FACT: agentic kernel synthesis, grounded in CUTLASS

Ask a language model for a fast CUDA kernel and it will hand back plausible code that often does not compile, let alone run quickly. FACT grounds the synthesis in CUTLASS C++ across a three-stage agentic workflow. Auto-tuned kernels beat PyTorch cuBLAS by 1.06x to 1.18x on GEMM workloads, and composing fused multi-head attention with a fused MLP GEMM+GELU reaches a 2.79x end-to-end speedup on a MiniGPT block. Preprint by Sina Heidari.

← All news