Comfyui-dgx-spark Secures NVIDIA AI Sessions By Triplany
The Comfyui-dgx-spark project provides scripts and source patches that stabilize ComfyUI on NVIDIA DGX Spark hardware. By adjusting memory handling and hardware parameters, the collection prevents system freezes during image and video generation workflows.
Creator Triplany assembled the toolkit to fix persistent memory leaks and silent output failures on the device. The configuration manages model loading and resource release automatically, removing the need for manual tweaks.
Core optimizations and automated patch system
- Fixes inaccurate memory reporting to stop premature unloading.
- Compiles attention libraries with native ARM64 kernels.
- Clamps invalid audio values that normally crash video generators.
- Launches ComfyUI with reserved memory space and disabled caching.
- Repairs missing text-encoder weights for custom workflows.
Operators running continuous generation batches maintain steady performance without resource exhaustion. The automatic clearing routine removes previous models completely, keeping the unified memory pool available for new tasks.
Practical notes on unified memory constraints
Exhausting available memory on this device causes a full system lockup requiring a physical restart. This setup prevents that outcome by turning off default memory hoarding and trusting the kernel to allocate space efficiently. Forcing low-precision modes globally should be avoided, as it breaks video rendering without generating clear error messages.
"Cold times are a little slower than other setups but this is stable and bullet proof for me,"
confirmed the developer in a Reddit post. The documentation also warns against incompatible attention libraries and notes that patches must be reapplied after core software updates. Download the Comfyui-dgx-spark repository to apply these targeted adjustments and verification checks.
