Starstruk Logo MarkStarstruk

v0.4.0

get-convex/convex-backendv0.4.0Nov 9, 2025by github-actions[bot]

AI Summary

Adds the PaddleOCR-VL backend (SigLIP + Ernie 0.9B) alongside DeepSeek-OCR, implements lazy loading to defer weight mmap, introduces model-aware prompts with bilingual feedback, and unifies sampling controls.

Key Highlights

  • New PaddleOCR-VL backend selectable alongside DeepSeek-OCR
  • Lazy loading feature to defer weight mmap and reduce startup time
  • Model-aware prompts with bilingual Markdown feedback for requests
  • Unified sampling controls (--do-sample, --temperature, --top-p, etc.) for CLI and Server

New Features

  • PaddleOCR-VL backend integration
  • Lazy loading implementation
  • Model-aware prompt construction
  • Unified sampling controls
  • Refactoring of common utilities into deepseek-ocr-core

Full Release Notes

## New Features

  - `PaddleOCR‑VL` backend (SigLIP + Ernie 0.9B with FlashAttention) is now selectable alongside `DeepSeek‑OCR`. Documentation covers model switching and architecture/memory differences.
  - Model-aware prompts with bilingual Markdown feedback when requests omit <image> placeholders.
  - Lazy loading: the server defers weight mmap until the first request, reducing startup time.
  - Server/CLI sampling controls: `--do-sample`, `--temperature`, `--top-p`, `--top-k`, `--repetition-penalty`, and `--no-repeat-ngram-size` are now recognized across both entry points.

  ## Improvements

  - Common utilities (token sampling, embedding gathers, transformer KV cache) moved into deepseek-ocr-core, trimming duplication between DeepSeek and Paddle crates.
  - Documentation clarifies multi-model selection, DeepSeek-only dynamic crop mode, bilingual terminology, and per-flag behavior.
  - CLI/Server prompt builders choose the correct format per model, improving output quality without manual tweaks.

  ## Bug Fixes

  - ModelScope provider now respects arbitrary repo IDs and exact file paths, fixing Paddle asset downloads that previously fetched the wrong config.json.
  - Requests without images no longer throw transport errors; both sync and streaming responses return a structured bilingual warning instead.
  - Prompt/image mismatches surface as normal assistant replies instead of opaque “prompt formatting failed” errors, keeping clients compatible with standard OpenAI flows.