll-in-python-not-a-valid-win32-application) and [here](https://stackoverflow.com/questions/57187566/python-ctypes-loading-dll-throws-oserror-winerror-193-1-is-not-a-valid-win) - but have not done such ports before. My initial thoughts are: - currently bitsandbytes loads libbitsandbytes.so which won't...

I've spent a bit more time looking at how [DeepSpeed](https://github.com/microsoft/DeepSpeed) solves building on Windows. They make use of `BuildExtension`, `CppExtension` and `CudaExtension` from `torch.utils.cpp_extension` and this build process is run...

To add one data point: I'm running unmodified `python train.py` with `--batch_size=8` on ~22gb vram.

> Note that it's using FP16 weights, not int8, so the memory requirements are 2x of the int8 quantized model. But personally I'm skeptical that the model can be quantized...