File size: 7,242 Bytes
881f9f2 d606d10 881f9f2 d606d10 881f9f2 d606d10 881f9f2 d606d10 881f9f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 | #!/bin/bash
# One-time setup for vLLM + Llama-3.1-8B-Instruct on Insomnia.
#
# The team operates out of a single shared checkout in edu scratch:
# /path/to/smartgridbench
# Only (re)run this script after coordinating with the team — the venv it
# creates is shared across everyone, and wiping it mid-job breaks in-flight
# work.
#
# --- FIRST-TIME CHECKLIST (do these in order; only the last step is this script) ---
#
# 1. Browser-approve access to the gated Llama 3.1 model:
# https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
# Meta's auto-approval is typically instant. Check status at
# https://huggingface.co/settings/gated-repos . Until approved, every
# `hf download` will fail with a 401.
#
# 2. Create a HF access token (Read scope only) at
# https://huggingface.co/settings/tokens
# and export it in your shell:
# export HF_TOKEN=<huggingface_token>
#
# 3. SSH into Insomnia and cd to the shared team checkout:
# cd /path/to/smartgridbench
#
# 4. Run this script.
#
# --- SETUP_MODE env var ---
#
# SETUP_MODE=all (default) venv + model
# SETUP_MODE=venv venv only; skip model download (use if the shared venv
# is broken but the model is already on disk)
# SETUP_MODE=model model only; skip venv (use if the model dir was wiped
# or you want a specific revision without touching the venv)
#
# --- MODEL_REVISION env var ---
#
# Leave unset to use the repo-standard resolved commit SHA. For
# reproducibility in benchmark runs, keep this explicit in logs, e.g.:
# export MODEL_REVISION=0e9e39f249a16976918f6564b8830bc894c89659
# The script prints the resolved SHA after the download completes so you
# can capture it for the next invocation. See
# docs/governance/model_registry.yaml for the current repo-level model
# contract around this pin.
#
# Usage:
# cd /path/to/smartgridbench
# export HF_TOKEN=<huggingface_token>
# bash scripts/setup_insomnia.sh
#
# See also: docs/insomnia_runbook.md (cluster gotchas), docs/runbook.md
# (reproducibility), docs/governance/model_registry.yaml (current model/runtime
# contract), and issue #111 (why this script is shaped like this).
set -euo pipefail
REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
VENV_DIR="$REPO_ROOT/.venv-insomnia"
MODEL_DIR="$REPO_ROOT/models"
REQUIREMENTS_FILE="$REPO_ROOT/requirements-insomnia.txt"
MODEL_NAME="meta-llama/Llama-3.1-8B-Instruct"
DEFAULT_MODEL_REVISION="0e9e39f249a16976918f6564b8830bc894c89659"
MODEL_REVISION="${MODEL_REVISION:-$DEFAULT_MODEL_REVISION}"
SETUP_MODE="${SETUP_MODE:-all}"
case "$SETUP_MODE" in
all|venv|model) ;;
*)
echo "ERROR: SETUP_MODE=$SETUP_MODE (expected all|venv|model)" >&2
exit 1
;;
esac
do_venv=0
do_model=0
case "$SETUP_MODE" in
all) do_venv=1; do_model=1 ;;
venv) do_venv=1 ;;
model) do_model=1 ;;
esac
echo "=== Insomnia Environment Setup ==="
echo "Repo root: $REPO_ROOT"
echo "Mode: $SETUP_MODE"
echo "Model: $MODEL_NAME @ $MODEL_REVISION"
for cmd in python3 uv; do
if ! command -v "$cmd" >/dev/null 2>&1; then
echo "ERROR: required command not found: $cmd" >&2
exit 1
fi
done
if [ "$do_model" = "1" ] && [ -z "${HF_TOKEN:-}" ]; then
echo "ERROR: HF_TOKEN must be exported for model download." >&2
echo "Accept the Llama license at https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct" >&2
echo "and create a Read-scope token at https://huggingface.co/settings/tokens" >&2
exit 1
fi
if [ "$do_venv" = "1" ] && [ ! -f "$REQUIREMENTS_FILE" ]; then
echo "ERROR: requirements file not found: $REQUIREMENTS_FILE" >&2
exit 1
fi
# --- Step 1: Create Python venv -----------------------------------------
if [ "$do_venv" = "1" ]; then
if [ ! -d "$VENV_DIR" ]; then
echo ""
echo "[1] Creating Python 3.11 virtual environment with uv..."
uv venv "$VENV_DIR" --python 3.11
else
echo "[1] Virtual environment already exists at $VENV_DIR"
fi
# --- Step 2: Install dependencies ---
echo ""
echo "[2] Installing pinned dependencies from requirements-insomnia.txt..."
uv pip install --python "$VENV_DIR/bin/python" -r "$REQUIREMENTS_FILE"
# Metadata-only version check (does NOT import torch / vllm — those imports
# are heavy enough to draw a warning email from RCS about login-node abuse;
# see docs/insomnia_runbook.md §"Login node etiquette").
echo ""
echo "Installed versions (from package metadata; no heavy imports):"
"$VENV_DIR/bin/python3" - <<'PY'
from importlib.metadata import PackageNotFoundError, version
for pkg in ("torch", "vllm", "transformers", "huggingface-hub", "nvidia-cudnn-cu12"):
try:
print(f" {pkg:20s} {version(pkg)}")
except PackageNotFoundError:
print(f" {pkg:20s} NOT INSTALLED")
PY
else
echo "[1/2] Skipped venv creation (SETUP_MODE=$SETUP_MODE)."
fi
# --- Step 3: HuggingFace login + download -------------------------------
if [ "$do_model" = "1" ]; then
# Prefer venv python if available so we pick up the pinned hf CLI,
# otherwise fall back to the system python3 + user-site install.
if [ -x "$VENV_DIR/bin/python3" ]; then
PYTHON_BIN="$VENV_DIR/bin/python3"
else
PYTHON_BIN="python3"
fi
echo ""
echo "[3] HuggingFace auth (needed for the gated Llama model)..."
# `python -m huggingface_hub.commands.huggingface_cli login` matches the
# actual CLI surface in the shared huggingface-hub 0.36.x environment.
# `huggingface-cli login` is still a thin backward-compat shim that works
# too, but this form keeps the script independent of an extra entry point.
"$PYTHON_BIN" -m huggingface_hub.commands.huggingface_cli login \
--token "$HF_TOKEN"
echo ""
echo "[4] Downloading $MODEL_NAME @ $MODEL_REVISION (~16 GB)..."
echo " This may take 10-30 minutes on the first run."
mkdir -p "$MODEL_DIR"
"$PYTHON_BIN" -m huggingface_hub.commands.huggingface_cli download \
"$MODEL_NAME" \
--revision "$MODEL_REVISION" \
--local-dir "$MODEL_DIR/Llama-3.1-8B-Instruct"
# Report the resolved commit SHA so reproducibility captures can pin it.
echo ""
echo "Resolved revision for reproducibility:"
"$PYTHON_BIN" - "$MODEL_NAME" "$MODEL_REVISION" <<'PY'
import sys
from huggingface_hub import HfApi
repo_id, revision = sys.argv[1], sys.argv[2]
info = HfApi().model_info(repo_id, revision=revision)
print(f" repo: {repo_id}")
print(f" revision: {info.sha} (requested: {revision})")
print(f" pin with: export MODEL_REVISION={info.sha}")
PY
else
echo "[3/4] Skipped model download (SETUP_MODE=$SETUP_MODE)."
fi
echo ""
echo "=== Setup Complete ==="
echo "Venv: $VENV_DIR"
echo "Model dir: $MODEL_DIR/Llama-3.1-8B-Instruct"
echo "Pinned deps: requirements-insomnia.txt"
echo ""
echo "Next steps:"
echo " - Submit the vLLM serving job from this directory:"
echo " sbatch --mail-type=BEGIN,END,FAIL --mail-user=\$MAIL_USER scripts/vllm_serve.sh"
echo " - Or run a benchmark cell:"
echo " sbatch scripts/run_experiment.sh configs/example_baseline.env"
|