Environment Variables

Single reference for every TINYLOOP_* environment variable. Organised by concern into subpages.

Most runtime flags are read once on first use and cached. Server env vars are read on uvicorn boot. Compile-time knobs are -D macros consumed by nvcc.

Subpages

• Runtime & Dispatch — log level, profiling, warmup, mmap, cuDNN, cuBLAS toggles
• Model Architecture — RoPE, GQA, post-norm, GeLU
• KV Cache — INT8, h-mode, INT4-h, paged attention
• Quantisation — error cancellation, Marlin, stochastic rounding
• FP16 Body & Fusion — FP16 body cache, fused LN+QKV
• Server & Testing — HTTP server config, test paths, β-informed deployment