mortred_model_server

Unsupported boundaries (no false expectations)

Mortred is sold and documented for a narrow, supported envelope. This page is the single list of what is out of scope. Source and CMake are ground truth; CHANGELOG is reference-only.

First-hour + full ops: deployment.md (mortredctl next).

Operating system

Supported	Not supported
Linux x64, Ubuntu 20.04 / 22.04, including WSL2 Ubuntu	Windows native, macOS, other CPU ISAs

Build scripts, packs, mortredctl, and release tarballs assume Linux. Do not expect a supported Windows/macOS binary line.

Models / catalog

Item	Status
HTTP catalog (mortred-model-server.out --list)	Served ids in README Model Zoo
RTDETR	Scaffold only — generated by scripts/new_model.py; hooks return MODEL_NOT_IMPLEMENTED; not registered in the factory / HTTP catalog. Do not treat as a shipped detector.
MOT	No MOT HTTP service
Bench-only ids	Listed under README “Bench-only”; not the SME HTTP path

GPU stack (gpu profile)

Supported	Not supported
CUDA 12.x	CUDA 11 / 13+ as the product line
TensorRT 10.x (libnvinfer.so.10, headers under TensorRT-10*)	TensorRT 8 and 9 (leftover libnvinfer.so.8* fails configure; pin is 10.3)
ORT 1.29 with matching headers (ORT_API_VERSION==29)	Mixing ORT 1.18 headers with 1.29 libs (configure fail-closed; see SME-10)

CMake and install_deps.sh refuse leftover TRT 8 / mismatched ORT. There is no supported “TRT 9 halfway” path — use TRT 10 or the cpu profile.

TensorRT engines

Engines are bound to this GPU + this TensorRT build. Unsupported as an out-of-box path:

• Copying .engine files from another machine / another TRT major
• Assuming a release tarball’s engines run on your card without convert
• Skipping mortredctl prepare --pack … when the pack lists TensorRT backends

Supported path: convert on the target machine for the current pack (mortredctl prepare, deployment §10). Zoo-wide auto-build stays opt-in.

Weight interchange (Hugging Face)

Hugging Face (MaybeShewill-CV/mortred_model_server) is the interchange store, not a runtime-engine store. The long-term artifact is ONNX. Product conf/model/**/*.toml stay type=mnn or type=tensorrt except already-ONNX ids (diffusion, MSOCRNET, SAM decoders) and LIBFACE (YuNet 2026may). CI overlays such as conf/ci/yolov8_onnx_hosted.toml are not product serving.

Unsupported as a product path:

• Treating a downloaded .engine as portable across GPU / TensorRT majors
• Switching other product tomls to type=onnx “because HF only has ONNX”
• Reverse-engineering ONNX from .mnn / .model / .engine
• Deleting CPU-profile .mnn files from HF before a local convert_mnn path can materialize the paths those tomls already name

Source of truth for per-id ONNX / contract / license: conf/onnx_sources.json. Export dual-file rules and remaining gaps: onnx-interchange.md. Fetch still uses conf/weights_manifest.json. During the cutover, HF still holds some .mnn / .model files; those are leftover runtime copies, not the target layout.

cpu profile

Supported	Not supported on cpu
MNN-CPU + ORT-CPU	Compiling / linking TensorRT into the cpu preset
Demo pack without TRT backends	Expecting type=tensorrt model configs to run (init fails clearly)

What this page is not

• Not a feature roadmap (RTDETR stays scaffold until registered).
• Not a substitute for mortredctl doctor --strict or install_deps --check.

---

不支持边界（摘要）

• Linux x64：原生 Ubuntu 与 WSL2 Ubuntu 使用同一套命令和二进制。无 Windows 原生 / macOS 产品线。
• RTDETR：脚手架，未进 HTTP catalog，勿当已上线检测器。
• GPU：CUDA 12.x + TensorRT 10.x only（8/9 不支持）。
• Engine：须在本机本卡用当前 TRT 重建；禁止跨机拷贝当开箱路径。
• HF 权重：长期只发 ONNX；产品 toml 不改成 type=onnx（LIBFACE YuNet 与已有 ONNX id 除外）；禁止从 MNN/engine 反推。
• cpu profile：无 TRT；tensorrt 配置会失败。

This site is open source. Improve this page.