Strip MiniCPM <think>...</think> reasoning tags from generated text

Pin transformers<5.0 so KV cache works with MiniCPM bundled code; re-enable use_cache

Revert to use_cache=False (eager attn also broken); hide all audio UI

Speed up: eager attn + KV cache; drop chat retries to 1; remove MiniCPM-o voice UI artifact

Cap zerogpu max_new_tokens at 256 (use_cache=False makes long generations O(n^2))

Disable KV cache: openbmb modeling_minicpm.py has a cache_utils API drift bug

Move .to('cuda') inside @spaces.GPU; background thread keeps model on CPU to avoid emulation bypass

Shim is_torch_fx_available so MiniCPM trust_remote_code import works on transformers >= 5.0

Use canonical .to('cuda') pattern + progress logs so container log shows what loader is doing

Load model in background thread so health/status endpoints don't block on 16GB download

Load model on cuda at module level (canonical ZeroGPU pattern)

Refactor: Docker+llama.cpp -> Gradio SDK + ZeroGPU transformers backend

Ship Phantom Grid Docker Space