Update README: duration 60→120s

Bump @spaces.GPU duration 60→120s (32B Q4_K_M first-call can exceed 60s)

update README: Gradio API endpoint, cu130/CUDA 13, cold-start notes, confirmed working

switch to cu130 wheel — native CUDA 13 build, no compat shim needed

production v5: patch all .so.13→.so.12 nvidia libs + RTLD_GLOBAL preload all

production v4: binary-patch soname libcudart.so.13→.so.12 + RTLD_GLOBAL preload

production v3: patchelf add-rpath + RTLD_GLOBAL preload for cuda abi

diagnostic v10: dlopen/RPATH debug inside @spaces.GPU

production v2: correct cuda compat path (nvidia/cu13/lib), n_gpu_layers=-1

diagnostic v9: find libcudart at startup (no GPU quota)

cuda compat inside worker + lazy model load + n_gpu_layers=-1

load model at startup, inference-only inside @spaces.GPU

cpu-only: use CPU llama-cpp wheel, n_gpu_layers=0, no CUDA dep

production v1: CUDA 13 compat shim + CPU fallback (n_gpu_layers=0)

diagnostic v8: find actual libcudart.so.13 path

diagnostic v7: full CUDA 12→13 shim + ldd output

CUDA compat shim: symlink libcudart.so.12→.so.13 for Blackwell ZeroGPU

build llama-cpp-python from source (CMAKE_ARGS=-DGGML_CUDA=on, CUDA 13.0)

diagnostic v6: CUDA version check

pre-download GGUF at startup (outside GPU lease) to fix ZeroGPU timeout

reduce @spaces.GPU duration to 60s to fit ZeroGPU free-tier limit

production: @spaces.GPU + Gradio call API (llama_cpp GGUF 32b)

diagnostic v5 with unique version marker

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

initial commit