MiniMax H3 V100 — Custom Node v0.1.4

English | 简体中文

Version 0.1.4 fixes TypeError: _patched_final_forward() takes 5 positional arguments but 8 were given in ComfyUI 0.34.5 while retaining the standalone v0.1.3 FP16 storage/branch profile and FP32 safety islands. This release has local structural/argument/dtype-flow validation.

Install / update

1. Fully stop ComfyUI.

2. Remove the previous H3 extension folder from custom_nodes.

3. Extract minimax-h3-v100-v0.1.4.zip into the active custom_nodes directory. The ZIP has one top-level folder:

   custom_nodes/minimax-h3-v100-l3-clean/__init__.py
   custom_nodes/minimax-h3-v100-l3-clean/runtime_patch.py
   

   In the supplied error report that directory is C:\Comfyui\custom_nodes.

4. Restart ComfyUI normally. No new workflow node or --fp16-unet flag is needed. Use an official, unmodified ComfyUI comfy/ldm/minimax/model.py.

5. Confirm the startup log contains v0.1.4 runtime profile installed, then model loading reports enabled v0.1.4 on the main DiT blocks.

Precision profile

• Native FP16 weights through ComfyUI loading, prefetch and offload.
• FP16 attention and MLP branches in main DiT blocks; FP32 target/reference-audio attention recomputation.
• Attention output scaling: /64 → FP16 projection → FP32 ×64.
• MLP: FP16 fc1, FP32 SwiGLU, /256 → FP16 fc2 → FP32 ×256.
• FP32 residual, normalization/modulation, condition input, Token Refiner and final heads.
• CUDA compute capability 7.0 restriction, duplicate-extension checks and no lazy weight.data conversion remain in place.

The attention and MLP compute functions retain their v0.1.3 AST hashes. The previous v0.1.3 documentation records successful ComfyUI 0.33.2 testing and reduced resident VRAM; that is historical evidence, not a measured result for v0.1.4.

Acknowledgement and license

Thanks to ComfyUI, MiniMax and Amduraznak/minimax-h3-fp16-fix for the Custom Node delivery pattern and related mixed-precision work. This is an independent community extension. SPDX: GPL-3.0-only; see LICENSE and NOTICE.md.