Restore full ASR model - all fixes applied

Restore ASR app with suppress_tokens fix (dev mode disabled)

Debug: minimal ZeroGPU+Gradio test (no model)

Fix 500: remove server_name param, keep default launch()

Fix 500: low_cpu_mem_usage, server_name=0.0.0.0, remove container=False

Fix AttributeError: remove generation_config.language from print statement

Fix IndexError: set suppress_tokens=None (was empty [] in generation_config)

Add error catching to display full traceback in UI for debugging

Fix ValueError: separate model.half() and model.to(cuda), remove language/task from generate

Fix IndexError: replace forced_decoder_ids with language/task in generate()

Fix ZeroGPU: load CPU at startup, move to GPU inside @spaces.GPU only

Fix ZeroGPU: load model to GPU at startup, create pipe globally

GPU: ZeroGPU A10G + float16 + @spaces.GPU decorator

Redesign: elegant dark-gradient UI, mic 5min, file upload 25min

Fix UI: combine mic+upload into single Audio component for Gradio 5.x

Fix: load processor from base model (adapter repo has no tokenizer files)

Fix: monkey-patch peft _torch_supports_distributed=False to bypass DTensor AttributeError

Fix: pre-import torch.distributed.tensor + pin peft==0.18.1 torch>=2.4.0

Fix: Load LoRA adapter correctly using PeftModel + base model

Fix: Remove WhisperTokenizer, use pipeline generate_kwargs for language/task

Fix: Remove spaces.GPU decorator and torch.cuda.amp.autocast for CPU compatibility

Fix: remove max_length param (gradio 5 compat)

Add app.py