Fix adapter loading: use set_peft_model_state_dict (plain load_state_dict silently no-opped the LoRA)

Fix: use pre-quantized int8 model + manual CPU adapter loading

doc: clarify duration=120 vs actual inference time

Fix: device_map=auto for base (ZeroGPU auto-transfer) + lazy adapter

Fix: base on CPU (device_map=cpu), adapter lazily inside @spaces.GPU

Fix: device_map=cpu for ZeroGPU compat (avoids CUDA emulation hooks with bitsandbytes+safetensors)

Fix: load LoRA adapter inside @spaces.GPU (safetensors fails in CUDA emulation mode)

Fix: re-upload correct app.py (ZeroGPU optimization)

Restore duration=120 for stability; keep replace-based im_end stripping

Use replace instead of removesuffix for im_end

Fix: strip newline before removesuffix

Use removesuffix instead of regex for im_end stripping

Reduce ZeroGPU duration to 30s; add HF_TOKEN for higher quota

Fix: strip <|im_end|> post-decode; eos_token_id is useless (model generates im_end as subword tokens)

Fix: pass eos_token_id to generate() and add safety truncation at im_start tokens

Load base in int8 to match training quantization

Remove Palmaitis/Klussis reference

Switch to Apertus-8B youtube adapter; XX→PR only; show fixed system prompt

Fall back to device_map=cpu when no GPU (bnb 4-bit CPU backend)

Fix generate(): return_dict=True + generate(**inputs) (BatchEncoding has no .shape)

Lazy-load model inside GPU fn; drop removed show_copy_button (gradio 6)

Add Old Prussian translator Gradio demo