depth_anything + dpt + zoedepth / depth-estimation: all models pass wmk perf

Summary

All depth estimation models fail. depth_anything and zoedepth fail because no OnnxConfig is registered for their model types. dpt models fail due to an import error (build_hf_model not found) and an incomplete export pipeline.

Eval Results (2026-03-11)

10/10 models fail — 0 pass.

Root Cause

1. depth_anything, zoedepth, depth_pro: ModelKit's modelkit/models/hf/depth_anything.py exists but imports only trigger registration of DepthAnythingIOConfig. Check whether the ONNX config is correctly registered for all sub-variants (depth_anything, depth_anything_v2, zoedepth, depth_pro).

2. dpt: cannot import name 'build_hf_model' from 'modelkit.build' indicates a stale import path — likely a refactor that moved build_hf_model without updating all import sites. Also has export pipeline stall for dpt-hybrid-midas.

Current State

• modelkit/models/hf/depth_anything.py — exists, registers DepthAnythingIOConfig, but may not cover all depth model types
• No dpt.py or zoedepth.py in modelkit/models/hf/
• modelkit/build/__init__.py — build_hf_model import broken for dpt

Desired State

All 10 depth estimation models above pass wmk perf.

Acceptance Criteria

• All 6 depth_anything (v1 + v2) models pass wmk perf
• Both dpt models pass wmk perf
• Intel/zoedepth-nyu-kitti passes wmk perf
• apple/DepthPro-hf passes wmk perf
• Fix is universal — no hardcoded model variant names (CLAUDE.md Cardinal Rule #1)
• uv run pytest tests/ passes (CLAUDE.md Cardinal Rule #3)
• Eval re-run confirms 10/10 pass

Technical Notes

• Check modelkit/models/hf/depth_anything.py — verify it registers configs for depth_anything, depth_anything_v2, and related model_type strings
• For zoedepth and depth_pro: add ONNX config registrations (they are ViT-based depth models similar to DPT/DepthAnything)
• For dpt import error: trace the build_hf_model import chain in modelkit/build/__init__.py and fix the broken import — likely a refactor artifact
• DPTForDepthEstimation uses a ViT backbone — check whether the ViT ONNX config is compatible

Related Files

• modelkit/models/hf/depth_anything.py — existing depth_anything config (needs audit)
• modelkit/build/__init__.py — broken build_hf_model import for dpt
• modelkit/models/hf/__init__.py:32 — depth_anything import line
• eval_results/2026-03-11/models/Intel__dpt-large__depth-estimation/result.json
• eval_results/2026-03-11/models/depth-anything__Depth-Anything-V2-Small-hf__depth-estimation/result.json

References

• CLAUDE.md Cardinal Rule #1: No hardcoded model type strings
• CLAUDE.md Cardinal Rule #3: Run uv run pytest tests/ after implementation

Start with modelkit/models/hf/depth_anything.py, modelkit/models/hf/init.py:32, and modelkit/build/init.py; trace the reported registration and build_hf_model failures, then inspect the referenced eval result files. Done means all 10 listed models pass wmk perf and uv run pytest tests/ passes without hardcoded model variants.