dinov2 / image-feature-extraction: all models pass wmk perf

Summary

facebook/dinov2-giant fails with Error parsing message with type 'onnx.ModelProto' — the giant DINOv2 model exceeds the protobuf 2GB limit. Smaller variants (small, base, large) all pass.

Eval Results (2026-03-11)

2/8 fail — only large/giant variants.

Root Cause

dinov2-giant (~1.1B params) generates an ONNX file exceeding 2GB. Same root cause as xlm-roberta (#429) — requires save_as_external_data=True during export.

dinov2-large (~307M params) times out — may be a compilation timeout on the larger attention graph.

Current State

• No dinov2.py in modelkit/models/hf/ — relies on Optimum defaults
• Small/base/xray/rad-dino all pass → export path is functional for smaller models
• External data format not enabled for large models

Desired State

All DINOv2 variants pass wmk perf, including dinov2-giant.

Acceptance Criteria

• facebook/dinov2-giant passes wmk perf
• facebook/dinov2-large passes wmk perf
• Existing small/base/dino-vitb16/dino-vits16 variants continue to pass
• Fix is size-based (universal) — not dinov2-specific (CLAUDE.md Cardinal Rule #1)
• uv run pytest tests/ passes (CLAUDE.md Cardinal Rule #3)

Technical Notes

• Fix: enable use_external_data_format=True for exports above ONNX size threshold — coordinate with #429 (xlm-roberta) as the same fix applies
• For dinov2-large TIMEOUT: may need an increased timeout for larger ViT models, or optimization to speed up QNN compilation for large attention graphs
• This issue will be automatically resolved once the shared external-data-format fix from #429 is implemented

Related Files

• modelkit/export/config.py — WinMLExportConfig, check for use_external_data_format field
• eval_results/2026-03-11/models/facebook__dinov2-giant__image-feature-extraction/result.json

References

• Blocked by / coordinate with: #429 (xlm-roberta ONNX size fix — same root cause and fix)

Start with modelkit/export/config.py and inspect WinMLExportConfig for external-data support, then review issue #429 and the referenced result.json. Run the DINOv2 wmk perf evaluations and uv run pytest tests/. Done means large and giant pass while the existing smaller variants remain passing and the fix is size-based.