v0.14.0

Features

• MTP (Multi-Token Prediction) support for llama.cpp (#102, #103): llama-server is now auto-launched with --spec-type draft-mtp --spec-draft-n-max 5 whenever the chosen model is an MTP variant. Detection runs in two passes:

  1. GGUF metadata probe — read the file's header KV table for an architecture-specific *.mtp.* key, a key ending in .nextn_predict_layers (the cross-arch convention used by Qwen, GLM, and DeepSeek MTP releases), or MTP in general.name / general.architecture. Large scalar arrays (tokenizer vocabularies, token-type tables) are skipped with a single seek so the KV walk stays aligned past them.
  2. Filename fallback — match *mtp* (word-bounded, case-insensitive) on the basename when the probe is inconclusive.

  Two env vars override the decision: LLAMACPP_MTP_ENABLED=0/1 forces off/on, and LLAMACPP_SPEC_DRAFT_N_MAX=N overrides the default 5 (valid range 1–16). The detector also recognizes flag combinations that llama.cpp does not yet pair with --spec-type draft-mtp (--mmproj, -np/--parallel > 1) and disables MTP with a warning when those flags are supplied. llamacpp_start_server() and build_llamacpp_server_args() now accept an extra_argv kwarg that flows into the conflict guard and lands at the tail of the spawned argv. An out-of-range or non-integer LLAMACPP_SPEC_DRAFT_N_MAX is surfaced as a notes entry on the MTP result. Reference: https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF.

• llamacpp-tuner skill for coding-agent performance tuning: new Claude Code skill that helps optimize llama.cpp server configuration for coding-agent workloads. Includes benchmark agent and configuration profiles.

Bug Fixes

• Resolve mypy and ruff failures from v0.13.1 CI.
• Apply ruff format and rename ambiguous l identifier in bench_agent.py.

Chore

• Level up llamacpp-tuner skill to A grade.

Full Changelog: v0.13.1...v0.14.0