# Dockerfile.finetune — Cover letter LoRA fine-tuner (QLoRA via unsloth)
# Large image (~12-15 GB after build). Built once, cached on rebuilds.
# GPU strongly recommended. CPU fallback works but training is very slow.
#
# Tested base: pytorch/pytorch:2.3.0-cuda12.1-cudnn8-runtime
# If your GPU requires a different CUDA version, change the FROM line and
# reinstall bitsandbytes for the matching CUDA (e.g. bitsandbytes-cuda121).
FROM pytorch/pytorch:2.3.0-cuda12.1-cudnn8-runtime

WORKDIR /app

# Build tools needed by bitsandbytes CUDA kernels and unsloth
RUN apt-get update && apt-get install -y --no-install-recommends \
    gcc g++ git libgomp1 \
    && rm -rf /var/lib/apt/lists/*

# Install training stack.
# unsloth detects CUDA version automatically from the base image.
RUN pip install --no-cache-dir \
    "unsloth @ git+https://github.com/unslothai/unsloth.git" \
    "datasets>=2.18" "trl>=0.8" peft transformers \
    "bitsandbytes>=0.43.0" accelerate sentencepiece \
    requests pyyaml

COPY scripts/ /app/scripts/
COPY config/  /app/config/

ENV PYTHONUNBUFFERED=1
# Pin to GPU 0; overridable at runtime with --env CUDA_VISIBLE_DEVICES=
ENV CUDA_VISIBLE_DEVICES=0

# Runtime env vars injected by compose.yml:
#   OLLAMA_URL              — Ollama API base (default: http://ollama:11434)
#   OLLAMA_MODELS_MOUNT     — finetune container's mount path for ollama models volume
#   OLLAMA_MODELS_OLLAMA_PATH — Ollama container's mount path for same volume
#   DOCS_DIR                — cover letters + training data root (default: /docs)

ENTRYPOINT ["python", "scripts/finetune_local.py"]