Inferact trains DSpark speculator for Muse Spark Open

AIvLLM says its NVIDIA work includes fused norm, gating, and residual kernels, plus fused MNNVL collectives and an FP32-norm-to-TF32-router path. It adds that AMD MI355X supports FP8 and MXFP4 with TP and PP on ROCm, and that initial TPU support uses TorchTPU. With DSpark at acceptance length 6, Muse Spark Open megakernels reach over 650 tokens/sec on 4 TPU v7 chips and 960 tokens/sec on 8 chips, versus over 270 and 350 tokens/sec without speculation. TPU code access is available through the team or Google.

AIInferact trained a DSpark speculator for Muse Spark Open, using block diffusion to propose up to seven draft tokens in one pass. The recipe supports CUDA and ROCm, and the draft model is published on Hugging Face.