inferna 0.4.0

Changes since the last Release

Added

• scripts/rwt.py checks the images the sd cases write -- exit 0 from inferna.sd only meant a file was written, so a NaN render (all black) passed. A case now also fails if its PNG is the wrong size or has a per-channel standard deviation below 2. The decoder is stdlib-only because neither the script nor the wheel venv has an image library.

• rwt.py clean --keep-images and run --keep-images -- leave z_turbo_*.png in the project root so a run's images can be inspected. Each sd case deletes its own image before it runs, so the next run replaces them.

• inferna server --api-key / --api-key-file require Authorization: Bearer <key> on both server types. /health and the static webui files stay public, so the webui can load; it sends the key configured in its settings. The flag names match llama-server. The embedded server's native bridge now passes request headers to Python; before, handle_http_request always received {}.

Changed

• llama.cpp sync to b10964 (v0.4.1, commit b29c606) -- bumped LLAMACPP_VERSION in scripts/manage.py (b10809 -> b10964). No binding changes were needed. The facade exports the new ggml_prec values GGML_PREC_UNDEFINED, _BF16, _F16, _Q8 and _Q4; upstream deprecates GGML_PREC_DEFAULT as an alias of UNDEFINED.

• stable-diffusion.cpp shares llama.cpp's ggml by default -- SD_USE_VENDORED_GGML now defaults to off in CMakeLists.txt and manage.py, matching cyllama. The macOS static wheel had linked SD's vendored ggml, which only the older Metal MSL patch covered. Local --dynamic builds that leave the flag unset now build llama.cpp from source, because the prebuilt release uses GGML_MAX_NAME=64. --sd-vendored-ggml or SD_USE_VENDORED_GGML=1 restores the old behaviour.

• whisper.cpp sync to v1.9.4 -- bumped WHISPERCPP_VERSION in scripts/manage.py (v1.9.2 -> v1.9.4). No changes on the wrapped API. Its ggml now takes ggml-metal-pin-msl-version-set-lang.patch.

• rwt.py sd cases use Z-Image-Turbo's sampling settings -- --steps 8 --cfg-scale 1.0, per upstream stable-diffusion.cpp docs/z_image.md, with a fixed --seed 42. They had run the CLI defaults, 20 steps at cfg 7.0: five times the diffusion passes of 8 steps at cfg 1.0. On an M1 the sd family dropped from 3065 s to 875 s, measured with cyllama's identical copy of the script. The fixed seed makes a backend's images comparable across releases.

• Ruff and mypy target Python 3.12, matching requires-python. Neither change adds lint findings, because UP rules are not selected. Fixed the mypy errors that already failed make typecheck. The SDContextParams.backend and params_backend annotations are for mypy only.

• New python-lint.yml PR workflow runs ruff check, ruff format --check and mypy against the tool versions in uv.lock, without building the extension. Seven files that ruff format had not been applied to were reformatted so the check starts green.

• PR CI runs tests/test_llama_cpp_surface.py. It needs no model and catches drift in the binding facade that the agent layer imports.

Removed

• scripts/patches/ggml-metal-pin-msl-version.patch and -perkind.patch -- no default build matches either. -set-lang now covers llama.cpp, whisper.cpp and the shared ggml sd.cpp links. An SD_USE_VENDORED_GGML=1 Metal build ships sd.cpp without the MSL pin.

Security

• EmbeddedServer now listens on exactly --host. A host of 127.0.0.1 or localhost was rewritten to 0.0.0.0. The embedded server is the default for inferna server, and the servers have no auth, so through 0.3.2 the default command exposed inference to every host that could reach the port. Pass --host 0.0.0.0 to expose the server deliberately.

• Both servers log a warning when --host is not a loopback address. Without --api-key, a non-loopback bind lets any reachable host run inference. No warning is logged when an API key is set. Warning rather than refusing keeps existing unauthenticated LAN setups working.

• Agent examples in the docs no longer call eval on tool input. Tool arguments come from the model, so a prompt-injected document could run arbitrary code through a copied example; eval(..., {"__builtins__": {}}) does not prevent this. The examples use inferna.agents.tools.calculator, and tests/test_docs_no_eval.py rejects new eval( calls in Markdown.