feat: add build turboocr

src/turboocr/README.md

@@ -52,7 +52,9 @@ Copy `.env.example` to `.env` and override only the variables you need to change
 docker compose up -d
 ```
 
-The first start builds TensorRT engines from ONNX. Build time depends on your GPU: roughly 5 minutes on high-end desktop GPUs and 20–30 minutes on laptop GPUs. The container may report `unhealthy` while compilation is in progress — this is normal. Once the build finishes the server starts and the container transitions to `healthy`. Subsequent restarts reuse the cached engines and start in seconds.
+The first start builds 4 TensorRT engines from ONNX. Measured build times on an RTX 3070 Laptop: det (~5 min) + rec (~30 min) + cls (~4 min) + layout (~28 min) = **~67–90 minutes total**. High-end desktop GPUs finish in ~15 minutes. The container reports `unhealthy` while compilation is in progress — this is expected. Once all engines are built the server starts and the container transitions to `healthy`. Subsequent restarts reuse the cached engines and start in seconds.
+
+> **Tip — faster first boot:** Set `TURBOOCR_DISABLE_LAYOUT=1` to skip the layout detection engine (~28 min on laptop GPUs). Only do this if you don't need the `?layout=1` PDF endpoint.
 
 ### Endpoints
 

src/turboocr/README.zh.md

@@ -52,7 +52,9 @@
 docker compose up -d
 ```
 
-首次启动需要从 ONNX 构建 TensorRT 引擎，耗时因 GPU 而异：高端桌面 GPU 约 5 分钟，笔记本 GPU 约 20–30 分钟。编译期间容器可能显示 `unhealthy`，这属于正常现象——构建完成后服务会自动启动并切换为 `healthy`。后续重启会复用缓存的引擎，几乎瞬间完成。
+首次启动需要编译 4 个 TensorRT 引擎。在 RTX 3070 Laptop 上的实测耗时：det（约 5 分钟）+ rec（约 30 分钟）+ cls（约 4 分钟）+ layout（约 28 分钟）= **总计约 67–90 分钟**。高端桌面 GPU 约 15 分钟完成。编译期间容器显示 `unhealthy` 属于正常现象——所有引擎构建完成后服务会自动启动并切换为 `healthy`。后续重启会复用缓存的引擎，几乎瞬间完成。
+
+> **提示——加快首次启动**：设置 `TURBOOCR_DISABLE_LAYOUT=1` 可跳过版面检测引擎的编译（笔记本 GPU 约节省 28 分钟）。仅在不需要 `?layout=1` PDF 端点时使用此选项。
 
 ### 访问端点
 

src/turboocr/docker-compose.yaml

@@ -14,7 +14,7 @@ services:
       - '${TURBOOCR_HTTP_PORT_OVERRIDE:-8000}:8000'
       - '${TURBOOCR_GRPC_PORT_OVERRIDE:-50051}:50051'
     volumes:
-      # Named volume caches TensorRT engines built from ONNX on first start (~90s).
+      # Named volume caches TensorRT engines built from ONNX on first start.
       # Must be a named volume - bind-mounting an empty host dir would shadow the
       # baked-in language bundles and prevent the server from loading models.
       - turboocr_trt_cache:/home/ocr/.cache/turbo-ocr
@@ -48,11 +48,13 @@ services:
       interval: 30s
       timeout: 10s
       retries: 5
-      # First start builds TensorRT engines from ONNX. Build time varies by GPU:
-      # ~5 min on high-end desktop GPUs, 20-30 min on laptop GPUs. The container
-      # may show "unhealthy" during compilation but will become healthy once done.
-      # Subsequent restarts reuse the cached engines and start in seconds.
-      start_period: 30m
+      # First start builds 4 TensorRT engines from ONNX. Measured build times:
+      #   det (~5 min) + rec (~30 min) + cls (~4 min) + layout (~28 min) ≈ 67-90 min
+      # on an RTX 3070 Laptop. High-end desktop GPUs finish in ~15 min.
+      # Set TURBOOCR_DISABLE_LAYOUT=1 to skip the layout engine and cut ~28 min.
+      # The container shows "unhealthy" while building but recovers once done.
+      # Subsequent restarts reuse cached engines and start in seconds.
+      start_period: 120m
     deploy:
       resources:
         limits: