Skip to content

CLAUDE.md

Latest commit

 

History

History
20 lines (14 loc) · 1.74 KB

CLAUDE.md

File metadata and controls

20 lines (14 loc) · 1.74 KB
Raw

CLAUDE.md

Project

Multi-platform AI inference runtime (Snapdragon / Hexagon focus). Languages: C/C++ (SDK), Go (CLI), Python (bindings), Java/JNI (Android). Build systems: Bazel (CLI) + CMake (SDK).

Hard constraints

  • Never move or reuse a published git tag. If the wrong tag shipped, cut a higher one.
  • Do not modify third-party code.
  • Follow CONTRIBUTING.md for branch naming, commit / PR title format, pre-commit checks, and the FFI-update rule when changing public SDK headers.
  • Compute-unit alias mapping lives in the SDK: sdk/src/device.cpp (geniex_resolve_device). CLI, pybind, and Android are all thin wrappers over that one function — do not re-implement the alias table in a binding. Aliases: cpu / gpu / npu / hybrid, with npu = pin HTP0 and hybrid = empty device_id (llama.cpp's per-tensor HTP+CPU scheduler, the fast path on Snapdragon). n_gpu_layers passes through the SDK unchanged for llama_cpp gpu / npu / hybrid; the CLI/binding default is -1, which llama.cpp reads as "all layers". cpu forces it to 0, and qairt forces it to 0 (its plugin rejects any non-zero value). Default when the user passes nothing: npu for both llama_cpp and qairt. QAIRT is NPU-only — other aliases are coerced with a warning, never an error. When changing the table, rebuild the SDK bridge (/build), then re-sync the table in notes/run.md § Compute-unit aliases. Any change to geniex.h is an FFI change — see CONTRIBUTING.md.

Workflows

  • Build anything (CLI / SDK bridge / release installer) → run /build.
  • Cut a release / bump the version → run /release.
  • Onboarding the AI setup itself → see notes/AI.md.