From 035b65dcc84e12797377d1bd71db7b842f22c5f5 Mon Sep 17 00:00:00 2001
From: dekun <dekun@local>
Date: Sun, 28 Jun 2026 15:18:42 +0800
Subject: [PATCH] =?UTF-8?q?OCR=20=E4=BF=AE=E5=A4=8D=EF=BC=9A=E6=97=A0=20cu?=
 =?UTF-8?q?DNN=20=E7=94=A8=20CPU=E3=80=81=E7=A6=81=E7=94=A8=20ir=5Foptim?=
 =?UTF-8?q?=20=E9=81=BF=E5=85=8D=20SIGILL=E3=80=82?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 deploy/ocr-common.sh         | 34 ++++++++++++++++++++++++++++++----
 deploy/ocr-worker/app.py     | 18 ++++--------------
 deploy/ocr-worker/install.sh | 10 +++++-----
 3 files changed, 39 insertions(+), 23 deletions(-)

diff --git a/deploy/ocr-common.sh b/deploy/ocr-common.sh
index 0004333..45216cc 100644
--- a/deploy/ocr-common.sh
+++ b/deploy/ocr-common.sh
@@ -7,6 +7,28 @@ ocr_worker_dir() {
   echo "${INSTALL_DIR}/deploy/ocr-worker"
 }
 
+detect_ocr_use_gpu() {
+  if [[ "${OCR_USE_GPU:-auto}" == "false" ]]; then
+    echo "false"
+    return
+  fi
+  if [[ "${OCR_USE_GPU:-auto}" == "true" ]]; then
+    if ldconfig -p 2>/dev/null | grep -q libcudnn; then
+      echo "true"
+    else
+      log_warn "未检测到 cuDNN 库，GPU OCR 不可用，改用 CPU"
+      echo "false"
+    fi
+    return
+  fi
+  # auto
+  if command -v nvidia-smi >/dev/null && ldconfig -p 2>/dev/null | grep -q libcudnn; then
+    echo "true"
+  else
+    echo "false"
+  fi
+}
+
 install_ocr_worker() {
   local worker_dir
   worker_dir="$(ocr_worker_dir)"
@@ -14,9 +36,11 @@ install_ocr_worker() {
     log_error "未找到 ${worker_dir}"
     return 1
   fi
-  log_info "安装/更新 OCR Worker (PaddleOCR GPU)…"
+  log_info "安装/更新 OCR Worker (PaddleOCR)…"
   chmod +x "${worker_dir}"/*.sh 2>/dev/null || true
-  OCR_PORT="${OCR_PORT}" bash "${worker_dir}/install.sh"
+  local use_gpu
+  use_gpu="$(detect_ocr_use_gpu)"
+  OCR_PORT="${OCR_PORT}" OCR_USE_GPU="${use_gpu}" bash "${worker_dir}/install.sh"
 }
 
 ocr_screen_running() {
@@ -35,7 +59,9 @@ start_ocr_screen() {
     return 1
   fi
 
-  log_info "启动 OCR Worker → screen 会话「${OCR_SCREEN_NAME}」(GPU 常驻, 端口 ${OCR_PORT})"
+  local use_gpu log_file
+  use_gpu="$(detect_ocr_use_gpu)"
+  log_info "启动 OCR Worker → screen「${OCR_SCREEN_NAME}」(GPU=${use_gpu}, 端口 ${OCR_PORT})"
   if ocr_screen_running; then
     screen -S "${OCR_SCREEN_NAME}" -X quit 2>/dev/null || true
     sleep 1
@@ -46,7 +72,7 @@ start_ocr_screen() {
 
   screen -dmS "${OCR_SCREEN_NAME}" bash -c "
     cd '${worker_dir}' &&
-    export OCR_USE_GPU=true OCR_PORT='${OCR_PORT}' OCR_HOST=0.0.0.0 &&
+    export OCR_USE_GPU='${use_gpu}' OCR_PORT='${OCR_PORT}' OCR_HOST=0.0.0.0 &&
     exec bash run.sh >> '${log_file}' 2>&1
   "
   sleep 2
diff --git a/deploy/ocr-worker/app.py b/deploy/ocr-worker/app.py
index c3bdb40..0604659 100644
--- a/deploy/ocr-worker/app.py
+++ b/deploy/ocr-worker/app.py
@@ -10,6 +10,7 @@ from fastapi import FastAPI, File, Header, HTTPException, UploadFile
 from PIL import Image
 
 os.environ.setdefault("OPENCV_IO_ENABLE_OPENEXR", "0")
+os.environ.setdefault("FLAGS_use_mkldnn", "0")
 
 logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
 logger = logging.getLogger("ocr-worker")
@@ -36,7 +37,8 @@ def _create_engine(use_gpu: bool):
         lang="ch",
         show_log=False,
         use_gpu=use_gpu,
-        enable_mkldnn=not use_gpu,
+        enable_mkldnn=False,
+        ir_optim=False,
         det_limit_side_len=min(OCR_MAX_SIDE, 1280),
         rec_batch_num=8,
     )
@@ -150,19 +152,7 @@ def _run_ocr_impl(content: bytes) -> dict:
 
 
 def run_ocr_on_bytes(content: bytes) -> dict:
-    try:
-        return _run_ocr_impl(content)
-    except Exception as exc:
-        err = str(exc).lower()
-        gpu_fail = _engine_mode == "gpu" and any(
-            x in err for x in ("cuda", "cudnn", "gpu", "out of memory", "resource exhausted", "precondition")
-        )
-        if gpu_fail and OCR_USE_GPU:
-            logger.warning("GPU OCR runtime failed, retry CPU: %s", exc)
-            _reset_engine()
-            get_engine(force_cpu=True)
-            return _run_ocr_impl(content)
-        raise
+    return _run_ocr_impl(content)
 
 
 @app.get("/health")
diff --git a/deploy/ocr-worker/install.sh b/deploy/ocr-worker/install.sh
index cef346b..904641d 100644
--- a/deploy/ocr-worker/install.sh
+++ b/deploy/ocr-worker/install.sh
@@ -25,23 +25,23 @@ source "${VENV}/bin/activate"
 pip install -U pip wheel -i "${PIP_MIRROR}"
 
 install_paddle() {
-  if command -v nvidia-smi >/dev/null 2>&1; then
+  local use_gpu="${OCR_USE_GPU:-false}"
+  if [[ "${use_gpu}" == "true" ]] && command -v nvidia-smi >/dev/null 2>&1 && ldconfig -p 2>/dev/null | grep -q libcudnn; then
     local cuda_major
     cuda_major="$(nvidia-smi 2>/dev/null | sed -n 's/.*CUDA Version: \([0-9]*\)\.[0-9]*/\1/p' | head -1)"
     cuda_major="${cuda_major:-11}"
-    echo "==> 检测到 NVIDIA GPU，CUDA 主版本: ${cuda_major}"
+    echo "==> 安装 paddlepaddle-gpu (CUDA ${cuda_major}.x)…"
     if [[ "${cuda_major}" -ge 12 ]]; then
-      echo "==> 安装 paddlepaddle-gpu (CUDA 12.x)…"
       pip install paddlepaddle-gpu==2.6.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu123/ \
         || pip install paddlepaddle-gpu==2.6.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/
     else
-      echo "==> 安装 paddlepaddle-gpu (CUDA 11.x)…"
       pip install paddlepaddle-gpu==2.6.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/
     fi
   else
-    echo "==> 未检测到 GPU，安装 CPU 版 paddlepaddle…"
+    echo "==> 安装 paddlepaddle CPU（无 cuDNN 或 OCR_USE_GPU=false）…"
     pip install paddlepaddle==2.6.2 -i "${PIP_MIRROR}"
   fi
+  pip install 'protobuf>=3.20,<4' -q
 }
 
 install_paddle