Update README.md

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

Relax huggingface_hub pin to <1.0 to satisfy transformers 4.55-4.57

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

Log + surface RuntimeErrors from witness chat too (still 503 for those)

Surface witness chat failures with traceback + error class in 500 detail

Add the 4 map layer PNGs missing from the initial Docker-era ship (ship_space.sh excluded data/*.png)

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

Surface zerogpu backend load_error / load detail in setup status

Eagerly import zerogpu_backend on Spaces so @spaces.GPU is registered before startup scan

Force demo.launch to bind 0.0.0.0:$PORT on HF Spaces (CLI hot-reload ignores GRADIO_SERVER_NAME)

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

Expose `demo` at module scope so Gradio SDK runner can launch the gr.Server app

Default provider to zerogpu_transformers on HF Spaces; drop bogus README env block

Gate setup/llama subprocess paths behind provider check; allow zerogpu_transformers

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

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

Fix port collision: scope PORT env to llama.cpp subprocess

Add Space README

Ship Phantom Grid Docker Space

initial commit