HT-compat compatibility matrix

Sibling to the OpenAI compatibility matrix. Rows are the HT-compat extension endpoints from the HT-compat spec; columns are the major OSS implementations that have started to converge on the canonical signatures.

Run aioc probe URL --profile ht to populate the data for a new server. PRs that update a cell should link the report.

Legend: ✅ pass · ⚠️ pass-with-deviation · ❌ not implemented · — out of scope

HT-compat-1.0 endpoints

Endpoint	vLLM omni	vanilla llama.cpp	comfy-openai shim	OpenAI
/v1/reranking	⚠️	⚠️	❌	—
/v1/segmentations	❌	❌	❌	—
/v1/audio/segmentations	❌	❌	❌	—
/v1/chat/completions (omni)	✅	❌	❌	—
/v1/images/decompositions	❌	❌	❌	—
/v1/3d/generations	❌	❌	✅	—
/v1/videos	❌	❌	✅	—

HT-compat-1.1 endpoints (encoder tasks)

Endpoint	TEI	HF Inference API	vanilla llama.cpp	vLLM	OpenAI
/v1/qa	❌	⚠️ (/inputs+parameters, no envelope)	❌	❌	—
/v1/ner	❌	⚠️ (/inputs+parameters, no envelope)	❌	❌	—
/v1/classifications	⚠️ (/predict, no envelope)	⚠️ (/inputs+parameters, no envelope)	❌	❌	—

Reference implementations

The HT-compat spec aligns to one reference implementation per endpoint. These are the upstreams we cribbed signatures from; the matrix above tracks which servers have adopted the canonical shape.

Endpoint	Reference implementation
/v1/reranking	Cohere Rerank v2 · Jina Reranker · vLLM Cohere-compat
/v1/segmentations	Meta SAM3 (paper + reference Python)
/v1/audio/segmentations	Meta SAM-Audio (paper + reference Python)
/v1/chat/completions[omni]	vLLM-Omni serving Qwen2.5-Omni
/v1/images/decompositions	Qwen-Image-Layered via fal.ai
/v1/3d/generations	TRELLIS-2 + Hunyuan3D via ComfyUI workflow shim
/v1/videos	OpenAI Sora signature (HT-implemented; no OSS impls yet)
/v1/qa (v1.1)	HF transformers question-answering pipeline · HF Inference API
/v1/ner (v1.1)	HF transformers token-classification pipeline · HF Inference API
/v1/classifications (v1.1)	HF transformers text-classification + zero-shot-classification · TEI /predict

Scope per fork

HT-compat is a buffet, not a checklist — most forks will only plausibly cover a subset.

Server family	Plausible HT-compat surface
llama.cpp + extensions	/v1/reranking (already routed; needs --reranking boot flag and a rerank-class model).
vLLM omni-capable builds	/v1/chat/completions[omni] is the only known reference impl today.
audio-segmentation servers	/v1/audio/segmentations once a SAM-Audio HTTP wrapper exists.
TEI deployments	/v1/classifications (matches /predict semantics); /v1/reranking (matches /rerank). /v1/qa and /v1/ner need a thin adapter — TEI doesn't currently expose those task types as HTTP endpoints.
ComfyUI workflow shims	/v1/images/{generations,edits,decompositions}, /v1/videos, /v1/3d/generations — anything whose pipeline already lives as a ComfyUI workflow.

A fork running aioc probe --profile ht on a server that only implements a subset should set fail-on: none (discovery mode); the report renders without failing the build. Flip to fail-on: FAIL once the server has wired up the endpoints it claims.

Caveats

• Wider than typical compat matrix. HT-compat targets model classes OpenAI doesn't have endpoints for, so most cells start ❌ by definition — the table tracks adoption rather than current parity.
• The OpenAI column is — throughout. HT-compat sits in OpenAI's gaps; if OpenAI ships a /v1/segmentations we'll re-evaluate.
• ⚠️ for vLLM rerank because vLLM's rerank endpoint is Cohere-compatible (no /v1/ prefix); the response shape matches.
• ✅ for comfy-openai 3D generations — first HT-compat endpoint with a working OSS-stack implementation. Hunyuan3D-2 served via a ComfyUI workflow shim; HTTP 202 on submission, GLB returned via /v1/3d/generations/{id}/content.
• ✅ for comfy-openai videos — handler accepts both image and image_url (Pydantic alias) so payloads from either convention validate cleanly.
• ⚠️ for llama.cpp rerank because llama-server ships the /v1/reranking route from upstream and, when booted without --reranking, returns 501 with the canonical OpenAI error envelope — the HT-compat-1.0 capability-gating contract. A server booted with --reranking and a rerank-class model would flip to ✅; first probe report against that config is welcome.
• ⚠️ for TEI /predict and HF Inference API tasks in the v1.1 table because the underlying engines do the work, but the wire shape differs from HT-compat: bare arrays instead of the {id, model, ..., usage} envelope, and inputs+parameters nesting instead of flat field names. A two-file FastAPI adapter in front of either engine is enough to flip to ✅.