drowzeys/Keys-NVIDIA-Two-Tower-Diffusion--dual-dgx-spark: keys-Nvidia — TRUE Two-Tower diffusion, NVFP4 APPLIED: Nemotron-Labs-TwoTower-30B split across two DGX Sparks over 200G fabric; NVFP4 experts + grouped Triton W4A16 GEMM = 38.9 tok/s, beats AR 1.4-1.8x.

keys-Nvidia — TRUE Two-Tower diffusion, NVFP4 APPLIED: Nemotron-Labs-TwoTower-30B split across two DGX Sparks over 200G fabric; NVFP4 experts + grouped Triton W4A16 GEMM = 38.9 tok/s, beats AR 1.4-1.8x.

Read Original

Related