Commit History
-
fix: critical T5 conditioner key sanitization and metadata
Verified signature- extract_t5.py: sanitize .layer.0. → .layer_0. and .layer.1. → .layer_1. for MLX ModuleParameters.unflattened() compatibility. Without this fix, all 24 T5 transformer block weights remain at random initialization. Also adds automatic T5 variant detection (small/base/large). - config.json: correct T5 metadata from t5-small to t5-large (d_model=1024, 24 layers, 16 heads, d_ff=4096). - README.md: fix text encoder reference, add T5 extraction section, MLX key sanitization note, T5 unscaled attention note, update Swift usage. - t5/model.safetensors: regenerated T5-large weights with sanitized keys (.layer.0. → .layer_0.) for MLX compatibility. - verify_t5.py: new file — Python MLX reference implementation for verifying T5 encoder output against the Swift implementation.9b56e00
