vLLM-Omni runs MAGI-2 Preview on Moore Threads MUSA, cutting step time from 15.6 s to 565 ms
AIThe vLLM team reports running MAGI-2 Preview, a 114B audio-video MoE DiT, on Moore Threads MTT S5000 GPUs using torchada to map CUDA-style calls to MUSA. Per-step time fell from 15.6 s in the first working run to 565 ms with all listed pull requests applied on 8 GPUs at 272p and 100 steps. The post pairs each change with an A/B measurement and an output hash check, and states that several pull requests remain open.
AIWhy it matters.
The post traces a step-by-step optimization path with paired measurements and bit-identical checks, useful for engineers porting CUDA-first models to other accelerators.
Source details
vLLM Blog · Moore Threads Team and vLLM-Omni Team · vllm.aiPublished · added here
Original title: MAGI-2 Preview on vLLM-Omni: Bringing a 114B Audio-Video MoE DiT to Moore Threads MUSA with torchada