llama.cpp makes prompt-lookup drafting 42x faster, and lossless
Four cache tweaks slash draft latency ~100x and halve memory, merged upstream. Plus DSPy lands on the BEAM: Imp v0.5 brings agent optimization to Elixir.

Copy markdown
A free, lossless speedup for local inference
Four n-gram-cache rewrites to llama.cpp's prompt-lookup drafting cut per-token draft latency from up to 165µs down to under 4µs — and up to ~140x with Daniel Lemire's follow-up threshold check — while dropping peak memory from 3.55GB to 1.31GB on a 541MB corpus. Acceptance rate is identical (fully lossless) and the PRs are merged upstream, so anyone running speculative decoding in llama.cpp (or Ollama on top of it) gets the win just by updating.
DSPy comes to the BEAM: Imp v0.5
Imp v0.5.0 ports DSPy-style declarative, self-optimizing LLM programs to Elixir. The headline is agent optimization: GEPA now reads whole agent runs — tool calls, results, and answers — then rewrites the agent's instructions, with both DSPy-compatible and native :beam_native profiles. It also adds deadline propagation across parallel work, a leaner ReActV2 loop, and MCP via ExMCP 1.5 with pooled concurrent calls. Install with {:imp, "~> 0.5"}.