Skip to content

v0.2.0 — Model Hub, Variant Selection & Vision/Audio Fix

Latest
Latest

Choose a tag to compare

View all tags
@ayushh0110 ayushh0110 released this 05 Aug 22:12
· 5 commits to main since this release
v0.2.0
45b208f

Fixes the <unused49> regression that broke screenshot analysis and voice transcription. Adds a Model Hub for downloading and switching between model variants.

⚠️ Breaking: Model Source Change

ScreenMind now downloads models from ggml-org (official) instead of Unsloth, and stores them in ~/.screenmind/models/ instead of the HuggingFace cache. Existing users need to re-download their model from the new Model Hub.

See Discussion #14 for background on the mmproj incompatibility.

Migration: Clean up old cached models

The old Unsloth models in ~/.cache/huggingface/ are no longer used and can be safely deleted to free disk space (~3-5 GB):

# Windows
rmdir /s /q "%USERPROFILE%\.cache\huggingface\hub\models--unsloth--gemma-4-E2B-it-GGUF"
rmdir /s /q "%USERPROFILE%\.cache\huggingface\hub\models--unsloth--gemma-4-E4B-it-GGUF"

# macOS / Linux
rm -rf ~/.cache/huggingface/hub/models--unsloth--gemma-4-*

✨ Features

  • Model Hub with variant selection — pick your preferred quantization (Q4_0, Q8_0, BF16, etc.) per model from the dashboard or settings. Download, switch, and delete variants independently. (model_manager.py, core.js, settings.js)

  • External server detection — if llama-server is already running on port 5809, ScreenMind adopts it instead of launching its own. Known models show as active; unknown models show an amber warning. (main.py)

  • Voice memo resilience — memos are always saved to DB even if transcription fails or returns empty. Audio file is preserved for playback. (main.py)

  • Parallel analysis pipeline — embedding + developer context detection now run concurrently via asyncio.gather (~2-3s faster per coding screenshot). (analysis_worker.py)

🐛 Bug Fixes

  • Fix vision & audio producing <unused49> garbage — switched from Unsloth to ggml-org models; their mmproj files stay compatible with llama.cpp updates
  • <unusedN> token stripping — safety net regex filter in llm_client.py strips garbage tokens from audio output
  • Fix: Voice recorder crash (SRDC) — ScreenCapture (mss) uses Windows GDI handles that have thread affinity. The capture object was being created on the main thread but used from the hotkey listener thread, causing intermittent crashes when taking a screenshot during voice memo recording. Now creates a fresh ScreenCapture() per invocation with a try/except so a screenshot failure doesn't kill the voice memo. (voice_recorder.py)
  • Embedder API rename — get_sentence_embedding_dimension() → get_embedding_dimension() for newer sentence-transformers (embedder.py)

📦 Models

Model Variants VRAM
Gemma 4 E2B (2B) Q4_0 · Q8_0 · BF16 ~4 GB
Gemma 4 E4B (4B) Q4_0 · Q8_0 · BF16 ~6 GB
Gemma 4 12B IQ3_M · Q4_K_M · Q5_K_M · Q6_K · Q8_0 ~10 GB

New users default to Gemma 4 E2B Q4_0 (~1.5 GB download).

🧪 Tests

  • test_model_manager.py rewritten: 56 → 80 tests (variant lifecycle, external server detection)
  • test_llm_client.py: +3 tests for token stripping
  • 473 tests passing

Upgrade

pip install --upgrade screenmind

After upgrading, open Model Hub and download your preferred variant.

Full diff: v0.1.2...v0.2.0
PyPI: screenmind 0.2.0

Assets 2
Loading