Model-level forecast patch on both flow DiTs (sparse-structure + SLaT) via hicache-pp. GPU-validated on TRELLIS-image-large/RTX 5090: interval=2 gives 2.1x at 0.59% Chamfer (near-lossless), interval=3 gives 2.5x. Handles TRELLIS' 1->0 schedule and split guidance-interval CFG (two parallel forecast states); SLaT SparseTensor forecast on .feats with layout rebuild. 8 standalone unit tests + an end-to-end GPU validation harness.
8 lines
310 B
Python
8 lines
310 B
Python
"""ComfyUI-TRELLIS-HiCache: training-free HiCache acceleration for TRELLIS."""
|
|
try:
|
|
from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
|
|
except ImportError:
|
|
from nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
|
|
|
|
__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]
|