File size: 9,563 Bytes
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 | #!/bin/bash
#SBATCH --job-name=lane2-int8-smoke
#SBATCH --account=edu
#SBATCH --partition=short
#SBATCH --qos=short
#SBATCH --gres=gpu:A6000:1
#SBATCH --mem=64G
#SBATCH --cpus-per-task=8
#SBATCH --time=01:00:00
#SBATCH --output=logs/lane2_int8_smoke_%j.out
#
# Lane 2 / #29 INT8 quantization startup + reachability smoke.
# Validates that vLLM 0.19.0 can serve a CompressedTensors INT8 (W8A8) variant
# of Llama-3.1-8B-Instruct and that the served model is reachable + responds
# to a one-shot completion. NOT a full benchmark — just startup + /v1/models
# + one /v1/completions round-trip.
#
# Status: deferred per docs/lane2_int8_kv_status.md. Run this only after team
# decision to revive INT8 for Cell C v2 / a model-scaling cell.
#
# Prerequisites (one-time):
# 1. HuggingFace gating approval for the INT8 checkpoint (typically
# RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8 or equivalent).
# 2. HF_TOKEN exported in the team .env (or in this shell).
# 3. ~16 GB free in models/ for the INT8 checkpoint (separate from the
# existing FP16 checkpoint).
#
# Usage:
# sbatch --mail-type=BEGIN,END,FAIL --mail-user=$USER@example.edu \
# scripts/test_int8_smoke.sh
#
# Override the model + revision via env if you want to try a different INT8 build:
# INT8_MODEL_REPO=other/llama-int8-repo \
# INT8_MODEL_REVISION=<sha> \
# sbatch ... scripts/test_int8_smoke.sh
set -euo pipefail
REPO_ROOT="${SLURM_SUBMIT_DIR:-$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)}"
cd "$REPO_ROOT"
umask 0002
mkdir -p logs
JOB="${SLURM_JOB_ID:?expected to be running inside a Slurm job}"
INT8_MODEL_REPO="${INT8_MODEL_REPO:-RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8}"
INT8_MODEL_REVISION="${INT8_MODEL_REVISION:-main}"
INT8_LOCAL_DIR="${INT8_LOCAL_DIR:-models/Llama-3.1-8B-Instruct-int8}"
PORT="${PORT:-8001}" # Different from default 8000 so doesn't collide with a live vLLM
OUT_DIR="benchmarks/lane2/int8_smoke/${JOB}"
mkdir -p "$OUT_DIR"
echo "=== Lane 2 / #29 INT8 Startup + Reachability Smoke ==="
echo "Node: $(hostname)"
echo "Slurm job: $JOB"
echo "INT8 model: $INT8_MODEL_REPO @ $INT8_MODEL_REVISION"
echo "Local dir: $INT8_LOCAL_DIR"
echo "Port: $PORT"
echo "Out dir: $OUT_DIR"
echo "Started: $(date -u +%Y-%m-%dT%H:%M:%SZ)"
echo ""
# CUDA setup (don't use module load cuda, broken on Insomnia)
export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH:-}
# Activate the team-shared venv
source .venv-insomnia/bin/activate
# Source the repo-root .env (if present) so HF_TOKEN / WANDB_API_KEY land in
# the environment without the caller having to remember to export them. The
# main runner (scripts/run_experiment.sh) does this; this standalone smoke
# previously didn't, which made the "HF_TOKEN can live in the team .env"
# usage doc misleading.
if [ -f "$REPO_ROOT/.env" ]; then
set -a
# shellcheck disable=SC1091
source "$REPO_ROOT/.env"
set +a
fi
# Step 1: download INT8 checkpoint if not already present
if [ ! -d "$INT8_LOCAL_DIR" ] || [ -z "$(ls -A "$INT8_LOCAL_DIR" 2>/dev/null)" ]; then
echo "--- Downloading $INT8_MODEL_REPO @ $INT8_MODEL_REVISION ---"
if [ -z "${HF_TOKEN:-}" ]; then
echo "ERROR: HF_TOKEN not set. Either add HF_TOKEN=... to the repo-root .env or export it before sbatch." >&2
exit 1
fi
mkdir -p "$INT8_LOCAL_DIR"
.venv-insomnia/bin/python -m huggingface_hub.commands.huggingface_cli \
download "$INT8_MODEL_REPO" \
--revision "$INT8_MODEL_REVISION" \
--local-dir "$INT8_LOCAL_DIR" \
--token "$HF_TOKEN"
else
echo "--- INT8 checkpoint already present at $INT8_LOCAL_DIR ---"
fi
# Step 2: launch vLLM with --quantization compressed-tensors
echo ""
echo "--- Starting vLLM with --quantization compressed-tensors ---"
VLLM_LOG="$OUT_DIR/vllm.log"
setsid python3 -u -m vllm.entrypoints.openai.api_server \
--model "$INT8_LOCAL_DIR" \
--served-model-name "Llama-3.1-8B-Instruct-int8" \
--host 127.0.0.1 \
--port "$PORT" \
--max-model-len 8192 \
--quantization compressed-tensors \
--generation-config vllm \
>"$VLLM_LOG" 2>&1 &
VLLM_PID=$!
# `setsid` puts the leader and all its children in a new process group whose
# pgid equals the leader's PID. Killing the pgid (-PID) reaps vLLM's worker
# processes too; killing only the leader leaves workers pinned to the GPU
# until Slurm hard-kills the job at the --time= boundary. Mirrors the cleanup
# pattern in scripts/run_experiment.sh:698-710.
VLLM_PGID="$VLLM_PID"
cleanup_vllm() {
if [ -n "${VLLM_PGID:-}" ] && kill -0 -- "-$VLLM_PGID" 2>/dev/null; then
kill -TERM -- "-$VLLM_PGID" 2>/dev/null || true
sleep 2
kill -KILL -- "-$VLLM_PGID" 2>/dev/null || true
wait "$VLLM_PID" 2>/dev/null || true
fi
}
trap cleanup_vllm EXIT
# Step 3: wait for /health
echo "--- Waiting for /health ---"
for i in $(seq 1 600); do
if curl -s "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
echo " ready after ${i}s"
break
fi
if ! kill -0 "$VLLM_PID" 2>/dev/null; then
echo "ERROR: vLLM died during startup. Last 50 lines of vllm.log:" >&2
tail -50 "$VLLM_LOG" >&2
exit 1
fi
sleep 1
done
if ! curl -s "http://127.0.0.1:$PORT/health" >/dev/null 2>&1; then
echo "ERROR: vLLM did not become ready within 600s" >&2
tail -50 "$VLLM_LOG" >&2
exit 1
fi
# Step 4: verify /v1/models lists the INT8 served name. Fail closed: use
# `--fail-with-body` so curl exits non-zero on HTTP 4xx/5xx instead of
# silently writing the OpenAI error payload to disk and continuing to
# "smoke complete". (curl rejects `-f` and `--fail-with-body` together
# as mutually exclusive; we want the body-preserving form.) Then assert
# that the served name is actually present in the response — vLLM
# happily lists the path basename if the --served-model-name flag is
# misread.
echo ""
echo "--- /v1/models ---"
if ! curl -sS --fail-with-body "http://127.0.0.1:$PORT/v1/models" \
-o "$OUT_DIR/models.json"; then
echo "ERROR: /v1/models returned non-200" >&2
cat "$OUT_DIR/models.json" >&2 || true
exit 1
fi
cat "$OUT_DIR/models.json"
echo ""
python3 - "$OUT_DIR/models.json" <<'PY' || exit 1
import json, sys
data = json.loads(open(sys.argv[1]).read())
ids = [m.get("id") for m in data.get("data", [])]
expected = "Llama-3.1-8B-Instruct-int8"
if expected not in ids:
print(f"ERROR: served-model-name '{expected}' not in /v1/models list: {ids}", file=sys.stderr)
sys.exit(1)
print(f"OK: /v1/models contains '{expected}'")
PY
# Step 5: one-shot completion smoke. Same fail-closed treatment: curl errors
# out on HTTP 4xx, then we assert no top-level `error` and a non-empty
# `choices[0].text`. Without these checks an OpenAI-compatible error
# (model-name typo, context-window exceeded, etc.) would still end with
# `=== INT8 smoke complete ===` and a meta.json claiming success.
echo ""
echo "--- Test completion ---"
TEST_PROMPT="A power transformer with elevated H2 and C2H2 in DGA indicates"
if ! curl -sS --fail-with-body "http://127.0.0.1:$PORT/v1/completions" \
-H "Content-Type: application/json" \
-d "{\"model\":\"Llama-3.1-8B-Instruct-int8\",\"prompt\":\"$TEST_PROMPT\",\"max_tokens\":80,\"temperature\":0.1}" \
-o "$OUT_DIR/completion.json"; then
echo "ERROR: /v1/completions returned non-200" >&2
cat "$OUT_DIR/completion.json" >&2 || true
exit 1
fi
cat "$OUT_DIR/completion.json"
echo ""
python3 - "$OUT_DIR/completion.json" <<'PY' || exit 1
import json, sys
data = json.loads(open(sys.argv[1]).read())
if "error" in data:
print(f"ERROR: completion returned error payload: {data['error']}", file=sys.stderr)
sys.exit(1)
choices = data.get("choices", [])
if not choices:
print(f"ERROR: completion returned empty choices array", file=sys.stderr)
sys.exit(1)
text = choices[0].get("text", "")
if not text.strip():
print(f"ERROR: completion choices[0].text is empty / whitespace", file=sys.stderr)
sys.exit(1)
print(f"OK: completion returned {len(text)} chars of non-empty text")
PY
# Step 6: GPU memory snapshot (proves INT8 actually reduced memory vs FP16)
echo ""
echo "--- nvidia-smi snapshot ---"
# Filter to the GPU Slurm allocated. nvidia-smi does NOT honor CUDA_VISIBLE_DEVICES
# on its own, so on multi-GPU nodes a bare query returns all GPUs and the
# memory line for the wrong device.
nvidia-smi --id="${CUDA_VISIBLE_DEVICES:-0}" \
--query-gpu=name,memory.used,memory.free,memory.total --format=csv \
| tee "$OUT_DIR/nvidia_smi.csv"
# Step 7: write meta summary
echo ""
python3 -c "
import json, pathlib, datetime
meta = {
'slurm_job_id': '$JOB',
'int8_model_repo': '$INT8_MODEL_REPO',
'int8_model_revision': '$INT8_MODEL_REVISION',
'int8_local_dir': '$INT8_LOCAL_DIR',
'vllm_port': $PORT,
'started_at': datetime.datetime.utcnow().isoformat() + 'Z',
'served_model_name': 'Llama-3.1-8B-Instruct-int8',
'quantization_flag': 'compressed-tensors',
'vllm_log': '$VLLM_LOG',
'completion_path': '$OUT_DIR/completion.json',
'nvidia_smi_path': '$OUT_DIR/nvidia_smi.csv',
}
pathlib.Path('$OUT_DIR/meta.json').write_text(json.dumps(meta, indent=2) + '\n')
print('meta.json written to $OUT_DIR')
"
echo ""
echo "=== INT8 smoke complete ==="
echo "Outputs in $OUT_DIR/"
echo "Update docs/lane2_int8_kv_status.md with the result."
echo "Finished: $(date -u +%Y-%m-%dT%H:%M:%SZ)"
# vLLM auto-stops on exit via the trap
|