Fixes#5: Add compatibility shim for find_pruneable_heads_and_indices
which was removed in transformers 4.40.0. All 12 GPT2 model files now
use a try/except fallback.
Fixes#6: Add ultra low memory mode for 12GB VRAM systems.
- New memory_mode dropdown: auto, normal, low, ultra
- Auto mode detects available VRAM and selects best mode
- Ultra mode uses ~6GB peak VRAM with aggressive cleanup
- Added VRAM monitoring and logging during generation
Bump version to 1.3.0
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Remove unused flashy imports from codeclm/utils/utils.py
- Add SequenceSummary fallback stub for transformers 4.50+ compatibility
- Remove transformers version assertion in levo.py (bundled llama works with newer versions)
- Add conf/vocab.yaml with pinyin tokens for OmegaConf resolver
- Fix OmegaConf load_yaml resolver to use package-relative paths
- Add embedding layer resizing to handle tokenizer vocabulary mismatches
- Fix relative path resolution for model checkpoints and third_party paths
- Remove flashy from requirements.txt
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
