Fu DaiandClaude Opus 4.8
1a585693be
FastAPI service wrapping the Surya-OCR-2 model (datalab-to) served through vLLM: legacy /v1/api/ai/* endpoints, an OpenAI-compatible /v1/chat/completions endpoint, a coalescing request batcher, a local OCR CLI, Docker packaging, multilingual example outputs, and quantization/concurrency benchmarks. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
78 lines
2.9 KiB
Bash
78 lines
2.9 KiB
Bash
#!/usr/bin/env bash
|
|
# vLLM serving-parameter tuning sweep for the OCR latency report.
|
|
# Serves the BF16 model under several vLLM flag combinations (CUDA graph vs eager,
|
|
# prefix caching on/off, chunked prefill on/off), captures a fixed page subset
|
|
# through the OCR API for each, and records per-config latency.
|
|
#
|
|
# Run INSIDE the bench container: bash scripts/quant/tune_vllm.sh
|
|
# Requires: api.py already running on :5002 pointed at :8000; compat libs on path.
|
|
set -uo pipefail
|
|
|
|
MODEL="${MODEL:-datalab-to/surya-ocr-2}"
|
|
OCR_URL="http://127.0.0.1:5002/v1/api/ai/suya_ocr_vllm/"
|
|
N_PAGES="${N_PAGES:-3}"
|
|
OUT_ROOT="results/quant/tuning"
|
|
LOG=/tmp/tune_vllm.log
|
|
export LD_LIBRARY_PATH="/usr/local/cuda/compat:${LD_LIBRARY_PATH:-}"
|
|
|
|
mapfile -t ALL < <(sed '/^#/d;/^$/d' eval_set/manifest.txt)
|
|
IMAGES=("${ALL[@]:0:$N_PAGES}")
|
|
|
|
# config_name | extra vLLM args
|
|
CONFIGS=(
|
|
"default|--enable-prefix-caching"
|
|
"eager|--enable-prefix-caching --enforce-eager"
|
|
"no_prefix_cache|--no-enable-prefix-caching"
|
|
"no_chunked_prefill|--enable-prefix-caching --no-enable-chunked-prefill"
|
|
)
|
|
|
|
base_args() {
|
|
echo "--host 127.0.0.1 --port 8000 --model $MODEL --served-model-name $MODEL \
|
|
--max-model-len 18000 --max-num-seqs 16 --gpu-memory-utilization 0.85 \
|
|
--mm-processor-kwargs {\"min_pixels\":3136,\"max_pixels\":6291456}"
|
|
}
|
|
|
|
stop_server() {
|
|
local pid
|
|
pid=$(ss -ltnp 2>/dev/null | grep ":8000 " | grep -oP 'pid=\K[0-9]+' | head -1)
|
|
[ -n "$pid" ] && kill -9 "$pid" 2>/dev/null
|
|
pkill -9 -f "vllm.entrypoints" 2>/dev/null
|
|
# The EngineCore worker holds the GPU and outlives the API server; kill every
|
|
# remaining CUDA compute process, then poll until the memory is actually freed
|
|
# (kill returns immediately, GPU release lags).
|
|
for p in $(nvidia-smi --query-compute-apps=pid --format=csv,noheader 2>/dev/null); do
|
|
kill -9 "$p" 2>/dev/null
|
|
done
|
|
for _ in $(seq 1 30); do
|
|
used=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -1)
|
|
[ "${used:-99999}" -lt 2000 ] && break
|
|
sleep 2
|
|
done
|
|
}
|
|
|
|
for entry in "${CONFIGS[@]}"; do
|
|
name="${entry%%|*}"
|
|
extra="${entry#*|}"
|
|
echo "=== tuning config: $name ($extra) ===" | tee -a "$LOG"
|
|
stop_server
|
|
# shellcheck disable=SC2046
|
|
nohup python3 -m vllm.entrypoints.openai.api_server $(base_args) $extra \
|
|
> "/tmp/tune_${name}.log" 2>&1 &
|
|
# wait for health or crash (max 7 min)
|
|
ok=0
|
|
for _ in $(seq 1 84); do
|
|
if curl -fs http://127.0.0.1:8000/health >/dev/null 2>&1; then ok=1; break; fi
|
|
if grep -qiE "Engine core initialization failed|RuntimeError" "/tmp/tune_${name}.log" 2>/dev/null; then break; fi
|
|
sleep 5
|
|
done
|
|
if [ "$ok" -ne 1 ]; then
|
|
echo "$name: FAILED to start" | tee -a "$LOG"
|
|
continue
|
|
fi
|
|
python3 -m scripts.quant.capture --url "$OCR_URL" --out-dir "$OUT_ROOT/$name" "${IMAGES[@]}" \
|
|
>> "$LOG" 2>&1
|
|
echo "$name: captured $(ls "$OUT_ROOT/$name"/*.json 2>/dev/null | wc -l) pages" | tee -a "$LOG"
|
|
done
|
|
stop_server
|
|
echo "TUNE_DONE" | tee -a "$LOG"
|