Dockerfile

# Use CUDA 11.8 with cuDNN 8
FROM nvidia/cuda:11.8.0-cudnn8-devel-ubuntu20.04

ENV DEBIAN_FRONTEND=noninteractive

# Install basic utilities
RUN apt-get update && apt-get install -y \
    tzdata \
    software-properties-common \
    curl && \
    ln -fs /usr/share/zoneinfo/Etc/UTC /etc/localtime && \
    echo "Etc/UTC" > /etc/timezone

# Install Python 3.11
RUN add-apt-repository ppa:deadsnakes/ppa && apt-get update && \
    apt-get install -y python3.11 python3.11-dev

# Install pip
RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python3.11
RUN update-alternatives --install /usr/bin/python python /usr/bin/python3.11 1

# Install Python packages (torch, xFormers, Transformers, etc.)
RUN python -m pip install --no-cache-dir \
    torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 && \
    python -m pip install --no-cache-dir \
    xformers==0.0.27 \
        --extra-index-url https://download.pytorch.org/whl/nightly/cu118/torch_nightly.html && \
    python -m pip install --no-cache-dir \
    transformers \
    accelerate \
    trl \
    unsloth \
    pandas \
    datasets \
    huggingface_hub \
    safetensors \
    bitsandbytes

# Helps reduce CUDA memory fragmentation
ENV PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

# Set your cache directories
ENV HF_HOME=/workspace/.huggingface
ENV TRANSFORMERS_CACHE=/workspace/.huggingface
ENV DATASETS_CACHE=/workspace/.cache
ENV TORCH_HOME=/workspace/.cache
ENV XDG_CACHE_HOME=/workspace/.cache
ENV TRITON_CACHE_DIR=/workspace/.cache/triton

# Create directories and set permissions
RUN mkdir -p /workspace/outputs \
    && mkdir -p /workspace/.huggingface \
    && mkdir -p /workspace/.cache/triton \
    && chmod -R 777 /workspace

WORKDIR /workspace

# Copy training script and data
COPY finetune_script.py /workspace/
COPY train.csv /workspace/train.csv
COPY valid.csv /workspace/valid.csv

# Preconfigure Accelerate for multi-GPU usage (4 GPUs, fp16)
RUN mkdir -p /root/.cache/huggingface/accelerate && \
    echo '{"compute_environment": "LOCAL_MACHINE", "distributed_type": "MULTI_GPU", "num_processes": 4, "mixed_precision": "fp16", "machine_rank": 0, "main_training_function": "main", "use_cpu": false, "num_machines": 1}' \
    > /root/.cache/huggingface/accelerate/default_config.yaml

# Launch your training script on 4 GPUs with fp16
CMD ["accelerate", "launch", "--num_processes=4", "--mixed_precision=fp16", "finetune_script.py"]