SageAttention's quantized kernels produce NaN in the Hunyuan3D v2.1
diffusion transformer, causing the downstream VoxelToMesh to generate
zero vertices and crash in save_glb.
Add low_precision_attention=False to both optimized_attention calls in
the v2.1 DiT (CrossAttention and Attention classes), following the same
pattern used by ACE (ace_step15.py). This makes SageAttention fall back
to pytorch attention for Hunyuan3D only, while all other models keep
the SageAttention speedup.
Root cause: the 3D occupancy/SDF prediction requires higher numerical
precision at voxel boundaries than SageAttention's quantized kernels
provide. Image and video diffusion tolerate this precision loss.
Fixes: comfyanonymous/ComfyUI#10943
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
124300b732