Summary - DistilBERT fill-mask/question-answering and BLIP image-to-text e2e regressions Root causes - SDPA attention mask dtype mismatch - decoder kwargs path incompatibility - BLIP decoder mask shape mismatch Fixes - export IO bool normalization for binary attention_mask - robust transformers compat patching - decoder_wrapper kwargs support - BLIP 2D mask pass-through Validation - focused unit tests passed - e2e passed: question-answering, fill-mask, image-to-text-fp16
Summary
Root causes
Fixes
Validation