HELM: Compiler-Guided Heterogeneous Execution for Large Language Models on Memory-Constrained Systems (Artifact)
Artifact for the CASES/ESWEEK paper 'HELM: Compiler-Guided Heterogeneous Execution for Large Language Models on Memory-Constrained Systems'. Contains the HELM compiler and runtime (roofline-based CPU-GPU partition search, static per-device FX graphs, AVX2+F16C GEMV kernel, paged KV cache with streaming attention), benc...