ZSE inference engine: 30x faster cold start than vLLM, no PyTorch needed

tom_doerr · x · 2026-08-14

ZSE is a zero-dependency LLM inference engine that eliminates PyTorch, Triton, bitsandbytes, and transformers, using pure Python, ctypes, and a kernel compiler to emit CUDA, ROCm, and Metal directly. Verified on Modal (T4, L4, A10G, A100), DigitalOcean (MI300X), and Apple M1, ZSE INT4 achieves dramatically faster cold starts compared to vLLM AWQ: for Qwen2.5-7B on T4, 7.25s vs 218.96s (30.2x speedup), and on L4, 5.58s vs 145.22s (26x). Models load in seconds with minimal memory footprint.

Original post →

More from Infra

Infra channel →