Drop attention_mask at inference: the checkpoint was trained without one, and generate()'s auto-mask degenerates the output

Serve with greedy decoding: sampling hallucinates on this lightly-tuned checkpoint

fix: unwrap BatchEncoding from apply_chat_template before generate()

diagnostics: full traceback in gen_error

diagnostics: surface gen_error + vram in /api/health

serving: manual load + GenerationMixin + use_cache=False + streaming

serving: manual load + GenerationMixin + use_cache=False + streaming

add app.py