Support for NVIDIA Spark DGX GB10

[t3ch9]

Hello, can you add support for running this model on NVIDIA Spark DGX GB10 128GB GPU?

[Dai-Wenxun]

@t3ch9 Honest answer in two parts, because "support" means two different things here:

Will it run? Probably. Nothing in the stack is architecturally incompatible — the GB10 is Blackwell-family (sm_121; we build and validate against sm_100/sm_120), and with 128 GB of unified memory, capacity is a non-issue: you wouldn't even need the JOYOMNI_LOW_VRAM=1 path we just shipped for the 32 GB RTX 5090. We don't have a DGX Spark to test on, so treat it as unvalidated.

Will it be real-time? No — and that's physics, not software. The pipeline is bound by memory bandwidth and tensor throughput: every denoising pass streams the ~16 GB of FP8 DiT weights. An RTX 5090 (1.79 TB/s GDDR7) is already right at the edge of the 24 FPS budget at 480p; the GB10's unified LPDDR5x runs at 273 GB/s — about 6.5× less bandwidth, with roughly 3× lower FP8 tensor throughput on top. The model would run, but at a small fraction of real time, which defeats the point of a streaming editor. So DGX Spark isn't on the roadmap as a real-time target.

What I'd suggest instead:

• Feel the real-time system with zero hardware — the live browser demo runs the full pipeline at 840 × 480 @ 24 FPS: https://huggingface.co/spaces/wxDai/joyai-video-edit
• If local real-time is the goal, the cheapest validated config as of today is a single RTX 5090 (32 GB) — see the Deployment Guide (RTX 5090 section) and #7.

That said, if you run the standard install on your Spark anyway, I'd genuinely like to see the logs and throughput — please post them here.