LTX2 Mem Eff Sage Attention Patch(LTX2MemoryEfficientSageAttentionPatch)

Description

EXPERIMENTAL! This node activates a custom sageattention patch to reduce peak VRAM usage. It overrides the current attention mode in the model, replacing it with memory-efficient sageattention kernels. Requires the latest version of the sageattention library.

Inputs

• model (MODEL, required)
  The diffusion model to which the memory-efficient attention patch will be applied.

• triton_kernels (BOOLEAN, required, default: true)
  When enabled, uses Triton fused RoPE kernels for the self-attention Q/K projections. This can further reduce memory and improve performance but requires Triton to be installed.

Outputs

• MODEL
  The same model with its attention mode patched to use sageattention, ready for inference with lower VRAM usage.

Usage Notes

• This node is experimental — behavior may change in future versions.
• Ensure you have the latest sageattention package installed (e.g., pip install sageattention --upgrade).
• Enabling triton_kernels requires Triton to be available in your environment. If Triton is not installed, set this input to false.
• The patch overrides the model's built-in attention mechanism; it will affect all attention layers in the model.
• Typically used in combination with LTX2‑family workflows to reduce VRAM peaks during sampling.