For others with a DGX Spark thought I would share what works for me and how I got here.
Love your feedback and suggestions for improvements.
After reading a lot of forums, trying settings other posted I kept bumping into one issue or another depending if I was doing video gen, image, switching workflows.
From double memory usage, to not seeing all the free vram and aborting (Wan 2.2 and Flux1 at full quant would randomly do this).
Not unloading models from vram when switching model/workflow
Opposite unloading after every run (so every run was cold).
Huge memory spikes when loading.
OOMS that brick it and force a hard reboot.
These are just a few I encountered trying to get it to run right.
Here is a install script that compiles and updates what is needed, script to start comfyui with the settings I use, and patches I use. https://github.com/Triplany/comfyui-dgx-spark
Cold times are a little slower than other setups but this is stable and bullet proof for me. whether I am doing a whole bunch of pictures or jumping to ltx or wan.
Memory usage stays low and consistent, can easily run flux2 at full quants
For each benchmark I restarted comfyui did cold, warm. stop/start whole thing
Stock Comfyui Templates:
zimage t2i (bf16) Memory use: 43.5gb. Speed: res 1024x1024 cold: 96.17s warm: 43.73s
Flux2.dev t2i (fp8mixed) Memory use: 68gb res 1024x1024 cold: 300.38s warm: 50.14s
LTX 2.3 t2v (fp8) Memory use: 44.73gb. Speed: res 1280x720, duration 5. cold: 179.55s warm 81.83s
Wan2.2 14b t2i (fp8) Memory use: 18gb. Speed: res 640x640, duration 5. cold: 644.75s warn: 565.24
Full Quants:
Flux2-dev (full) w mistal3_small at bf16 Memory Use: 93.80gb Speed: 1024x1024 cold: 407.52s warm: 80.25s
Flux1-dev (full) w t5xxl at fp16 Memory Use: 32.16gb Speed: 1024x1024 image cold: 113.17s warm: 32.61s
There is no crazy spikes on memory use. what I posted above stays constant. rerunning workflow (warm) shows it stays loaded and each additional run is much fast.
Switching Workflow (say flux2 to Wan2.2) properly evicts the old models freeing up memory
Hope this helps another spark user not waste as much time as I did lol.
