Multi Gpu Instance Provisioning With Heterogeneous Gpu Configurations

via “multi-gpu model distribution and memory management”

LTX-Video Support for ComfyUI

Unique: Implements GPU-aware model partitioning through LTXVGemmaCLIPModelLoaderMGPU that automatically detects available GPUs and distributes text encoder, DiT, and VAE components based on VRAM availability. Integrates with ComfyUI's device management system for seamless multi-GPU workflows.

vs others: More granular control than simple data parallelism; enables model parallelism for components that don't fit on single GPU, unlike standard ComfyUI which requires manual device specification.