๊ฐ€์ž… ํ›„ ์ดˆ๋Œ€ ๋งํฌ๋ฅผ ๊ณต์œ ํ•˜๋ฉด ๋™์˜์ƒ ์žฌ์ƒ ๋ฐ ์ดˆ๋Œ€ ๋ณด์ƒ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Zbigniew Majewski
@majewskizby
๊ฐ€์ž… May 2014
53 ํŒ”๋กœ์ž‰ ์ค‘    508 ํŒฌ
MiniMax H3 (Hailuo 3) video on 2ร— DGX Spark ๐ŸŽฌ โ€ข 5 s 1344ร—768 clip with audio: 643 s (one Spark: 1399 s) โ€ข 10 s: 1690 s, 15 s: 3080 s (one Spark: out of memory) โ€ข 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) โ€ข Text-encoder TP2 + Ulysses SP2 over one ConnectX-7 cable (RoCE, 19.5 GB/s) โ€ข Online FP8 DiT, cuDNN attention, 24 fps, native audio โ€ข 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 โ€ข 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:
๋” ๋ณด๊ธฐ