Native CUDA transformer helper kernels from FlashRT for fused activation, layout, RoPE, argmax/spec-accept, and router top-k work.
See README.md for the public function list.
README.md
The public surface includes relu2_quantize_fp8_static_bf16, a fused BF16 ReLU-squared to FP8 producer.
relu2_quantize_fp8_static_bf16