mortred_model_server

ONNX interchange export

Hugging Face (MaybeShewill-CV/mortred_model_server) is the ONNX interchange store, not a portable-engine store. Per-id contract, source, and license live in conf/onnx_sources.json. Bytes and sha256 live in conf/weights_manifest.json.

This page is the export/adopt rulebook and the remaining-gap list. Serving and fetch: deployment.md §9. Out of scope: unsupported-boundaries.md.

Policy

• Dual files on disk: {stem}.static_bs1.onnx (batch=1) and {stem}.dyn.onnx (symbolic batch on dim0). Spatial dims stay as the product graph needs them (frozen H/W, or height/width when the C++ path is dynamic).
• Keep the original {stem}.onnx when it already existed (product toml, trt_engines.json, hosted HF names).
• Do not reverse .mnn / .model / .engine into ONNX.
• Product conf/model/**/*.toml stay type=mnn or type=tensorrt, except:
  • already-ONNX product ids: diffusion (DDPM / DDIM / CLS_COND_DDIM / LDM), MSOCRNET, SAM decoders
  • LIBFACE after the YuNet 2026may cutover (type=onnx, 12-head decode)
• CI-only overlays under conf/ci/*_onnx_hosted.toml prove ORT CPU. Do not point conf/server at those files.
• HF upload of the 39 product *.static_bs1.onnx graphs is the fetch_weights.py --kind onnx set (kind=onnx in conf/weights_manifest.json). .engine files are not that download. on_hf stays false until those bytes are uploaded to MaybeShewill-CV/mortred_model_server.

How to export or adopt

From the repo root. Prefer official checkpoints / official ONNX, then write_pair_from_static.

Job	Script
Copy a static graph into the dual pair (batch dim only)	scripts/export_onnx/adopt_official_onnx.py
Same, plus Reshape [1,-1,…] → [0,-1,…] on the dyn file	scripts/export_onnx/adopt_legacy_pairs.py (deep: True)
Freeze rank-4 H/W then pair	scripts/export_onnx/freeze_spatial.py
Dump IO	scripts/export_onnx/dump_io.py
Stamp conf/onnx_sources.json	scripts/export_onnx/update_onnx_sources_p2.py
Rewrite CI overlays to static_bs1	scripts/export_onnx/write_ci_overlays.py
Shared helpers	scripts/export_onnx/common.py (write_pair_from_static, set_batch_dim, apply_reshape_batch_passthrough)

Family exporters (from-scratch): export_classification.py, export_bisenetv2.py, export_attentive_gan.py, export_enlighten.py, export_ppmatting.py, export_pphuman.py, export_realesrgan.py, export_superpoint.py, export_dinov2_timm.py, export_clip.py, plus YOLO / NanoDet / DBNet adopt paths recorded in onnx_sources.json onnx_source.script.

LibFace check (ORT + the C++ YuNet decoder):

python3 scripts/verify_libface_yunet.py

Inventory (2026-09-23)

Counted from conf/onnx_sources.json against files under weights/. 50 graphs: 48 have a dual pair on disk (36 local + 12 hosted); 2 blocked; 0 missing.

HTTP product graphs with dual files include classification (MobileNetV2 / ResNet / DenseNet / DINOv2), YOLO v5/v6/v7/v8s, NanoDet, LibFace YuNet, CenterFace, DBNet, BiSeNetV2, PP-HumanSeg v2-mobile, HRNet, MODNet, PP-Matting 512, EnlightenGAN, AttentiveGAN, Real-ESRGAN, SuperPoint, Depth Anything, Metric3D, SAM AMG (MobileSAM encoder + sm86 decoder), and the diffusion UNets / KL decoder.

Bench / sibling dual files (not HTTP product rows): LightGlue extractor/matcher, OpenAI CLIP visual/textual, MSOCRNET fp16, FastSAM s/x, nano_sam, vit_l decoder, YOLOv8 n/l/x, YOLOv7x, PP-HumanSeg lite/server, PP-Matting 1024 / resnet34 / v2.

Still open

Graph	Status	Why
SAM_PREDICTOR vit_l encoder	blocked	Decoder already has dual ONNX. Encoder is sam_vit_l_encoder.mnn (~1.2 GB). No official encoder ONNX adopted. Do not reverse the MNN.
RTDETR	blocked	Scaffold only (MODEL_NOT_IMPLEMENTED). Not in the HTTP catalog. Skip until registered.

No remaining onnx_status=missing rows. The previous LibFace loc/conf block is closed: C++ reads YuNet cls_* / obj_* / bbox_* / kps_*; interchange is face_detection_yunet_2026may.{static_bs1,dyn}.onnx (original 2026may kept).

Honest caveats (not “missing”)

• SAM AMG sm61 engine has no matching ONNX; interchange uses the sm86 decoder pair.
• Classification trio is torchvision ImageNet-1K V1 + NHWC wrap; C++ still applies caffe mean/std (golden may drift vs caffe MNN).
• DBNet used PaddleOCR ch_ppocr_server_v2.0_det_infer frozen 544×960; backbone may differ from db_model_large.mnn.
• DINOv2 pairs interpolate official 518 pos-embed to 224.
• CLIP vocab (bpe_simple_vocab_16e6.txt) is not ONNX.
• LLM weights/llm/embedding/ is out of this CV catalog.

Adding a new interchange graph

1. Export or adopt so both {stem}.static_bs1.onnx and {stem}.dyn.onnx exist and match the C++ IO names/shapes in onnx_sources.json contract.
2. Do not change product type to onnx unless the runtime weight is gone and the detector was rewritten for that graph (LibFace is the documented case).
3. Stamp onnx_source via update_onnx_sources_p2.py (or the legacy stamper).
4. Add a conf/ci/*_onnx_hosted.toml overlay if the id should prove ORT CPU.
5. Leave HF upload off until the remaining blocked graphs are adopted or waived.

This site is open source. Improve this page.