Fused Triton kernels for Transformer inference: RMSNorm+RoPE, Gated MLP, FP8 GEMM — CPU-testable references, autotuning, and benchmarking
python acceleration deep-learning high-performance cuda inference pytorch nvidia triton gpu-computing quantization gpu-kernels kernel-fusion llm fp8 rotary-position-embedding rmsnorm llm-inference operator-fusion transformer-inference
-
Updated
May 25, 2026 - Python