Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Adjust used dtypes for hunyuan video VAE and diffusion model.

comfyanonymous39b1fc4cccc8
diff --git a/comfy/sd.py b/comfy/sd.py
index 0d498fef4..89f0eb24e 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -323,6 +323,7 @@ class VAE:
self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1])
self.memory_used_decode = lambda shape, dtype: (1500 * shape[2] * shape[3] * shape[4] * (4 * 8 * 8)) * model_management.dtype_size(dtype)
self.memory_used_encode = lambda shape, dtype: (900 * max(shape[2], 2) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
+ self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
elif "decoder.layers.1.layers.0.beta" in sd:
self.first_stage_model = AudioOobleckVAE()
diff --git a/comfy/supported_models.py b/comfy/supported_models.py
index df232e83f..ed3af9d1c 100644
--- a/comfy/supported_models.py
+++ b/comfy/supported_models.py
@@ -753,7 +753,7 @@ class HunyuanVideo(supported_models_base.BASE):
memory_usage_factor = 2.0 #TODO
- supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+ supported_inference_dtypes = [torch.bfloat16, torch.float32]
vae_key_prefix = ["vae."]
text_encoder_key_prefix = ["text_encoders."]