- axil module:
libaxil-qllm is now an axil module rather than a
standalone daemon; xy_install() registers the ask/chat telnet commands
and the OpenAI-compatible HTTP endpoints.
- OpenAI-compatible embeddings: new
POST /v1/embeddings endpoint —
single-string input, standard OpenAI JSON shape (data[0].embedding).
- Worker-thread inference: LLM inference runs on the module's worker
thread, so it never blocks axil's single-threaded event loop.
- Sliding-window context rework: ported the sliding-window branch — model
cache with refcounting, multi-context accounting (n_contexts),
qllm_compress/qllm_anchor_*, qllm_n_ctx, history management
(qllm_prime, qllm_chat, qllm_render, qllm_reset), and the
sampler/grammar API (qllm_set_grammar, qllm_sampler_*).
- Mock-based test suite: unit (
tests/unit/core), edge (tests/edge)
and stress (tests/stress) tests compiled against llama/gguf/corm/vulkan
mocks — make test.
- Tools: the
qllm-chat client plus the qllm-list and qllm-path
helpers.
- Dependencies updated; README shows how to download a GGUF model with
curl
(no huggingface-cli needed).
Verification
gpg --verify CHECKSUMS.asc
sha256sum -c CHECKSUMS
