# Imagen base CUDA 12.1 *runtime*: trae las libs CUDA para que el wheel # precompilado de llama-cpp-python offloadee a la T4. NO compilamos (compilar # en el builder de HF muere por OOM); usamos wheel ya hecho del indice cu121. FROM nvidia/cuda:12.1.1-runtime-ubuntu22.04 ENV DEBIAN_FRONTEND=noninteractive # Python + Stockfish + libs de OpenCV (sin toolchain: no compilamos nada) RUN apt-get update && apt-get install -y --no-install-recommends \ python3 python3-pip \ stockfish \ libgl1 libglib2.0-0 \ && rm -rf /var/lib/apt/lists/* WORKDIR /app # Requirements primero para aprovechar layer cache COPY requirements.txt . # numpy ANTES que torch (evita downgrade implícito) RUN pip3 install --no-cache-dir --upgrade pip \ && pip3 install --no-cache-dir numpy==1.26.0 # torch con CUDA 12.1 RUN pip3 install --no-cache-dir torch torchvision \ --index-url https://download.pytorch.org/whl/cu121 # llama-cpp-python: SOLO wheel precompilado con CUDA (--only-binary). Sin fijar # version → toma la más nueva que tenga wheel cu121 para py3.10. Si no hay wheel, # falla en SEGUNDOS (no intenta compilar 25 min). La API que usa qwen.py es estable. RUN pip3 install --no-cache-dir --only-binary llama-cpp-python \ llama-cpp-python \ --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121 # Resto de dependencias de la app RUN pip3 install --no-cache-dir -r requirements.txt # Pre-descarga del coach (LoRA fine-tuned, GGUF) en build → arranque rápido en # runtime. Repo público en HF Hub; ajústalo con --build-arg si cambias el nombre. ARG QWEN_COACH_REPO=SamilG2/qwen-coach-checkscan ARG QWEN_COACH_FILENAME=qwen-coach-q4_k_m.gguf RUN mkdir -p /app/models && \ QWEN_COACH_REPO="$QWEN_COACH_REPO" QWEN_COACH_FILENAME="$QWEN_COACH_FILENAME" \ python3 -c "from huggingface_hub import hf_hub_download; import os; hf_hub_download(repo_id=os.environ['QWEN_COACH_REPO'], filename=os.environ['QWEN_COACH_FILENAME'], local_dir='/app/models')" # Código de la app COPY . /app ENV STOCKFISH_PATH=/usr/games/stockfish \ QWEN_MODEL_PATH=/app/models/qwen-coach-q4_k_m.gguf \ COACH_MODEL=qwen \ STOCKFISH_THREADS=4 \ STOCKFISH_HASH_MB=256 \ PYTHONUNBUFFERED=1 EXPOSE 7860 CMD ["gunicorn", "-w", "1", "-b", "0.0.0.0:7860", "--timeout", "600", "app:app"]