Repository navigation
Compitability with latest ComfyUI #5
Description
Works great with ComfyUI 0.30.0 as stated, but the following incompatibility showed up with latest Comfy:
[rank1]: Traceback (most recent call last):
[rank1]: File "/home/u13/comfy/ComfyUI/custom_nodes/buqi-minimax-h3-multigpu/sp_worker.py", line 136, in
[rank1]: main()
[rank1]: File "/home/u13/comfy/ComfyUI/custom_nodes/buqi-minimax-h3-multigpu/sp_worker.py", line 128, in main
[rank1]: spf.sp_forward(dit, [video, audio], timestep, context, dict(meta["options"]),
[rank1]: File "/home/u13/comfy/ComfyUI/custom_nodes/buqi-minimax-h3-multigpu/minimax_sp/sp_forward.py", line 470, in sp_forward
[rank1]: all_video_rows[~img_update] = cond_video_rows
[rank1]: ~~~~~~~~~~~~~~^^^^^^^^^^^^^
[rank1]: RuntimeError: shape mismatch: value tensor of shape [390, 96] cannot be broadcast to indexing result of shape [0, 96]
Activity
gajeke commentedon Sep 18, 2026
最新的comfyui针对h3做了优化，如果可以，期待作者更新以适配新版comfyui
buqi-code commentedon Sep 28, 2026
Fixed by #7 for the current MiniMax H3 API introduced at ComfyUI commit
8d534945and validated against7fbcfa8b.The SP path now follows the current
PackedLayout, arbitrary image/audio guides, denoise masks, audio carry, PDD final heads, andAttentionTensorContainercontract. The old condition-layout mismatch ([390, 96]into[0, 96]) is covered by T2V/I2V/reference/audio-guide/mask parity tests.Real SP2 server tests passed for T2V, first-frame I2V, first+last I2V, AddGuide and R2V. The main branch intentionally no longer claims compatibility with the old ComfyUI 0.30 H3 API.