Issues

• Peak-RSS and streamed-speed numbers ignore the OS page cache (add a cache-dropping mode; report cold-cache runs)

  enhancement
  ankit1057/sipllm#82 · Filter by author ankit1057 opened on Oct 23, 2026

• --ram-budget can be exceeded: KV-cache growth transient and unpriced semantic cache

  bug
  ankit1057/sipllm#81 · Filter by author ankit1057 opened on Oct 23, 2026

• Long-prompt prefill and decode far behind llama.cpp and LiteRT-LM on a phone (SmolLM2-135M, ~1000-token prompt)

  enhancement
  ankit1057/sipllm#80 · Filter by author ankit1057 opened on Oct 23, 2026

• CLI: -t sets temperature, not threads; -t 8 is silently taken as temperature 8.0

  bug
  ankit1057/sipllm#79 · Filter by author ankit1057 opened on Oct 23, 2026

• With no --ram-budget the CLI streams every layer even when the model fits in RAM (slowest mode by default)

  bugquestion
  ankit1057/sipllm#78 · Filter by author ankit1057 opened on Oct 23, 2026

• LAN pooled inference: split layers across machines, each donating a bounded RAM slice

  enhancementsipmesh
  ankit1057/sipllm#70 · Filter by author ankit1057 opened on Oct 19, 2026
  2

• Metal compute backend for Apple GPUs: zero-copy over mmap'd weights, prefill first

  architectureenhancement
  ankit1057/sipllm#69 · Filter by author ankit1057 opened on Oct 19, 2026

• Portable binaries: runtime CPU feature dispatch (x86 baseline/AVX2, ARMv8.0/dotprod) — release bundles SIGILL or miss fast paths today

  architectureenhancement
  ankit1057/sipllm#68 · Filter by author ankit1057 opened on Oct 19, 2026
  1

• Vulkan compute backend: prove or kill GPU prefill on mobile GPUs (pre-registered bar: beat llama.cpp CPU prefill)

  architectureenhancement
  ankit1057/sipllm#67 · Filter by author ankit1057 opened on Oct 19, 2026
  2