transformer-fused-ops

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.

The public surface includes relu2_quantize_fp8_static_bf16, a fused BF16 ReLU-squared to FP8 producer.

The SM110 BF16 producer surface also includes quantize_fp8_static_bf16, layer_norm_quant_fp8_static_bf16, and gate_geglu_merged_quant_fp8_static_bf16 for PI0.5 prefill and SigLIP MLP pipelines.

The SM110 surface also includes rms_norm_fp16, layer_norm_fp16, layer_norm_quant_fp8_static_fp16, rope_rotate_half_fp16_, the explicit GROOT vec aliases (rms_norm_fp16_vec, layer_norm_fp16_vec, layer_norm_fp8_static_fp16_vec, rope_rotate_half_fp16_vec, quantize_fp8_static_fp16_vec, residual_add_fp16_vec, gpu_repeat_interleave_heads_vec), quantize_fp8_static_fp16, residual_add_fp16_, and repeat_interleave_heads_fp16 for GROOT N1.7 Thor static-buffer pipelines.

The MoE surface includes model-neutral router_topk_bf16 and moe_weighted_sum_bf16_to_fp32; the older nexn2_router_topk_bf16 name remains an exact compatibility alias.