Performance portable LLM inference in Fortran
The author demonstrates that Fortran can be used for performance-portable LLM inference using native array parallelism and GPU offloading without external pragmas or directives. In benchmarks using the Qwen3.8-27B model on an NVIDIA H200, an LFortran-compiled Fortran implementation achieved 102 tokens per second, outperforming established frameworks like vLLM and llama.cpp. The author argues that standard Fortran constructs are sufficient for modern AI workloads, though compiler optimizations and code simplification remain necessary. Open article →