setup.sh

#!/usr/bin/env bash
# =============================================================================
# setup.sh — reproducible environment for the ensemble-pipeline repo
#
# Usage (fresh VM):
#   hf download MikeGreen2710/ensemble-pipeline setup.sh --repo-type dataset --local-dir .
#     (or: wget https://huggingface.co/datasets/MikeGreen2710/ensemble-pipeline/resolve/main/setup.sh)
#   bash setup.sh                # installs miniconda if needed + creates env 'avm'
#   conda activate avm
#
# Version policy:
#   - If requirements_train.txt exists next to this script, it is installed
#     verbatim (exact VM1 parity — preferred).
#   - Otherwise falls back to the known-good pins below (the stack every fix
#     in the July 2026 debugging session was validated against).
#   - transformers MUST stay 4.x: the generator/inference scripts use 4.x
#     Trainer APIs (custom compute_loss KL objective, custom optimizer
#     injection). transformers 5.x breaks these, sometimes silently.
# =============================================================================
set -euo pipefail

ENV_NAME="${1:-avm}"
PYTHON_VERSION="3.12"
MINICONDA_DIR="/content/miniconda3"

echo "== ensemble-pipeline setup: env='$ENV_NAME' python=$PYTHON_VERSION =="

# ── 1. conda: use existing install if any, else miniconda ────────────────────
if command -v conda >/dev/null 2>&1; then
    echo "-- conda already present: $(conda --version)"
elif [ -x "$MINICONDA_DIR/bin/conda" ]; then
    echo "-- miniconda present at $MINICONDA_DIR"
else
    echo "-- installing miniconda -> $MINICONDA_DIR"
    ARCH=$(uname -m)
    curl -fsSL "https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-${ARCH}.sh" -o /tmp/miniconda.sh
    bash /tmp/miniconda.sh -b -p "$MINICONDA_DIR"
    rm /tmp/miniconda.sh
    "$MINICONDA_DIR/bin/conda" init bash
    echo "-- NOTE: conda init written to ~/.bashrc (takes effect in new shells)"
fi
# make conda usable in THIS shell regardless of init state
source "$(conda info --base 2>/dev/null || echo "$MINICONDA_DIR")/etc/profile.d/conda.sh"

# ── 1b. accept default-channel ToS (required by Miniconda 2025+; no-op on old) ─
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main 2>/dev/null || true
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r    2>/dev/null || true

# ── 2. env: create if missing ────────────────────────────────────────────────
if conda env list | grep -qE "^${ENV_NAME}[[:space:]]"; then
    echo "-- env '$ENV_NAME' exists — reusing (packages will be updated to pins)"
else
    conda create -y -n "$ENV_NAME" "python=$PYTHON_VERSION"
fi
conda activate "$ENV_NAME"
echo "-- python: $(python --version) @ $(which python)"

# ── 3. packages ──────────────────────────────────────────────────────────────
python -m pip install --upgrade pip

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
if [ -f "$SCRIPT_DIR/requirements_train.txt" ]; then
    echo "-- installing from requirements_train.txt (exact parity)"
    pip install -r "$SCRIPT_DIR/requirements_train.txt"
else
    echo "-- requirements_train.txt not found; installing known-good pins"
    # torch: default wheel bundles CUDA 12.x runtime; override with
    #   TORCH_SPEC='torch==2.x.y --index-url https://download.pytorch.org/whl/cu121'
    pip install ${TORCH_SPEC:-torch}
    pip install \
        "transformers==4.56.2" \
        "huggingface_hub==0.34.4" \
        "accelerate==1.10.1" \
        "datasets>=2.19,<4" \
        "safetensors>=0.4" \
        "tokenizers>=0.19" \
        "sentencepiece" \
        "protobuf" \
        "scikit-learn==1.7.2" \
        "lightgbm" \
        "pandas" \
        "numpy<3" \
        "pyarrow" \
        "joblib" \
        "matplotlib" \
        "tqdm"
fi

# ── 4. verification gate — fails loudly if the stack is wrong ────────────────
python - << 'PYEOF'
import sys
import torch, transformers, accelerate, safetensors, datasets, sklearn
import huggingface_hub as hub
print("torch        :", torch.__version__, "| cuda:", torch.cuda.is_available())
print("transformers :", transformers.__version__)
print("huggingface  :", hub.__version__)
print("accelerate   :", accelerate.__version__)
print("sklearn      :", sklearn.__version__)
major = int(transformers.__version__.split(".")[0])
assert major == 4, (
    "transformers %s is not 4.x — the pipeline's Trainer customisations "
    "(KL compute_loss, custom optimizer) are 4.x APIs" % transformers.__version__)
print("\nOK — environment ready. Activate with: conda activate " + "${ENV_NAME}".strip() if False else "")
PYEOF

echo ""
echo "== done. next steps =="
echo "   conda activate $ENV_NAME"
echo "   python scripts/hf_sync.py pull --repo MikeGreen2710/ensemble-pipeline   # or selective hf download"