Suya OCR API — vLLM-backed, OpenAI-compatible OCR service

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>
2026-06-17 10:20:02 +04:00
commit 1a585693be
147 changed files with 13827 additions and 0 deletions
@@ -0,0 +1,60 @@
+#!/usr/bin/env bash
+# Serve and score compressor-built checkpoints (int8/gptq/awq) over a page subset.
+# Each method: hardened server teardown -> serve checkpoint -> health/fail watch ->
+# capture N_PAGES through the OCR API. Run INSIDE the bench container:
+#   bash scripts/quant/score_methods.sh int8 gptq awq
+# Requires api.py running on :5002 pointed at :8000.
+set -uo pipefail
+
+export LD_LIBRARY_PATH="/usr/local/cuda/compat:${LD_LIBRARY_PATH:-}"
+N_PAGES="${N_PAGES:-3}"
+OCR_URL="http://127.0.0.1:5002/v1/api/ai/suya_ocr_vllm/"
+LOG=/tmp/score_methods.log
+
+mapfile -t ALL < <(sed '/^#/d;/^$/d' eval_set/manifest.txt)
+IMAGES=("${ALL[@]:0:$N_PAGES}")
+
+stop_server() {
+  local pid
+  pid=$(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
+  # EngineCore outlives the API server and holds the GPU; kill compute procs and
+  # poll until memory actually frees (kill returns before GPU release).
+  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)
+    [ "${used:-99999}" -lt 2000 ] && break
+    sleep 2
+  done
+}
+
+for m in "$@"; do
+  echo "=== scoring $m ===" | tee -a "$LOG"
+  stop_server
+  # VLLM_EXTRA_ARGS lets a caller add e.g. --dtype float16 (Exllama W4A16 kernel
+  # on Ampere only supports float16 activations).
+  nohup python3 -m vllm.entrypoints.openai.api_server --host 127.0.0.1 --port 8000 \
+    --model "results/quant/models/$m" --served-model-name datalab-to/surya-ocr-2 \
+    --max-model-len 18000 --max-num-seqs 16 --gpu-memory-utilization 0.85 \
+    --enable-prefix-caching --mm-processor-kwargs '{"min_pixels":3136,"max_pixels":6291456}' \
+    ${VLLM_EXTRA_ARGS:-} \
+    > "/tmp/serve_${m}.log" 2>&1 &
+  ok=0
+  for _ in $(seq 1 84); do
+    if curl -fs http://127.0.0.1:8000/health >/dev/null 2>&1; then ok=1; break; fi
+    if grep -qiE "Engine core initialization failed" "/tmp/serve_${m}.log" 2>/dev/null; then break; fi
+    sleep 5
+  done
+  if [ "$ok" -ne 1 ]; then
+    echo "$m: FAILED to start" | tee -a "$LOG"
+    continue
+  fi
+  python3 -m scripts.quant.capture --url "$OCR_URL" --out-dir "results/quant/results/$m" "${IMAGES[@]}" \
+    >> "$LOG" 2>&1
+  echo "$m: captured $(ls "results/quant/results/$m"/*.json 2>/dev/null | wc -l) pages" | tee -a "$LOG"
+done
+stop_server
+echo "SCORE_DONE" | tee -a "$LOG"