MiniMax H3 (Hailuo 3) video on 2ร DGX Spark ๐ฌ
โข 5 s 1344ร768 clip with audio: 643 s (one Spark: 1399 s)
โข 10 s: 1690 s, 15 s: 3080 s (one Spark: out of memory)
โข 20 s: 6740 s, beyond the model's 15 s spec and still coherent to the last frame (clip below, one continuous generation, seed 1101, 20 steps)
โข Text-encoder TP2 + Ulysses SP2 over one ConnectX-7 cable (RoCE, 19.5 GB/s)
โข Online FP8 DiT, cuDNN attention, 24 fps, native audio
โข Frames leave the GPU as uint8, VAE tiles decoded on both GPUs: same bytes, 15 s fits with 25 GB to spare on each host
โข H3-Base open weights (768p), vLLM-Omni, no Ray and no changes to the vLLM-Omni tree
vllm serve cannot span two hosts, so the recipe adds a cross-host executor: rank 1 runs on the second Spark and is driven over the link through vLLM's own message queues, behind the unchanged single-node API. Quality checks, memory traces and
the open levers are in the repo.
Full recipe, measurements and credits: