Convert 8bit model to fp16 mac silicon?

KSamplerAdvanced

Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype.

Just want to use Image to video for a project Wan2.2 How can I convert 8bit to FP16 on Mac M4pro 64GB Unified Ram.

Well the short answer is - You cannot !

But there is a way around. Follow this:

Understanding error = Mac do not support any model that runs on fp8. You would need to look for a model that works on fp16 or bf16. These work on Mac. Alternatively you can look for any model via MLX runner.

Solution to your current problem = Wan 2.2 is used in Image-to-Video I2V or Text-to-Video T2V. You can achieve same results with another model called LTX-2. LTX-2 works on a Mac plus has added audio capabilities. To use I2V using LTX-2 just open ComfyUI > Update to latest version > Restart > Go to templates > search LTX-2 > download all files > IMPORTANT : Make sure you download a seperate checkpoint that will make sure LTX-2 runs on Mac (The way you wanted to convert wan 2.2). The file is a checkpoint named = ‘ltx-2-19b-dev.safetensors‘. Here is the link to download the file = https://huggingface.co/Lightricks/LTX-2/resolve/main/ltx-2-19b-dev.safetensors . > Place this file on your Mac’s ComfyUI directory inside checkpoint folder = ComfyUI/models/checkpoints/ltx-2-19b-dev.safetensors. > Restart ComfyUI > Load the new checkpoint in node (non fp8) > Load your workflow/LORAs/VAEs and hit run > Done.

Alternatively (Rare case) - If you still run into error, You would need to run via MLX runner. This runs the same model using a different method than ComfyUI. This is FULL NATIVE METHOD TO MAC & is most stable while FULLY UTILIZING Apple Silicone Capabilities. Downside - It runs inside Mac’s Terminal - You would have to code…… something like python scripts/generate.py … . You can still look for custom GUIs using this method if you feel like. But the summary is this method is most reliable, Full use of Apple Silicone but is less flexible.