The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Add a --bf16-unet to test running the unet in bf16.
comfyanonymousfd4c5f07e7b4
diff --git a/comfy/cli_args.py b/comfy/cli_args.pyindex 35d44164f..d86557646 100644--- a/comfy/cli_args.py+++ b/comfy/cli_args.py@@ -53,6 +53,8 @@ fp_group = parser.add_mutually_exclusive_group()fp_group.add_argument("--force-fp32", action="store_true", help="Force fp32 (If this makes your GPU work better please report it).")fp_group.add_argument("--force-fp16", action="store_true", help="Force fp16.")+parser.add_argument("--bf16-unet", action="store_true", help="Run the UNET in bf16. This should only be used for testing stuff.")+fpvae_group = parser.add_mutually_exclusive_group()fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.")fpvae_group.add_argument("--fp32-vae", action="store_true", help="Run the VAE in full precision fp32.")diff --git a/comfy/model_management.py b/comfy/model_management.pyindex 1161c2447..c24c7b27e 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -449,6 +449,8 @@ def unet_inital_load_device(parameters, dtype):return cpu_devdef unet_dtype(device=None, model_params=0):+ if args.bf16_unet:+ return torch.bfloat16if should_use_fp16(device=device, model_params=model_params):return torch.float16return torch.float32