macos implementation notes - fp32 on m1, maybe fp8 should just raise, use watermark env vars

[doctorpangloss]

ComfyUI/comfy/model_management.py

Line 1033 in 19a64d6

if (device is not None and is_device_mps(device)) or mps_mode():

it's worth reading https://news.ycombinator.com/item?id=36576673

test fp32 versus fp16 versus bf16 on M1 macs, because there's no hardware bf16 support on M1. i don't know if apple's emulation on m1 actually provides memory or performance gains.

if the user tries to run an fp8 workflow, maybe raise because it is not supported on macos at all

[doctorpangloss]

reserve vram should use the watermark envs - https://github.com/pytorch/pytorch/blob/bf8da4c145f0b11ad19daa451d923ab3594ff6ce/docs/source/mps_environment_variables.rst#L27

you can also consider the same for CUDA configuration on windows

[comfyanonymous]

On an M4 mac at least it does reduce memory usage.