feat(moe): add BF16 MoE definitions, workloads, and CUTLASS SM90 solutions for Qwen3-235B
#9
by omeng-nvidia - opened
Add BF16 MoE benchmark assets for three parallelism configurations of Qwen3-235B (E=128, H=4096): data-parallel (I=1536), tensor-parallel TP=8 (I=192/rank), and expert-parallel EP=8 (E=16, I=1536). Validated: 14/14 seq_len workloads PASSED (max_abs <= 0.002).