Skip to content

v3.5.0 — CPU accuracy parity, strict request validation, source restructure

Choose a tag to compare

View all tags
@nataell95 nataell95 released this 20 Jul 13:10
· 9 commits to main since this release
Immutable release. Only release title and notes can be modified.
v3.5.0
6cbf057

Highlights

  • CPU accuracy parity with GPU: FUNSD-50 F1 82.28% → 85.52% (GPU: 85.33%), medium model 91.03% → 91.88%. The CPU recognizer no longer stretches small crops to a 320px floor (single-warp crop at the shared 32px floor) and detection preprocessing is BGR per model spec.
  • Request validation on a declarative per-endpoint spec. Three silent failures are now loud 400s (ignoring them falsified responses):
    • route_table/route_formula overrides the endpoint or build cannot honor
    • text=0 (layout-only) where no layout-only path exists
    • /ocr/markdown?embed=0 (file-ref assets were written server-side, unreachable over HTTP)
  • Zero breaking changes otherwise: other unsupported query params behave exactly as v3.4.0 (ignored) and now return X-Ignored-Params + X-Deprecation: unsupported-param-tolerance; removed-in=v4 headers. Set TURBO_OCR_STRICT_QUERY_PARAMS=1 to reject them today; v4 will reject them by default.
  • New capabilities: per-request backend routing on /ocr/pixels, /ocr/batch, and over gRPC (route_table/route_formula request fields); gRPC layout_only, per-request PDF reading_order, and Health.capabilities_json (the exact /capabilities document); parameter validation reads the URL query string only (POST bodies are no longer parsed as parameters).
  • Source restructure: modules split into single-concern translation units grouped by domain; duplicated logic unified; dead code removed. Unit tests grow to 294 cases.

Images

  • GPU (TensorRT, x86_64): ghcr.io/aiptimizer/turboocr:v3.5.0 (also :latest)
  • CPU (ONNX Runtime, amd64+arm64): ghcr.io/aiptimizer/turboocr-cpu:v3.5.0 (also :latest, and ghcr.io/aiptimizer/turboocr:v3.5.0-cpu)

Models are unchanged from v3.4.0 — existing model releases apply.

Verification

Benches at or above the v3.4.0 baseline: GPU tiny 85.33% @ ~590 img/s, batch 85.25% @ ~90 img/s, medium 92.33% @ ~92 img/s, CPU tiny 85.52%. Full endpoint × parameter matrix diffed against a live v3.4.0 container (status-identical outside the bug-fix class above); all endpoints validated over HTTP and gRPC on both images pulled fresh from GHCR, including cold-start engine builds and clean-clone native builds (CPU and GPU).

Assets 3
Loading