Nodes/CRT-Nodes/Ideogram 4 FlashAttention (CRT)
ComfyUI Node

Ideogram 4 FlashAttention (CRT)

Patches Ideogram 4 attention to use FlashAttention 2, including its padding mask through the variable-length kernel. Optionally swaps the last transformer blocks between CPU and GPU to reduce VRAM use.

By PGCRT·Created 2 years ago·Updated 7 days ago· 129
Ideogram 4 FlashAttention (CRT)
  • model
  • MODEL
blocks_to_swap0
CategoryCRT/Model Patches

Inputs (2)

NameTypeDefaultDescription
modelMODEL
blocks_to_swapINT00–34Number of trailing Ideogram transformer blocks to keep on CPU and load one at a time during inference.

Outputs (1)

NameTypeDescription
MODELMODEL