A simple diff to add mochi support to the clip loader

[nkavassalis]

Hey all, here is a simple diff to add mochi support to the clip loader. Makes it very viable on dual 3090/4090 setups to do mochi with the full model and full t5 large:

diff --git a/__init__.py b/__init__.py
index 8a3ee41..6bc939f 100644
--- a/__init__.py
+++ b/__init__.py
@@ -203,7 +203,7 @@ class CLIPLoaderMultiGPU:
             "required": {
                 "clip_name": (folder_paths.get_filename_list("clip"),),
                 "type": (
-                    ["stable_diffusion", "stable_cascade", "sd3", "stable_audio"],
+                    ["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi"],
                 ),
                 "device": ([f"cuda:{i}" for i in range(torch.cuda.device_count())],),
             }
@@ -223,6 +223,8 @@ class CLIPLoaderMultiGPU:
             clip_type = comfy.sd.CLIPType.SD3
         elif type == "stable_audio":
             clip_type = comfy.sd.CLIPType.STABLE_AUDIO
+        elif type == "mochi":
+            clip_type = comfy.sd.CLIPType.MOCHI
         else:
             clip_type = comfy.sd.CLIPType.STABLE_DIFFUSION

[RageshAntonyHM]

@nkavassalis

Is it possible to load VAE in separate GPU when using Mochi ?

[pollockjj]

nkavassalis,

I’ve added mochi support right into my fork that addresses this feature.

It’s already tested and working on my end, so you won’t be wasting your time giving it a spin. It’s on ComfyUI‑Manager and should remain backward‑compatible with the existing nodes.

If you’d like to test or contribute, I’d love your feedback over at pollockjj’s fork of MultiGPU. I’m actively maintaining and expanding it, so any bug reports or suggestions are more than welcome.