text-generation-inference

c9bdaa8b - feat(server): reduce mlp and attn in one op for flash neox (#145)