This directory is the shared documentation root for public release material. It may contain:
- GitHub-facing Markdown documents linked directly from the root
README.md. - MkDocs site content under MkDocs-specific source trees. None is present here; see "The MkDocs manual" below.
- Git/GitHub workflow documents for maintainers and contributors.
The current top-level files are GitHub-facing documents linked from the root README. They should remain readable as standalone Markdown files even if richer MkDocs pages are added later.
Current GitHub-facing docs:
installation.md- platform prerequisites and build flags.environment-variables.md-MLXCEL_*runtime, build, downloader, cache, and diagnostic knobs.benchmarks.md- benchmark methodology and the requirements for future raw result tables.supported-models.md- maintained architecture/checkpoint support matrix.architecture.md- runtime architecture and major components.distributed.md- tensor/pipeline parallel setup and limitations.turbo-kv-cache.md- TurboQuant modes, the unified paged KV cache, quality/performance trade-offs, and flags.CONTINUOUS_BATCHING.md- continuous-batching scheduler, paged decode, and disaggregated prefill/decode/router serving.responses-api.md- implemented/v1/responsessubset and gaps.server-features.md- operator guide to server routes, model-store and router semantics, LoRA, settings, streaming, observability, transport, and security boundaries.audio-api.md- implemented/v1/audioendpoints: audio-capable chat and Whisper STT setup, container and streaming behavior, request/response reference, WAV output, and validation order.adding-models.md- contribution guide for new model architectures.block-diffusion.md- DiffusionGemma block-diffusion generation: canvas denoising vs autoregressive, CLI flags, throughput, and phase 1 limitations.python-client.md- themlxcelPython client over the OpenAI-compatible server, covering managed and connect modes, streaming, chat, structured output, theopenai_clientescape hatch, async usage, and troubleshooting.audio-preprocessing.md- shared WAV/MP3/FLAC model-input normalization, loaded family policy, resource limits, cancellation, metrics, and current XLA capability boundary.speculative-acceptance.md- speculative-decoding acceptance rules, which rule each code path runs, the distribution-preservation guarantee and its RNG dependency, the kill switch, and how to read the active rule off a log.mla-absorbed-decode.md- DeepSeek-family matrix-absorbed MLA decode over a compressed-latent KV cache: the identity, the cache layout, the flags, and what is and is not verified.cascade-attention.md- shared-prompt-prefix (cascade) decode: computing a prefix shared by several concurrent sequences once per step instead of once per sequence, the two-level decomposition, the flags, and how to tell which path a launch took.sparse-paged-decode.md- sparse attention reduced to page indirection over the fused v2 decode kernel: the addressing argument, the MiniMax-M3 routing, why DeepSeek Sparse Attention is not routed, where the dispatch floor sits, the kill switch, and the benchmark harness.mtp-policy-api.md- the supported read interface for the adaptive B=1 MTP verdict (GET /v1/internal/mtp-policy): the response body, the four states, the unavailable reasons, and the schema versioning and compatibility policy.code-guidelines.md- the file-size and module-split thresholds, including when to extract a<name>_helpers.rsand when inline tests move to a sibling_tests.rs, the// Used by:annotation convention for shared functions, the JIT kernel rule that every varying input dtype must appear intemplate_argsbecause CUDA keys the compiled-module cache on it, and theHashMapiteration-order rule covering what counts as an order-sensitive consumer, why a stable sort on a non-total key does not pin the result, the fresh-map-per-iteration testing requirement, and why no static check enforces it.embeddings.md- the/v1/embeddingsand/v1/rerankendpoints plusmlxcel embedandmlxcel rerank: embedding and reranker detection, pooling and scoring, multimodal and late-interaction inputs, request/response schemas, server flags, error codes, and family-registration guidance.llama-server-compat.md- the verified llama-server b10621 compatibility boundary: how to read the machine-readable manifest undercompat/llama-server/b10621/, the compatibility-policy states, the area sharding used by epic #1431, the CI gates, and the deterministic regeneration procedure.
adr/ holds numbered Architecture Decision Records, one significant decision per file, immutable once Accepted. See adr/README.md for the index.
The manual published at https://mlxcel.lablup.ai/en/manual/ is not built from
this directory. Its sources (docs/en, docs/ko, docs/shared,
docs/requirements.txt, docs/scripts) are maintained in a separate
documentation tree and are not part of this repository.
The four mkdocs configs at the repository root (mkdocs.yml, mkdocs.ko.yml,
mkdocs.pdf.yml, mkdocs.ko.pdf.yml) belong to that tree and are kept here in
sync with it. Read their docs_dir, custom_dir, and nav: entries as paths
into that tree: none of them names a file listed above, and none of the files
listed above appears in a nav:. That is deliberate, not drift. The
GitHub-facing documents here are meant to read as plain Markdown on GitHub, and
the manual is a separately authored artifact.
The docs-* Makefile targets build the manual from those sources. In this
repository they stop immediately with an explanation rather than failing partway
through a uv, symlink, or site build. Read the published manual instead.
Expected future layout examples:
docs/github/...for GitHub issue/PR/release workflow notes.docs/git/...for branch, commit, tag, and mirroring procedures.
Keep root README links stable unless the corresponding top-level document is intentionally replaced with a redirect-style index page.