Skip to content

feat: complete v0.4.0 drop-in readiness work - #13

Merged
espetro merged 23 commits into
mainfrom
feat/batched-inference
Sep 21, 2026
Merged

espetro merged 23 commits into
mainfrom
feat/batched-inference

Conversation

@espetro

@espetro espetro commented Sep 21, 2026

Copy link
Copy Markdown
Owner

Summary

Closes the remaining evidence and compatibility gaps identified by the original DINOv2 swap audit.

  • Strict numeric CLI parsing: rejects nonnumeric, partial, whitespace, overflow, and invalid-range values for all numeric flags with clean stderr and exit 1.
  • Preview binary embeddings: adds --embeddings-binary with a versioned little-endian D2EMB v1 format, CLS + pooled defaults, optional patch vectors, L2 normalization, and file/directory output via -o.
  • Parity evidence: commits reproducible regular and register-token small-model Hugging Face comparisons with final-SHA provenance, checksums, thresholds, and 8/8 passing gates.
  • Ubuntu benchmark evidence: replaces placeholder data with checksum-backed Ubuntu CPU measurements and preserves multi-platform benchmarking as a deliberate follow-up.
  • Backbone-only DINOv2: supports feature-mode conversion/loading, validates register metadata and tensor shapes, and rejects classifier requests cleanly when classifier heads are absent.
  • Converter compatibility: supports direct DINOv2 backbone and classifier-wrapper model structures.
  • Workflow hardening: uses the supported hf download command and preserves benchmark provenance artifacts.

Evidence

  • Regular small parity: CLS 0.999911, pooled 0.999928, flat patches 0.999571, top-1 match, probability delta 0.000039.
  • Register small parity: CLS 0.999984, pooled 0.999984, flat patches 0.999847, top-1 match, probability delta 0.006114.
  • Ubuntu small f16 CPU benchmark: mean 255.0 ms, stddev 9.1 ms, peak RSS 104 MB, 2 threads, 1 warmup, 5 runs.
  • Backbone-only regular and register small conversions and feature/PCA/binary smoke tests passed.

Test plan

  • Debug CTest: 3/3
  • ASAN CTest: 3/3
  • UBSAN CTest: 3/3
  • Strict parser regression tests
  • Benchmark parser regression tests
  • Final regular/register parity runs: 8/8 each
  • Binary D2EMB payload independently compared with JSON
  • Backbone-only conversion and feature-mode smoke tests
  • Classifier-only backbone rejection
  • hf download workflow/actionlint validation

Deferred by decision

  • Multi-platform benchmarks until adoption justifies the CI cost
  • Quantization-specific benchmark mode
  • DINOv3 and other task-specific heads
  • Stable binary format guarantees, since the D2EMB output is explicitly preview

No release was created by this PR. The branch is ready for a v0.4.0 release decision after merge.

@espetro
espetro merged commit ce9429b into main Sep 21, 2026
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant