&2\n+ echo \"Pass --image <path> to select one explicitly.\" >&2\n+ return 1\n+ fi\n+ IMAGE_NAME=\"${found[0]}\"\n+ IMAGE_SRC=\"${FIXTURES_DIR}/${IMAGE_NAME}\"\n+ IMAGE_SOURCE=\"fixture\"\n+ echo \"Using fixture: ${IMAGE_NAME}\"\n+}\n+run_step resolve-image resolve_image_impl || { mark_remaining_skipped \"resolve-image failed: ${HALT_REASON}\"; exit 1; }\n+\n+# ---------- Step: resolve-destination ----------\n+resolve_destination_impl() {\n+ if [[ -n \"${UDID}\" ]]; then\n+ DEST=\"platform=iOS,id=${UDID}\"\n+ IS_SIMULATOR=0\n+ elif [[ -n \"${SIMULATOR}\" ]]; then\n+ DEST=\"platform=iOS Simulator,name=${SIMULATOR}\"\n+ IS_SIMULATOR=1\n+ else\n+ DEST=\"platform=iOS Simulator,name=${DEFAULT_SIMULATOR}\"\n+ IS_SIMULATOR=1\n+ fi\n+ echo \"Destination: ${DEST}\"\n+}\n+run_step resolve-destination resolve_destination_impl || { mark_remaining_skipped \"resolve-destination failed: ${HALT_REASON}\"; exit 1; }\n+\n+# ---------- Step: accuracy-precheck (skipped unless --accuracy-check is set) ----------\n+accuracy_precheck_impl() {\n+ ACCURACY_STATUS=\"ERROR\"\n+ ACCURACY_REASON=\"\"\n+ ACCURACY_SUMMARY_USED=\"${ACCURACY_SUMMARY_PATH}\"\n+ rm -rf \"${ACCURACY_RESULT_PATH}\"\n+ rm -f \"${ACCURACY_IOS_EXPORT_PATH}\" \"${ACCURACY_SUMMARY_PATH}\"\n+\n+ if [[ -n \"${ACCURACY_REFERENCE_JSON}\" ]]; then\n+ ACCURACY_REFERENCE_USED=\"${ACCURACY_REFERENCE_JSON}\"\n+ else\n+ ACCURACY_REFERENCE_USED=\"${ACCURACY_REFERENCE_OUT}\"\n+ if ! python3 \"${SCRIPT_DIR}/ocr_reference_run.py\" \\\n+ --image \"${IMAGE_SRC}\" \\\n+ --output \"${ACCURACY_REFERENCE_OUT}\" \\\n+ --device cpu \\\n+ --align-ios-defaults; then\n+ ACCURACY_STATUS=\"ERROR\"\n+ ACCURACY_REASON=\"reference generation failed\"\n+ echo \"Accuracy precheck: ${ACCURACY_STATUS} (${ACCURACY_REASON})\" >&2\n+ return 1\n+ fi\n+ fi\n+\n+ local runenv=(\n+ env \"TEST_RUNNER_PADDLEOCR_BENCHMARK_IMAGE_NAME=${IMAGE_NAME}\"\n+ \"TEST_RUNNER_PADDLEOCR_BENCHMARK_BUILD_CONFIGURATION=${BUILD_CONFIGURATION}\"\n+ )\n+ if [[ -n \"${ORT_EP_CANON}\" ]]; then\n+ runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_ORT_EXECUTION_PROVIDER=${ORT_EP_CANON}\")\n+ fi\n+\n+ if ! run_xcodebuild \"${runenv[@]}\" xcodebuild test \\\n+ -workspace \"${IOS_DEMO_ROOT}/PaddleOCRDemo.xcworkspace\" \\\n+ -scheme PaddleOCRDemo \\\n+ -configuration \"${BUILD_CONFIGURATION}\" \\\n+ -destination \"${DEST}\" \\\n+ -resultBundlePath \"${ACCURACY_RESULT_PATH}\" \\\n+ \"${XCODE_TEST_SINGLE_RUN_FLAGS[@]}\" \\\n+ -only-testing:PaddleOCRDemoTests/OCRBenchmarkTests/testOCRExportJSONSchema; then\n+ ACCURACY_STATUS=\"ERROR\"\n+ ACCURACY_REASON=\"accuracy XCTest export failed\"\n+ echo \"Accuracy precheck: ${ACCURACY_STATUS} (${ACCURACY_REASON})\" >&2\n+ return 1\n+ fi\n+\n+ if ! python3 \"${SCRIPT_DIR}/extract_xcresult_attachments.py\" \\\n+ --result \"${ACCURACY_RESULT_PATH}\" \\\n+ --output-dir \"${OUT_DIR}\" \\\n+ --name ios-ocr-export.json; then\n+ ACCURACY_STATUS=\"ERROR\"\n+ ACCURACY_REASON=\"accuracy attachment extraction failed\"\n+ echo \"Accuracy precheck: ${ACCURACY_STATUS} (${ACCURACY_REASON})\" >&2\n+ return 1\n+ fi\n+\n+ set +e\n+ python3 \"${SCRIPT_DIR}/compare_ocr_json.py\" \\\n+ \"${ACCURACY_REFERENCE_USED}\" \\\n+ \"${ACCURACY_IOS_EXPORT_PATH}\" \\\n+ --json-summary-out \"${ACCURACY_SUMMARY_PATH}\"\n+ local compare_rc=$?\n+ set -e\n+ if [[ \"${compare_rc}\" -eq 0 ]]; then\n+ ACCURACY_STATUS=\"PASS\"\n+ elif [[ -f \"${ACCURACY_SUMMARY_PATH}\" ]]; then\n+ ACCURACY_STATUS=\"FAIL\"\n+ ACCURACY_REASON=\"accuracy thresholds were not met\"\n+ else\n+ ACCURACY_STATUS=\"ERROR\"\n+ ACCURACY_REASON=\"accuracy comparison failed\"\n+ fi\n+ echo \"Accuracy precheck: ${ACCURACY_STATUS}${ACCURACY_REASON:+ (${ACCURACY_REASON})}\"\n+ [[ \"${ACCURACY_STATUS}\" != \"ERROR\" ]]\n+}\n+\n+if [[ \"${ACCURACY_CHECK}\" -eq 1 ]]; then\n+ if ! run_step accuracy-precheck accuracy_precheck_impl; then\n+ HALT_REASON=\"accuracy precheck error: ${ACCURACY_REASON:-${HALT_REASON}}\"\n+ step_reason_set accuracy-precheck \"${HALT_REASON}\"\n+ mark_remaining_skipped \"${HALT_REASON}\"\n+ write_status_json \"ERROR\" 1\n+ generate_report_impl || true\n+ exit 1\n+ fi\n+ if [[ \"${ACCURACY_STATUS}\" == \"FAIL\" ]]; then\n+ printf \"Accuracy precheck: %s (%s)\\n\" \"${ACCURACY_STATUS}\" \"${ACCURACY_REASON:-accuracy thresholds were not met}\"\n+ fi\n+ if [[ \"${ACCURACY_STOP_ON_FAILURE}\" -eq 1 && \"${ACCURACY_STATUS}\" == \"FAIL\" ]]; then\n+ HALTED=1\n+ HALT_REASON=\"accuracy precheck did not pass: ${ACCURACY_STATUS}\"\n+ step_status_set accuracy-precheck fail\n+ step_exit_set accuracy-precheck 1\n+ step_reason_set accuracy-precheck \"${HALT_REASON}\"\n+ mark_remaining_skipped \"${HALT_REASON}\"\n+ write_status_json \"ERROR\" 1\n+ generate_report_impl || true\n+ exit 1\n+ fi\n+else\n+ step_status_set accuracy-precheck skipped\n+ step_reason_set accuracy-precheck \"accuracy precheck not requested\"\n+ ACCURACY_STATUS=\"SKIPPED\"\n+fi\n+\n+# ---------- Step: xcodebuild-test ----------\n+xcodebuild_test_impl() {\n+ rm -rf \"${LATENCY_RESULT_PATH}\"\n+ rm -rf \"${MEMORY_RESULT_PATH}\"\n+ rm -f \"${XCTEST_METRICS_PATH}\"\n+ echo \"xcodebuild: configuration=${BUILD_CONFIGURATION} benchmark image name=${IMAGE_NAME} warmup=${WARMUP_MERGED:-} measured=${MEASURED_MERGED:-} ort_ep=${ORT_EP_CANON:-default} ort_profiling=${ORT_PROFILING_MERGED:+on} scope=${TESTING_SCOPE}\"\n+ local runenv=(\n+ env \"TEST_RUNNER_PADDLEOCR_BENCHMARK_IMAGE_NAME=${IMAGE_NAME}\"\n+ \"TEST_RUNNER_PADDLEOCR_BENCHMARK_BUILD_CONFIGURATION=${BUILD_CONFIGURATION}\"\n+ )\n+ [[ -n \"${WARMUP_MERGED}\" ]] && runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_WARMUP_ITERATIONS=${WARMUP_MERGED}\")\n+ [[ -n \"${MEASURED_MERGED}\" ]] && runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_MEASURED_ITERATIONS=${MEASURED_MERGED}\")\n+ if [[ -n \"${ORT_EP_CANON}\" ]]; then\n+ runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_ORT_EXECUTION_PROVIDER=${ORT_EP_CANON}\")\n+ fi\n+ [[ -n \"${ORT_PROFILING_MERGED:-}\" ]] && runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_ORT_PROFILING=${ORT_PROFILING_MERGED}\")\n+ [[ -n \"${THREADS_MERGED}\" ]] && runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_INTRA_OP_THREADS=${THREADS_MERGED}\")\n+ [[ -n \"${XNNPACK_THREADS_MERGED}\" ]] && runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_XNNPACK_THREADS=${XNNPACK_THREADS_MERGED}\")\n+ [[ -n \"${REC_BATCH_SIZE_MERGED}\" ]] && runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_REC_BATCH_SIZE=${REC_BATCH_SIZE_MERGED}\")\n+ [[ -n \"${COREML_OPTIONS_MERGED}\" ]] && runenv+=(\"TEST_RUNNER_PADDLEOCR_BENCHMARK_COREML_OPTIONS=${COREML_OPTIONS_MERGED}\")\n+ local scopes=() scope result_path\n+ IFS=',' read -r -a scopes <<< \"${TESTING_SCOPE}\"\n+ for scope in \"${scopes[@]}\"; do\n+ scope=\"$(printf '%s' \"$scope\" | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')\"\n+ [[ -n \"$scope\" ]] || continue\n+ result_path=\"${LATENCY_RESULT_PATH}\"\n+ if [[ \"${scope}\" == *\"/testOnDeviceMemoryBenchmark\" ]]; then\n+ result_path=\"${MEMORY_RESULT_PATH}\"\n+ fi\n+ rm -rf \"${result_path}\"\n+ echo \"xcodebuild: running scope=${scope} result=${result_path}\"\n+ if ! run_xcodebuild \"${runenv[@]}\" xcodebuild test \\\n+ -workspace \"${IOS_DEMO_ROOT}/PaddleOCRDemo.xcworkspace\" \\\n+ -scheme PaddleOCRDemo \\\n+ -configuration \"${BUILD_CONFIGURATION}\" \\\n+ -destination \"${DEST}\" \\\n+ -resultBundlePath \"${result_path}\" \\\n+ \"${XCODE_TEST_SINGLE_RUN_FLAGS[@]}\" \\\n+ -only-testing:\"${scope}\"; then\n+ return 1\n+ fi\n+ done\n+ collect_app_binary_size\n+}\n+run_step xcodebuild-test xcodebuild_test_impl || { mark_remaining_skipped \"xcodebuild-test failed: ${HALT_REASON}\"; exit 1; }\n+\n+# ---------- Step: extract-artifacts ----------\n+extract_artifacts_impl() {\n+ local metrics_result=\"${MEMORY_RESULT_PATH}\"\n+ [[ -d \"${metrics_result}\" ]] || metrics_result=\"${LATENCY_RESULT_PATH}\"\n+ python3 \"${SCRIPT_DIR}/extract_xcresult_attachments.py\" \\\n+ --result \"${LATENCY_RESULT_PATH}\" \\\n+ --output-dir \"${OUT_DIR}\" \\\n+ --name on-device-performance.json \\\n+ --optional-name ios-ocr-export.json \\\n+ --optional-name ort_profile_detection \\\n+ --optional-name ort_profile_recognition\n+ python3 \"${SCRIPT_DIR}/extract_xctest_metrics.py\" \\\n+ --result \"${metrics_result}\" \\\n+ --output \"${XCTEST_METRICS_PATH}\"\n+}\n+run_step extract-artifacts extract_artifacts_impl || { mark_remaining_skipped \"extract-artifacts failed: ${HALT_REASON}\"; exit 1; }\n+\n+# ---------- Step: report ----------\n+# First pass captures report step timing; write_status_json persists final step state;\n+# second pass refreshes the Markdown so **Overall** and the step table match this run.\n+report_impl() {\n+ generate_report_impl\n+}\n+report_refresh() {\n+ generate_report_impl\n+}\n+run_step report report_impl || { write_status_json \"ERROR\" 1; exit 1; }\n+write_status_json \"COMPLETED\" 0\n+report_refresh || true\n+\n+exit 0\ndiff --git a/deploy/ios_demo/scripts/utils.py b/deploy/ios_demo/scripts/utils.py\nnew file mode 100644\nindex 00000000000..41a29560da6\n--- /dev/null\n+++ b/deploy/ios_demo/scripts/utils.py\n@@ -0,0 +1,89 @@\n+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+#\n+# Licensed under the Apache License, Version 2.0 (the \"License\");\n+# you may not use this file except in compliance with the License.\n+# You may obtain a copy of the License at\n+#\n+# http://www.apache.org/licenses/LICENSE-2.0\n+#\n+# Unless required by applicable law or agreed to in writing, software\n+# distributed under the License is distributed on an \"AS IS\" BASIS,\n+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+# See the License for the specific language governing permissions and\n+# limitations under the License.\n+\n+\"\"\"Shared helpers (calibration I/O) for ``quantize_onnx_model`` / ``debug_onnx_qdq`` / ``build_onnx_calib_npy``.\n+\n+Entry scripts prepend this file's directory to ``sys.path`` before ``from utils import ...`` so\n+``import utils`` does not depend on the process working directory. The top-level name ``utils`` is\n+generic; the scripts resolve ``utils.py`` in this same ``scripts/`` directory first.\n+\"\"\"\n+\n+from __future__ import annotations\n+\n+import sys\n+from pathlib import Path\n+from typing import Any\n+\n+__all__ = [\n+ \"die\",\n+ \"user_input_names\",\n+ \"build_npy_dir_reader\",\n+]\n+\n+\n+def die(msg: str) -> None:\n+ print(f\"error: {msg}\", file=sys.stderr)\n+ raise SystemExit(1)\n+\n+\n+def user_input_names(model_path: Path) -> list[str]:\n+ \"\"\"Return non-initializer graph input names (user-visible inputs).\"\"\"\n+ import onnx\n+\n+ m = onnx.load(str(model_path), load_external_data=True)\n+ init = {t.name for t in m.graph.initializer}\n+ return [i.name for i in m.graph.input if i.name not in init]\n+\n+\n+def build_npy_dir_reader(\n+ model_path: Path, data_dir: Path, *, max_samples: int | None = None\n+) -> Any:\n+ \"\"\"ONNX Runtime ``CalibrationDataReader``: one float32 ``(NCHW)`` ``.npy`` per sample; sorted by filename.\n+\n+ *model_path* is used to discover the single graph input name (``user_input_names``); it may be\n+ the original or an augmented model as long as the first user input is unchanged.\n+ \"\"\"\n+ from onnxruntime.quantization import CalibrationDataReader\n+\n+ class NpyDirectoryDataReader(CalibrationDataReader):\n+ def __init__(self) -> None:\n+ names = user_input_names(model_path)\n+ if len(names) != 1:\n+ die(\n+ \"expected exactly one non-initializer graph input; \"\n+ f\"found {len(names)}: {names}\"\n+ )\n+ self._input_name = names[0]\n+ files = sorted(data_dir.glob(\"*.npy\"))\n+ if not files:\n+ die(f\"no .npy files under {data_dir}\")\n+ if max_samples is not None:\n+ files = files[: max(0, max_samples)]\n+ self._it = iter(files)\n+\n+ def get_next(self) -> Any:\n+ import numpy as np\n+\n+ try:\n+ path = next(self._it)\n+ except StopIteration:\n+ return None\n+ arr = np.load(str(path))\n+ if not isinstance(arr, np.ndarray):\n+ die(f\"expected ndarray in {path}, got {type(arr)}\")\n+ if arr.dtype != np.float32:\n+ arr = arr.astype(np.float32, copy=False)\n+ return {self._input_name: arr}\n+\n+ return NpyDirectoryDataReader()\ndiff --git a/docs/version3.x/inference_deployment/cross_platform/ios_deployment.en.md b/docs/version3.x/inference_deployment/cross_platform/ios_deployment.en.md\nnew file mode 100644\nindex 00000000000..3c371be698a\n--- /dev/null\n+++ b/docs/version3.x/inference_deployment/cross_platform/ios_deployment.en.md\n@@ -0,0 +1,171 @@\n+---\n+comments: true\n+---\n+\n+# iOS Deployment\n+\n+PaddleOCR ships an iOS OCR sample app under [`deploy/ios_demo`](https://github.com/PaddlePaddle/PaddleOCR/tree/{{PADDLEOCR_GITHUB_REF}}/deploy/ios_demo). The **SwiftUI** demo runs exported **ONNX** models on device using the [ONNX Runtime Objective-C API](https://onnxruntime.ai/docs/tutorials/mobile/) for detection and recognition.\n+\n+This guide covers environment requirements, model asset setup, building and running in Xcode, and optional ORT format conversion and benchmarking.\n+\n+## Project layout\n+\n+App sources, bundled resources, and third-party **source** vendored for this demo live under **`PaddleOCRDemo/`**. Unit tests are in **`PaddleOCRDemoTests/`** next to the Xcode project. The project root also contains `Podfile`, **`scripts/`**, `README.md`, and `NOTICE`.\n+\n+## Prerequisites\n+\n+- **macOS** with **Xcode 16.0+** (deployment target **iOS 16+**)\n+- **CocoaPods** (`gem install cocoapods` or Homebrew)\n+- `curl`, `tar`\n+\n+## Supported model presets\n+\n+`scripts/fetch_ios_demo_models.sh` downloads ONNX bundles into **`PaddleOCRDemo/Models/`**. Supported presets:\n+\n+- **`PP-OCRv6_small`** (default)\n+- **`PP-OCRv6_tiny`**\n+- **`PP-OCRv5_mobile`**\n+\n+## Quick start\n+\n+From the **project root** (`deploy/ios_demo`):\n+\n+```bash\n+pod install\n+./scripts/fetch_ios_demo_models.sh\n+```\n+\n+Optionally pass the **model preset** (bundle name such as `PP-OCRv6_small`) as a positional argument:\n+\n+```bash\n+./scripts/fetch_ios_demo_models.sh PP-OCRv6_small\n+```\n+\n+Intermediate `.tar` caches are stored under **`.fetch_ios_demo_models_work/`** at the project root.\n+\n+After `pod install`, open the workspace:\n+\n+```bash\n+open PaddleOCRDemo.xcworkspace\n+```\n+\n+Build the **PaddleOCRDemo** scheme. Ensure **`PaddleOCRDemo/Models/`** and **`PaddleOCRDemo/Resources/SampleImages/`** are included in the app target via folder references / **Copy Bundle Resources**, and **`PaddleOCRDemoTests/Fixtures/`** in the test target (as in the checked-in project). The built-in picker sample is **`general_ocr_002.jpg`**.\n+\n+### ONNX Runtime execution providers (EPs)\n+\n+The **CPU** / **XNNPACK** / **Core ML** segmented control chooses which ONNX Runtime EP is preferred first when creating the session. ONNX Runtime may still execute some operators on the default **CPU** EP when another EP lacks an implementation. Selecting **CPU** does not register the XNNPACK or Core ML EPs.\n+\n+## Convert to ORT model format (optional)\n+\n+To produce [ORT format](https://onnxruntime.ai/docs/performance/model-optimizations/ort-format-models.html) weights:\n+\n+```bash\n+python3 -m pip install -r requirements-onnx-convert-ort.txt\n+./scripts/convert_onnx_to_ort.sh\n+```\n+\n+By default, conversion writes `inference*.ort` next to each `inference.onnx` under `PaddleOCRDemo/Models/`. The demo loads `inference*.ort` when present, so you do not need to ship ONNX in the bundle. To keep only one weight file in the app, delete the `inference.onnx` files manually, or use **`--output-dir`** and replace files under `PaddleOCRDemo/Models/`. Use **`--input-dir`** to point at any ONNX tree on the host; it defaults to `PaddleOCRDemo/Models` when omitted.\n+\n+By default the script uses **`--optimization-style Fixed`** (one `.ort` per ONNX file). The raw Python tool defaults to both Fixed and Runtime (`*.ort` and `*.with_runtime_opt.ort`). For Runtime-style ORT files (e.g. Core ML–oriented exports), pass **`--optimization-style Runtime`** (alias: `--optimization_style`). You can pass both: **`--optimization-style Fixed Runtime`**.\n+\n+The converter emits **`required_operators*.config`** for minimal ORT builds; **by default the shell script deletes them** after a successful run. Pass **`--keep-operator-config`** to retain them. They are **not** read at inference time. The iOS demo loads only **`.ort`** (or `.onnx`) weights.\n+\n+Example (Runtime conversion to a separate folder):\n+\n+```bash\n+./scripts/convert_onnx_to_ort.sh --output-dir ./out/ort_bundles --optimization-style Runtime\n+```\n+\n+## Benchmark\n+\n+This demo provides a benchmark pipeline for measuring on-device OCR latency and memory.\n+\n+### Prerequisites\n+\n+Complete [Quick start](#quick-start) first. For benchmark runs on a **physical device** you also need:\n+\n+- **Code signing (real device only):** `xcodebuild` requires a valid Apple Development certificate and provisioning profile. Open `PaddleOCRDemo.xcworkspace` in Xcode once, select your Development Team under *Signing & Capabilities*, and let Xcode manage provisioning; later `run_benchmark.sh` invocations reuse the cached profile. On first deploy, trust the developer certificate on the device under **Settings → General → VPN & Device Management** (one-time per device). Simulator runs do not require signing.\n+- Optional accuracy precheck: PaddleOCR (ONNX Runtime engine) on the host for reference generation, and `python3 -m pip install -r requirements-accuracy.txt`.\n+\n+### Full pipeline\n+\n+From the project root, use **`./scripts/run_benchmark.sh`**. The script resolves the input image, optionally runs an accuracy precheck, runs the XCTest benchmark on a simulator or device, extracts artifacts, and writes the Markdown report. It always invokes `xcodebuild test` with `-configuration Release`.\n+\n+Prefer script flags. `--image` copies an arbitrary image into test fixtures for this run. `--fixture` selects an existing file under `PaddleOCRDemoTests/Fixtures/`. If neither is set, the script requires exactly one non-`local-*` image under `Fixtures/` (e.g. `ios_ocr_benchmark_reference.jpg`).\n+\n+| Intent | Flags |\n+| --- | --- |\n+| Destination | `--udid <id>` for a real device, or `--simulator <name>` |\n+| Image from an arbitrary path | `--image <path>` |\n+| Image already under `Fixtures/` | `--fixture <name>` |\n+| Benchmark intensity | `--warmup <n>`, `--measured-iterations <n>` |\n+| Preferred ONNX Runtime EP | `--ort-execution-provider CPU`, `XNNPACK`, or `CORE_ML` |\n+| ONNX Runtime profiling JSON | `--ort-profiling` |\n+| Optional accuracy precheck | `--accuracy-check`, optionally `--accuracy-reference-json <path>` |\n+| Gate benchmark on accuracy `FAIL` | `--accuracy-check --stop-on-accuracy-failure` |\n+| Output directory | `--output-dir <dir>` (default: `out/`) |\n+| Clean previous artifacts | `--clean` |\n+\n+```bash\n+./scripts/run_benchmark.sh --udid <device-udid> --warmup 5 --measured-iterations 30\n+./scripts/run_benchmark.sh --udid <udid> --image /path/to/photo.png --ort-execution-provider CPU\n+./scripts/run_benchmark.sh --fixture ios_ocr_benchmark_reference --warmup 2 --measured-iterations 20\n+./scripts/run_benchmark.sh\n+PADDLEOCR_BENCHMARK_MEASURED_ITERATIONS=30 ./scripts/run_benchmark.sh --warmup 0\n+PADDLEOCR_BENCHMARK_ORT_EXECUTION_PROVIDER=XNNPACK ./scripts/run_benchmark.sh --fixture ios_ocr_benchmark_reference --warmup 1\n+./scripts/run_benchmark.sh --accuracy-check --udid <device-udid> --measured-iterations 30\n+./scripts/run_benchmark.sh --accuracy-check --stop-on-accuracy-failure --udid <device-udid>\n+./scripts/run_benchmark.sh --output-dir ./benchmark-out --measured-iterations 30\n+./scripts/run_benchmark.sh --udid <device-udid> --ort-profiling\n+```\n+\n+`--accuracy-check` runs before the benchmark in a separate XCTest invocation. Results are `PASS`, `FAIL`, or `ERROR`. By default, `FAIL` records a mismatch but continues; `ERROR` stops the pipeline. `--stop-on-accuracy-failure` also skips the benchmark on `FAIL`.\n+\n+**ORT session profiling** (attachments `ort_profile_detection`, `ort_profile_recognition`): use `--ort-profiling`. Run profiling in a **separate** run from clean latency measurements.\n+\n+After the benchmark **completes**, the script overwrites `run-status.json`, `on-device-performance.json`, `xctest-memory-metrics.json`, and `benchmark-report.md` under the output directory (`out/` by default). On **ERROR**, those files may be missing or partial.\n+\n+The report includes model input tensor shape distributions, first measured run line count, inferred model preset, actual model format (`onnx` or `ort`), det/rec/total weight sizes, app executable size when resolvable, cold model load time, measured latency, memory, etc.\n+\n+| Artifact | Producer | Purpose |\n+|---|---|---|\n+| `accuracy-reference.json` | `ocr_reference_run.py` | Optional reference JSON |\n+| `accuracy-result.xcresult` | `xcodebuild test` | Optional accuracy precheck |\n+| `ios-ocr-export.json` | `extract_xcresult_attachments.py` | Optional iOS accuracy payload |\n+| `accuracy-summary.json` | `compare_ocr_json.py` | Optional accuracy summary |\n+| `latency-result.xcresult` | `xcodebuild test` | Latency benchmark |\n+| `memory-result.xcresult` | `xcodebuild test` | Memory benchmark |\n+| `on-device-performance.json` | `extract_xcresult_attachments.py` | Latency stats from `latency-result.xcresult` |\n+| `ort_profile_detection`, `ort_profile_recognition` | `extract_xcresult_attachments.py` | Optional ORT profiling JSON |\n+| `xctest-memory-metrics.json` | `extract_xctest_metrics.py` | XCTest memory metrics |\n+| `logs/*.log` | `run_benchmark.sh` | Per-step logs |\n+| `run-status.json` | `run_benchmark.sh` | Step outcomes and metadata |\n+| `benchmark-report.md` | `generate_benchmark_report.py` | Human-readable report |\n+\n+Exit **`0`** when the pipeline reaches **`COMPLETED`**; non-zero on **`ERROR`**. Accuracy `FAIL` changes the exit code only with `--stop-on-accuracy-failure`; `ERROR` always stops the pipeline.\n+\n+### Advanced: XCTest environment\n+\n+Most users should use `run_benchmark.sh` flags instead of setting XCTest environment variables directly. This section is for debugging the `PaddleOCRDemoTests` target from Xcode or a custom `xcodebuild test` command.\n+\n+Benchmark tests read **`PADDLEOCR_BENCHMARK_*`** variables on the XCTest runner process.\n+\n+| How you launch tests | What to configure |\n+| --- | --- |\n+| **Xcode** | Scheme **PaddleOCRDemo** → **Test** → **Arguments** → **Environment Variables** (`PADDLEOCR_BENCHMARK_…` as-is). |\n+| **Custom `xcodebuild test`** | **`TEST_RUNNER_` + variable name** (e.g. `TEST_RUNNER_PADDLEOCR_BENCHMARK_IMAGE_NAME=…`). |\n+\n+| Variable | If unset | Role |\n+| --- | --- | --- |\n+| `PADDLEOCR_BENCHMARK_IMAGE_NAME` | **`ios_ocr_benchmark_reference`** | Bundled test image under **`Fixtures/`**. |\n+| `PADDLEOCR_BENCHMARK_WARMUP_ITERATIONS` | **3** | Untimed full OCR runs before timing. |\n+| `PADDLEOCR_BENCHMARK_MEASURED_ITERATIONS` | **10** | Timed runs. |\n+| `PADDLEOCR_BENCHMARK_ORT_EXECUTION_PROVIDER` | **CPU** | **`CPU`**, **`XNNPACK`**, or **`CORE_ML`**. |\n+| `PADDLEOCR_BENCHMARK_ORT_PROFILING` | (unset) | Enable ORT session profiling (`1` / `true` / `yes` / `on`). |\n+| `PADDLEOCR_BENCHMARK_ONLY_TESTING_SCOPE` | latency + memory benchmark tests | `run_benchmark.sh` only: comma-separated `-only-testing` scope. |\n+\n+Non-negative integers for the two iteration variables.\n+\n+## Third-party licenses\n+\n+Bundled **Clipper** (polyclipping 6.4.2) is under the [Boost Software License 1.0](https://www.boost.org/LICENSE_1_0.txt); see `NOTICE` and `PaddleOCRDemo/ThirdParty/Clipper1/LICENSE`. CocoaPods pods are governed by their respective licenses (see `Podfile.lock` after `pod install`).\ndiff --git a/docs/version3.x/inference_deployment/cross_platform/ios_deployment.md b/docs/version3.x/inference_deployment/cross_platform/ios_deployment.md\nnew file mode 100644\nindex 00000000000..5c86ccffe08\n--- /dev/null\n+++ b/docs/version3.x/inference_deployment/cross_platform/ios_deployment.md\n@@ -0,0 +1,171 @@\n+---\n+comments: true\n+---\n+\n+# iOS 部署\n+\n+PaddleOCR 在 [`deploy/ios_demo`](https://github.com/PaddlePaddle/PaddleOCR/tree/{{PADDLEOCR_GITHUB_REF}}/deploy/ios_demo) 提供基于 **SwiftUI** 的 iOS OCR 示例应用。应用在设备上使用导出的 **ONNX** 模型，通过 [ONNX Runtime Objective-C API](https://onnxruntime.ai/docs/tutorials/mobile/) 完成检测与识别推理。\n+\n+本指南介绍该 Demo 的环境要求、模型资源准备、在 Xcode 中构建运行，以及可选的 ORT 格式转换与基准测试流程。\n+\n+## 工程结构\n+\n+应用源码、打包资源及为本 Demo 引入的第三方**源码**均位于 **`PaddleOCRDemo/`** 下。单元测试位于与 Xcode 工程同级的 **`PaddleOCRDemoTests/`**。工程根目录还包含 `Podfile`、`scripts/`、`README.md` 与 `NOTICE`。\n+\n+## 环境要求\n+\n+- 搭载 **macOS** 的开发机，**Xcode 16.0+**（部署目标 **iOS 16+**）\n+- **CocoaPods**（`gem install cocoapods` 或 Homebrew）\n+- `curl`、`tar`\n+\n+## 支持的模型预设\n+\n+`scripts/fetch_ios_demo_models.sh` 会将 ONNX 模型包下载到 **`PaddleOCRDemo/Models/`**。当前支持的预设为：\n+\n+- **`PP-OCRv6_small`**（默认）\n+- **`PP-OCRv6_tiny`**\n+- **`PP-OCRv5_mobile`**\n+\n+## 快速开始 {#快速开始}\n+\n+在 **工程根目录**（`deploy/ios_demo`）执行：\n+\n+```bash\n+pod install\n+./scripts/fetch_ios_demo_models.sh\n+```\n+\n+可选：通过位置参数指定模型预设（即 bundle 名称，如 `PP-OCRv6_small`）：\n+\n+```bash\n+./scripts/fetch_ios_demo_models.sh PP-OCRv6_small\n+```\n+\n+中间 `.tar` 缓存目录为工程根下的 **`.fetch_ios_demo_models_work/`**。\n+\n+使用 CocoaPods 时，需先执行 `pod install` 再打开工作区：\n+\n+```bash\n+open PaddleOCRDemo.xcworkspace\n+```\n+\n+构建 **PaddleOCRDemo** scheme。请确认 **`PaddleOCRDemo/Models/`** 与 **`PaddleOCRDemo/Resources/SampleImages/`** 已通过文件夹引用 / **Copy Bundle Resources** 加入 App target；**`PaddleOCRDemoTests/Fixtures/`** 已加入测试 target（与仓库中工程配置一致）。内置样例图为 **`general_ocr_002.jpg`**。\n+\n+### ONNX Runtime 执行提供程序（EP）\n+\n+界面上的 **CPU** / **XNNPACK** / **Core ML** 分段控件用于在创建 Session 时指定优先使用的 EP。当某一 EP 无法覆盖部分算子时，ONNX Runtime 仍可能回退到默认 **CPU** EP。选择 **CPU** 时不会注册 XNNPACK 或 Core ML EP。\n+\n+## 转换为 ORT 模型格式（可选）\n+\n+生成 [ORT 格式](https://onnxruntime.ai/docs/performance/model-optimizations/ort-format-models.html) 权重：\n+\n+```bash\n+python3 -m pip install -r requirements-onnx-convert-ort.txt\n+./scripts/convert_onnx_to_ort.sh\n+```\n+\n+默认在 `PaddleOCRDemo/Models/` 下每个 `inference.onnx` 旁生成 `inference*.ort`；Demo 在存在 `.ort` 时会优先加载，因此可不将 ONNX 打入 App 包。若只希望保留一种权重文件，可手动删除 `inference.onnx`，或用 **`--output-dir`** 将 `.ort` 输出到单独目录后再替换 `PaddleOCRDemo/Models/` 中的文件。省略 **`--input-dir`** 时默认输入目录为 `PaddleOCRDemo/Models`。\n+\n+默认使用 **`--optimization-style Fixed`**（每个 ONNX 对应一个 `.ort`）。底层 Python 工具若未指定，会同时生成 Fixed 与 Runtime 两种（`*.ort` 与 `*.with_runtime_opt.ort`）。需要 Runtime 风格（如面向 Core ML 的导出）时，传入 **`--optimization-style Runtime`**（别名 `--optimization_style`）；可同时传入 **`--optimization-style Fixed Runtime`**。\n+\n+转换会生成 **`required_operators*.config`**（用于裁剪最小 ORT 构建）；**默认在脚本成功结束后删除**。使用 **`--keep-operator-config`** 可保留。这些文件**不参与** Demo 推理，仅加载 **`.ort`**（或 `.onnx`）权重。\n+\n+示例（Runtime，输出到独立目录）：\n+\n+```bash\n+./scripts/convert_onnx_to_ort.sh --output-dir ./out/ort_bundles --optimization-style Runtime\n+```\n+\n+## 基准测试（Benchmark）\n+\n+Demo 提供在真机或模拟器上测量 OCR 时延与内存的流水线。\n+\n+### 前置条件\n+\n+先完成 [快速开始](#快速开始) 中的资源准备。真机基准测试还需要：\n+\n+- **代码签名（仅真机）**：`xcodebuild` 部署到物理设备需要有效的 Apple Development 证书与描述文件。建议在 Xcode 中打开 `PaddleOCRDemo.xcworkspace`，在 *Signing & Capabilities* 中选择 Development Team 并启用自动管理；之后 `run_benchmark.sh` 等命令可复用已缓存的配置。首次在设备上启动时，须在 **设置 → 通用 → VPN 与设备管理** 中信任开发者证书（每台设备一次）。模拟器无需签名。\n+- 可选精度预检：主机安装 PaddleOCR（ONNX Runtime 引擎）用于生成参考结果，并执行 `python3 -m pip install -r requirements-accuracy.txt`。\n+\n+### 完整流水线\n+\n+在工程根目录执行 **`./scripts/run_benchmark.sh`**。脚本会解析输入图像、可选执行精度预检、在模拟器或真机上运行 XCTest 基准、提取产物并生成 Markdown 报告。始终以 **Release** 配置调用 `xcodebuild test`。\n+\n+**配置方式**：优先使用脚本参数。`--image` 将任意路径图像复制到本次运行的 `Fixtures/local-*`；`--fixture` 选择 `PaddleOCRDemoTests/Fixtures/` 下已有文件。若两者均未指定，则要求 `Fixtures/` 下恰好有一张非 `local-*` 的图像（常见为 `ios_ocr_benchmark_reference.jpg`）。\n+\n+| 用途 | 参数 |\n+| --- | --- |\n+| 运行目标 | `--udid <id>`（真机）或 `--simulator <name>` |\n+| 任意路径图像 | `--image <path>` |\n+| 已有 Fixture | `--fixture <name>`（stem 或 `stem.ext`） |\n+| 强度 | `--warmup <n>`、`--measured-iterations <n>` |\n+| 优先 ORT EP | `--ort-execution-provider CPU`、`XNNPACK` 或 `CORE_ML` |\n+| ORT profiling JSON | `--ort-profiling` |\n+| 可选精度预检 | `--accuracy-check`，可选 `--accuracy-reference-json <path>` |\n+| 精度 FAIL 时中止 | `--accuracy-check --stop-on-accuracy-failure` |\n+| 输出目录 | `--output-dir <dir>`（默认 `out/`） |\n+| 清理旧产物 | `--clean` |\n+\n+```bash\n+./scripts/run_benchmark.sh --udid <device-udid> --warmup 5 --measured-iterations 30\n+./scripts/run_benchmark.sh --udid <udid> --image /path/to/photo.png --ort-execution-provider CPU\n+./scripts/run_benchmark.sh --fixture ios_ocr_benchmark_reference --warmup 2 --measured-iterations 20\n+./scripts/run_benchmark.sh\n+PADDLEOCR_BENCHMARK_MEASURED_ITERATIONS=30 ./scripts/run_benchmark.sh --warmup 0\n+PADDLEOCR_BENCHMARK_ORT_EXECUTION_PROVIDER=XNNPACK ./scripts/run_benchmark.sh --fixture ios_ocr_benchmark_reference --warmup 1\n+./scripts/run_benchmark.sh --accuracy-check --udid <device-udid> --measured-iterations 30\n+./scripts/run_benchmark.sh --accuracy-check --stop-on-accuracy-failure --udid <device-udid>\n+./scripts/run_benchmark.sh --output-dir ./benchmark-out --measured-iterations 30\n+./scripts/run_benchmark.sh --udid <device-udid> --ort-profiling\n+```\n+\n+`--accuracy-check` 在基准测试前以独立 XCTest 运行。结果为 `PASS`、`FAIL` 或 `ERROR`。默认 `FAIL` 仅记录并继续基准；`ERROR` 会中止流水线。`--stop-on-accuracy-failure` 会在 `FAIL` 时跳过基准并返回非零退出码。\n+\n+**ORT Session profiling**（附件 `ort_profile_detection`、`ort_profile_recognition`）：使用 `--ort-profiling`。Profiling 会改变运行时行为，应与干净时延测量**分开**运行。\n+\n+基准**成功完成**后，脚本会在输出目录（默认 `out/`）覆盖写入 `run-status.json`、`on-device-performance.json`、`xctest-memory-metrics.json`、`benchmark-report.md`。若流水线提前 **ERROR**，这些文件可能缺失或不完整。\n+\n+报告内容包括模型输入张量形状分布、首次测量运行的行数、推断的模型预设、实际模型格式（`onnx` 或 `ort`）、det/rec/总权重大小、在可解析时的 App 可执行文件大小、冷启动模型加载时间、测量时延与内存等。形状统计在测量循环内按模型调用计数：每次完整 OCR 检测贡献 1 次，识别按 batch 各贡献 1 次。\n+\n+| 产物 | 来源 | 用途 |\n+|---|---|---|\n+| `accuracy-reference.json` | `ocr_reference_run.py` | 可选参考 JSON（除非提供 `--accuracy-reference-json`） |\n+| `accuracy-result.xcresult` | `xcodebuild test` | 可选精度预检 |\n+| `ios-ocr-export.json` | `extract_xcresult_attachments.py` | 从精度预检提取的 iOS 精度 payload |\n+| `accuracy-summary.json` | `compare_ocr_json.py` | 精度预检摘要 |\n+| `latency-result.xcresult` | `xcodebuild test` | 时延基准 |\n+| `memory-result.xcresult` | `xcodebuild test` | 内存基准 |\n+| `on-device-performance.json` | `extract_xcresult_attachments.py` | 从 `latency-result.xcresult` 提取的时延统计 |\n+| `ort_profile_detection`、`ort_profile_recognition` | `extract_xcresult_attachments.py` | 可选 ORT profiling JSON |\n+| `xctest-memory-metrics.json` | `extract_xctest_metrics.py` | XCTest 内存指标 |\n+| `logs/*.log` | `run_benchmark.sh` | 各步骤日志 |\n+| `run-status.json` | `run_benchmark.sh` | 步骤状态与元数据 |\n+| `benchmark-report.md` | `generate_benchmark_report.py` | 可读报告 |\n+\n+流水线状态为 **`COMPLETED`** 时退出码 **`0`**；**`ERROR`** 时非零。精度预检 `FAIL` 仅在使用 `--stop-on-accuracy-failure` 时改变退出码；`ERROR` 始终中止。\n+\n+### 进阶：XCTest 环境变量\n+\n+多数场景应使用 `run_benchmark.sh` 参数，而非直接设置环境变量。本节适用于在 Xcode 或自定义 `xcodebuild test` 中调试 **PaddleOCRDemoTests**。\n+\n+基准测试通过 **`PADDLEOCR_BENCHMARK_*`** 读取配置（作用于 XCTest runner，不会自动传入交互式 shell）。\n+\n+| 启动方式 | 配置方法 |\n+| --- | --- |\n+| **Xcode** | Scheme **PaddleOCRDemo** → **Test** → **Arguments** → **Environment Variables**，变量名如下。 |\n+| **自定义 `xcodebuild test`** | 使用 **`TEST_RUNNER_` + 变量名**（如 `TEST_RUNNER_PADDLEOCR_BENCHMARK_IMAGE_NAME=…`）。 |\n+\n+| 变量 | 未设置时 | 作用 |\n+| --- | --- | --- |\n+| `PADDLEOCR_BENCHMARK_IMAGE_NAME` | **`ios_ocr_benchmark_reference`** | `Fixtures/` 下 bundled 图像（stem 或 `stem.ext`）。 |\n+| `PADDLEOCR_BENCHMARK_WARMUP_ITERATIONS` | **3** | 计时时前的预热完整 OCR 次数。 |\n+| `PADDLEOCR_BENCHMARK_MEASURED_ITERATIONS` | **10** | 计时迭代次数。 |\n+| `PADDLEOCR_BENCHMARK_ORT_EXECUTION_PROVIDER` | **CPU** | **`CPU`**、**`XNNPACK`** 或 **`CORE_ML`**。 |\n+| `PADDLEOCR_BENCHMARK_ORT_PROFILING` | （未设置） | 启用 ORT session profiling（`1` / `true` / `yes` / `on`）。 |\n+| `PADDLEOCR_BENCHMARK_ONLY_TESTING_SCOPE` | latency + memory 基准测试 | 仅 `run_benchmark.sh`：传给多次 `xcodebuild -only-testing` 的逗号分隔列表。 |\n+\n+两个迭代相关变量须为非负整数。\n+\n+## 第三方许可\n+\n+随 Demo 打包的 **Clipper**（polyclipping 6.4.2）采用 [Boost Software License 1.0](https://www.boost.org/LICENSE_1_0.txt)；详见 `NOTICE` 与 `PaddleOCRDemo/ThirdParty/Clipper1/LICENSE`。CocoaPods 依赖的许可以 `pod install` 后的 `Podfile.lock` 及各自仓库为准。\ndiff --git a/docs/version3.x/inference_deployment/cross_platform/on_device_deployment.en.md b/docs/version3.x/inference_deployment/cross_platform/on_device_deployment.en.md\ndeleted file mode 100644\nindex e59cc6311c5..00000000000\n--- a/docs/version3.x/inference_deployment/cross_platform/on_device_deployment.en.md\n+++ /dev/null\n@@ -1,479 +0,0 @@\n----\n-comments: true\n----\n-\n-# OCR On-Device Deployment Demo Usage Guide\n-\n-- [Quick Start](#quick-start)\n- - [Environment Preparation](#environment-preparation)\n- - [Deployment Steps](#deployment-steps)\n-- [Code Introduction](#code-introduction)\n-- [Project Explanation](#engineering-details)\n-- [Advanced Usage](#advanced-usage)\n- - [Update Prediction Library](#updating-the-prediction-library)\n- - [Convert NB Model](#converting-nb-models)\n- - [Update Model, Label File, and Prediction Image](#updating-models-label-files-and-prediction-images)\n- - [Update Model](#updating-models)\n- - [Update Label File](#updating-label-files)\n- - [Update Prediction Image](#updating-prediction-images)\n- - [Update Input/Output Preprocessing](#updating-inputoutput-preprocessing)\n-\n-This guide mainly introduces how to run the PaddleX on-device deployment demo for OCR text recognition on an Android shell.\n-\n-The following OCR models are supported in this guide:\n-\n-- PP-OCRv3_mobile (cpu)\n-- PP-OCRv4_mobile (cpu)\n-- PP-OCRv5_mobile (cpu)\n-\n-## Quick Start\n-\n-### Environment Preparation\n-\n-1. Install the CMAKE compilation tool in your local environment and download an NDK package for your current system from the [Android NDK official website](https://developer.android.google.cn/ndk/downloads). For example, if developing on a Mac, download the NDK package for the Mac platform from the Android NDK official website.\n-\n- **Environment Requirements**\n-\n- - `CMake >= 3.10` (the minimum version has not been verified; 3.20 or above is recommended)\n- - `Android NDK >= r17c` (the minimum version has not been verified; r20b or above is recommended)\n-\n- **Test Environment Used in This Guide**:\n-\n- - `cmake == 3.20.0`\n- - `android-ndk == r20b`\n-\n-2. Prepare an Android phone and enable USB debugging mode. Method: `Phone Settings -> Find Developer Options -> Enable Developer Options and USB Debugging Mode`\n-\n-3. Install the ADB tool on your computer for debugging. The installation methods for ADB are as follows:\n-\n- 3.1. Install ADB on a Mac:\n-\n- ```shell\n- brew cask install android-platform-tools\n- ```\n-\n- 3.2. Install ADB on Linux:\n-\n- ```shell\n- sudo apt update\n- sudo apt install -y wget adb\n- ```\n-\n- 3.3. Install ADB on Windows:\n-\n- For installation on Windows, download and install the ADB software package from Google's Android platform: [Link](https://developer.android.com/studio)\n-\n- Open a terminal, connect your phone to the computer, and enter the following command in the terminal:\n-\n- ```shell\n- adb devices\n- ```\n-\n- If there is a `device` output, the installation is successful.\n-\n- ```shell\n- List of devices attached\n- 744be294 device\n- ```\n-\n-### Material Preparation\n-\n-1. Clone the `feature/paddle-x` branch of the `Paddle-Lite-Demo` repository to the `PaddleX-Lite-Deploy` directory.\n-\n- ```shell\n- git clone -b feature/paddle-x https://github.com/PaddlePaddle/Paddle-Lite-Demo.git PaddleX-Lite-Deploy\n- ```\n-\n-2. Fill out the [questionnaire](https://paddle.wjx.cn/vm/eaaBo0H.aspx#) to download the compressed package. Place the compressed package in the specified extraction directory, switch to the specified extraction directory, and execute the extraction command.\n-\n- ```shell\n- # 1. Switch to the specified extraction directory\n- cd PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo\n-\n- # 2. Execute the extraction command\n- unzip ocr.zip\n- ```\n-\n-### Deployment Steps\n-\n-1. Switch the working directory to `PaddleX-Lite-Deploy/libs` and run the `download.sh` script to download the required Paddle Lite prediction library. This step only needs to be executed once to support each demo.\n-\n-2. Switch the working directory to `PaddleX-Lite-Deploy/ocr/assets` and run the `download.sh` script to download the [paddle_lite_opt tool](https://www.paddlepaddle.org.cn/lite/v2.10/user_guides/model_optimize_tool.html)-optimized NB model file, prediction images, dictionary files, and other materials.\n-\n-3. Switch the working directory to `PaddleX-Lite-Deploy/ocr/android/shell/cxx/ppocr_demo` and run the `build.sh` script to complete the compilation of the executable file.\n-\n-4. Switch the working directory to `PaddleX-Lite-Deploy/ocr/android/shell/cxx/ppocr_demo` and run the `run.sh` script to complete the on-device prediction.\n-\n-**Notes**:\n-\n-- Before running the `build.sh` script, change the path specified by `NDK_ROOT` to the actual installation path of NDK.\n-- On Windows systems, you can use Git Bash to execute the deployment steps.\n-- If compiling on a Windows system, set `CMAKE_SYSTEM_NAME` to `windows` in `CMakeLists.txt`.\n-- If compiling on a Mac system, set `CMAKE_SYSTEM_NAME` to `darwin` in `CMakeLists.txt`.\n-- Keep the ADB connection active when running the `run.sh` script.\n-- The `download.sh` and `run.sh` scripts support passing parameters to specify the model. If no model is specified, the `PP-OCRv5_mobile` model is used by default. The following models are currently supported:\n- - `PP-OCRv3_mobile`\n- - `PP-OCRv4_mobile`\n- - `PP-OCRv5_mobile`\n-\n-Here is an example of the actual operation:\n-\n-```shell\n-# 1. Download the required Paddle Lite prediction library\n-cd PaddleX-Lite-Deploy/libs\n-sh download.sh\n-\n-# 2. Download the paddle_lite_opt tool-optimized NB model file, prediction images, dictionary files, and other materials\n-cd ../ocr/assets\n-sh download.sh PP-OCRv5_mobile\n-\n-# 3. Complete the compilation of the executable file\n-cd ../android/shell/ppocr_demo\n-sh build.sh\n-\n-# 4. Prediction\n-sh run.sh PP-OCRv5_mobile\n-```\n-\n-The output is as follows:\n-\n-```text\n-The detection visualized image saved in ./test_img_result.jpg\n-0 纯臻营养护发素 0.998541\n-1 产品信息/参数 0.999094\n-2 (45元/每公斤，100公斤起订） 0.948841\n-3 每瓶22元，1000瓶起订) 0.961245\n-4 【品牌】：代加工方式/OEMODM 0.970401\n-5 【品名】：纯臻营养护发素 0.977496\n-6 ODMOEM 0.955396\n-7 【产品编号】：YM-X-3011 0.977864\n-8 【净含量】：220ml 0.970538\n-9 【适用人群】：适合所有肤质 0.995907\n-10 【主要成分】：鲸蜡硬脂醇、燕麦β-葡聚 0.975813\n-11 糖、椰油酰胺丙基甜菜碱、泛醌 0.964397\n-12 （成品包材） 0.97298\n-13 【主要功能】：可紧致头发磷层，从而达到 0.989097\n-14 即时持久改善头发光泽的效果，给干燥的头 0.990088\n-15 发足够的滋养 0.998037\n-``` \n-\n-\n-\n-## Code Introduction\n-\n-```\n-.\n-├── ...\n-├── ocr \n-│ ├── ...\n-│ ├── android\n-│ │ ├── ...\n-│ │ └── shell\n-│ │ └── ppocr_demo\n-│ │ ├── src # Contains prediction code\n-│ │ | ├── cls_process.cc # Full inference process for orientation classifier, including preprocessing, prediction, and postprocessing\n-│ │ | ├── rec_process.cc # Full inference process for recognition model CRNN, including preprocessing, prediction, and postprocessing\n-│ │ | ├── det_process.cc # Full inference process for detection model CRNN, including preprocessing, prediction, and postprocessing\n-│ │ | ├── det_post_process.cc # Postprocessing file for detection model DB\n-│ │ | ├── pipeline.cc # Full inference process code for OCR text recognition demo\n-│ │ | └── MakeFile # MakeFile file for prediction code\n-│ │ | \n-│ │ ├── CMakeLists.txt # CMake file that defines the compilation method for the executable\n-│ │ ├── README.md\n-│ │ ├── build.sh # Used for compiling the executable\n-│ │ └── run.sh # Used for prediction\n-│ └── assets # Stores models, test images, label files, and config files\n-│ ├── images # Stores test images\n-│ ├── labels # Stores dictionary files (see remarks below for details)\n-│ ├── models # Stores nb models\n-│ ├── config.txt\n-│ └── download.sh # Download script for paddle_lite_opt tool-optimized models\n-└── libs # Stores prediction libraries and OpenCV libraries for different platforms.\n- ├── ...\n- └── download.sh # Download script for Paddle Lite prediction libraries and OpenCV libraries\n-```\n-\n-**Remarks**:\n-\n- - The `PaddleX-Lite-Deploy/ocr/assets/labels/` directory contains the dictionary files `ppocr_keys_v1.txt` for PP-OCRv3 and PP-OCRv4 models, and `ppocr_keys_ocrv5.txt` for the PP-OCRv5 model. The appropriate dictionary file is automatically selected during inference based on the model name, so no manual intervention is required.\n- - If you are using an English/numeric or other language model, you need to replace it with the corresponding language dictionary. The PaddleOCR repository provides [some dictionary files](../../../ppocr/utils).\n-\n-```shell\n-# Parameters of the executable in run.sh script:\n-adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/test.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/${LABEL_FILE} \\\n- ./config.txt\"\n-\n-First parameter: ppocr_demo executable\n-Second parameter: ./models/${MODEL_NAME}_det.nb Detection model .nb file\n-Third parameter: ./models/${MODEL_NAME}_rec.nb Recognition model .nb file\n-Fourth parameter: ./models/${CLS_MODEL_FILE} Text line orientation classification model .nb file (automatically selected based on model name by default)\n-Fifth parameter: ./images/test.jpg Test image\n-Sixth parameter: ./test_img_result.jpg Result save file\n-Seventh parameter: ./labels/${LABEL_FILE} Label file (automatically selected based on model name by default)\n-Eighth parameter: ./config.txt Configuration file containing hyperparameters for the detection and classification models\n-```\n-\n-```shell\n-# List of Specific Parameters in config.txt:\n-max_side_len 960 # When the width or height of the input image is greater than 960, the image is scaled proportionally so that the longest side of the image is 960.\n-det_db_thresh 0.3 # Used to filter the binarized images predicted by DB; setting it to 0.3 has no significant impact on the results.\n-det_db_box_thresh 0.5 # Threshold for filtering boxes in the DB post-processing; if there are missing boxes in detection, you may reduce this value.\n-det_db_unclip_ratio 1.6 # Represents the compactness of the text box; the smaller the value, the closer the text box is to the text.\n-use_direction_classify 0 # Whether to use a direction classifier: 0 means not using it, 1 means using it.\n-```\n-\n-## Engineering Details\n-\n-The OCR text recognition demo accomplishes the OCR text recognition function using three models collaboratively. First, the input image undergoes detection processing via the `${MODEL_NAME}_det.nb` model, followed by text direction classification using the `ch_ppocr_mobile_v2.0_cls_slim_opt.nb` model, and finally, text recognition with the `${MODEL_NAME}_rec.nb` model.\n-\n-1. `pipeline.cc`: Full-process prediction code for the OCR text recognition demo\n- This file handles the entire process control for serial inference of the three models, including scheduling for the entire processing flow.\n-\n- - The `Pipeline::Pipeline(...)` method initializes the three model class constructors, accomplishes model loading, thread count, core binding, and predictor creation.\n- - The `Pipeline::Process(...)` method manages the entire process control for serial inference of the three models.\n- \n-2. `cls_process.cc`: Prediction file for the direction classifier\n- This file handles the preprocessing, prediction, and postprocessing for the direction classifier.\n-\n- - The `ClsPredictor::ClsPredictor()` method initializes model loading, thread count, core binding, and predictor creation.\n- - The `ClsPredictor::Preprocess()` method handles model preprocessing.\n- - The `ClsPredictor::Postprocess()` method handles model postprocessing.\n-\n-3. `rec_process.cc`: Prediction file for the CRNN recognition model\n- This file handles the preprocessing, prediction, and postprocessing for the CRNN recognition model.\n-\n- - The `RecPredictor::RecPredictor()` method initializes model loading, thread count, core binding, and predictor creation.\n- - The `RecPredictor::Preprocess()` method handles model preprocessing.\n- - The `RecPredictor::Postprocess()` method handles model postprocessing.\n-\n-4. `det_process.cc`: Prediction file for the DB detection model\n- This file handles the preprocessing, prediction, and postprocessing for the DB detection model.\n-\n- - The `DetPredictor::DetPredictor()` method initializes model loading, thread count, core binding, and predictor creation.\n- - The `DetPredictor::Preprocess()` method handles model preprocessing.\n- - The `DetPredictor::Postprocess()` method handles model postprocessing.\n-\n-5. `db_post_process`: Postprocessing functions for the DB detection model, including calls to the clipper library\n- This file implements third-party library calls and other postprocessing methods for the DB detection model.\n-\n- - The `std::vector<std::vector<std::vector<int>>> BoxesFromBitmap(...)` method retrieves detection boxes from a Bitmap.\n- - The `std::vector<std::vector<std::vector<int>>> FilterTagDetRes(...)` method retrieves target box positions based on recognition results.\n-\n-## Advanced Usage\n-\n-If the quick start section does not meet your needs, refer to this section for custom modifications to the demo.\n-\n-This section mainly includes four parts:\n-\n-- Updating the prediction library;\n-- Converting `.nb` models;\n-- Updating models, label files, and prediction images;\n-- Updating input/output preprocessing.\n-\n-### Updating the Prediction Library\n-\n-The prediction library used in this guide is the latest version (214rc), and manual updates are not recommended.\n-\n-If you need to use a different version, follow these steps to update the prediction library:\n-\n-* Paddle Lite project: https://github.com/PaddlePaddle/Paddle-Lite\n- * Refer to the [Paddle Lite Source Code Compilation Documentation](https://www.paddlepaddle.org.cn/lite/develop/source_compile/compile_env.html) to compile the Android prediction library.\n- * The final compilation output is located in `build.lite.xxx.xxx.xxx` under `inference_lite_lib.xxx.xxx`.\n- * Replace the C++ library:\n- * Header files:\n- Replace the `PaddleX-Lite-Deploy/libs/android/cxx/include` folder in the demo with the generated `build.lite.android.xxx.gcc/inference_lite_lib.android.xxx/cxx/include` folder.\n- * armeabi-v7a:\n- Replace the `PaddleX-Lite-Deploy/libs/android/cxx/libs/armeabi-v7a/libpaddle_lite_api_shared.so` library in the demo with the generated `build.lite.android.armv7.gcc/inference_lite_lib.android.armv7/cxx/libs/libpaddle_lite_api_shared.so` library.\n- * arm64-v8a:\n- Replace the `PaddleX-Lite-Deploy/libs/android/cxx/libs/arm64-v8a/libpaddle_lite_api_shared.so` library in the demo with the generated `build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/cxx/libs/libpaddle_lite_api_shared.so` library.\n-\n-### Converting .nb Models\n-\n-If you want to use your own trained models, follow the process below to obtain `.nb` models.\n-\n-#### Terminal Command Method (Supports Mac/Ubuntu)\n-\n-1. Navigate to the [release interface](https://github.com/PaddlePaddle/Paddle-Lite/releases) of the Paddle-Lite GitHub repository and download the corresponding conversion tool, opt, for the desired version (the latest version is recommended).\n-\n-2. After downloading the opt tool, execute the following command (using the 2.14rc version of the linux_x86 opt tool to convert the PP-OCRv5_mobile_det model as an example):\n-\n- ```bash\n- ./opt_linux_x86 \\\n- --model_file=PP-OCRv5_mobile_det/inference.pdmodel \\\n- --param_file=PP-OCRv5_mobile_det/inference.pdiparams \\\n- --optimize_out=PP-OCRv5_mobile_det \\\n- --valid_targets=arm\n- ```\n-\n-For detailed instructions on converting `.nb` models using the terminal command method, refer to the [Using the Executable opt](https://www.paddlepaddle.org.cn/lite/v2.12/user_guides/opt/opt_bin.html) section in the Paddle-Lite repository.\n-\n-#### Python Script Method (Supports Windows/Mac/Ubuntu)\n-\n-1. Install the latest version of the paddlelite wheel package.\n-\n- ```bash\n- pip install --pre paddlelite\n- ```\n-\n-2. Use the Python script to convert the model. Below is an example code snippet for converting the PP-OCRv5_mobile_det model:\n-\n- ```python\n- from paddlelite.lite import Opt\n-\n- # 1. Create an Opt instance\n- opt = Opt()\n- # 2. Specify the input model paths \n- opt.set_model_file(\"./PP-OCRv5_mobile_det/inference.pdmodel\")\n- opt.set_param_file(\"./PP-OCRv5_mobile_det/inference.pdiparams\")\n- # 3. Specify the target platform for optimization\n- opt.set_valid_places(\"arm\")\n- # 4. Specify the output path for the optimized model\n- opt.set_optimize_out(\"./PP-OCRv5_mobile_det\")\n- # 5. Execute model optimization\n- opt.run()\n- ```\n-\n-For detailed instructions on converting `.nb` models using the Python script method, refer to the [Python Script opt Usage](https://www.paddlepaddle.org.cn/lite/v2.12/api_reference/python_api/opt.html) section in the Paddle-Lite repository.\n-\n-**Notes**\n-\n-- For detailed information about the model optimization tool `opt`, refer to Paddle-Lite's [Model Optimization Tool opt](https://www.paddlepaddle.org.cn/lite/v2.12/user_guides/model_optimize_tool.html).\n-- Currently, only static graph models in `.pdmodel` format can be converted to `.nb` format. In PaddlePaddle version 3.0 and above, the default exported model format is `.json`. If you want to export the model in `.pdmodel` format, simply add `-o Global.export_with_pir=False` during export.\n-\n-### Updating Models, Label Files, and Prediction Images\n-\n-#### Updating Models\n-\n-This guide has only validated the `PP-OCRv3_mobile`, `PP-OCRv4_mobile`, and `PP-OCRv5_mobile` models. Other models may not be compatible.\n-\n-If you fine-tune the `PP-OCRv5_mobile` model and generate a new model named `PP-OCRv5_mobile_ft`, follow these steps to replace the original model with your fine-tuned model:\n-\n-1. Place the `.nb` models of `PP-OCRv5_mobile_ft` into the directory `PaddleX-Lite-Deploy/ocr/assets/models/`. The resulting file structure should be:\n-\n- ```text\n- .\n- ├── ocr \n- │ ├── ...\n- │ └── assets \n- │ ├── models\n- │ │ ├── ...\n- │ │ ├── PP-OCRv5_mobile_ft_det.nb \n- │ │ └── PP-OCRv5_mobile_ft_rec.nb \n- │ └── ...\n- └── ...\n- ```\n-\n-2. Add the model name to the `MODEL_LIST` in the `run.sh` script.\n-\n- ```shell\n- MODEL_LIST=\"PP-OCRv3_mobile PP-OCRv4_mobile PP-OCRv5_mobile PP-OCRv5_mobile_ft\" # Models are separated by spaces\n- ```\n-\n-3. Specify the model directory name when running the `run.sh` script.\n-\n- ```shell\n- sh run.sh PP-OCRv5_mobile_ft\n- ```\n-\n-**Notes**:\n-\n-- If the input Tensor, Shape, or Dtype of the model is updated:\n-\n- - For the text direction classifier model, update the `ClsPredictor::Preprocess` function in `ppocr_demo/src/cls_process.cc`.\n- - For the detection model, update the `DetPredictor::Preprocess` function in `ppocr_demo/src/det_process.cc`.\n- - For the recognition model, update the `RecPredictor::Preprocess` function in `ppocr_demo/src/rec_process.cc`.\n-\n-- If the output Tensor or Dtype of the model is updated:\n-\n- - For the text direction classifier model, update the `ClsPredictor::Postprocess` function in `ppocr_demo/src/cls_process.cc`.\n- - For the detection model, update the `DetPredictor::Postprocess` function in `ppocr_demo/src/det_process.cc`.\n- - For the recognition model, update the `RecPredictor::Postprocess` function in `ppocr_demo/src/rec_process.cc`.\n-\n-#### Updating Label Files\n-\n-To update the label file, place the new label file in the directory `PaddleX-Lite-Deploy/ocr/assets/labels/` and update the execution command in `PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo/run.sh` following the model update method.\n-\n-For example, to update to `new_labels.txt`:\n-\n- ```shell\n- # File: `PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo/run.sh`\n- # Original command\n- adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/test.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/${LABEL_FILE} \\\n- ./config.txt\"\n- # Updated command\n- adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/test.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/new_labels.txt \\\n- ./config.txt\"\n- ```\n- \n-#### Updating Prediction Images\n-\n-If you need to update the prediction images, place the updated images in the `PaddleX-Lite-Deploy/ocr/assets/images/` directory and update the execution command in the `PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo/run.sh` file.\n-\n-Here is an example of updating to `new_pics.jpg`:\n-\n- ```shell\n- # File: `PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo/run.sh`\n- ## Original command\n- adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/test.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/${LABEL_FILE} \\\n- ./config.txt\"\n- # Updated command\n- adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/new_pics.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/${LABEL_FILE} \\\n- ./config.txt\"\n- ```\n-\n-### Updating Input/Output Preprocessing\n-\n-- Updating Input Preprocessing\n- - For the text direction classifier model, update the `ClsPredictor::Preprocess` function in `ppocr_demo/src/cls_process.cc`.\n- - For the detection model, update the `DetPredictor::Preprocess` function in `ppocr_demo/src/det_process.cc`.\n- - For the recognition model, update the `RecPredictor::Preprocess` function in `ppocr_demo/src/rec_process.cc`.\n-\n-- Updating Output Preprocessing\n- - For the text direction classifier model, update the `ClsPredictor::Postprocess` function in `ppocr_demo/src/cls_process.cc`.\n- - For the detection model, update the `DetPredictor::Postprocess` function in `ppocr_demo/src/det_process.cc`.\n- - For the recognition model, update the `RecPredictor::Postprocess` function in `ppocr_demo/src/rec_process.cc`.\ndiff --git a/docs/version3.x/inference_deployment/cross_platform/on_device_deployment.md b/docs/version3.x/inference_deployment/cross_platform/on_device_deployment.md\ndeleted file mode 100644\nindex 47b8f887406..00000000000\n--- a/docs/version3.x/inference_deployment/cross_platform/on_device_deployment.md\n+++ /dev/null\n@@ -1,479 +0,0 @@\n----\n-comments: true\n----\n-\n-# OCR 端侧部署 demo 使用指南\n-\n-- [快速开始](#_1)\n- - [环境准备](#_2)\n- - [部署步骤](#_4)\n-- [代码介绍](#_5)\n-- [工程详解](#_6)\n-- [进阶使用](#_7)\n- - [更新预测库](#_8)\n- - [转换 nb 模型](#nb)\n- - [更新模型、标签文件和预测图片](#_9)\n- - [更新模型](#_10)\n- - [更新标签文件](#_11)\n- - [更新预测图片](#_12)\n- - [更新输入/输出预处理](#_13)\n-\n-本指南主要介绍 PaddleX 端侧部署——OCR文字识别 demo 在 Android shell 上的运行方法。\n-\n-本指南适配了以下 OCR 模型：\n-\n-- PP-OCRv3_mobile（cpu）\n-- PP-OCRv4_mobile（cpu）\n-- PP-OCRv5_mobile（cpu）\n-\n-## 快速开始\n-\n-### 环境准备\n-\n-1. 在本地环境安装好 CMAKE 编译工具，并在 [Android NDK 官网](https://developer.android.google.cn/ndk/downloads)下载当前系统的某个版本的 NDK 软件包。例如，在 Mac 上开发，需要在 Android NDK 官网下载 Mac 平台的 NDK 软件包\n-\n- **环境要求**\n-\n- - `CMake >= 3.10`（最低版本未经验证，推荐 3.20 及以上）\n- - `Android NDK >= r17c`（最低版本未经验证，推荐 r20b 及以上）\n-\n- **本指南所使用的测试环境：**\n-\n- - `cmake == 3.20.0`\n- - `android-ndk == r20b`\n-\n-2. 准备一部 Android 手机，并开启 USB 调试模式。开启方法: `手机设置 -> 查找开发者选项 -> 打开开发者选项和 USB 调试模式`\n-\n-3. 电脑上安装 ADB 工具，用于调试。ADB 安装方式如下：\n-\n- 3.1. Mac 电脑安装 ADB:\n-\n- ```shell\n- brew cask install android-platform-tools\n- ```\n-\n- 3.2. Linux 安装 ADB\n-\n- ```shell\n- sudo apt update\n- sudo apt install -y wget adb\n- ```\n-\n- 3.3. Windows 安装 ADB\n-\n- win 上安装需要去谷歌的安卓平台下载 ADB 软件包进行安装：[链接](https://developer.android.com/studio)\n-\n- 打开终端，手机连接电脑，在终端中输入\n-\n- ```shell\n- adb devices\n- ```\n-\n- 如果有 device 输出，则表示安装成功。\n-\n- ```shell\n- List of devices attached\n- 744be294 device\n- ```\n-\n-### 物料准备\n-\n-1. 克隆 `Paddle-Lite-Demo` 仓库的 `feature/paddle-x` 分支到 `PaddleX-Lite-Deploy` 目录。\n-\n- ```shell\n- git clone -b feature/paddle-x https://github.com/PaddlePaddle/Paddle-Lite-Demo.git PaddleX-Lite-Deploy\n- ```\n-\n-2. 填写 [问卷](https://paddle.wjx.cn/vm/eaaBo0H.aspx#) 下载压缩包，将压缩包放到指定解压目录，切换到指定解压目录后执行解压命令。\n-\n- ```shell\n- # 1. 切换到指定解压目录\n- cd PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo\n-\n- # 2. 执行解压命令\n- unzip ocr.zip\n- ```\n-\n-### 部署步骤\n-\n-1. 将工作目录切换到 `PaddleX-Lite-Deploy/libs` 目录，运行 `download.sh` 脚本，下载需要的 Paddle Lite 预测库。此步骤只需执行一次，即可支持每个 demo 使用。\n-\n-2. 将工作目录切换到 `PaddleX-Lite-Deploy/ocr/assets` 目录，运行 `download.sh` 脚本，下载 [paddle_lite_opt 工具](https://www.paddlepaddle.org.cn/lite/v2.10/user_guides/model_optimize_tool.html) 优化后的 nb 模型文件及预测图片、字典文件等物料。\n-\n-3. 将工作目录切换到 `PaddleX-Lite-Deploy/ocr/android/shell/cxx/ppocr_demo` 目录，运行 `build.sh` 脚本，完成可执行文件的编译。\n-\n-4. 将工作目录切换到 `PaddleX-Lite-Deploy/ocr/android/shell/cxx/ppocr_demo`，运行 `run.sh` 脚本，完成在端侧的预测。\n-\n-**注意事项：**\n-\n-- 在运行 `build.sh` 脚本前，需要更改 `NDK_ROOT` 指定的路径为实际安装的 NDK 路径。\n-- 在 Windows 系统上可以使用 Git Bash 执行部署步骤。\n-- 若在 Windows 系统上编译，需要将 `CMakeLists.txt` 中的 `CMAKE_SYSTEM_NAME` 设置为 `windows`。\n-- 若在 Mac 系统上编译，需要将 `CMakeLists.txt` 中的 `CMAKE_SYSTEM_NAME` 设置为 `darwin`。\n-- 在运行 `run.sh` 脚本时需保持 ADB 连接。\n-- `download.sh` 和 `run.sh` 支持传入参数来指定模型，若不指定则默认使用 `PP-OCRv5_mobile` 模型。目前适配了以下模型：\n- - `PP-OCRv3_mobile`\n- - `PP-OCRv4_mobile`\n- - `PP-OCRv5_mobile`\n-\n-以下为实际操作时的示例：\n-\n-```shell\n-# 1. 下载需要的 Paddle Lite 预测库\n-cd PaddleX-Lite-Deploy/libs\n-sh download.sh\n-\n-# 2. 下载 paddle_lite_opt 工具优化后的 nb 模型文件及预测图片、字典文件等物料\n-cd ../ocr/assets\n-sh download.sh PP-OCRv5_mobile\n-\n-# 3. 完成可执行文件的编译\n-cd ../android/shell/ppocr_demo\n-sh build.sh\n-\n-# 4. 预测\n-sh run.sh PP-OCRv5_mobile\n-```\n-\n-运行结果如下所示：\n-\n-```text\n-The detection visualized image saved in ./test_img_result.jpg\n-0 纯臻营养护发素 0.998541\n-1 产品信息/参数 0.999094\n-2 (45元/每公斤，100公斤起订） 0.948841\n-3 每瓶22元，1000瓶起订) 0.961245\n-4 【品牌】：代加工方式/OEMODM 0.970401\n-5 【品名】：纯臻营养护发素 0.977496\n-6 ODMOEM 0.955396\n-7 【产品编号】：YM-X-3011 0.977864\n-8 【净含量】：220ml 0.970538\n-9 【适用人群】：适合所有肤质 0.995907\n-10 【主要成分】：鲸蜡硬脂醇、燕麦β-葡聚 0.975813\n-11 糖、椰油酰胺丙基甜菜碱、泛醌 0.964397\n-12 （成品包材） 0.97298\n-13 【主要功能】：可紧致头发磷层，从而达到 0.989097\n-14 即时持久改善头发光泽的效果，给干燥的头 0.990088\n-15 发足够的滋养 0.998037\n-``` \n-\n-\n-\n-## 代码介绍\n-\n-```\n-.\n-├── ...\n-├── ocr \n-│ ├── ...\n-│ ├── android\n-│ │ ├── ...\n-│ │ └── shell\n-│ │ └── ppocr_demo\n-│ │ ├── src # 存放预测代码\n-│ │ | ├── cls_process.cc # 方向分类器的推理全流程，包含预处理、预测和后处理三部分\n-│ │ | ├── rec_process.cc # 识别模型 CRNN 的推理全流程，包含预处理、预测和后处理三部分\n-│ │ | ├── det_process.cc # 检测模型 CRNN 的推理全流程，包含预处理、预测和后处理三部分\n-│ │ | ├── det_post_process.cc # 检测模型 DB 的后处理文件\n-│ │ | ├── pipeline.cc # OCR 文字识别 demo 推理全流程代码\n-│ │ | └── MakeFile # 预测代码的 MakeFile 文件\n-│ │ | \n-│ │ ├── CMakeLists.txt # CMake 文件，约束可执行文件的编译方法\n-│ │ ├── README.md\n-│ │ ├── build.sh # 用于可执行文件的编译\n-│ │ └── run.sh # 用于预测\n-│ └── assets # 存放模型、测试图片、标签文件、config 文件\n-│ ├── images # 存放测试图片\n-│ ├── labels # 存放字典文件，更多详情可参考下文备注\n-│ ├── models # 存放 nb 模型\n-│ ├── config.txt\n-│ └── download.sh # 下载脚本，用于下载 paddle_lite_opt 工具优化后的模型\n-└── libs # 存放不同端的预测库和 OpenCV 库。\n- ├── ...\n- └── download.sh # 下载脚本，用于下载 Paddle Lite 预测库和 OpenCV 库\n-```\n-\n-**备注：**\n-\n- - `PaddleX-Lite-Deploy/ocr/assets/labels/` 目录下存放了 PP-OCRv3、PP-OCRv4 模型的字典文件 `ppocr_keys_v1.txt` 以及 PP-OCRv5 模型的字典文件 `ppocr_keys_ocrv5.txt`。在实际推理过程中，会根据模型名称自动选择相应的字典文件，因此无需手动干预。\n- - 如果使用的 nb 模型是英文数字或其他语言的模型，需要更换为对应语言的字典。PaddleOCR 仓库提供了[部分字典文件](../../../ppocr/utils)。\n-\n-```shell\n-# run.sh 脚本中可执行文件的参数含义：\n-adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/test.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/${LABEL_FILE} \\\n- ./config.txt\"\n-\n-第一个参数：ppocr_demo 可执行文件\n-第二个参数：./models/${MODEL_NAME}_det.nb 检测模型的.nb文件\n-第三个参数：./models/${MODEL_NAME}_rec.nb 识别模型的.nb文件\n-第四个参数：./models/${CLS_MODEL_FILE} 文本行方向分类模型的.nb文件，默认根据模型名自动选择\n-第五个参数：./images/test.jpg 测试图片\n-第六个参数：./test_img_result.jpg 结果保存文件\n-第七个参数：./labels/${LABEL_FILE} label 文件，默认根据模型名自动选择\n-第八个参数：./config.txt 配置文件，模型的超参数配置文件，包含了检测器、分类器的超参数\n-```\n-\n-```shell\n-# config.txt 具体参数 List：\n-max_side_len 960 # 输入图像长宽大于 960 时，等比例缩放图像，使得图像最长边为 960\n-det_db_thresh 0.3 # 用于过滤 DB 预测的二值化图像，设置为 0.3 对结果影响不明显\n-det_db_box_thresh 0.5 # DB 后处理过滤 box 的阈值，如果检测存在漏框情况，可酌情减小\n-det_db_unclip_ratio 1.6 # 表示文本框的紧致程度，越小则文本框更靠近文本\n-use_direction_classify 0 # 是否使用方向分类器，0 表示不使用，1 表示使用\n-```\n-\n-## 工程详解\n-\n-OCR 文字识别 demo 由三个模型一起完成 OCR 文字识别功能，对输入图片先通过 `${MODEL_NAME}_det.nb` 模型做检测处理，然后通过 `ch_ppocr_mobile_v2.0_cls_slim_opt.nb` 模型做文字方向分类处理，最后通过 `${MODEL_NAME}_rec.nb` 模型完成文字识别处理。\n-\n-1. `pipeline.cc` : OCR 文字识别 demo 预测全流程代码\n- 该文件完成了三个模型串行推理的全流程控制处理，包含整个处理过程的调度处理。\n-\n- - `Pipeline::Pipeline(...)` 方法完成调用三个模型类构造函数，完成模型加载和线程数、绑核处理及 predictor 创建处理\n- - `Pipeline::Process(...)` 方法用于完成这三个模型串行推理的全流程控制处理\n- \n-2. `cls_process.cc` 方向分类器的预测文件\n- 该文件完成了方向分类器的预处理、预测和后处理过程\n-\n- - `ClsPredictor::ClsPredictor()` 方法用于完成模型加载和线程数、绑核处理及 predictor 创建处理\n- - `ClsPredictor::Preprocess()` 方法用于模型的预处理\n- - `ClsPredictor::Postprocess()` 方法用于模型的后处理\n-\n-3. `rec_process.cc` 识别模型 CRNN 的预测文件\n- 该文件完成了识别模型 CRNN 的预处理、预测和后处理过程\n-\n- - `RecPredictor::RecPredictor()` 方法用于完成模型加载和线程数、绑核处理及 predictor 创建处理\n- - `RecPredictor::Preprocess()` 方法用于模型的预处理\n- - `RecPredictor::Postprocess()` 方法用于模型的后处理\n-\n-4. `det_process.cc` 检测模型 DB 的预测文件\n- 该文件完成了检测模型 DB 的预处理、预测和后处理过程\n-\n- - `DetPredictor::DetPredictor()` 方法用于完成模型加载和线程数、绑核处理及 predictor 创建处理\n- - `DetPredictor::Preprocess()` 方法用于模型的预处理\n- - `DetPredictor::Postprocess()` 方法用于模型的后处理\n-\n-5. `db_post_process` 检测模型 DB 的后处理函数，包含 clipper 库的调用\n- 该文件完成了检测模型 DB 的第三方库调用和其他后处理方法实现\n-\n- - `std::vector<std::vector<std::vector<int>>> BoxesFromBitmap(...)` 方法从 Bitmap 图中获取检测框\n- - `std::vector<std::vector<std::vector<int>>> FilterTagDetRes(...)` 方法根据识别结果获取目标框位置\n-\n-## 进阶使用\n-\n-如果快速开始部分无法满足你的需求，可以参考本节对 demo 进行自定义修改。\n-\n-本节主要包含四部分： \n-\n-- 更新预测库；\n-- 转换 `.nb` 模型；\n-- 更新模型、标签文件和预测图片；\n-- 更新输入/输出预处理。\n-\n-### 更新预测库\n-\n-本指南所使用的预测库为最新版本（214rc），不推荐自行更新预测库。\n-\n-若有使用其他版本的需求，可参考如下步骤更新预测库：\n-\n-* Paddle Lite 项目：https://github.com/PaddlePaddle/Paddle-Lite\n- * 参考 [Paddle Lite 源码编译文档](https://www.paddlepaddle.org.cn/lite/develop/source_compile/compile_env.html)，编译 Android 预测库\n- * 编译最终产物位于 `build.lite.xxx.xxx.xxx` 下的 `inference_lite_lib.xxx.xxx`\n- * 替换 c++ 库\n- * 头文件\n- 将生成的 `build.lite.android.xxx.gcc/inference_lite_lib.android.xxx/cxx/include` 文件夹替换 demo 中的 `PaddleX-Lite-Deploy/libs/android/cxx/include`\n- * armeabi-v7a\n- 将生成的 `build.lite.android.armv7.gcc/inference_lite_lib.android.armv7/cxx/libs/libpaddle_lite_api_shared.so` 库替换 demo 中的 `PaddleX-Lite-Deploy/libs/android/cxx/libs/armeabi-v7a/libpaddle_lite_api_shared.so`\n- * arm64-v8a\n- 将生成的 `build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/cxx/libs/libpaddle_lite_api_shared.so` 库替换 demo 中的 `PaddleX-Lite-Deploy/libs/android/cxx/libs/arm64-v8a/libpaddle_lite_api_shared.so`\n-\n-### 转换 .nb 模型\n-\n-若想使用自己训练的模型，可先参考以下流程得到 `.nb` 模型。\n-\n-#### 终端命令方法（支持Mac/Ubuntu）\n-\n-1. 进入Paddle-Lite Github仓库的[release界面](https://github.com/PaddlePaddle/Paddle-Lite/releases)，选择所需版本下载对应的转化工具opt（推荐使用最新版本）。\n-\n-2. 下载 opt 工具后，执行以下命令（此处以 2.14rc 版本的 linux_x86 opt 工具转换 PP-OCRv5_mobile_det 模型为例）：\n-\n- ```bash\n- ./opt_linux_x86 \\\n- --model_file=PP-OCRv5_mobile_det/inference.pdmodel \\\n- --param_file=PP-OCRv5_mobile_det/inference.pdiparams \\\n- --optimize_out=PP-OCRv5_mobile_det \\\n- --valid_targets=arm\n- ```\n-\n-有关使用终端命令方法转换 `.nb` 模型的详细介绍，可参考 Paddle-Lite 仓库的[使用可执行文件 opt](https://www.paddlepaddle.org.cn/lite/v2.12/user_guides/opt/opt_bin.html)。\n-\n-#### python 脚本方法（支持Windows/Mac/Ubuntu）\n-\n-1. 安装最新版本的 paddlelite wheel 包。\n-\n- ```bash\n- pip install --pre paddlelite\n- ```\n-\n-2. 使用 python 脚本进行模型转换。以下为转换 PP-OCRv5_mobile_det 模型的示例代码：\n-\n- ```python\n- from paddlelite.lite import Opt\n-\n- # 1. 创建opt实例\n- opt = Opt()\n- # 2. 指定输入模型地址 \n- opt.set_model_file(\"./PP-OCRv5_mobile_det/inference.pdmodel\")\n- opt.set_param_file(\"./PP-OCRv5_mobile_det/inference.pdiparams\")\n- # 3. 指定转化类型\n- opt.set_valid_places(\"arm\")\n- # 4. 指定输出模型地址\n- opt.set_optimize_out(\"./PP-OCRv5_mobile_det\")\n- # 5. 执行模型优化\n- opt.run()\n- ```\n-\n-有关使用 python 脚本方法转换 `.nb` 模型的详细介绍，可参考 Paddle-Lite 仓库的[使用 Python 脚本 opt](https://www.paddlepaddle.org.cn/lite/v2.12/api_reference/python_api/opt.html)。\n-\n-**注意**\n-\n-- 有关模型优化工具 opt 的详细介绍，可参考 Paddle-Lite 仓库的[模型优化工具 opt](https://www.paddlepaddle.org.cn/lite/v2.12/user_guides/model_optimize_tool.html)\n-- 目前仅支持将 `.pdmodel` 格式的静态图模型转换为 `.nb` 格式。在 PaddlePaddle 3.0 以上的版本，默认导出为 `.json` 的模型，如果希望导出为 `.pdmodel` 格式的模型，导出时加上 `-o Global.export_with_pir=False` 即可。\n-\n-### 更新模型、标签文件和预测图片\n-\n-#### 更新模型\n-\n-本指南只对 `PP-OCRv3_mobile`、`PP-OCRv4_mobile`、`PP-OCRv5_mobile` 模型进行了验证，其他模型不保证适用性。\n-\n-如果你对 `PP-OCRv5_mobile` 模型进行了微调，并生成了一个名为 `PP-OCRv5_mobile_ft` 的新模型，可以按照以下步骤将原有模型替换为你的微调模型：\n-\n-1. 将 `PP-OCRv5_mobile_ft` 的 nb 模型存放到目录 `PaddleX-Lite-Deploy/ocr/assets/models/` 下，最终得到的文件结构如下：\n-\n- ```text\n- .\n- ├── ocr \n- │ ├── ...\n- │ └── assets \n- │ ├── models\n- │ │ ├── ...\n- │ │ ├── PP-OCRv5_mobile_ft_det.nb \n- │ │ └── PP-OCRv5_mobile_ft_rec.nb \n- │ └── ...\n- └── ...\n- ```\n-\n-2. 将模型名加入到 `run.sh` 脚本中的 `MODEL_LIST`。\n-\n- ```shell\n- MODEL_LIST=\"PP-OCRv3_mobile PP-OCRv4_mobile PP-OCRv5_mobile PP-OCRv5_mobile_ft\" # 模型之间以单个空格为间隔\n- ```\n-\n-3. 运行 `run.sh` 脚本时使用模型目录名。\n-\n- ```shell\n- sh run.sh PP-OCRv5_mobile_ft\n- ```\n-\n-**注意：**\n-\n-- 如果更新模型中的输入 Tensor、Shape、和 Dtype 发生更新:\n-\n- - 更新文字方向分类器模型，则需要更新 `ppocr_demo/src/cls_process.cc` 中 `ClsPredictor::Preprocss` 函数\n- - 更新检测模型，则需要更新 `ppocr_demo/src/det_process.cc` 中 `DetPredictor::Preprocss` 函数\n- - 更新识别器模型，则需要更新 `ppocr_demo/src/rec_process.cc` 中 `RecPredictor::Preprocss` 函数\n-\n-- 如果更新模型中的输出 Tensor 和 Dtype 发生更新:\n-\n- - 更新文字方向分类器模型，则需要更新 `ppocr_demo/src/cls_process.cc` 中 `ClsPredictor::Postprocss` 函数\n- - 更新检测模型，则需要更新 `ppocr_demo/src/det_process.cc` 中 `DetPredictor::Postprocss` 函数\n- - 更新识别器模型，则需要更新 `ppocr_demo/src/rec_process.cc` 中 `RecPredictor::Postprocss` 函数\n-\n-#### 更新标签文件\n-\n-如果需要更新标签文件，则需要将新的标签文件存放在目录 `PaddleX-Lite-Deploy/ocr/assets/labels/` 下，并参考模型更新方法更新 `PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo/run.sh` 中执行命令；\n-\n-以更新 `new_labels.txt` 为例：\n-\n- ```shell\n- # 代码文件 `PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo/run.sh`\n- # old\n- adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/test.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/${LABEL_FILE} \\\n- ./config.txt\"\n- # update\n- adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/test.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/new_labels.txt \\\n- ./config.txt\"\n- ```\n-\n-#### 更新预测图片\n-\n-如果需要更新预测图片，将更新的图片存放在 `PaddleX-Lite-Deploy/ocr/assets/images/` 下，更新文件 `PaddleX-Lite-Deploy/ocr/android/shell/ppocr_demo/rush.sh` 中执行命令；\n-\n-以更新 `new_pics.jpg` 为例：\n-\n- ```shell\n- # 代码文件 `PaddleX-Lite-Deploy/ocr/assets/images/run.sh`\n- ## old\n- adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"./models/${MODEL_NAME}_rec.nb\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/test.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/${LABEL_FILE} \\\n- ./config.txt\"\n- # update\n- adb shell \"cd ${ppocr_demo_path} \\\n- && chmod +x ./ppocr_demo \\\n- && export LD_LIBRARY_PATH=${ppocr_demo_path}:${LD_LIBRARY_PATH} \\\n- && ./ppocr_demo \\\n- \\\"./models/${MODEL_NAME}_det.nb\\\" \\\n- \\\"\\\"./models/${MODEL_NAME}_rec.nb\\\"\\\" \\\n- ./models/${CLS_MODEL_FILE} \\\n- ./images/new_pics.jpg \\\n- ./test_img_result.jpg \\\n- ./labels/${LABEL_FILE} \\\n- ./config.txt\"\n- ```\n-\n-### 更新输入/输出预处理\n-\n-- 更新输入预处理\n- - 更新文字方向分类器模型，则需要更新 `ppocr_demo/src/cls_process.cc` 中 `ClsPredictor::Preprocss` 函数\n- - 更新检测模型，则需要更新 `ppocr_demo/src/det_process.cc` 中 `DetPredictor::Preprocss` 函数\n- - 更新识别器模型，则需要更新 `ppocr_demo/src/rec_process.cc` 中 `RecPredictor::Preprocss` 函数\n-\n-- 更新输出预处理\n- - 更新文字方向分类器模型，则需要更新 `ppocr_demo/src/cls_process.cc` 中 `ClsPredictor::Postprocss` 函数\n- - 更新检测模型，则需要更新 `ppocr_demo/src/det_process.cc` 中 `DetPredictor::Postprocss` 函数\n- - 更新识别器模型，则需要更新 `ppocr_demo/src/rec_process.cc` 中 `RecPredictor::Postprocss` 函数\ndiff --git a/docs/version3.x/inference_deployment/others/benchmark.en.md b/docs/version3.x/inference_deployment/others/benchmark.en.md\nindex 7c9219347ca..6443f99c707 100644\n--- a/docs/version3.x/inference_deployment/others/benchmark.en.md\n+++ b/docs/version3.x/inference_deployment/others/benchmark.en.md\n@@ -60,6 +60,10 @@ The following table describes the methods and parameters related to pipeline inf\n Create a `test_infer.py` script:\n \n ```python\n+import os\n+\n+os.environ[\"PADDLE_PDX_PIPELINE_BENCHMARK\"] = \"True\"\n+\n from paddleocr import PaddleOCR, benchmark\n \n pipeline = PaddleOCR()\n@@ -80,7 +84,7 @@ benchmark.save_pipeline_data(\"./benchmark\") # Save benchmark data to the benchma\n Execute the script:\n \n ```bash\n-PADDLE_PDX_PIPELINE_BENCHMARK=True python test_infer.py\n+python test_infer.py\n ```\n \n The benchmark results obtained from running the example program are as follows:\ndiff --git a/docs/version3.x/inference_deployment/others/benchmark.md b/docs/version3.x/inference_deployment/others/benchmark.md\nindex 2dbdb966079..e9b4dfdbe9b 100644\n--- a/docs/version3.x/inference_deployment/others/benchmark.md\n+++ b/docs/version3.x/inference_deployment/others/benchmark.md\n@@ -60,6 +60,10 @@ Benchmark 功能会统计产线在端到端推理过程中，所有操作的平\n 创建 `test_infer.py` 脚本：\n \n ```python\n+import os\n+\n+os.environ[\"PADDLE_PDX_PIPELINE_BENCHMARK\"] = \"True\"\n+\n from paddleocr import PaddleOCR, benchmark\n \n pipeline = PaddleOCR()\n@@ -80,7 +84,7 @@ benchmark.save_pipeline_data(\"./benchmark\") # 将benchmark数据保存至benchma\n 执行脚本：\n \n ```bash\n-PADDLE_PDX_PIPELINE_BENCHMARK=True python test_infer.py\n+python test_infer.py\n ```\n \n 运行示例程序所得到的 benchmark 结果如下：\ndiff --git a/docs/version3.x/installation.en.md b/docs/version3.x/installation.en.md\nindex f270f151ab9..ce30445655a 100644\n--- a/docs/version3.x/installation.en.md\n+++ b/docs/version3.x/installation.en.md\n@@ -8,7 +8,7 @@ comments: true\n \n This section explains how to install, as needed, the `paddleocr` distribution package, optional dependency groups by capability domain, and the inference engine. This path covers running pretrained pipelines for inference locally, as well as auxiliary features such as document format conversion. **Model training and model export** are covered in Section 2 and are independent of the installation path above.\n \n-**Python version requirement**: `paddleocr` itself and the `doc2md` dependency group support Python 3.8 and later. The other optional dependency groups (`doc-parser`, `ie`, `trans`, `all`) require Python 3.9 or later due to upstream dependencies.\n+**Python version requirement**: `paddleocr` itself and the `doc2md` dependency group support Python 3.8 and later. The other optional dependency groups (`doc-parser`, `ie`, `trans`, `all`, etc.) require Python 3.9 or later due to upstream dependencies.\n \n ### 1.1 Install paddleocr\n \ndiff --git a/docs/version3.x/installation.md b/docs/version3.x/installation.md\nindex bd417034b5f..5508431ea6b 100644\n--- a/docs/version3.x/installation.md\n+++ b/docs/version3.x/installation.md\n@@ -8,7 +8,7 @@ comments: true\n \n 本节说明如何按需安装`paddleocr` 分发包，按能力域启用的可选依赖组，以及安装推理引擎。这条路径适用于在本地调用预训练产线完成推理，或是使用文档格式转换等辅助功能。**模型训练与模型导出**见第 2 节，与上述安装路径相互独立。\n \n-**Python 版本要求**：`paddleocr` 本体与 `doc2md` 依赖组支持 Python 3.8 及以上；其他可选依赖组（`doc-parser`、`ie`、`trans`、`all`）受上游依赖限制，需要 Python 3.9 及以上。\n+**Python 版本要求**：`paddleocr` 本体与 `doc2md` 依赖组支持 Python 3.8 及以上；其他可选依赖组（`doc-parser`、`ie`、`trans`、`all` 等）受上游依赖限制，需要 Python 3.9 及以上。\n \n ### 1.1 安装 `paddleocr` {#install-paddleocr}\n \ndiff --git a/docs/version3.x/integrations/mcp_server.en.md b/docs/version3.x/integrations/mcp_server.en.md\nindex e2ea5b7a5c9..3106dc20270 100644\n--- a/docs/version3.x/integrations/mcp_server.en.md\n+++ b/docs/version3.x/integrations/mcp_server.en.md\n@@ -19,8 +19,6 @@ PaddleOCR provides a lightweight [Model Context Protocol (MCP)](https://modelcon\n | `PP-StructureV3` | `pp_structurev3` | Identifies and extracts text blocks, titles, paragraphs, images, tables, and other layout elements from images or PDF files, converting the input into Markdown documents. |\n | `PaddleOCR-VL Series` | `paddleocr_vl` | Performs layout parsing with a VLM-based approach and converts the input into Markdown documents. Includes versions: PaddleOCR-VL, PaddleOCR-VL-1.5, and PaddleOCR-VL-1.6. |\n \n- > Each MCP server instance exposes exactly one MCP tool.\n-\n - **Supported Inference Methods**\n - **Local Inference**: Runs PaddleOCR pipelines directly on the local machine. This method has certain requirements for the local environment and hardware performance, and is suitable for offline use and scenarios with strict data privacy requirements.\n - **Official API**: Invokes the PaddleOCR Official API. This method is suitable for quickly trying out features, validating solutions, and other no-code development scenarios.\ndiff --git a/docs/version3.x/integrations/mcp_server.md b/docs/version3.x/integrations/mcp_server.md\nindex d02a59fd97d..d21bbca8880 100644\n--- a/docs/version3.x/integrations/mcp_server.md\n+++ b/docs/version3.x/integrations/mcp_server.md\n@@ -19,8 +19,6 @@ PaddleOCR 提供轻量级的 [Model Context Protocol（MCP）](https://modelcont\n | `PP-StructureV3` | `pp_structurev3` | 从图像或 PDF 文件中识别和提取文本块、标题、段落、图片、表格以及其他版面元素，将输入转换为 Markdown 文档。 |\n | `PaddleOCR-VL 系列` | `paddleocr_vl` | 使用基于多模态大模型的方案进行版面解析，将输入转换为 Markdown 文档。包含版本：PaddleOCR-VL、PaddleOCR-VL-1.5、PaddleOCR-VL-1.6。 |\n \n- > 每个 MCP 服务器实例对外只暴露一个 MCP 工具。\n-\n - **支持的推理方式**\n - **本地推理**：在本机直接运行 PaddleOCR 产线。此方式对本地环境与计算机性能有一定要求，适用于需要离线使用、对数据隐私有严格要求的场景。\n - **官方 API**：调用 PaddleOCR 官方提供的 API。此方式适合快速体验功能、快速验证方案等，也适用于零代码开发场景。\ndiff --git a/mkdocs.yml b/mkdocs.yml\nindex 3e9ae264e4c..2c39919a9ac 100644\n--- a/mkdocs.yml\n+++ b/mkdocs.yml\n@@ -138,8 +138,8 @@ plugins:\n 高性能推理: High-Performance Inference\n 打包 PaddleOCR 项目: Package PaddleOCR Projects\n 获取ONNX模型: Obtaining ONNX Models\n- 端侧部署: On-Device Deployment\n Android 部署: Android Deployment\n+ iOS 部署: iOS Deployment\n 服务化部署: Serving\n 浏览器端部署: Browser Deployment\n 模块列表: Module List\n@@ -195,7 +195,6 @@ plugins:\n 基于C++预测引擎推理: CPP Inference\n Visual Studio 2019 Community CMake 编译指南: Visual Studio 2019 Community CMake Compilation Guide\n 服务化部署: Server Deployment\n- 端侧部署: On-Device Deployment\n Android部署: Android Deployment\n Paddle2ONNX模型转化与预测: Paddle2ONNX\n 云上飞桨部署工具: Paddle Cloud\n@@ -362,8 +361,8 @@ nav:\n - Go SDK: version3.x/inference_deployment/serving/paddleocr_official_api/go.md\n - CLI: version3.x/inference_deployment/serving/paddleocr_official_api/cli.md\n - 跨端部署:\n- - 端侧部署: version3.x/inference_deployment/cross_platform/on_device_deployment.md\n - Android 部署: version3.x/inference_deployment/cross_platform/android_deployment.md\n+ - iOS 部署: version3.x/inference_deployment/cross_platform/ios_deployment.md\n - 浏览器端部署: version3.x/inference_deployment/cross_platform/browser.md\n - 其它:\n - 获取ONNX模型: version3.x/inference_deployment/others/obtaining_onnx_models.md\ndiff --git a/paddleocr/_common_args.py b/paddleocr/_common_args.py\nindex 46bbee015d2..2a99bb699eb 100644\n--- a/paddleocr/_common_args.py\n+++ b/paddleocr/_common_args.py\n@@ -31,6 +31,7 @@\n \"paddle_static\",\n \"paddle_dynamic\",\n \"transformers\",\n+ \"onnxruntime\",\n ]\n \n \n"},"changed_files":{"kind":"list like","value":["RELEASING.md","RELEASING_cn.md","deploy/ios_demo/.gitignore","deploy/ios_demo/NOTICE","deploy/ios_demo/PaddleOCRDemo.xcodeproj/project.pbxproj","deploy/ios_demo/PaddleOCRDemo.xcodeproj/xcshareddata/xcschemes/PaddleOCRDemo.xcscheme","deploy/ios_demo/PaddleOCRDemo.xcworkspace/contents.xcworkspacedata","deploy/ios_demo/PaddleOCRDemo/App/Bundle+BundledImage.swift","deploy/ios_demo/PaddleOCRDemo/App/ContentView.swift","deploy/ios_demo/PaddleOCRDemo/App/DemoAssets.swift","deploy/ios_demo/PaddleOCRDemo/App/DemoSectionChrome.swift","deploy/ios_demo/PaddleOCRDemo/App/PaddleOCRDemoApp.swift","deploy/ios_demo/PaddleOCRDemo/Engine/BoxSorter.swift","deploy/ios_demo/PaddleOCRDemo/Engine/CTCDecoder.swift","deploy/ios_demo/PaddleOCRDemo/Engine/DBPostProcess.swift","deploy/ios_demo/PaddleOCRDemo/Engine/DetectionEngine.swift","deploy/ios_demo/PaddleOCRDemo/Engine/EncodedImageCodec.swift","deploy/ios_demo/PaddleOCRDemo/Engine/InferenceConfig.swift","deploy/ios_demo/PaddleOCRDemo/Engine/ModelConfig.swift","deploy/ios_demo/PaddleOCRDemo/Engine/OCREngine.swift","deploy/ios_demo/PaddleOCRDemo/Engine/OCRParameterResolver.swift","deploy/ios_demo/PaddleOCRDemo/Engine/ORTProfilingBridge.h","deploy/ios_demo/PaddleOCRDemo/Engine/ORTProfilingBridge.mm","deploy/ios_demo/PaddleOCRDemo/Engine/ORTSessionManager.swift","deploy/ios_demo/PaddleOCRDemo/Engine/OpenCVDBBridge.h","deploy/ios_demo/PaddleOCRDemo/Engine/OpenCVDBBridge.mm","deploy/ios_demo/PaddleOCRDemo/Engine/OpenCVImageBridge.h","deploy/ios_demo/PaddleOCRDemo/Engine/OpenCVImageBridge.mm","deploy/ios_demo/PaddleOCRDemo/Engine/PDBPolygonOffsetBridge.h","deploy/ios_demo/PaddleOCRDemo/Engine/PDBPolygonOffsetBridge.mm","deploy/ios_demo/PaddleOCRDemo/Engine/Preprocessing.swift","deploy/ios_demo/PaddleOCRDemo/Engine/QuadTextCrop.swift","deploy/ios_demo/PaddleOCRDemo/Engine/RecPreprocessor.swift","deploy/ios_demo/PaddleOCRDemo/Engine/RecognitionEngine.swift","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AccentColor.colorset/Contents.json","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Contents.json","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-1024.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-20@2x.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-20@3x.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-29@2x.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-29@3x.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-40@2x.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-40@3x.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-60@2x.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-60@3x.png","deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/Contents.json","deploy/ios_demo/PaddleOCRDemo/Resources/SampleImages/general_ocr_002.jpg","deploy/ios_demo/PaddleOCRDemo/Supporting/Info.plist","deploy/ios_demo/PaddleOCRDemo/Supporting/PaddleOCRDemo-Bridging-Header.h","deploy/ios_demo/PaddleOCRDemo/ThirdParty/Clipper1/LICENSE","deploy/ios_demo/PaddleOCRDemo/ThirdParty/Clipper1/clipper.cpp","deploy/ios_demo/PaddleOCRDemo/ThirdParty/Clipper1/clipper.hpp","deploy/ios_demo/PaddleOCRDemo/ViewModels/OCRViewModel.swift","deploy/ios_demo/PaddleOCRDemo/Views/EngineSettingsSheet.swift","deploy/ios_demo/PaddleOCRDemo/Views/ErrorView.swift","deploy/ios_demo/PaddleOCRDemo/Views/ImagePickerSection.swift","deploy/ios_demo/PaddleOCRDemo/Views/OCRParametersPanel.swift","deploy/ios_demo/PaddleOCRDemo/Views/OnnxRuntimeSettingsPanel.swift","deploy/ios_demo/PaddleOCRDemo/Views/ResultImageView.swift","deploy/ios_demo/PaddleOCRDemo/Views/ResultsListView.swift","deploy/ios_demo/PaddleOCRDemo/Views/TimingView.swift","deploy/ios_demo/PaddleOCRDemoTests/BenchmarkFixtures.swift","deploy/ios_demo/PaddleOCRDemoTests/BenchmarkPayloads.swift","deploy/ios_demo/PaddleOCRDemoTests/Fixtures/.gitignore","deploy/ios_demo/PaddleOCRDemoTests/Fixtures/ios_ocr_validation_reference.jpg","deploy/ios_demo/PaddleOCRDemoTests/OCRBenchmarkTests.swift","deploy/ios_demo/PaddleOCRDemoTests/OCRParameterResolverTests.swift","deploy/ios_demo/PaddleOCRDemoTests/ORTSessionManagerTests.swift","deploy/ios_demo/Podfile","deploy/ios_demo/Podfile.lock","deploy/ios_demo/README.md","deploy/ios_demo/README_en.md","deploy/ios_demo/requirements-accuracy.txt","deploy/ios_demo/requirements-build-calib.txt","deploy/ios_demo/requirements-onnx-convert-ort.txt","deploy/ios_demo/requirements-onnx-quantize.txt","deploy/ios_demo/scripts/build_onnx_calib_npy.py","deploy/ios_demo/scripts/compare_ocr_json.py","deploy/ios_demo/scripts/convert_onnx_to_ort.sh","deploy/ios_demo/scripts/debug_onnx_qdq.py","deploy/ios_demo/scripts/extract_xcresult_attachments.py","deploy/ios_demo/scripts/extract_xctest_metrics.py","deploy/ios_demo/scripts/fetch_ios_demo_models.sh","deploy/ios_demo/scripts/generate_benchmark_report.py","deploy/ios_demo/scripts/ocr_reference_run.py","deploy/ios_demo/scripts/quantize_onnx_model.py","deploy/ios_demo/scripts/run_benchmark.sh","deploy/ios_demo/scripts/utils.py","docs/version3.x/inference_deployment/cross_platform/ios_deployment.en.md","docs/version3.x/inference_deployment/cross_platform/ios_deployment.md","docs/version3.x/inference_deployment/cross_platform/on_device_deployment.en.md","docs/version3.x/inference_deployment/cross_platform/on_device_deployment.md","docs/version3.x/inference_deployment/others/benchmark.en.md","docs/version3.x/inference_deployment/others/benchmark.md","docs/version3.x/installation.en.md","docs/version3.x/installation.md","docs/version3.x/integrations/mcp_server.en.md","docs/version3.x/integrations/mcp_server.md","mkdocs.yml","paddleocr/_common_args.py"],"string":"[\n \"RELEASING.md\",\n \"RELEASING_cn.md\",\n \"deploy/ios_demo/.gitignore\",\n \"deploy/ios_demo/NOTICE\",\n \"deploy/ios_demo/PaddleOCRDemo.xcodeproj/project.pbxproj\",\n \"deploy/ios_demo/PaddleOCRDemo.xcodeproj/xcshareddata/xcschemes/PaddleOCRDemo.xcscheme\",\n \"deploy/ios_demo/PaddleOCRDemo.xcworkspace/contents.xcworkspacedata\",\n \"deploy/ios_demo/PaddleOCRDemo/App/Bundle+BundledImage.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/App/ContentView.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/App/DemoAssets.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/App/DemoSectionChrome.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/App/PaddleOCRDemoApp.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/BoxSorter.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/CTCDecoder.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/DBPostProcess.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/DetectionEngine.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/EncodedImageCodec.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/InferenceConfig.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/ModelConfig.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/OCREngine.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/OCRParameterResolver.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/ORTProfilingBridge.h\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/ORTProfilingBridge.mm\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/ORTSessionManager.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/OpenCVDBBridge.h\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/OpenCVDBBridge.mm\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/OpenCVImageBridge.h\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/OpenCVImageBridge.mm\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/PDBPolygonOffsetBridge.h\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/PDBPolygonOffsetBridge.mm\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/Preprocessing.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/QuadTextCrop.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/RecPreprocessor.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Engine/RecognitionEngine.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AccentColor.colorset/Contents.json\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Contents.json\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-1024.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-20@2x.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-20@3x.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-29@2x.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-29@3x.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-40@2x.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-40@3x.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-60@2x.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/AppIcon.appiconset/Icon-60@3x.png\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/Assets.xcassets/Contents.json\",\n \"deploy/ios_demo/PaddleOCRDemo/Resources/SampleImages/general_ocr_002.jpg\",\n \"deploy/ios_demo/PaddleOCRDemo/Supporting/Info.plist\",\n \"deploy/ios_demo/PaddleOCRDemo/Supporting/PaddleOCRDemo-Bridging-Header.h\",\n \"deploy/ios_demo/PaddleOCRDemo/ThirdParty/Clipper1/LICENSE\",\n \"deploy/ios_demo/PaddleOCRDemo/ThirdParty/Clipper1/clipper.cpp\",\n \"deploy/ios_demo/PaddleOCRDemo/ThirdParty/Clipper1/clipper.hpp\",\n \"deploy/ios_demo/PaddleOCRDemo/ViewModels/OCRViewModel.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Views/EngineSettingsSheet.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Views/ErrorView.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Views/ImagePickerSection.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Views/OCRParametersPanel.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Views/OnnxRuntimeSettingsPanel.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Views/ResultImageView.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Views/ResultsListView.swift\",\n \"deploy/ios_demo/PaddleOCRDemo/Views/TimingView.swift\",\n \"deploy/ios_demo/PaddleOCRDemoTests/BenchmarkFixtures.swift\",\n \"deploy/ios_demo/PaddleOCRDemoTests/BenchmarkPayloads.swift\",\n \"deploy/ios_demo/PaddleOCRDemoTests/Fixtures/.gitignore\",\n \"deploy/ios_demo/PaddleOCRDemoTests/Fixtures/ios_ocr_validation_reference.jpg\",\n \"deploy/ios_demo/PaddleOCRDemoTests/OCRBenchmarkTests.swift\",\n \"deploy/ios_demo/PaddleOCRDemoTests/OCRParameterResolverTests.swift\",\n \"deploy/ios_demo/PaddleOCRDemoTests/ORTSessionManagerTests.swift\",\n \"deploy/ios_demo/Podfile\",\n \"deploy/ios_demo/Podfile.lock\",\n \"deploy/ios_demo/README.md\",\n \"deploy/ios_demo/README_en.md\",\n \"deploy/ios_demo/requirements-accuracy.txt\",\n \"deploy/ios_demo/requirements-build-calib.txt\",\n \"deploy/ios_demo/requirements-onnx-convert-ort.txt\",\n \"deploy/ios_demo/requirements-onnx-quantize.txt\",\n \"deploy/ios_demo/scripts/build_onnx_calib_npy.py\",\n \"deploy/ios_demo/scripts/compare_ocr_json.py\",\n \"deploy/ios_demo/scripts/convert_onnx_to_ort.sh\",\n \"deploy/ios_demo/scripts/debug_onnx_qdq.py\",\n \"deploy/ios_demo/scripts/extract_xcresult_attachments.py\",\n \"deploy/ios_demo/scripts/extract_xctest_metrics.py\",\n \"deploy/ios_demo/scripts/fetch_ios_demo_models.sh\",\n \"deploy/ios_demo/scripts/generate_benchmark_report.py\",\n \"deploy/ios_demo/scripts/ocr_reference_run.py\",\n \"deploy/ios_demo/scripts/quantize_onnx_model.py\",\n \"deploy/ios_demo/scripts/run_benchmark.sh\",\n \"deploy/ios_demo/scripts/utils.py\",\n \"docs/version3.x/inference_deployment/cross_platform/ios_deployment.en.md\",\n \"docs/version3.x/inference_deployment/cross_platform/ios_deployment.md\",\n \"docs/version3.x/inference_deployment/cross_platform/on_device_deployment.en.md\",\n \"docs/version3.x/inference_deployment/cross_platform/on_device_deployment.md\",\n \"docs/version3.x/inference_deployment/others/benchmark.en.md\",\n \"docs/version3.x/inference_deployment/others/benchmark.md\",\n \"docs/version3.x/installation.en.md\",\n \"docs/version3.x/installation.md\",\n \"docs/version3.x/integrations/mcp_server.en.md\",\n \"docs/version3.x/integrations/mcp_server.md\",\n \"mkdocs.yml\",\n \"paddleocr/_common_args.py\"\n]","isConversational":false},"review_threads":{"kind":"list like","value":[],"string":"[]","isConversational":false},"test_patch":{"kind":"string","value":"diff --git a/deploy/ios_demo/scripts/extract_xctest_metrics.py b/deploy/ios_demo/scripts/extract_xctest_metrics.py\nnew file mode 100755\nindex 00000000000..24b9e375513\n--- /dev/null\n+++ b/deploy/ios_demo/scripts/extract_xctest_metrics.py\n@@ -0,0 +1,129 @@\n+#!/usr/bin/env python3\n+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+#\n+# Licensed under the Apache License, Version 2.0 (the \"License\");\n+# you may not use this file except in compliance with the License.\n+# You may obtain a copy of the License at\n+#\n+# http://www.apache.org/licenses/LICENSE-2.0\n+#\n+# Unless required by applicable law or agreed to in writing, software\n+# distributed under the License is distributed on an \"AS IS\" BASIS,\n+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+# See the License for the specific language governing permissions and\n+# limitations under the License.\n+\n+\"\"\"Export XCTest performance metrics from an xcresult bundle.\"\"\"\n+\n+from __future__ import annotations\n+\n+import argparse\n+import json\n+import subprocess\n+import sys\n+from pathlib import Path\n+from typing import Any, Dict, List, Optional, Sequence\n+\n+\n+def _die(message: str, code: int = 1) -> int:\n+ sys.stderr.write(f\"[extract_xctest_metrics] {message}\\n\")\n+ return code\n+\n+\n+def _metric_summary(values: List[float]) -> Dict[str, float]:\n+ if not values:\n+ return {}\n+ return {\n+ \"count\": len(values),\n+ \"min\": min(values),\n+ \"max\": max(values),\n+ \"mean\": sum(values) / len(values),\n+ }\n+\n+\n+def export_metrics(result: Path) -> Dict[str, Any]:\n+ proc = subprocess.run(\n+ [\n+ \"xcrun\",\n+ \"xcresulttool\",\n+ \"get\",\n+ \"test-results\",\n+ \"metrics\",\n+ \"--path\",\n+ str(result),\n+ \"--compact\",\n+ ],\n+ text=True,\n+ capture_output=True,\n+ check=False,\n+ )\n+ if proc.returncode != 0:\n+ raise RuntimeError(proc.stderr.strip() or \"xcresulttool metrics export failed\")\n+ raw = json.loads(proc.stdout or \"[]\")\n+ tests = raw if isinstance(raw, list) else []\n+ memory_metrics: List[Dict[str, Any]] = []\n+ for test in tests:\n+ if not isinstance(test, dict):\n+ continue\n+ for run in test.get(\"testRuns\") or []:\n+ if not isinstance(run, dict):\n+ continue\n+ device = run.get(\"device\") if isinstance(run.get(\"device\"), dict) else {}\n+ config = (\n+ run.get(\"testPlanConfiguration\")\n+ if isinstance(run.get(\"testPlanConfiguration\"), dict)\n+ else {}\n+ )\n+ for metric in run.get(\"metrics\") or []:\n+ if not isinstance(metric, dict):\n+ continue\n+ display_name = str(metric.get(\"displayName\") or \"\")\n+ identifier = str(metric.get(\"identifier\") or \"\")\n+ unit = str(metric.get(\"unitOfMeasurement\") or \"\")\n+ haystack = f\"{display_name} {identifier} {unit}\".lower()\n+ if \"memory\" not in haystack:\n+ continue\n+ values = [\n+ float(v)\n+ for v in metric.get(\"measurements\") or []\n+ if isinstance(v, (int, float))\n+ ]\n+ memory_metrics.append(\n+ {\n+ \"testIdentifier\": test.get(\"testIdentifier\"),\n+ \"testIdentifierURL\": test.get(\"testIdentifierURL\"),\n+ \"device\": device,\n+ \"configuration\": config,\n+ \"displayName\": display_name,\n+ \"identifier\": identifier or None,\n+ \"unitOfMeasurement\": unit,\n+ \"measurements\": values,\n+ \"summary\": _metric_summary(values),\n+ }\n+ )\n+ return {\n+ \"schemaVersion\": 1,\n+ \"source\": \"xcresulttool get test-results metrics\",\n+ \"metrics\": memory_metrics,\n+ }\n+\n+\n+def main(argv: Optional[Sequence[str]] = None) -> int:\n+ parser = argparse.ArgumentParser(description=__doc__)\n+ parser.add_argument(\"--result\", type=Path, required=True)\n+ parser.add_argument(\"--output\", type=Path, required=True)\n+ args = parser.parse_args(list(argv) if argv is not None else None)\n+ try:\n+ doc = export_metrics(args.result)\n+ except (OSError, RuntimeError, json.JSONDecodeError) as exc:\n+ return _die(str(exc))\n+ args.output.parent.mkdir(parents=True, exist_ok=True)\n+ args.output.write_text(\n+ json.dumps(doc, indent=2, ensure_ascii=False), encoding=\"utf-8\"\n+ )\n+ print(f\"Wrote {args.output}\")\n+ return 0\n+\n+\n+if __name__ == \"__main__\":\n+ raise SystemExit(main())\n"},"merged":{"kind":"bool","value":true,"string":"true"}}},{"rowIdx":8997,"cells":{"repo":{"kind":"string","value":"PaddlePaddle/PaddleOCR"},"pr_number":{"kind":"number","value":18133,"string":"18,133"},"task_type":{"kind":"string","value":"issue_to_patch"},"issue_text":{"kind":"string","value":""},"pr_title":{"kind":"string","value":"add onnxruntime installation desc"},"pr_body":{"kind":"string","value":""},"base_sha":{"kind":"string","value":"71c7fefbf5e8478bc7f14b2d9b095bb80580178b"},"head_sha":{"kind":"string","value":"062b5f9b208d471bf5e9850c461b01fa5e9f2442"},"gold_diff":{"kind":"string","value":"diff --git a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md b/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md\nindex 45c741a1a4d..cfd0821b5b8 100644\n--- a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md\n+++ b/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md\n@@ -55,6 +55,8 @@ When using ONNX Runtime as the inference engine, install the ONNX Runtime packag\n python -m pip install onnxruntime-gpu\n ```\n \n+The command above applies to NVIDIA GPUs in CUDA 12.x environments. For installation instructions for other devices or environments, please refer to the [ONNX Runtime official documentation](https://onnxruntime.ai/docs/install/).\n+\n ## 4. Configuration and Supported Values of `engine` and `engine_config`\n \n ### 4.1 `engine`\ndiff --git a/docs/version3.x/inference_deployment/local_inference/inference_engine.md b/docs/version3.x/inference_deployment/local_inference/inference_engine.md\nindex dcbebe02a08..5d03cc865aa 100644\n--- a/docs/version3.x/inference_deployment/local_inference/inference_engine.md\n+++ b/docs/version3.x/inference_deployment/local_inference/inference_engine.md\n@@ -55,6 +55,8 @@ python -m pip install \"transformers>=5.10.0\"\n python -m pip install onnxruntime-gpu\n ```\n \n+以上命令适用于 CUDA 12.x 环境下的 NVIDIA GPU，其他设备或环境的安装方式请参考 [ONNX Runtime 官方文档](https://onnxruntime.ai/docs/install/)。\n+\n ## 4. `engine` 和 `engine_config` 的设置与取值\n \n ### 4.1 `engine`\n"},"changed_files":{"kind":"list like","value":["docs/version3.x/inference_deployment/local_inference/inference_engine.en.md","docs/version3.x/inference_deployment/local_inference/inference_engine.md"],"string":"[\n \"docs/version3.x/inference_deployment/local_inference/inference_engine.en.md\",\n \"docs/version3.x/inference_deployment/local_inference/inference_engine.md\"\n]","isConversational":false},"review_threads":{"kind":"list like","value":[],"string":"[]","isConversational":false},"test_patch":{"kind":"string","value":""},"merged":{"kind":"bool","value":true,"string":"true"}}},{"rowIdx":8998,"cells":{"repo":{"kind":"string","value":"PaddlePaddle/PaddleOCR"},"pr_number":{"kind":"number","value":18124,"string":"18,124"},"task_type":{"kind":"string","value":"issue_to_patch"},"issue_text":{"kind":"string","value":""},"pr_title":{"kind":"string","value":"add speed data"},"pr_body":{"kind":"string","value":""},"base_sha":{"kind":"string","value":"6d2e6fb9c7df029f9f5d626952f179d4c9e49344"},"head_sha":{"kind":"string","value":"e02cc271977066ac7974fd2f6ae8e567b9559204"},"gold_diff":{"kind":"string","value":"diff --git a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md b/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md\nindex 9501d5ed48b..45c741a1a4d 100644\n--- a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md\n+++ b/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md\n@@ -21,11 +21,13 @@ In PaddleOCR, an inference engine refers to the underlying runtime used to execu\n | - | - | - |\n | PaddlePaddle framework | `paddle`, `paddle_static`, `paddle_dynamic` | Runs on the PaddlePaddle framework. |\n | Transformers | `transformers` | Runs on Hugging Face Transformers. |\n+| ONNX Runtime | `onnxruntime` | Runs on ONNX Runtime. |\n \n - `paddle`: The unified entry point for the PaddlePaddle framework. It selects `paddle_static` or `paddle_dynamic` according to the model type and files in the model directory. If both are available, `paddle_static` is preferred.\n - `paddle_static`: PaddlePaddle static-graph inference, suitable for scenarios that require better inference performance or more fine-grained performance tuning.\n - `paddle_dynamic`: PaddlePaddle dynamic-graph inference, which is more flexible and easier to debug compared to static graph.\n - `transformers`: Hugging Face Transformers inference, making it convenient to integrate with the Hugging Face ecosystem.\n+- `onnxruntime`: ONNX Runtime inference, used to load and execute ONNX-format models.\n \n ## 3. Installation by Inference Engine\n \n@@ -37,14 +39,22 @@ When using the PaddlePaddle framework for inference, install PaddlePaddle first.\n \n ### 3.2 Transformers\n \n-When using Transformers as the inference engine, you need to install Hugging Face Transformers (`>=5.8.0`). Example command:\n+When using Transformers as the inference engine, you need to install Hugging Face Transformers (`>=5.10.0`). Example command:\n \n ```bash\n-python -m pip install \"transformers>=5.8.0\"\n+python -m pip install \"transformers>=5.10.0\"\n ```\n \n In many cases, you also need to install the underlying inference framework. For details, see the [Transformers official documentation](https://huggingface.co/docs/transformers/installation).\n \n+### 3.3 ONNX Runtime\n+\n+When using ONNX Runtime as the inference engine, install the ONNX Runtime package. Example command:\n+\n+```bash\n+python -m pip install onnxruntime-gpu\n+```\n+\n ## 4. Configuration and Supported Values of `engine` and `engine_config`\n \n ### 4.1 `engine`\n@@ -58,6 +68,7 @@ In many cases, you also need to install the underlying inference framework. For\n | `paddle_static` | Static-graph inference | Uses Paddle static-graph inference. |\n | `paddle_dynamic` | Dynamic-graph inference | Uses Paddle dynamic-graph inference. |\n | `transformers` | Transformers inference | Uses Hugging Face Transformers inference. |\n+| `onnxruntime` | ONNX Runtime inference | Uses ONNX Runtime inference. |\n \n ### 4.2 `engine_config`\n \n@@ -102,12 +113,18 @@ Common fields include:\n - `processor_kwargs`: extra arguments passed to the processor / image processor loading API;\n - `tokenizer_kwargs`: a compatibility-preserved field that is merged with `processor_kwargs`.\n \n+#### `onnxruntime`\n+\n+Common fields include:\n+\n+- `device_type` / `device_id`: inference device type and device index.\n+\n #### 4.2.1 Flat vs. bucketed `engine_config`\n \n At the same level `engine_config` may be:\n \n - **Flat**: a dict whose keys are only those required by the **resolved** engine (for example, when using static graph only, top-level keys such as `run_mode` and `cpu_threads`).\n-- **Bucketed**: top-level keys are **only** registered engine names (e.g. `paddle_static`, `paddle_dynamic`, `transformers`), each mapping to a nested dict. You **must not** mix bucket keys with flat keys at the same level (e.g. `{\"paddle_static\": {...}, \"run_mode\": \"paddle\"}` is invalid).\n+- **Bucketed**: top-level keys are **only** registered engine names (e.g. `paddle_static`, `paddle_dynamic`, `transformers`, `onnxruntime`), each mapping to a nested dict. You **must not** mix bucket keys with flat keys at the same level (e.g. `{\"paddle_static\": {...}, \"run_mode\": \"paddle\"}` is invalid).\n \n When an engine is resolved, only the corresponding config is used: flat configs are validated as a whole; bucketed configs take the entry for that engine.\n \n@@ -128,6 +145,7 @@ When an engine is resolved, only the corresponding config is used: flat configs\n \n ```bash\n paddleocr text_detection -i general_ocr_001.png --engine transformers\n+paddleocr text_detection -i general_ocr_001.png --engine onnxruntime\n ```\n \n ### 5.2 Individual model (Python): explicitly specify `transformers`\n@@ -143,7 +161,20 @@ model = TextDetection(\n result = model.predict(\"general_ocr_001.png\")\n ```\n \n-### 5.3 Individual model (Python): specify `paddle_static` and `engine_config`\n+### 5.3 Individual model (Python): explicitly specify `onnxruntime`\n+\n+```python\n+from paddleocr import TextDetection\n+\n+model = TextDetection(\n+ model_name=\"PP-OCRv5_server_det\",\n+ engine=\"onnxruntime\",\n+)\n+\n+result = model.predict(\"general_ocr_001.png\")\n+```\n+\n+### 5.4 Individual model (Python): specify `paddle_static` and `engine_config`\n \n ```python\n from paddleocr import TextDetection\n@@ -161,13 +192,13 @@ model = TextDetection(\n result = model.predict(\"general_ocr_001.png\")\n ```\n \n-### 5.4 Pipeline (CLI): select the engine with `--engine`\n+### 5.5 Pipeline (CLI): select the engine with `--engine`\n \n ```bash\n paddleocr ocr -i general_ocr_001.png --engine paddle_static\n ```\n \n-### 5.5 Pipeline (Python API): configure the inference engine for a specific module\n+### 5.6 Pipeline (Python API): configure the inference engine for a specific module\n \n If you want to specify `engine` and `engine_config` for a specific module inside a pipeline, you can first export the configuration file, modify the corresponding module configuration, and then load it. For how to export, edit, and load the configuration file, see [Using PaddleX Pipeline Configuration Files](../../paddleocr_and_paddlex.en.md#3-using-paddlex-pipeline-configuration-files). Example:\n \ndiff --git a/docs/version3.x/inference_deployment/local_inference/inference_engine.md b/docs/version3.x/inference_deployment/local_inference/inference_engine.md\nindex 35cfedce7d7..dcbebe02a08 100644\n--- a/docs/version3.x/inference_deployment/local_inference/inference_engine.md\n+++ b/docs/version3.x/inference_deployment/local_inference/inference_engine.md\n@@ -21,11 +21,13 @@ PaddleOCR 3.5 引入了统一的推理引擎配置方式：使用 `engine` 选\n | - | - | - |\n | 飞桨框架 | `paddle`、`paddle_static`、`paddle_dynamic` | 基于飞桨框架运行。 |\n | Transformers | `transformers` | 基于 Hugging Face Transformers 运行。 |\n+| ONNX Runtime | `onnxruntime` | 基于 ONNX Runtime 运行。 |\n \n - `paddle`：飞桨框架统一入口。根据模型类型和模型目录中的文件选择 `paddle_static` 或 `paddle_dynamic`，在二者都可用的情况下偏好 `paddle_static`。\n - `paddle_static`：飞桨静态图推理，适合对推理性能有一定要求或者需要进行精细化推理性能调优的场景。\n - `paddle_dynamic`：飞桨动态图推理，相比静态图更加灵活、易于调试。\n - `transformers`：Hugging Face Transformers 推理，便于与 Hugging Face 生态集成。\n+- `onnxruntime`：ONNX Runtime 推理，用于加载并执行 ONNX 格式模型。\n \n ## 3. 各推理引擎安装方式\n \n@@ -37,14 +39,22 @@ PaddleOCR 3.5 引入了统一的推理引擎配置方式：使用 `engine` 选\n \n ### 3.2 Transformers\n \n-当您使用 Transformers 作为推理引擎时，需要安装 Hugging Face Transformers（`>=5.8.0`）。示例命令如下：\n+当您使用 Transformers 作为推理引擎时，需要安装 Hugging Face Transformers（`>=5.10.0`）。示例命令如下：\n \n ```bash\n-python -m pip install \"transformers>=5.8.0\"\n+python -m pip install \"transformers>=5.10.0\"\n ```\n \n 通常，您还需要安装底层推理框架，详情可参考 [Transformers 官方文档](https://huggingface.co/docs/transformers/installation)。\n \n+### 3.3 ONNX Runtime\n+\n+当您使用 ONNX Runtime 作为推理引擎时，需要安装 ONNX Runtime。示例命令如下：\n+\n+```bash\n+python -m pip install onnxruntime-gpu\n+```\n+\n ## 4. `engine` 和 `engine_config` 的设置与取值\n \n ### 4.1 `engine`\n@@ -58,6 +68,7 @@ python -m pip install \"transformers>=5.8.0\"\n | `paddle_static` | 静态图推理 | 使用飞桨静态图推理。 |\n | `paddle_dynamic` | 飞桨动态图推理 | 使用飞桨动态图推理。 |\n | `transformers` | Transformers 推理 | 使用 Hugging Face Transformers 推理。 |\n+| `onnxruntime` | ONNX Runtime 推理 | 使用 ONNX Runtime 推理。 |\n \n ### 4.2 `engine_config`\n \n@@ -102,12 +113,18 @@ python -m pip install \"transformers>=5.8.0\"\n - `processor_kwargs`：传给 processor / image processor 加载接口的额外参数；\n - `tokenizer_kwargs`：兼容保留字段，会与 `processor_kwargs` 合并使用。\n \n+#### `onnxruntime`\n+\n+常见字段包括：\n+\n+- `device_type` / `device_id`：推理设备类型和设备编号。\n+\n #### 4.2.1 扁平与分桶 `engine_config`\n \n 同一层级的 `engine_config` 可以是：\n \n - **扁平**：只包含**当前解析得到的引擎**所需的字段（例如仅使用静态图时，顶层直接是 `run_mode`、`cpu_threads` 等）。\n-- **分桶**：顶层键**仅**为 PaddleX 已注册的引擎名（如 `paddle_static`、`paddle_dynamic`、`transformers` 等），每个键对应一个嵌套字典。**不得**在同一层级混用「分桶键」与扁平字段（例如 `{\"paddle_static\": {...}, \"run_mode\": \"paddle\"}` 会报错）。\n+- **分桶**：顶层键**仅**为 PaddleX 已注册的引擎名（如 `paddle_static`、`paddle_dynamic`、`transformers`、`onnxruntime` 等），每个键对应一个嵌套字典。**不得**在同一层级混用「分桶键」与扁平字段（例如 `{\"paddle_static\": {...}, \"run_mode\": \"paddle\"}` 会报错）。\n \n 解析为某一引擎时，只会使用与该引擎对应的一份配置：扁平形式直接参与校验；分桶形式则取出对应键下的字典。\n \n@@ -128,6 +145,7 @@ python -m pip install \"transformers>=5.8.0\"\n \n ```bash\n paddleocr text_detection -i general_ocr_001.png --engine transformers\n+paddleocr text_detection -i general_ocr_001.png --engine onnxruntime\n ```\n \n ### 5.2 单模型（Python）：显式指定 `transformers`\n@@ -143,7 +161,20 @@ model = TextDetection(\n result = model.predict(\"general_ocr_001.png\")\n ```\n \n-### 5.3 单模型（Python）：指定 `paddle_static` 与 `engine_config`\n+### 5.3 单模型（Python）：显式指定 `onnxruntime`\n+\n+```python\n+from paddleocr import TextDetection\n+\n+model = TextDetection(\n+ model_name=\"PP-OCRv5_server_det\",\n+ engine=\"onnxruntime\",\n+)\n+\n+result = model.predict(\"general_ocr_001.png\")\n+```\n+\n+### 5.4 单模型（Python）：指定 `paddle_static` 与 `engine_config`\n \n ```python\n from paddleocr import TextDetection\n@@ -161,13 +192,13 @@ model = TextDetection(\n result = model.predict(\"general_ocr_001.png\")\n ```\n \n-### 5.4 产线（CLI）：通过 `--engine` 选择引擎\n+### 5.5 产线（CLI）：通过 `--engine` 选择引擎\n \n ```bash\n paddleocr ocr -i general_ocr_001.png --engine paddle_static\n ```\n \n-### 5.5 产线（Python API）：为某个模块单独配置推理引擎\n+### 5.6 产线（Python API）：为某个模块单独配置推理引擎\n \n 如需为产线中的某一个模块单独指定 `engine`、`engine_config`，可先导出配置文件，修改对应模块配置后，再通过加载配置文件。配置文件的导出、编辑与加载方式可参见 [使用 PaddleX 产线配置文件](../../paddleocr_and_paddlex.md#3-paddlex)。示例如下：\n \ndiff --git a/docs/version3.x/module_usage/chart_parsing.en.md b/docs/version3.x/module_usage/chart_parsing.en.md\nindex 655cc713f70..76412e8dade 100644\n--- a/docs/version3.x/module_usage/chart_parsing.en.md\n+++ b/docs/version3.x/module_usage/chart_parsing.en.md\n@@ -52,6 +52,14 @@ paddleocr chart_parsing -i \"{'image': 'https://paddle-model-ecology.bj.bcebos.co\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr chart_parsing -i \"{'image': 'https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/chart_parsing_02.png'}\" \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_dynamic` inference engine delivers better inference performance and is the recommended first choice.\n \n **Note:** By default, PaddleOCR retrieves models from HuggingFace. If HuggingFace access is restricted in your environment, you can switch the model source to BOS by setting the environment variable: `PADDLE_PDX_MODEL_SOURCE=\"BOS\"`. Support for more mainstream sources is planned.\n@@ -89,6 +97,23 @@ for res in results:\n res.save_to_json(f\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import ChartParsing\n+model = ChartParsing(\n+ model_name=\"PP-Chart2Table\",\n+ engine=\"onnxruntime\",\n+)\n+results = model.predict(\n+ input={\"image\": \"chart_parsing_02.png\"},\n+ batch_size=1\n+)\n+for res in results:\n+ res.print()\n+ res.save_to_json(f\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_dynamic` inference engine delivers better inference performance and is the recommended first choice.\n \n The output result will be:\n@@ -155,7 +180,7 @@ Defaults to GPU 0 if available; otherwise falls back to CPU.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_dynamic</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_dynamic</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -315,7 +340,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/medal_table.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/medal_table.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -325,7 +350,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\ndiff --git a/docs/version3.x/module_usage/chart_parsing.md b/docs/version3.x/module_usage/chart_parsing.md\nindex 41b8d82c786..689ebe8dfc3 100644\n--- a/docs/version3.x/module_usage/chart_parsing.md\n+++ b/docs/version3.x/module_usage/chart_parsing.md\n@@ -52,6 +52,14 @@ paddleocr chart_parsing -i \"{'image': 'https://paddle-model-ecology.bj.bcebos.co\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr chart_parsing -i \"{'image': 'https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/chart_parsing_02.png'}\" \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_dynamic` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -89,6 +97,23 @@ for res in results:\n res.save_to_json(f\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import ChartParsing\n+model = ChartParsing(\n+ model_name=\"PP-Chart2Table\",\n+ engine=\"onnxruntime\",\n+)\n+results = model.predict(\n+ input={\"image\": \"chart_parsing_02.png\"},\n+ batch_size=1\n+)\n+for res in results:\n+ res.print()\n+ res.save_to_json(f\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_dynamic` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 运行后，得到的结果为：\n@@ -153,7 +178,7 @@ for res in results:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_dynamic</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_dynamic</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -311,7 +336,7 @@ for res in results:\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/medal_table.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/medal_table.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -321,7 +346,7 @@ for res in results:\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\ndiff --git a/docs/version3.x/module_usage/doc_img_orientation_classification.en.md b/docs/version3.x/module_usage/doc_img_orientation_classification.en.md\nindex 5f49463d771..b45b48f5c38 100644\n--- a/docs/version3.x/module_usage/doc_img_orientation_classification.en.md\n+++ b/docs/version3.x/module_usage/doc_img_orientation_classification.en.md\n@@ -104,6 +104,14 @@ paddleocr doc_img_orientation_classification -i https://paddle-model-ecology.bj.\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr doc_img_orientation_classification -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/img_rot180_demo.jpg \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable `PADDLE_PDX_MODEL_SOURCE=\"BOS\"` to change the model source to BOS. In the future, more model sources will be supported.\n@@ -139,6 +147,22 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import DocImgOrientationClassification\n+\n+model = DocImgOrientationClassification(\n+ model_name=\"PP-LCNet_x1_0_doc_ori\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"img_rot180_demo.jpg\", batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_img(\"./output/demo.png\")\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#52-weight-conversion) section in the [Inference Engine](#5-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -201,7 +225,7 @@ By default, GPU 0 is used if available; otherwise, CPU is used.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -407,7 +431,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-LCNet_x1_0_doc_ori</td>\n+ <td rowspan=\"4\">PP-LCNet_x1_0_doc_ori</td>\n <td>paddle_static</td>\n <td>2.21</td>\n <td>3.36</td>\n@@ -428,12 +452,19 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>0.14</td>\n <td>8.36</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>2.02</td>\n+ <td>0.87</td>\n+ <td>0.05</td>\n+ <td>3.03</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/img_rot180_demo.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/img_rot180_demo.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -443,13 +474,13 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Text Image Orientation Classification Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/doc_img_orientation_classification.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Text Image Orientation Classification Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/doc_img_orientation_classification.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\n \n ## 6. FAQ\ndiff --git a/docs/version3.x/module_usage/doc_img_orientation_classification.md b/docs/version3.x/module_usage/doc_img_orientation_classification.md\nindex 22f90700162..fe4b5ed4e60 100644\n--- a/docs/version3.x/module_usage/doc_img_orientation_classification.md\n+++ b/docs/version3.x/module_usage/doc_img_orientation_classification.md\n@@ -106,6 +106,14 @@ paddleocr doc_img_orientation_classification -i https://paddle-model-ecology.bj.\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr doc_img_orientation_classification -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/img_rot180_demo.jpg \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 如需了解推理引擎的介绍、速度数据及权重转换等详细信息，请参考后文 [推理引擎](#五推理引擎) 章节。\n@@ -134,6 +142,17 @@ model = DocImgOrientationClassification(\n )\n ```\n \n+如果希望使用 `onnxruntime` 引擎进行推理，可以参考如下示例：\n+\n+```python\n+from paddleocr import DocImgOrientationClassification\n+\n+model = DocImgOrientationClassification(\n+ model_name=\"PP-LCNet_x1_0_doc_ori\",\n+ engine=\"onnxruntime\",\n+)\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 随后继续按上文方式调用 `predict()` 即可。\n@@ -197,7 +216,7 @@ model = DocImgOrientationClassification(\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -403,7 +422,7 @@ model = DocImgOrientationClassification(\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-LCNet_x1_0_doc_ori</td>\n+ <td rowspan=\"4\">PP-LCNet_x1_0_doc_ori</td>\n <td>paddle_static</td>\n <td>2.21</td>\n <td>3.36</td>\n@@ -424,12 +443,19 @@ model = DocImgOrientationClassification(\n <td>0.14</td>\n <td>8.36</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>2.02</td>\n+ <td>0.87</td>\n+ <td>0.05</td>\n+ <td>3.03</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/img_rot180_demo.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/img_rot180_demo.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -439,13 +465,13 @@ model = DocImgOrientationClassification(\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 权重转换 {#52-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 文本图像方向分类模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/doc_img_orientation_classification.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 文本图像方向分类模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/doc_img_orientation_classification.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n \n ## 六、FAQ\ndiff --git a/docs/version3.x/module_usage/layout_analysis.en.md b/docs/version3.x/module_usage/layout_analysis.en.md\nindex a9dbfbe275e..dc332232aa0 100644\n--- a/docs/version3.x/module_usage/layout_analysis.en.md\n+++ b/docs/version3.x/module_usage/layout_analysis.en.md\n@@ -109,6 +109,13 @@ paddleocr layout_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+paddleocr layout_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg --model_name PP-DocLayoutV3 \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable <code>PADDLE_PDX_MODEL_SOURCE=\"BOS\"</code> to change the model source to BOS. In the future, more model sources will be supported.\n@@ -144,6 +151,22 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import LayoutDetection\n+\n+model = LayoutDetection(\n+ model_name=\"PP-DocLayoutV2\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"layout.jpg\", batch_size=1, layout_nms=True)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#42-weight-conversion) section in the [Inference Engine](#4-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -209,7 +232,7 @@ By default, GPU 0 is used if available; otherwise, CPU is used.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -513,7 +536,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-DocLayoutV3</td>\n+ <td rowspan=\"4\">PP-DocLayoutV3</td>\n <td>paddle_static</td>\n <td>10.95</td>\n <td>47.99</td>\n@@ -535,7 +558,14 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>78.97</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-DocLayoutV2</td>\n+ <td>onnxruntime</td>\n+ <td>9.26</td>\n+ <td>39.11</td>\n+ <td>10.23</td>\n+ <td>58.80</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-DocLayoutV2</td>\n <td>paddle_static</td>\n <td>10.48</td>\n <td>30.94</td>\n@@ -556,12 +586,19 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>2.43</td>\n <td>69.30</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>9.25</td>\n+ <td>15.13</td>\n+ <td>1.25</td>\n+ <td>25.82</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -571,11 +608,11 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 4.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Layout Analysis Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/layout_analysis.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Layout Analysis Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/layout_analysis.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\ndiff --git a/docs/version3.x/module_usage/layout_analysis.md b/docs/version3.x/module_usage/layout_analysis.md\nindex 2d319fe3582..75b37a947e5 100644\n--- a/docs/version3.x/module_usage/layout_analysis.md\n+++ b/docs/version3.x/module_usage/layout_analysis.md\n@@ -68,6 +68,13 @@ paddleocr layout_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+paddleocr layout_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg --model_name PP-DocLayoutV3 \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 您可以将版面区域检测模块中的模型推理集成到您的项目中。运行以下代码前，请您下载[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg)到本地。\n@@ -101,6 +108,22 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import LayoutDetection\n+\n+model = LayoutDetection(\n+ model_name=\"PP-DocLayoutV3\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"layout.jpg\", batch_size=1, layout_nms=True)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 训练后的模型如果想使用 `paddle_dynamic` 或 `transformers` 引擎，请参考后文 [推理引擎](#四推理引擎) 中的 [权重转换](#42-权重转换) 部分将模型由 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式。\n@@ -165,7 +188,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -462,7 +485,7 @@ for res in output:\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-DocLayoutV3</td>\n+ <td rowspan=\"4\">PP-DocLayoutV3</td>\n <td>paddle_static</td>\n <td>10.95</td>\n <td>47.99</td>\n@@ -484,7 +507,14 @@ for res in output:\n <td>78.97</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-DocLayoutV2</td>\n+ <td>onnxruntime</td>\n+ <td>9.26</td>\n+ <td>39.11</td>\n+ <td>10.23</td>\n+ <td>58.80</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-DocLayoutV2</td>\n <td>paddle_static</td>\n <td>10.48</td>\n <td>30.94</td>\n@@ -505,12 +535,19 @@ for res in output:\n <td>2.43</td>\n <td>69.30</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>9.25</td>\n+ <td>15.13</td>\n+ <td>1.25</td>\n+ <td>25.82</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -520,11 +557,11 @@ for res in output:\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 4.2 权重转换 {#42-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 版面分析模块权重转换](https://paddlepaddle.github.io/PaddleX/main/module_usage/tutorials/ocr_modules/layout_analysis.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 版面分析模块权重转换](https://paddlepaddle.github.io/PaddleX/main/module_usage/tutorials/ocr_modules/layout_analysis.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\ndiff --git a/docs/version3.x/module_usage/layout_detection.en.md b/docs/version3.x/module_usage/layout_detection.en.md\nindex e24679ad3a1..d895c44a492 100644\n--- a/docs/version3.x/module_usage/layout_detection.en.md\n+++ b/docs/version3.x/module_usage/layout_detection.en.md\n@@ -328,6 +328,14 @@ paddleocr layout_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr layout_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable <code>PADDLE_PDX_MODEL_SOURCE=\"BOS\"</code> to change the model source to BOS. In the future, more model sources will be supported.\n@@ -363,6 +371,22 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import LayoutDetection\n+\n+model = LayoutDetection(\n+ model_name=\"PP-DocLayout_plus-L\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"layout.jpg\", batch_size=1, layout_nms=True)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#52-weight-conversion) section in the [Inference Engine](#5-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -432,7 +456,7 @@ By default, GPU 0 is used if available; otherwise, CPU is used.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -744,7 +768,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-DocLayout_plus-L</td>\n+ <td rowspan=\"4\">PP-DocLayout_plus-L</td>\n <td>paddle_static</td>\n <td>10.92</td>\n <td>26.11</td>\n@@ -766,7 +790,14 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>52.24</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-DocBlockLayout</td>\n+ <td>onnxruntime</td>\n+ <td>9.14</td>\n+ <td>12.35</td>\n+ <td>0.14</td>\n+ <td>21.81</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-DocBlockLayout</td>\n <td>paddle_static</td>\n <td>9.51</td>\n <td>27.59</td>\n@@ -787,12 +818,19 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>0.75</td>\n <td>50.96</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>7.53</td>\n+ <td>9.22</td>\n+ <td>0.06</td>\n+ <td>16.97</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -802,13 +840,13 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Layout Detection Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/layout_detection.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Layout Detection Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/layout_detection.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\n \n ## 6. FAQ\ndiff --git a/docs/version3.x/module_usage/layout_detection.md b/docs/version3.x/module_usage/layout_detection.md\nindex 7452c06cf89..f0766ed4581 100644\n--- a/docs/version3.x/module_usage/layout_detection.md\n+++ b/docs/version3.x/module_usage/layout_detection.md\n@@ -328,6 +328,14 @@ paddleocr layout_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr layout_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -363,6 +371,22 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import LayoutDetection\n+\n+model = LayoutDetection(\n+ model_name=\"PP-DocLayout_plus-L\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"layout.jpg\", batch_size=1, layout_nms=True)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 训练后的模型如果想使用 `paddle_dynamic` 或 `transformers` 引擎，请参考后文 [推理引擎](#五推理引擎) 中的 [权重转换](#52-权重转换) 部分将模型由 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式。\n@@ -433,7 +457,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -737,7 +761,7 @@ for res in output:\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-DocLayout_plus-L</td>\n+ <td rowspan=\"4\">PP-DocLayout_plus-L</td>\n <td>paddle_static</td>\n <td>10.92</td>\n <td>26.11</td>\n@@ -759,7 +783,14 @@ for res in output:\n <td>52.24</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-DocBlockLayout</td>\n+ <td>onnxruntime</td>\n+ <td>9.14</td>\n+ <td>12.35</td>\n+ <td>0.14</td>\n+ <td>21.81</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-DocBlockLayout</td>\n <td>paddle_static</td>\n <td>9.51</td>\n <td>27.59</td>\n@@ -780,12 +811,19 @@ for res in output:\n <td>0.75</td>\n <td>50.96</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>7.53</td>\n+ <td>9.22</td>\n+ <td>0.06</td>\n+ <td>16.97</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/layout.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -795,13 +833,13 @@ for res in output:\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 权重转换 {#52-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 版面区域检测模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/layout_detection.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 版面区域检测模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/layout_detection.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n \n ## 六、FAQ\ndiff --git a/docs/version3.x/module_usage/table_cells_detection.en.md b/docs/version3.x/module_usage/table_cells_detection.en.md\nindex 25f40e63f86..f01f5fe86e4 100644\n--- a/docs/version3.x/module_usage/table_cells_detection.en.md\n+++ b/docs/version3.x/module_usage/table_cells_detection.en.md\n@@ -107,6 +107,14 @@ paddleocr table_cells_detection -i https://paddle-model-ecology.bj.bcebos.com/pa\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr table_cells_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable `PADDLE_PDX_MODEL_SOURCE=\"BOS\"` to change the model source to BOS. In the future, more model sources will be supported.\n@@ -140,6 +148,21 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import TableCellsDetection\n+model = TableCellsDetection(\n+ model_name=\"RT-DETR-L_wired_table_cell_det\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"table_recognition.jpg\", threshold=0.3, batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_img(\"./output/\")\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#52-weight-conversion) section in the [Inference Engine](#5-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -206,7 +229,7 @@ By default, GPU 0 is used if available; otherwise, CPU is used.</td>\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -430,7 +453,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">RT-DETR-L_wired_table_cell_det</td>\n+ <td rowspan=\"4\">RT-DETR-L_wired_table_cell_det</td>\n <td>paddle_static</td>\n <td>3.59</td>\n <td>23.11</td>\n@@ -452,7 +475,14 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>42.10</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">RT-DETR-L_wireless_table_cell_det</td>\n+ <td>onnxruntime</td>\n+ <td>2.70</td>\n+ <td>9.10</td>\n+ <td>0.12</td>\n+ <td>12.07</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">RT-DETR-L_wireless_table_cell_det</td>\n <td>paddle_static</td>\n <td>3.77</td>\n <td>23.44</td>\n@@ -473,12 +503,19 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>0.71</td>\n <td>41.91</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>2.77</td>\n+ <td>9.11</td>\n+ <td>0.12</td>\n+ <td>12.14</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -488,13 +525,13 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Table Cell Detection Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/table_cells_detection.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Table Cell Detection Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/table_cells_detection.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\n \n ## 6. FAQ\ndiff --git a/docs/version3.x/module_usage/table_cells_detection.md b/docs/version3.x/module_usage/table_cells_detection.md\nindex 45bcd6e2d04..743515c26c3 100644\n--- a/docs/version3.x/module_usage/table_cells_detection.md\n+++ b/docs/version3.x/module_usage/table_cells_detection.md\n@@ -105,6 +105,14 @@ paddleocr table_cells_detection -i https://paddle-model-ecology.bj.bcebos.com/pa\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr table_cells_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -138,6 +146,21 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import TableCellsDetection\n+model = TableCellsDetection(\n+ model_name=\"RT-DETR-L_wired_table_cell_det\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"table_recognition.jpg\", threshold=0.3, batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_img(\"./output/\")\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 训练后的模型如果想使用 `paddle_dynamic` 或 `transformers` 引擎，请参考后文 [推理引擎](#五推理引擎) 中的 [权重转换](#52-权重转换) 部分将模型由 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式。\n@@ -205,7 +228,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -441,7 +464,7 @@ for res in output:\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">RT-DETR-L_wired_table_cell_det</td>\n+ <td rowspan=\"4\">RT-DETR-L_wired_table_cell_det</td>\n <td>paddle_static</td>\n <td>3.59</td>\n <td>23.11</td>\n@@ -463,7 +486,14 @@ for res in output:\n <td>42.10</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">RT-DETR-L_wireless_table_cell_det</td>\n+ <td>onnxruntime</td>\n+ <td>2.70</td>\n+ <td>9.10</td>\n+ <td>0.12</td>\n+ <td>12.07</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">RT-DETR-L_wireless_table_cell_det</td>\n <td>paddle_static</td>\n <td>3.77</td>\n <td>23.44</td>\n@@ -484,12 +514,19 @@ for res in output:\n <td>0.71</td>\n <td>41.91</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>2.77</td>\n+ <td>9.11</td>\n+ <td>0.12</td>\n+ <td>12.14</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -499,13 +536,13 @@ for res in output:\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 权重转换 {#52-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 表格单元格检测模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/table_cells_detection.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 表格单元格检测模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/table_cells_detection.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n \n ## 六、FAQ\ndiff --git a/docs/version3.x/module_usage/table_classification.en.md b/docs/version3.x/module_usage/table_classification.en.md\nindex 2e7c64052a7..2feb28fa444 100644\n--- a/docs/version3.x/module_usage/table_classification.en.md\n+++ b/docs/version3.x/module_usage/table_classification.en.md\n@@ -100,6 +100,14 @@ paddleocr table_classification -i https://paddle-model-ecology.bj.bcebos.com/pad\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr table_classification -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable `PADDLE_PDX_MODEL_SOURCE=\"BOS\"` to change the model source to BOS. In the future, more model sources will be supported.\n@@ -131,6 +139,20 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import TableClassification\n+model = TableClassification(\n+ model_name=\"PP-LCNet_x1_0_table_cls\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"table_recognition.jpg\", batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#52-weight-conversion) section in the [Inference Engine](#5-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -194,7 +216,7 @@ By default, GPU 0 is used if available; otherwise, CPU is used.</td>\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -386,7 +408,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-LCNet_x1_0_table_cls</td>\n+ <td rowspan=\"4\">PP-LCNet_x1_0_table_cls</td>\n <td>paddle_static</td>\n <td>3.56</td>\n <td>3.38</td>\n@@ -407,12 +429,19 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>0.15</td>\n <td>14.05</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>3.23</td>\n+ <td>0.88</td>\n+ <td>0.05</td>\n+ <td>4.25</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_image_classification_001.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_image_classification_001.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -422,13 +451,13 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Table Classification Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/table_classification.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Table Classification Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/table_classification.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\n \n ## 6. FAQ\ndiff --git a/docs/version3.x/module_usage/table_classification.md b/docs/version3.x/module_usage/table_classification.md\nindex ec83132fd33..9c9ba9c1a0a 100644\n--- a/docs/version3.x/module_usage/table_classification.md\n+++ b/docs/version3.x/module_usage/table_classification.md\n@@ -98,6 +98,14 @@ paddleocr table_classification -i https://paddle-model-ecology.bj.bcebos.com/pad\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr table_classification -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -129,6 +137,20 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import TableClassification\n+model = TableClassification(\n+ model_name=\"PP-LCNet_x1_0_table_cls\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"table_recognition.jpg\", batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 训练后的模型如果想使用 `paddle_dynamic` 或 `transformers` 引擎，请参考后文 [推理引擎](#五推理引擎) 中的 [权重转换](#52-权重转换) 部分将模型由 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式。\n@@ -192,7 +214,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -387,7 +409,7 @@ for res in output:\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-LCNet_x1_0_table_cls</td>\n+ <td rowspan=\"4\">PP-LCNet_x1_0_table_cls</td>\n <td>paddle_static</td>\n <td>3.56</td>\n <td>3.38</td>\n@@ -408,12 +430,19 @@ for res in output:\n <td>0.15</td>\n <td>14.05</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>3.23</td>\n+ <td>0.88</td>\n+ <td>0.05</td>\n+ <td>4.25</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_image_classification_001.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_image_classification_001.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -423,13 +452,13 @@ for res in output:\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 权重转换 {#52-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 表格分类模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/table_classification.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 表格分类模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/table_classification.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n \n ## 六、FAQ\ndiff --git a/docs/version3.x/module_usage/table_structure_recognition.en.md b/docs/version3.x/module_usage/table_structure_recognition.en.md\nindex 308d7567098..7e8da27cc3a 100644\n--- a/docs/version3.x/module_usage/table_structure_recognition.en.md\n+++ b/docs/version3.x/module_usage/table_structure_recognition.en.md\n@@ -124,6 +124,14 @@ paddleocr table_structure_recognition -i https://paddle-model-ecology.bj.bcebos.\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr table_structure_recognition -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable `PADDLE_PDX_MODEL_SOURCE=\"BOS\"` to change the model source to BOS. In the future, more model sources will be supported.\n@@ -155,6 +163,20 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import TableStructureRecognition\n+model = TableStructureRecognition(\n+ model_name=\"SLANet\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(input=\"table_recognition.jpg\", batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#52-weight-conversion) section in the [Inference Engine](#5-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -213,7 +235,7 @@ By default, GPU 0 is used if available; otherwise, CPU is used.</td>\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -476,7 +498,23 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">SLANeXt_wired</td>\n+ <td rowspan=\"1\">SLANet</td>\n+ <td>onnxruntime</td>\n+ <td>1.42</td>\n+ <td>22.02</td>\n+ <td>0.22</td>\n+ <td>23.81</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"1\">SLANet_plus</td>\n+ <td>onnxruntime</td>\n+ <td>1.40</td>\n+ <td>21.64</td>\n+ <td>0.22</td>\n+ <td>23.42</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">SLANeXt_wired</td>\n <td>paddle_static</td>\n <td>1.50</td>\n <td>30.91</td>\n@@ -498,7 +536,14 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>51.12</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">SLANeXt_wireless</td>\n+ <td>onnxruntime</td>\n+ <td>1.60</td>\n+ <td>30.96</td>\n+ <td>0.25</td>\n+ <td>32.99</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">SLANeXt_wireless</td>\n <td>paddle_static</td>\n <td>1.67</td>\n <td>30.49</td>\n@@ -519,12 +564,19 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>0.75</td>\n <td>51.76</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.62</td>\n+ <td>31.01</td>\n+ <td>0.26</td>\n+ <td>33.06</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -534,13 +586,13 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Table Structure Recognition Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/table_structure_recognition.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Table Structure Recognition Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/table_structure_recognition.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\n \n ## 6. FAQ\ndiff --git a/docs/version3.x/module_usage/table_structure_recognition.md b/docs/version3.x/module_usage/table_structure_recognition.md\nindex eb5277ac9a8..12c6845ede9 100644\n--- a/docs/version3.x/module_usage/table_structure_recognition.md\n+++ b/docs/version3.x/module_usage/table_structure_recognition.md\n@@ -124,6 +124,14 @@ paddleocr table_structure_recognition -i https://paddle-model-ecology.bj.bcebos.\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr table_structure_recognition -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -155,6 +163,20 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import TableStructureRecognition\n+model = TableStructureRecognition(\n+ model_name=\"SLANet\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(input=\"table_recognition.jpg\", batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 训练后的模型如果想使用 `paddle_dynamic` 或 `transformers` 引擎，请参考后文 [推理引擎](#五推理引擎) 中的 [权重转换](#52-权重转换) 部分将模型由 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式。\n@@ -213,7 +235,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -480,7 +502,23 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \\\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">SLANeXt_wired</td>\n+ <td rowspan=\"1\">SLANet</td>\n+ <td>onnxruntime</td>\n+ <td>1.42</td>\n+ <td>22.02</td>\n+ <td>0.22</td>\n+ <td>23.81</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"1\">SLANet_plus</td>\n+ <td>onnxruntime</td>\n+ <td>1.40</td>\n+ <td>21.64</td>\n+ <td>0.22</td>\n+ <td>23.42</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">SLANeXt_wired</td>\n <td>paddle_static</td>\n <td>1.50</td>\n <td>30.91</td>\n@@ -502,7 +540,14 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \\\n <td>51.12</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">SLANeXt_wireless</td>\n+ <td>onnxruntime</td>\n+ <td>1.60</td>\n+ <td>30.96</td>\n+ <td>0.25</td>\n+ <td>32.99</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">SLANeXt_wireless</td>\n <td>paddle_static</td>\n <td>1.67</td>\n <td>30.49</td>\n@@ -523,12 +568,19 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \\\n <td>0.75</td>\n <td>51.76</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.62</td>\n+ <td>31.01</td>\n+ <td>0.26</td>\n+ <td>33.06</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/table_recognition.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -538,13 +590,13 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \\\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 权重转换 {#52-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 表格结构识别模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/table_structure_recognition.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 表格结构识别模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/table_structure_recognition.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n \n ## 六、FAQ\ndiff --git a/docs/version3.x/module_usage/text_detection.en.md b/docs/version3.x/module_usage/text_detection.en.md\nindex 5f4ae5b7463..aa2cfce1b00 100644\n--- a/docs/version3.x/module_usage/text_detection.en.md\n+++ b/docs/version3.x/module_usage/text_detection.en.md\n@@ -130,6 +130,14 @@ paddleocr text_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/i\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr text_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.png \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable `PADDLE_PDX_MODEL_SOURCE=\"BOS\"` to change the model source to BOS. In the future, more model sources will be supported.\n@@ -163,6 +171,21 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import TextDetection\n+model = TextDetection(\n+ model_name=\"PP-OCRv5_server_det\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"general_ocr_001.png\", batch_size=1)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#52-weight-conversion) section in the [Inference Engine](#5-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -233,7 +256,7 @@ By default, GPU 0 is used if available; otherwise, CPU is used.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -599,7 +622,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-OCRv5_mobile_det</td>\n+ <td rowspan=\"4\">PP-OCRv5_mobile_det</td>\n <td>paddle_static</td>\n <td>11.43</td>\n <td>13.80</td>\n@@ -621,7 +644,14 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>37.54</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-OCRv5_server_det</td>\n+ <td>onnxruntime</td>\n+ <td>9.98</td>\n+ <td>5.70</td>\n+ <td>2.04</td>\n+ <td>17.90</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv5_server_det</td>\n <td>paddle_static</td>\n <td>13.24</td>\n <td>26.91</td>\n@@ -642,12 +672,106 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>7.44</td>\n <td>36.76</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>10.01</td>\n+ <td>13.76</td>\n+ <td>1.92</td>\n+ <td>25.86</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_medium_det</td>\n+ <td>paddle_static</td>\n+ <td>13.89</td>\n+ <td>16.02</td>\n+ <td>2.49</td>\n+ <td>33.14</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>11.42</td>\n+ <td>26.23</td>\n+ <td>2.30</td>\n+ <td>40.10</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>11.40</td>\n+ <td>8.57</td>\n+ <td>8.35</td>\n+ <td>29.57</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>10.80</td>\n+ <td>13.06</td>\n+ <td>2.19</td>\n+ <td>26.18</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_small_det</td>\n+ <td>paddle_static</td>\n+ <td>10.91</td>\n+ <td>10.97</td>\n+ <td>2.41</td>\n+ <td>24.45</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>11.56</td>\n+ <td>22.17</td>\n+ <td>2.66</td>\n+ <td>36.55</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>11.70</td>\n+ <td>7.34</td>\n+ <td>3.87</td>\n+ <td>23.89</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>11.32</td>\n+ <td>7.46</td>\n+ <td>2.54</td>\n+ <td>21.49</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_tiny_det</td>\n+ <td>paddle_static</td>\n+ <td>11.14</td>\n+ <td>10.71</td>\n+ <td>2.84</td>\n+ <td>24.85</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>11.52</td>\n+ <td>21.70</td>\n+ <td>2.94</td>\n+ <td>36.31</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>10.90</td>\n+ <td>6.99</td>\n+ <td>4.13</td>\n+ <td>23.00</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>11.19</td>\n+ <td>6.35</td>\n+ <td>2.79</td>\n+ <td>20.49</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -657,14 +781,14 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Text Detection Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/text_detection.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Text Detection Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/text_detection.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\n \n ## 6. FAQ\n \ndiff --git a/docs/version3.x/module_usage/text_detection.md b/docs/version3.x/module_usage/text_detection.md\nindex b65dacfc9c7..3de5793780a 100644\n--- a/docs/version3.x/module_usage/text_detection.md\n+++ b/docs/version3.x/module_usage/text_detection.md\n@@ -165,6 +165,14 @@ paddleocr text_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/i\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr text_detection -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.png \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -200,6 +208,21 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import TextDetection\n+model = TextDetection(\n+ model_name=\"PP-OCRv5_server_det\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"general_ocr_001.png\", batch_size=1)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 训练后的模型如果想使用 `paddle_dynamic` 或 `transformers` 引擎，请参考后文 [推理引擎](#五推理引擎) 中的 [权重转换](#52-权重转换) 部分将模型由 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式。\n@@ -269,7 +292,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -642,7 +665,7 @@ python3 tools/export_model.py -c configs/det/PP-OCRv5/PP-OCRv5_server_det.yml -o\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-OCRv5_mobile_det</td>\n+ <td rowspan=\"4\">PP-OCRv5_mobile_det</td>\n <td>paddle_static</td>\n <td>11.43</td>\n <td>13.80</td>\n@@ -664,7 +687,14 @@ python3 tools/export_model.py -c configs/det/PP-OCRv5/PP-OCRv5_server_det.yml -o\n <td>37.54</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-OCRv5_server_det</td>\n+ <td>onnxruntime</td>\n+ <td>9.98</td>\n+ <td>5.70</td>\n+ <td>2.04</td>\n+ <td>17.90</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv5_server_det</td>\n <td>paddle_static</td>\n <td>13.24</td>\n <td>26.91</td>\n@@ -685,12 +715,106 @@ python3 tools/export_model.py -c configs/det/PP-OCRv5/PP-OCRv5_server_det.yml -o\n <td>7.44</td>\n <td>36.76</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>10.01</td>\n+ <td>13.76</td>\n+ <td>1.92</td>\n+ <td>25.86</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_medium_det</td>\n+ <td>paddle_static</td>\n+ <td>13.89</td>\n+ <td>16.02</td>\n+ <td>2.49</td>\n+ <td>33.14</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>11.42</td>\n+ <td>26.23</td>\n+ <td>2.30</td>\n+ <td>40.10</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>11.40</td>\n+ <td>8.57</td>\n+ <td>8.35</td>\n+ <td>29.57</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>10.80</td>\n+ <td>13.06</td>\n+ <td>2.19</td>\n+ <td>26.18</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_small_det</td>\n+ <td>paddle_static</td>\n+ <td>10.91</td>\n+ <td>10.97</td>\n+ <td>2.41</td>\n+ <td>24.45</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>11.56</td>\n+ <td>22.17</td>\n+ <td>2.66</td>\n+ <td>36.55</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>11.70</td>\n+ <td>7.34</td>\n+ <td>3.87</td>\n+ <td>23.89</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>11.32</td>\n+ <td>7.46</td>\n+ <td>2.54</td>\n+ <td>21.49</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_tiny_det</td>\n+ <td>paddle_static</td>\n+ <td>11.14</td>\n+ <td>10.71</td>\n+ <td>2.84</td>\n+ <td>24.85</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>11.52</td>\n+ <td>21.70</td>\n+ <td>2.94</td>\n+ <td>36.31</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>10.90</td>\n+ <td>6.99</td>\n+ <td>4.13</td>\n+ <td>23.00</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>11.19</td>\n+ <td>6.35</td>\n+ <td>2.79</td>\n+ <td>20.49</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -700,14 +824,14 @@ python3 tools/export_model.py -c configs/det/PP-OCRv5/PP-OCRv5_server_det.yml -o\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 权重转换 {#52-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 文本检测模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/text_detection.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 文本检测模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/text_detection.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n \n ## 六、常见问题与解决方案\n \ndiff --git a/docs/version3.x/module_usage/text_image_unwarping.en.md b/docs/version3.x/module_usage/text_image_unwarping.en.md\nindex 71a1fd4f734..ed826c0795a 100644\n--- a/docs/version3.x/module_usage/text_image_unwarping.en.md\n+++ b/docs/version3.x/module_usage/text_image_unwarping.en.md\n@@ -106,6 +106,14 @@ paddleocr text_image_unwarping -i https://paddle-model-ecology.bj.bcebos.com/pad\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr text_image_unwarping -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/doc_test.jpg \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable `PADDLE_PDX_MODEL_SOURCE=\"BOS\"` to change the model source to BOS. In the future, more model sources will be supported.\n@@ -139,6 +147,21 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import TextImageUnwarping\n+model = TextImageUnwarping(\n+ model_name=\"UVDoc\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"doc_test.jpg\", batch_size=1)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n After running, the result obtained is:\n@@ -195,7 +218,7 @@ By default, GPU 0 will be used if available; otherwise, the CPU will be used.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -404,7 +427,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">UVDoc</td>\n+ <td rowspan=\"4\">UVDoc</td>\n <td>paddle_static</td>\n <td>14.96</td>\n <td>18.60</td>\n@@ -425,12 +448,19 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>0.91</td>\n <td>33.07</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>10.60</td>\n+ <td>8.44</td>\n+ <td>1.75</td>\n+ <td>21.30</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/doc_test.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/doc_test.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -440,7 +470,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\ndiff --git a/docs/version3.x/module_usage/text_image_unwarping.md b/docs/version3.x/module_usage/text_image_unwarping.md\nindex b2a166d2d04..d3e8028e611 100644\n--- a/docs/version3.x/module_usage/text_image_unwarping.md\n+++ b/docs/version3.x/module_usage/text_image_unwarping.md\n@@ -105,6 +105,14 @@ paddleocr text_image_unwarping -i https://paddle-model-ecology.bj.bcebos.com/pad\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr text_image_unwarping -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/doc_test.jpg \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -138,6 +146,21 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import TextImageUnwarping\n+model = TextImageUnwarping(\n+ model_name=\"UVDoc\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"doc_test.jpg\", batch_size=1)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 运行后，得到的结果为：\n@@ -196,7 +219,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -399,7 +422,7 @@ for res in output:\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">UVDoc</td>\n+ <td rowspan=\"4\">UVDoc</td>\n <td>paddle_static</td>\n <td>14.96</td>\n <td>18.60</td>\n@@ -420,12 +443,19 @@ for res in output:\n <td>0.91</td>\n <td>33.07</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>10.60</td>\n+ <td>8.44</td>\n+ <td>1.75</td>\n+ <td>21.30</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/doc_test.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/doc_test.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -435,7 +465,7 @@ for res in output:\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\ndiff --git a/docs/version3.x/module_usage/text_recognition.en.md b/docs/version3.x/module_usage/text_recognition.en.md\nindex f96f39b109e..c7d65fe4773 100644\n--- a/docs/version3.x/module_usage/text_recognition.en.md\n+++ b/docs/version3.x/module_usage/text_recognition.en.md\n@@ -552,6 +552,14 @@ paddleocr text_recognition -i https://paddle-model-ecology.bj.bcebos.com/paddlex\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr text_recognition -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_rec_001.png \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note:</b> The official PaddleOCR models are downloaded from HuggingFace by default. If you cannot access HuggingFace, you can change the model source to BOS by setting the environment variable `PADDLE_PDX_MODEL_SOURCE=\"BOS\"`. More mainstream model sources will be supported in the future.\n@@ -585,6 +593,21 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import TextRecognition\n+model = TextRecognition(\n+ model_name=\"PP-OCRv5_server_rec\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(input=\"general_ocr_rec_001.png\", batch_size=1)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#52-weight-conversion) section in the [Inference Engine](#5-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -643,7 +666,7 @@ By default, GPU 0 is used; if unavailable, CPU is used.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -906,7 +929,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-OCRv5_mobile_rec</td>\n+ <td rowspan=\"4\">PP-OCRv5_mobile_rec</td>\n <td>paddle_static</td>\n <td>1.94</td>\n <td>6.69</td>\n@@ -928,7 +951,14 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>21.68</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-OCRv5_server_rec</td>\n+ <td>onnxruntime</td>\n+ <td>1.82</td>\n+ <td>2.05</td>\n+ <td>0.91</td>\n+ <td>4.91</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv5_server_rec</td>\n <td>paddle_static</td>\n <td>1.98</td>\n <td>11.37</td>\n@@ -949,12 +979,106 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>0.51</td>\n <td>16.36</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.80</td>\n+ <td>3.15</td>\n+ <td>0.90</td>\n+ <td>5.98</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_medium_rec</td>\n+ <td>paddle_static</td>\n+ <td>1.74</td>\n+ <td>5.38</td>\n+ <td>0.84</td>\n+ <td>8.08</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>1.76</td>\n+ <td>13.38</td>\n+ <td>0.85</td>\n+ <td>16.10</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>2.58</td>\n+ <td>7.04</td>\n+ <td>0.41</td>\n+ <td>10.19</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.74</td>\n+ <td>2.28</td>\n+ <td>0.84</td>\n+ <td>4.97</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_small_rec</td>\n+ <td>paddle_static</td>\n+ <td>1.74</td>\n+ <td>4.73</td>\n+ <td>0.82</td>\n+ <td>7.41</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>1.76</td>\n+ <td>12.43</td>\n+ <td>0.87</td>\n+ <td>15.18</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>2.55</td>\n+ <td>6.70</td>\n+ <td>0.41</td>\n+ <td>9.82</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.73</td>\n+ <td>1.79</td>\n+ <td>0.83</td>\n+ <td>4.46</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_tiny_rec</td>\n+ <td>paddle_static</td>\n+ <td>1.76</td>\n+ <td>2.77</td>\n+ <td>0.40</td>\n+ <td>5.04</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>1.75</td>\n+ <td>6.96</td>\n+ <td>0.36</td>\n+ <td>9.19</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>2.45</td>\n+ <td>3.12</td>\n+ <td>0.40</td>\n+ <td>6.12</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.73</td>\n+ <td>0.92</td>\n+ <td>0.36</td>\n+ <td>3.12</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_rec_001.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_rec_001.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -964,13 +1088,13 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Text Image Orientation Classification Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/text_recognition.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Text Image Orientation Classification Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/text_recognition.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\n \n ## 6. FAQ\ndiff --git a/docs/version3.x/module_usage/text_recognition.md b/docs/version3.x/module_usage/text_recognition.md\nindex 00e882fc7c6..6350757bea0 100644\n--- a/docs/version3.x/module_usage/text_recognition.md\n+++ b/docs/version3.x/module_usage/text_recognition.md\n@@ -561,6 +561,14 @@ paddleocr text_recognition -i https://paddle-model-ecology.bj.bcebos.com/paddlex\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr text_recognition -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_rec_001.png \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -594,6 +602,21 @@ for res in output:\n res.save_to_json(save_path=\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import TextRecognition\n+model = TextRecognition(\n+ model_name=\"PP-OCRv5_server_rec\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(input=\"general_ocr_rec_001.png\", batch_size=1)\n+for res in output:\n+ res.print()\n+ res.save_to_img(save_path=\"./output/\")\n+ res.save_to_json(save_path=\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 训练后的模型如果想使用 `paddle_dynamic` 或 `transformers` 引擎，请参考后文 [推理引擎](#五推理引擎) 中的 [权重转换](#52-权重转换) 部分将模型由 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式。\n@@ -655,7 +678,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -929,7 +952,7 @@ Global.save_inference_dir=\"./PP-OCRv5_server_rec_infer/\"\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-OCRv5_mobile_rec</td>\n+ <td rowspan=\"4\">PP-OCRv5_mobile_rec</td>\n <td>paddle_static</td>\n <td>1.94</td>\n <td>6.69</td>\n@@ -951,7 +974,14 @@ Global.save_inference_dir=\"./PP-OCRv5_server_rec_infer/\"\n <td>21.68</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-OCRv5_server_rec</td>\n+ <td>onnxruntime</td>\n+ <td>1.82</td>\n+ <td>2.05</td>\n+ <td>0.91</td>\n+ <td>4.91</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv5_server_rec</td>\n <td>paddle_static</td>\n <td>1.98</td>\n <td>11.37</td>\n@@ -972,12 +1002,106 @@ Global.save_inference_dir=\"./PP-OCRv5_server_rec_infer/\"\n <td>0.51</td>\n <td>16.36</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.80</td>\n+ <td>3.15</td>\n+ <td>0.90</td>\n+ <td>5.98</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_medium_rec</td>\n+ <td>paddle_static</td>\n+ <td>1.74</td>\n+ <td>5.38</td>\n+ <td>0.84</td>\n+ <td>8.08</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>1.76</td>\n+ <td>13.38</td>\n+ <td>0.85</td>\n+ <td>16.10</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>2.58</td>\n+ <td>7.04</td>\n+ <td>0.41</td>\n+ <td>10.19</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.74</td>\n+ <td>2.28</td>\n+ <td>0.84</td>\n+ <td>4.97</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_small_rec</td>\n+ <td>paddle_static</td>\n+ <td>1.74</td>\n+ <td>4.73</td>\n+ <td>0.82</td>\n+ <td>7.41</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>1.76</td>\n+ <td>12.43</td>\n+ <td>0.87</td>\n+ <td>15.18</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>2.55</td>\n+ <td>6.70</td>\n+ <td>0.41</td>\n+ <td>9.82</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.73</td>\n+ <td>1.79</td>\n+ <td>0.83</td>\n+ <td>4.46</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-OCRv6_tiny_rec</td>\n+ <td>paddle_static</td>\n+ <td>1.76</td>\n+ <td>2.77</td>\n+ <td>0.40</td>\n+ <td>5.04</td>\n+ </tr>\n+ <tr>\n+ <td>paddle_dynamic</td>\n+ <td>1.75</td>\n+ <td>6.96</td>\n+ <td>0.36</td>\n+ <td>9.19</td>\n+ </tr>\n+ <tr>\n+ <td>transformers</td>\n+ <td>2.45</td>\n+ <td>3.12</td>\n+ <td>0.40</td>\n+ <td>6.12</td>\n+ </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>1.73</td>\n+ <td>0.92</td>\n+ <td>0.36</td>\n+ <td>3.12</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_rec_001.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_rec_001.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -987,13 +1111,13 @@ Global.save_inference_dir=\"./PP-OCRv5_server_rec_infer/\"\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 权重转换 {#52-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 文字识别模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/text_recognition.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 文字识别模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/text_recognition.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n \n ## 六、FAQ\ndiff --git a/docs/version3.x/module_usage/textline_orientation_classification.en.md b/docs/version3.x/module_usage/textline_orientation_classification.en.md\nindex 1f079e689ee..b5ace71016e 100644\n--- a/docs/version3.x/module_usage/textline_orientation_classification.en.md\n+++ b/docs/version3.x/module_usage/textline_orientation_classification.en.md\n@@ -112,6 +112,14 @@ paddleocr textline_orientation_classification -i https://paddle-model-ecology.bj\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr textline_orientation_classification -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/textline_rot180_demo.jpg \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <b>Note: </b>The official models would be download from HuggingFace by default. If can't access to HuggingFace, please set the environment variable <code>PADDLE_PDX_MODEL_SOURCE=\"BOS\"</code> to change the model source to BOS. In the future, more model sources will be supported.\n@@ -143,6 +151,21 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import TextLineOrientationClassification\n+model = TextLineOrientationClassification(\n+ model_name=\"PP-LCNet_x0_25_textline_ori\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"textline_rot180_demo.jpg\", batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_img(\"./output/demo.png\")\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n If you want to use the trained model with the `paddle_dynamic` or `transformers` engine, refer to the [Weight Conversion](#52-weight-conversion) section in the [Inference Engine](#5-inference-engine) section below to convert the model from the `pdparams` format to the `safetensors` format using PaddleX.\n@@ -206,7 +229,7 @@ By default, GPU 0 is used if available; otherwise, CPU is used.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_static</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_static</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -454,7 +477,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-LCNet_x0_25_textline_ori</td>\n+ <td rowspan=\"4\">PP-LCNet_x0_25_textline_ori</td>\n <td>paddle_static</td>\n <td>0.30</td>\n <td>2.89</td>\n@@ -476,7 +499,14 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>5.36</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-LCNet_x1_0_textline_ori</td>\n+ <td>onnxruntime</td>\n+ <td>0.27</td>\n+ <td>0.76</td>\n+ <td>0.05</td>\n+ <td>1.16</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-LCNet_x1_0_textline_ori</td>\n <td>paddle_static</td>\n <td>0.33</td>\n <td>3.20</td>\n@@ -497,12 +527,19 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <td>0.14</td>\n <td>5.04</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>0.27</td>\n+ <td>0.77</td>\n+ <td>0.05</td>\n+ <td>1.16</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/textline_rot180_demo.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/textline_rot180_demo.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -512,11 +549,11 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 Weight Conversion\n \n-When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Text Line Orientation Classification Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/textline_orientation_classification.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference.\n+When using the inference engine, the system will automatically download the official pre-trained model. If you need to use a self-trained model with the `paddle_dynamic` or `transformers` engine, please refer to the [PaddleX Text Line Orientation Classification Module Weight Conversion](https://paddlepaddle.github.io/PaddleX/latest/en/module_usage/tutorials/ocr_modules/textline_orientation_classification.html#442) section to convert the model from the `pdparams` format to the `safetensors` format using PaddleX. This allows seamless integration into the PaddleOCR API for inference. If you need to use a self-trained model with the `onnxruntime` engine, refer to [PaddleX Obtain ONNX Models](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html) to obtain the ONNX model, so it can be seamlessly integrated into the PaddleOCR API for inference.\ndiff --git a/docs/version3.x/module_usage/textline_orientation_classification.md b/docs/version3.x/module_usage/textline_orientation_classification.md\nindex d23271fc29f..bc86483c54f 100644\n--- a/docs/version3.x/module_usage/textline_orientation_classification.md\n+++ b/docs/version3.x/module_usage/textline_orientation_classification.md\n@@ -117,6 +117,14 @@ paddleocr textline_orientation_classification -i https://paddle-model-ecology.bj\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr textline_orientation_classification -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/textline_rot180_demo.jpg \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <b>注：</b>PaddleOCR 官方模型默认从 HuggingFace 获取，如运行环境访问 HuggingFace 不便，可通过环境变量修改模型源为 BOS：`PADDLE_PDX_MODEL_SOURCE=\"BOS\"`，未来将支持更多主流模型源；\n@@ -150,6 +158,21 @@ for res in output:\n res.save_to_json(\"./output/res.json\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请确保已配置 ONNX Runtime 环境，然后执行如下代码：\n+\n+```python\n+from paddleocr import TextLineOrientationClassification\n+model = TextLineOrientationClassification(\n+ model_name=\"PP-LCNet_x0_25_textline_ori\",\n+ engine=\"onnxruntime\",\n+)\n+output = model.predict(\"textline_rot180_demo.jpg\", batch_size=1)\n+for res in output:\n+ res.print(json_format=False)\n+ res.save_to_img(\"./output/demo.png\")\n+ res.save_to_json(\"./output/res.json\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n 训练后的模型如果想使用 `paddle_dynamic` 或 `transformers` 引擎，请参考后文 [推理引擎](#五推理引擎) 中的 [权重转换](#52-权重转换) 部分将模型由 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式。\n@@ -212,7 +235,7 @@ for res in output:\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n+<td><b>含义：</b>推理引擎。<br><b>说明：</b>支持 <code>None</code>（默认值）、<code>paddle</code>、<code>paddle_static</code>、<code>paddle_dynamic</code>、<code>transformers</code>、<code>onnxruntime</code>。保持为默认值 <code>None</code> 时，本地推理默认使用 <code>paddle_static</code> 引擎。详细说明、取值、兼容性规则与示例请参见 <a href=\"../inference_deployment/local_inference/inference_engine.md\">推理引擎与配置说明</a>。</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -418,7 +441,7 @@ for res in output:\n </thead>\n <tbody>\n <tr>\n- <td rowspan=\"3\">PP-LCNet_x0_25_textline_ori</td>\n+ <td rowspan=\"4\">PP-LCNet_x0_25_textline_ori</td>\n <td>paddle_static</td>\n <td>0.30</td>\n <td>2.89</td>\n@@ -440,7 +463,14 @@ for res in output:\n <td>5.36</td>\n </tr>\n <tr>\n- <td rowspan=\"3\">PP-LCNet_x1_0_textline_ori</td>\n+ <td>onnxruntime</td>\n+ <td>0.27</td>\n+ <td>0.76</td>\n+ <td>0.05</td>\n+ <td>1.16</td>\n+ </tr>\n+ <tr>\n+ <td rowspan=\"4\">PP-LCNet_x1_0_textline_ori</td>\n <td>paddle_static</td>\n <td>0.33</td>\n <td>3.20</td>\n@@ -461,12 +491,19 @@ for res in output:\n <td>0.14</td>\n <td>5.04</td>\n </tr>\n+ <tr>\n+ <td>onnxruntime</td>\n+ <td>0.27</td>\n+ <td>0.77</td>\n+ <td>0.05</td>\n+ <td>1.16</td>\n+ </tr>\n </tbody>\n </table>\n \n <strong>测试环境说明:</strong>\n <ul>\n- <li><strong>测试数据：</strong>[示例图片](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/textline_rot180_demo.jpg)</li>\n+ <li><strong>测试数据：</strong><a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/textline_rot180_demo.jpg\">示例图片</a></li>\n <li><strong>硬件配置：</strong>\n <ul>\n <li>GPU：NVIDIA A100 40G</li>\n@@ -476,11 +513,11 @@ for res in output:\n <li><strong>软件环境：</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n \n ### 5.2 权重转换 {#52-权重转换}\n \n-使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 文本行方向分类模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/textline_orientation_classification.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。\n+使用推理引擎时，系统会自动下载官方预训练模型。若需使用自训练模型配合 `paddle_dynamic` 或 `transformers` 引擎，请参考 [PaddleX 文本行方向分类模块权重转换](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/textline_orientation_classification.html#442) 部分，将 `pdparams` 格式通过 PaddleX 转换为 `safetensors` 格式，即可无缝集成到 PaddleOCR 的 API 中进行推理。若需使用自训练模型配合`onnxruntime`引擎，请参考[PaddleX 获取 ONNX 模型](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html)获取onnx模型，即可无缝集成到 PaddleOCR 的 API 中进行推理。\ndiff --git a/docs/version3.x/pipeline_usage/OCR.en.md b/docs/version3.x/pipeline_usage/OCR.en.md\nindex 48728e32a33..ac4041272aa 100644\n--- a/docs/version3.x/pipeline_usage/OCR.en.md\n+++ b/docs/version3.x/pipeline_usage/OCR.en.md\n@@ -705,6 +705,7 @@ Before using the general OCR pipeline locally, install the dependencies for your\n \n - If you use the local `paddle_static` inference engine, install PaddlePaddle by following [PaddlePaddle Framework Installation](../paddlepaddle_installation.en.md).\n - If you use the `transformers` inference engine, install the required dependencies by following [Inference Engine and Configuration](../inference_deployment/local_inference/inference_engine.en.md).\n+- If you use the `onnxruntime` inference engine, install the required dependencies by following [Inference Engine and Configuration](../inference_deployment/local_inference/inference_engine.en.md).\n \n #### 2.0.2 Install paddleocr\n \n@@ -770,6 +771,19 @@ paddleocr ocr -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_im\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured by following [Inference Engine and Configuration](../inference_deployment/local_inference/inference_engine.en.md), and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr ocr -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png \\\n+ --use_doc_orientation_classify False \\\n+ --use_doc_unwarping False \\\n+ --use_textline_orientation False \\\n+ --save_path ./output \\\n+ --device gpu:0 \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n <details><summary><b>Command line supports more parameter settings. Click to expand for detailed instructions on command line parameters.</b></summary>\n@@ -2577,6 +2591,35 @@ for res in result:\n res.save_to_json(\"output\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured by following [Inference Engine and Configuration](../inference_deployment/local_inference/inference_engine.en.md), and then run the following code:\n+\n+```python\n+from paddleocr import PaddleOCR\n+\n+ocr = PaddleOCR(\n+ use_doc_orientation_classify=False, # Disable document orientation classification\n+ use_doc_unwarping=False, # Disable document unwarping\n+ use_textline_orientation=False, # Disable textline orientation classification\n+ engine=\"onnxruntime\",\n+)\n+# ocr = PaddleOCR(lang=\"en\", engine=\"onnxruntime\") # Use the English model\n+# ocr = PaddleOCR(ocr_version=\"PP-OCRv4\", engine=\"onnxruntime\") # Use another PP-OCR version\n+# ocr = PaddleOCR(device=\"gpu\", engine=\"onnxruntime\") # Use GPU for inference\n+# ocr = PaddleOCR(\n+# text_detection_model_name=\"PP-OCRv5_server_det\",\n+# text_recognition_model_name=\"PP-OCRv5_server_rec\",\n+# use_doc_orientation_classify=False,\n+# use_doc_unwarping=False,\n+# use_textline_orientation=False,\n+# engine=\"onnxruntime\",\n+# ) # Switch to the PP-OCRv5_server models\n+result = ocr.predict(\"./general_ocr_002.png\")\n+for res in result:\n+ res.print()\n+ res.save_to_img(\"output\")\n+ res.save_to_json(\"output\")\n+```\n+\n In most scenarios, the default `paddle_static` inference engine delivers better inference performance and is the recommended first choice.\n \n #### 4.2.2 Specify the local model path through the configuration file\ndiff --git a/docs/version3.x/pipeline_usage/OCR.md b/docs/version3.x/pipeline_usage/OCR.md\nindex 8d1ff7bf759..97ad6cefede 100644\n--- a/docs/version3.x/pipeline_usage/OCR.md\n+++ b/docs/version3.x/pipeline_usage/OCR.md\n@@ -713,6 +713,7 @@ devanagari_PP-OCRv3_mobile_rec_infer.tar\">推理模型</a>/<a href=\"https://padd\n \n - 如需使用本地推理引擎 `paddle_static`，请参考[飞桨框架安装说明](../paddlepaddle_installation.md)安装 PaddlePaddle。\n - 如需使用 `transformers` 推理引擎，请参考[推理引擎文档](../inference_deployment/local_inference/inference_engine.md)安装相关依赖。\n+- 如需使用 `onnxruntime` 推理引擎，请参考[推理引擎文档](../inference_deployment/local_inference/inference_engine.md)安装相关依赖。\n \n #### 2.0.2 安装 paddleocr\n \n@@ -800,6 +801,19 @@ paddleocr ocr -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_im\n --engine transformers\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请先参考[推理引擎文档](../inference_deployment/local_inference/inference_engine.md)完成 ONNX Runtime 环境配置，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+paddleocr ocr -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png \\\n+ --use_doc_orientation_classify False \\\n+ --use_doc_unwarping False \\\n+ --use_textline_orientation False \\\n+ --save_path ./output \\\n+ --device gpu:0 \\\n+ --engine onnxruntime\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n \n <!-- Luxorion-12修改：表格参数描述 -->\n@@ -1257,6 +1271,35 @@ for res in result:\n res.save_to_json(\"output\")\n ```\n \n+如果选择 `onnxruntime` 作为推理引擎，请先参考[推理引擎文档](../inference_deployment/local_inference/inference_engine.md)完成 ONNX Runtime 环境配置，然后执行如下代码：\n+\n+```python\n+from paddleocr import PaddleOCR\n+\n+ocr = PaddleOCR(\n+ use_doc_orientation_classify=False, # 通过 use_doc_orientation_classify 参数指定不使用文档方向分类模型\n+ use_doc_unwarping=False, # 通过 use_doc_unwarping 参数指定不使用文本图像矫正模型\n+ use_textline_orientation=False, # 通过 use_textline_orientation 参数指定不使用文本行方向分类模型\n+ engine=\"onnxruntime\",\n+)\n+# ocr = PaddleOCR(lang=\"en\", engine=\"onnxruntime\") # 通过 lang 参数来使用英文模型\n+# ocr = PaddleOCR(ocr_version=\"PP-OCRv4\", engine=\"onnxruntime\") # 通过 ocr_version 参数来使用 PP-OCR 其他版本\n+# ocr = PaddleOCR(device=\"gpu\", engine=\"onnxruntime\") # 通过 device 参数使得在模型推理时使用 GPU\n+# ocr = PaddleOCR(\n+# text_detection_model_name=\"PP-OCRv5_server_det\",\n+# text_recognition_model_name=\"PP-OCRv5_server_rec\",\n+# use_doc_orientation_classify=False,\n+# use_doc_unwarping=False,\n+# use_textline_orientation=False,\n+# engine=\"onnxruntime\",\n+# ) # 更换 PP-OCRv5_server 模型\n+result = ocr.predict(\"./general_ocr_002.png\")\n+for res in result:\n+ res.print()\n+ res.save_to_img(\"output\")\n+ res.save_to_json(\"output\")\n+```\n+\n 在大多数场景下，默认的 `paddle_static` 推理引擎通常具备更好的推理性能，建议优先使用。\n <!-- Luxorion-12 修改：表格参数说明-->\n 在上述 Python 脚本中，执行了如下几个步骤：\ndiff --git a/docs/version3.x/pipeline_usage/PP-StructureV3.en.md b/docs/version3.x/pipeline_usage/PP-StructureV3.en.md\nindex bebb796e855..67604513331 100644\n--- a/docs/version3.x/pipeline_usage/PP-StructureV3.en.md\n+++ b/docs/version3.x/pipeline_usage/PP-StructureV3.en.md\n@@ -990,9 +990,17 @@ If you choose `transformers` as the inference engine, make sure the Transformers\n \n ```bash\n # Use the transformers engine for inference\n-# Some models are still being supported. For inference, please disable formula recognition and replace the wireless table structure recognition model using the following command:\n paddleocr pp_structurev3 -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/pp_structure_v3_demo.png \\\n- --engine transformers --use_formula_recognition False --wireless_table_structure_recognition_model_name SLANeXt_wireless\n+ --engine transformers\n+```\n+\n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured by following [Inference Engine and Configuration](../inference_deployment/local_inference/inference_engine.en.md), and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+# Some models are still being supported. For inference, please disable formula recognition using the following command:\n+paddleocr pp_structurev3 -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/pp_structure_v3_demo.png \\\n+ --engine onnxruntime --use_formula_recognition False\n ```\n \n <details><summary><b>Command line supports more parameters. Click to expand for detailed parameter descriptions</b></summary>\n@@ -1698,11 +1706,30 @@ If you choose `transformers` as the inference engine, make sure the Transformers\n ```python\n from paddleocr import PPStructureV3\n \n-# Some models are still being supported. For inference, please disable formula recognition and replace the wireless table structure recognition model using the following code:\n pipeline = PPStructureV3(\n engine=\"transformers\",\n+)\n+# pipeline = PPStructureV3(lang=\"en\") # Set the lang parameter to use the English text recognition model. For other supported languages, see Section 5: Appendix. By default, both Chinese and English text recognition models are enabled.\n+# pipeline = PPStructureV3(use_doc_orientation_classify=True) # Use use_doc_orientation_classify to enable/disable document orientation classification model\n+# pipeline = PPStructureV3(use_doc_unwarping=True) # Use use_doc_unwarping to enable/disable document unwarping module\n+# pipeline = PPStructureV3(use_textline_orientation=True) # Use use_textline_orientation to enable/disable textline orientation classification model\n+# pipeline = PPStructureV3(device=\"gpu\") # Use device to specify GPU for model inference\n+output = pipeline.predict(\"./pp_structure_v3_demo.png\")\n+for res in output:\n+ res.print() ## Print the structured prediction output\n+ res.save_to_json(save_path=\"output\") ## Save the current image's structured result in JSON format\n+ res.save_to_markdown(save_path=\"output\") ## Save the current image's result in Markdown format\n+```\n+\n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured by following [Inference Engine and Configuration](../inference_deployment/local_inference/inference_engine.en.md), and then run the following code:\n+\n+```python\n+from paddleocr import PPStructureV3\n+\n+# Some models are still being supported. For inference, please disable formula recognition using the following code:\n+pipeline = PPStructureV3(\n+ engine=\"onnxruntime\",\n use_formula_recognition=False,\n- wireless_table_structure_recognition_model_name=\"SLANeXt_wireless\",\n )\n # pipeline = PPStructureV3(lang=\"en\") # Set the lang parameter to use the English text recognition model. For other supported languages, see Section 5: Appendix. By default, both Chinese and English text recognition models are enabled.\n # pipeline = PPStructureV3(use_doc_orientation_classify=True) # Use use_doc_orientation_classify to enable/disable document orientation classification model\ndiff --git a/docs/version3.x/pipeline_usage/PP-StructureV3.md b/docs/version3.x/pipeline_usage/PP-StructureV3.md\nindex 0f6d9a3a173..93725f90ef6 100644\n--- a/docs/version3.x/pipeline_usage/PP-StructureV3.md\n+++ b/docs/version3.x/pipeline_usage/PP-StructureV3.md\n@@ -1069,10 +1069,18 @@ paddleocr pp_structurev3 -i ./pp_structure_v3_demo.png --device gpu\n 如果选择 `transformers` 作为推理引擎，请先参考[推理引擎文档](../inference_deployment/local_inference/inference_engine.md)完成 Transformers 环境配置，然后执行如下命令：\n \n ```bash\n-# 使用 transformers 引擎进行推理\n # 部分模型尚在支持中，推理时需关闭公式识别功能并更换无线表格结构识别模型，请使用以下命令：\n paddleocr pp_structurev3 -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/pp_structure_v3_demo.png \\\n- --engine transformers --use_formula_recognition False --wireless_table_structure_recognition_model_name SLANeXt_wireless\n+ --engine transformers\n+```\n+\n+如果选择 `onnxruntime` 作为推理引擎，请先参考[推理引擎文档](../inference_deployment/local_inference/inference_engine.md)完成 ONNX Runtime 环境配置，然后执行如下命令：\n+\n+```bash\n+# 使用 onnxruntime 引擎进行推理\n+# 部分模型尚在支持中，推理时需关闭公式识别功能，请使用以下命令：\n+paddleocr pp_structurev3 -i https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/pp_structure_v3_demo.png \\\n+ --engine onnxruntime --use_formula_recognition False\n ```\n \n <details><summary><b>命令行支持更多参数设置，点击展开以查看命令行参数的详细说明</b></summary>\n@@ -1670,11 +1678,30 @@ for res in output:\n ```python\n from paddleocr import PPStructureV3\n \n-# 部分模型尚在支持中，推理时需关闭公式识别功能并更换无线表格结构识别模型，请使用以下代码：\n pipeline = PPStructureV3(\n engine=\"transformers\",\n+)\n+# pipeline = PPStructureV3(lang=\"en\") # 将 lang 参数设置为使用英文文本识别模型。对于其他支持的语言，请参阅第5节：附录部分。默认配置为中英文模型。\n+# pipeline = PPStructureV3(use_doc_orientation_classify=True) # 通过 use_doc_orientation_classify 指定是否使用文档方向分类模型\n+# pipeline = PPStructureV3(use_doc_unwarping=True) # 通过 use_doc_unwarping 指定是否使用文本图像矫正模块\n+# pipeline = PPStructureV3(use_textline_orientation=True) # 通过 use_textline_orientation 指定是否使用文本行方向分类模型\n+# pipeline = PPStructureV3(device=\"gpu\") # 通过 device 指定模型推理时使用 GPU\n+output = pipeline.predict(\"./pp_structure_v3_demo.png\")\n+for res in output:\n+ res.print() ## 打印预测的结构化输出\n+ res.save_to_json(save_path=\"output\") ## 保存当前图像的结构化json结果\n+ res.save_to_markdown(save_path=\"output\") ## 保存当前图像的markdown格式的结果\n+```\n+\n+如果选择 `onnxruntime` 作为推理引擎，请先参考[推理引擎文档](../inference_deployment/local_inference/inference_engine.md)完成 ONNX Runtime 环境配置，然后执行如下代码：\n+\n+```python\n+from paddleocr import PPStructureV3\n+\n+# 部分模型尚在支持中，推理时需关闭公式识别功能，请使用以下代码：\n+pipeline = PPStructureV3(\n+ engine=\"onnxruntime\",\n use_formula_recognition=False,\n- wireless_table_structure_recognition_model_name=\"SLANeXt_wireless\",\n )\n # pipeline = PPStructureV3(lang=\"en\") # 将 lang 参数设置为使用英文文本识别模型。对于其他支持的语言，请参阅第5节：附录部分。默认配置为中英文模型。\n # pipeline = PPStructureV3(use_doc_orientation_classify=True) # 通过 use_doc_orientation_classify 指定是否使用文档方向分类模型\n"},"changed_files":{"kind":"list like","value":["docs/version3.x/inference_deployment/local_inference/inference_engine.en.md","docs/version3.x/inference_deployment/local_inference/inference_engine.md","docs/version3.x/module_usage/chart_parsing.en.md","docs/version3.x/module_usage/chart_parsing.md","docs/version3.x/module_usage/doc_img_orientation_classification.en.md","docs/version3.x/module_usage/doc_img_orientation_classification.md","docs/version3.x/module_usage/layout_analysis.en.md","docs/version3.x/module_usage/layout_analysis.md","docs/version3.x/module_usage/layout_detection.en.md","docs/version3.x/module_usage/layout_detection.md","docs/version3.x/module_usage/table_cells_detection.en.md","docs/version3.x/module_usage/table_cells_detection.md","docs/version3.x/module_usage/table_classification.en.md","docs/version3.x/module_usage/table_classification.md","docs/version3.x/module_usage/table_structure_recognition.en.md","docs/version3.x/module_usage/table_structure_recognition.md","docs/version3.x/module_usage/text_detection.en.md","docs/version3.x/module_usage/text_detection.md","docs/version3.x/module_usage/text_image_unwarping.en.md","docs/version3.x/module_usage/text_image_unwarping.md","docs/version3.x/module_usage/text_recognition.en.md","docs/version3.x/module_usage/text_recognition.md","docs/version3.x/module_usage/textline_orientation_classification.en.md","docs/version3.x/module_usage/textline_orientation_classification.md","docs/version3.x/pipeline_usage/OCR.en.md","docs/version3.x/pipeline_usage/OCR.md","docs/version3.x/pipeline_usage/PP-StructureV3.en.md","docs/version3.x/pipeline_usage/PP-StructureV3.md"],"string":"[\n \"docs/version3.x/inference_deployment/local_inference/inference_engine.en.md\",\n \"docs/version3.x/inference_deployment/local_inference/inference_engine.md\",\n \"docs/version3.x/module_usage/chart_parsing.en.md\",\n \"docs/version3.x/module_usage/chart_parsing.md\",\n \"docs/version3.x/module_usage/doc_img_orientation_classification.en.md\",\n \"docs/version3.x/module_usage/doc_img_orientation_classification.md\",\n \"docs/version3.x/module_usage/layout_analysis.en.md\",\n \"docs/version3.x/module_usage/layout_analysis.md\",\n \"docs/version3.x/module_usage/layout_detection.en.md\",\n \"docs/version3.x/module_usage/layout_detection.md\",\n \"docs/version3.x/module_usage/table_cells_detection.en.md\",\n \"docs/version3.x/module_usage/table_cells_detection.md\",\n \"docs/version3.x/module_usage/table_classification.en.md\",\n \"docs/version3.x/module_usage/table_classification.md\",\n \"docs/version3.x/module_usage/table_structure_recognition.en.md\",\n \"docs/version3.x/module_usage/table_structure_recognition.md\",\n \"docs/version3.x/module_usage/text_detection.en.md\",\n \"docs/version3.x/module_usage/text_detection.md\",\n \"docs/version3.x/module_usage/text_image_unwarping.en.md\",\n \"docs/version3.x/module_usage/text_image_unwarping.md\",\n \"docs/version3.x/module_usage/text_recognition.en.md\",\n \"docs/version3.x/module_usage/text_recognition.md\",\n \"docs/version3.x/module_usage/textline_orientation_classification.en.md\",\n \"docs/version3.x/module_usage/textline_orientation_classification.md\",\n \"docs/version3.x/pipeline_usage/OCR.en.md\",\n \"docs/version3.x/pipeline_usage/OCR.md\",\n \"docs/version3.x/pipeline_usage/PP-StructureV3.en.md\",\n \"docs/version3.x/pipeline_usage/PP-StructureV3.md\"\n]","isConversational":false},"review_threads":{"kind":"list like","value":[{"comment":"ONNXRuntime 官方名字两个单词是分开写的：ONNX Runtime\n\nhttps://onnxruntime.ai/docs/","path":"docs/version3.x/module_usage/chart_parsing.en.md","hunk":"@@ -52,6 +52,14 @@ paddleocr chart_parsing -i \"{'image': 'https://paddle-model-ecology.bj.bcebos.co\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNXRuntime environment is configured, and then run the following command:","resolving_sha":"e02cc271977066ac7974fd2f6ae8e567b9559204","resolving_diff":"diff --git a/docs/version3.x/module_usage/chart_parsing.en.md b/docs/version3.x/module_usage/chart_parsing.en.md\nindex 655cc713f70..76412e8dade 100644\n--- a/docs/version3.x/module_usage/chart_parsing.en.md\n+++ b/docs/version3.x/module_usage/chart_parsing.en.md\n@@ -52,6 +52,14 @@ paddleocr chart_parsing -i \"{'image': 'https://paddle-model-ecology.bj.bcebos.co\n --engine transformers\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\n+\n+```bash\n+# Use the onnxruntime engine for inference\n+paddleocr chart_parsing -i \"{'image': 'https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/chart_parsing_02.png'}\" \\\n+ --engine onnxruntime\n+```\n+\n In most scenarios, the default `paddle_dynamic` inference engine delivers better inference performance and is the recommended first choice.\n \n **Note:** By default, PaddleOCR retrieves models from HuggingFace. If HuggingFace access is restricted in your environment, you can switch the model source to BOS by setting the environment variable: `PADDLE_PDX_MODEL_SOURCE=\"BOS\"`. Support for more mainstream sources is planned.\n@@ -89,6 +97,23 @@ for res in results:\n res.save_to_json(f\"./output/res.json\")\n ```\n \n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\n+\n+```python\n+from paddleocr import ChartParsing\n+model = ChartParsing(\n+ model_name=\"PP-Chart2Table\",\n+ engine=\"onnxruntime\",\n+)\n+results = model.predict(\n+ input={\"image\": \"chart_parsing_02.png\"},\n+ batch_size=1\n+)\n+for res in results:\n+ res.print()\n+ res.save_to_json(f\"./output/res.json\")\n+```\n+\n In most scenarios, the default `paddle_dynamic` inference engine delivers better inference performance and is the recommended first choice.\n \n The output result will be:\n@@ -155,7 +180,7 @@ Defaults to GPU 0 if available; otherwise falls back to CPU.\n </tr>\n <tr>\n <td><code>engine</code></td>\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_dynamic</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_dynamic</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\"../inference_deployment/local_inference/inference_engine.en.md\">Inference Engine and Configuration</a>.</td>\n <td><code>str|None</code></td>\n <td><code>None</code></td>\n </tr>\n@@ -315,7 +340,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n \n <strong>Test Environment Description:</strong>\n <ul>\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/medal_table.jpg)</li>\n+ <li><strong>Test Data:</strong> <a href=\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/medal_table.jpg\">Sample Image</a></li>\n <li><strong>Hardware Configuration:</strong>\n <ul>\n <li>GPU: NVIDIA A100 40G</li>\n@@ -325,7 +350,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\n <li><strong>Software Environment:</strong>\n <ul>\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\n </ul>\n </li>\n </ul>\n"}],"string":"[\n {\n \"comment\": \"ONNXRuntime 官方名字两个单词是分开写的：ONNX Runtime\\n\\nhttps://onnxruntime.ai/docs/\",\n \"path\": \"docs/version3.x/module_usage/chart_parsing.en.md\",\n \"hunk\": \"@@ -52,6 +52,14 @@ paddleocr chart_parsing -i \\\"{'image': 'https://paddle-model-ecology.bj.bcebos.co\\n --engine transformers\\n ```\\n \\n+If you choose `onnxruntime` as the inference engine, make sure the ONNXRuntime environment is configured, and then run the following command:\",\n \"resolving_sha\": \"e02cc271977066ac7974fd2f6ae8e567b9559204\",\n \"resolving_diff\": \"diff --git a/docs/version3.x/module_usage/chart_parsing.en.md b/docs/version3.x/module_usage/chart_parsing.en.md\\nindex 655cc713f70..76412e8dade 100644\\n--- a/docs/version3.x/module_usage/chart_parsing.en.md\\n+++ b/docs/version3.x/module_usage/chart_parsing.en.md\\n@@ -52,6 +52,14 @@ paddleocr chart_parsing -i \\\"{'image': 'https://paddle-model-ecology.bj.bcebos.co\\n --engine transformers\\n ```\\n \\n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following command:\\n+\\n+```bash\\n+# Use the onnxruntime engine for inference\\n+paddleocr chart_parsing -i \\\"{'image': 'https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/chart_parsing_02.png'}\\\" \\\\\\n+ --engine onnxruntime\\n+```\\n+\\n In most scenarios, the default `paddle_dynamic` inference engine delivers better inference performance and is the recommended first choice.\\n \\n **Note:** By default, PaddleOCR retrieves models from HuggingFace. If HuggingFace access is restricted in your environment, you can switch the model source to BOS by setting the environment variable: `PADDLE_PDX_MODEL_SOURCE=\\\"BOS\\\"`. Support for more mainstream sources is planned.\\n@@ -89,6 +97,23 @@ for res in results:\\n res.save_to_json(f\\\"./output/res.json\\\")\\n ```\\n \\n+If you choose `onnxruntime` as the inference engine, make sure the ONNX Runtime environment is configured, and then run the following code:\\n+\\n+```python\\n+from paddleocr import ChartParsing\\n+model = ChartParsing(\\n+ model_name=\\\"PP-Chart2Table\\\",\\n+ engine=\\\"onnxruntime\\\",\\n+)\\n+results = model.predict(\\n+ input={\\\"image\\\": \\\"chart_parsing_02.png\\\"},\\n+ batch_size=1\\n+)\\n+for res in results:\\n+ res.print()\\n+ res.save_to_json(f\\\"./output/res.json\\\")\\n+```\\n+\\n In most scenarios, the default `paddle_dynamic` inference engine delivers better inference performance and is the recommended first choice.\\n \\n The output result will be:\\n@@ -155,7 +180,7 @@ Defaults to GPU 0 if available; otherwise falls back to CPU.\\n </tr>\\n <tr>\\n <td><code>engine</code></td>\\n-<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_dynamic</code>, and <code>transformers</code>. When left as <code>None</code>, local inference uses the <code>paddle_dynamic</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\\\"../inference_deployment/local_inference/inference_engine.en.md\\\">Inference Engine and Configuration</a>.</td>\\n+<td><b>Meaning:</b> Inference engine.<br/><b>Description:</b> Supports <code>None</code> (the default), <code>paddle</code>, <code>paddle_dynamic</code>, <code>transformers</code>, and <code>onnxruntime</code>. When left as <code>None</code>, local inference uses the <code>paddle_dynamic</code> engine by default. For detailed descriptions, supported values, compatibility rules, and examples, see <a href=\\\"../inference_deployment/local_inference/inference_engine.en.md\\\">Inference Engine and Configuration</a>.</td>\\n <td><code>str|None</code></td>\\n <td><code>None</code></td>\\n </tr>\\n@@ -315,7 +340,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\\n \\n <strong>Test Environment Description:</strong>\\n <ul>\\n- <li><strong>Test Data:</strong> [Sample Image](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/medal_table.jpg)</li>\\n+ <li><strong>Test Data:</strong> <a href=\\\"https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/medal_table.jpg\\\">Sample Image</a></li>\\n <li><strong>Hardware Configuration:</strong>\\n <ul>\\n <li>GPU: NVIDIA A100 40G</li>\\n@@ -325,7 +350,7 @@ For detailed descriptions, values, compatibility rules, and examples of the infe\\n <li><strong>Software Environment:</strong>\\n <ul>\\n <li>Ubuntu 22.04 / CUDA 12.6 / cuDNN 9.5</li>\\n- <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10</li>\\n+ <li>paddlepaddle-gpu 3.2.1 / paddleocr 3.5 / transformers 5.4.0 / torch 2.10 / onnxruntime-gpu 1.23.2</li>\\n </ul>\\n </li>\\n </ul>\\n\"\n }\n]","isConversational":false},"test_patch":{"kind":"string","value":""},"merged":{"kind":"bool","value":true,"string":"true"}}},{"rowIdx":8999,"cells":{"repo":{"kind":"string","value":"PaddlePaddle/PaddleOCR"},"pr_number":{"kind":"number","value":18121,"string":"18,121"},"task_type":{"kind":"string","value":"issue_to_patch"},"issue_text":{"kind":"string","value":""},"pr_title":{"kind":"string","value":"[Feat] Add PP-OCRv6 Android Demo"},"pr_body":{"kind":"string","value":""},"base_sha":{"kind":"string","value":"1e5aa0ad31bc8a82cd8e1daef7adc24e577d2534"},"head_sha":{"kind":"string","value":"35b1339167286d0a5e4872e620daf06b115c3ec9"},"gold_diff":{"kind":"string","value":"diff --git a/deploy/ppocr-android/.gitignore b/deploy/ppocr-android/.gitignore\nnew file mode 100644\nindex 00000000000..62d8c741289\n--- /dev/null\n+++ b/deploy/ppocr-android/.gitignore\n@@ -0,0 +1,93 @@\n+# OS generated files\n+.DS_Store\n+.DS_Store?\n+._*\n+.Spotlight-V100\n+.Trashes\n+ehthumbs.db\n+Thumbs.db\n+\n+# Built application files\n+*.apk\n+*.aar\n+*.ap_\n+*.aab\n+\n+# Files for the ART/Dalvik VM\n+*.dex\n+\n+# Java class files\n+*.class\n+\n+# Generated files\n+bin/\n+gen/\n+out/\n+release/\n+\n+# Gradle files\n+.gradle/\n+build/\n+\n+# Local configuration file (sdk path, etc)\n+local.properties\n+\n+# Proguard folder generated by Eclipse\n+proguard/\n+\n+# Log Files\n+*.log\n+\n+# Android Studio Navigation editor temp files\n+.navigation/\n+\n+# Android Studio captures folder\n+captures/\n+\n+# IntelliJ\n+*.iml\n+.idea/\n+\n+# Keystore files\n+# Uncomment the following lines if you do not want to check your keystore files in.\n+#*.jks\n+#*.keystore\n+\n+# External native build folder generated in Android Studio 2.2 and later\n+.externalNativeBuild\n+.cxx/\n+\n+# Google Services (e.g. APIs or Firebase)\n+google-services.json\n+\n+# Freeline\n+freeline.py\n+freeline/\n+freeline_project_description.json\n+\n+# fastlane\n+fastlane/report.xml\n+fastlane/Preview.html\n+fastlane/screenshots\n+fastlane/test_output\n+fastlane/readme.md\n+\n+# Version control\n+vcs.xml\n+\n+# lint\n+lint/intermediates/\n+lint/generated/\n+lint/outputs/\n+lint/tmp/\n+\n+# Secrets\n+*.env\n+secrets.properties\n+\n+# Kotlin\n+.kotlin/\n+\n+# Model\n+*.onnx\n+*.yml\ndiff --git a/deploy/ppocr-android/README.md b/deploy/ppocr-android/README.md\nnew file mode 100644\nindex 00000000000..254c048afda\n--- /dev/null\n+++ b/deploy/ppocr-android/README.md\n@@ -0,0 +1,5 @@\n+# PaddleOCR Android Demo\n+\n+中文 | [English](./README_en.md)\n+\n+请查看 [文档](../../docs/version3.x/inference_deployment/cross_platform/android_deployment.md)。\ndiff --git a/deploy/ppocr-android/README_en.md b/deploy/ppocr-android/README_en.md\nnew file mode 100644\nindex 00000000000..6e669ff5903\n--- /dev/null\n+++ b/deploy/ppocr-android/README_en.md\n@@ -0,0 +1,5 @@\n+# PaddleOCR Android Demo\n+\n+[中文](./README.md) | English\n+\n+Please refer to the [documentation](../../docs/version3.x/inference_deployment/cross_platform/android_deployment.en.md).\ndiff --git a/deploy/ppocr-android/app/build.gradle.kts b/deploy/ppocr-android/app/build.gradle.kts\nnew file mode 100644\nindex 00000000000..4d5b99179fd\n--- /dev/null\n+++ b/deploy/ppocr-android/app/build.gradle.kts\n@@ -0,0 +1,54 @@\n+plugins {\n+ alias(libs.plugins.android.application)\n+ alias(libs.plugins.kotlin.android)\n+ alias(libs.plugins.kotlin.compose)\n+}\n+\n+android {\n+ namespace = \"com.paddle.ocr.demo\"\n+ compileSdk = 35\n+\n+ defaultConfig {\n+ applicationId = \"com.paddle.ocr.demo\"\n+ minSdk = 26\n+ targetSdk = 35\n+ versionCode = 1\n+ versionName = \"1.0.0\"\n+ }\n+\n+ buildTypes {\n+ release {\n+ isMinifyEnabled = true\n+ proguardFiles(\n+ getDefaultProguardFile(\"proguard-android-optimize.txt\"),\n+ \"proguard-rules.pro\"\n+ )\n+ }\n+ }\n+ compileOptions {\n+ sourceCompatibility = JavaVersion.VERSION_17\n+ targetCompatibility = JavaVersion.VERSION_17\n+ }\n+ kotlin {\n+ compilerOptions {\n+ jvmTarget.set(org.jetbrains.kotlin.gradle.dsl.JvmTarget.JVM_17)\n+ }\n+ }\n+}\n+\n+dependencies {\n+ implementation(project(\":ppocr-sdk\"))\n+\n+ implementation(platform(libs.compose.bom))\n+ implementation(libs.compose.ui)\n+ implementation(libs.compose.ui.graphics)\n+ implementation(libs.compose.ui.tooling.preview)\n+ implementation(libs.compose.material3)\n+ implementation(libs.androidx.activity.compose)\n+ implementation(libs.androidx.lifecycle.runtime.ktx)\n+ implementation(libs.androidx.lifecycle.viewmodel.compose)\n+ implementation(libs.androidx.core.ktx)\n+ implementation(libs.coil.compose)\n+\n+ debugImplementation(libs.compose.ui.tooling)\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/AndroidManifest.xml b/deploy/ppocr-android/app/src/main/AndroidManifest.xml\nnew file mode 100644\nindex 00000000000..fc8c983326c\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/AndroidManifest.xml\n@@ -0,0 +1,27 @@\n+<?xml version=\"1.0\" encoding=\"utf-8\"?>\n+<manifest xmlns:android=\"http://schemas.android.com/apk/res/android\"\n+ xmlns:tools=\"http://schemas.android.com/tools\">\n+\n+ <uses-permission android:name=\"android.permission.READ_MEDIA_IMAGES\" />\n+ <uses-permission android:name=\"android.permission.READ_EXTERNAL_STORAGE\"\n+ android:maxSdkVersion=\"32\" />\n+\n+ <application\n+ android:name=\".OCRApplication\"\n+ android:allowBackup=\"true\"\n+ android:icon=\"@mipmap/ic_launcher\"\n+ android:label=\"PP-OCRv6 Demo\"\n+ android:supportsRtl=\"true\"\n+ android:theme=\"@style/Theme.PaddleOCR\"\n+ tools:targetApi=\"35\">\n+ <activity\n+ android:name=\".MainActivity\"\n+ android:exported=\"true\"\n+ android:theme=\"@style/Theme.PaddleOCR\">\n+ <intent-filter>\n+ <action android:name=\"android.intent.action.MAIN\" />\n+ <category android:name=\"android.intent.category.LAUNCHER\" />\n+ </intent-filter>\n+ </activity>\n+ </application>\n+</manifest>\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/MainActivity.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/MainActivity.kt\nnew file mode 100644\nindex 00000000000..c82dfbced1e\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/MainActivity.kt\n@@ -0,0 +1,34 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo\n+\n+import android.os.Bundle\n+import androidx.activity.ComponentActivity\n+import androidx.activity.compose.setContent\n+import androidx.activity.enableEdgeToEdge\n+import com.paddle.ocr.demo.ui.screen.HomeScreen\n+import com.paddle.ocr.demo.ui.theme.PPOCRTheme\n+\n+class MainActivity : ComponentActivity() {\n+ override fun onCreate(savedInstanceState: Bundle?) {\n+ super.onCreate(savedInstanceState)\n+ enableEdgeToEdge()\n+ setContent {\n+ PPOCRTheme {\n+ HomeScreen()\n+ }\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/OCRApplication.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/OCRApplication.kt\nnew file mode 100644\nindex 00000000000..22446f94aa1\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/OCRApplication.kt\n@@ -0,0 +1,109 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo\n+\n+import android.app.Application\n+import android.util.Log\n+import com.paddle.ocr.PaddleOCR\n+import com.paddle.ocr.PaddleOCRConfig\n+import com.paddle.ocr.util.OpenCVUtils\n+import kotlinx.coroutines.CancellationException\n+import kotlinx.coroutines.CoroutineScope\n+import kotlinx.coroutines.Dispatchers\n+import kotlinx.coroutines.SupervisorJob\n+import kotlinx.coroutines.flow.MutableStateFlow\n+import kotlinx.coroutines.flow.StateFlow\n+import kotlinx.coroutines.flow.asStateFlow\n+import kotlinx.coroutines.launch\n+\n+class OCRApplication : Application() {\n+\n+ sealed class ModelState {\n+ data object Loading : ModelState()\n+ data class Ready(val ocr: PaddleOCR) : ModelState()\n+ data class Error(val message: String) : ModelState()\n+ }\n+\n+ private val applicationScope = CoroutineScope(SupervisorJob() + Dispatchers.IO)\n+ private val _modelState = MutableStateFlow<ModelState>(ModelState.Loading)\n+\n+ val modelState: StateFlow<ModelState> = _modelState.asStateFlow()\n+\n+ val ocr: PaddleOCR?\n+ get() = (_modelState.value as? ModelState.Ready)?.ocr\n+\n+ val isModelLoaded: Boolean\n+ get() = _modelState.value is ModelState.Ready\n+\n+ override fun onCreate() {\n+ super.onCreate()\n+ instance = this\n+ loadModels()\n+ }\n+\n+ fun retryLoadModels() {\n+ if (_modelState.value is ModelState.Loading) return\n+ loadModels()\n+ }\n+\n+ private fun loadModels() {\n+ _modelState.value = ModelState.Loading\n+ applicationScope.launch {\n+ try {\n+ if (!OpenCVUtils.init(this@OCRApplication)) {\n+ throw IllegalStateException(\"Failed to initialize OpenCV native library\")\n+ }\n+ val loadedOcr = PaddleOCR.create(\n+ context = this@OCRApplication,\n+ config = PaddleOCRConfig(\n+ recScoreThresh = 0.0f,\n+ recBatchSize = 1,\n+ ),\n+ )\n+ _modelState.value = ModelState.Ready(loadedOcr)\n+ } catch (t: Throwable) {\n+ if (t is CancellationException) throw t\n+ val message = modelLoadErrorMessage(t)\n+ Log.e(TAG, message, t)\n+ _modelState.value = ModelState.Error(message)\n+ }\n+ }\n+ }\n+\n+ private fun modelLoadErrorMessage(error: Throwable): String {\n+ val details = listOfNotNull(error.message, error.cause?.message)\n+ .distinct()\n+ .joinToString(\": \")\n+ return if (details.isBlank()) {\n+ \"Failed to load OCR models\"\n+ } else {\n+ \"Failed to load OCR models: $details\"\n+ }\n+ }\n+\n+ override fun onTerminate() {\n+ super.onTerminate()\n+ applicationScope.launch {\n+ ocr?.release()\n+ }\n+ }\n+\n+ companion object {\n+ private const val TAG = \"OCRApplication\"\n+\n+ lateinit var instance: OCRApplication\n+ private set\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ErrorDialog.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ErrorDialog.kt\nnew file mode 100644\nindex 00000000000..e7429fbece6\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ErrorDialog.kt\n@@ -0,0 +1,58 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.component\n+\n+import androidx.compose.foundation.layout.Column\n+import androidx.compose.foundation.layout.Spacer\n+import androidx.compose.foundation.layout.fillMaxWidth\n+import androidx.compose.foundation.layout.height\n+import androidx.compose.material3.AlertDialog\n+import androidx.compose.material3.MaterialTheme\n+import androidx.compose.material3.Text\n+import androidx.compose.material3.TextButton\n+import androidx.compose.runtime.Composable\n+import androidx.compose.ui.Modifier\n+import androidx.compose.ui.unit.dp\n+\n+@Composable\n+fun ErrorDialog(\n+ message: String,\n+ onRetry: () -> Unit,\n+ onDismiss: () -> Unit,\n+) {\n+ AlertDialog(\n+ onDismissRequest = onDismiss,\n+ title = { Text(\"Error\") },\n+ text = {\n+ Column(modifier = Modifier.fillMaxWidth()) {\n+ Text(\n+ text = message,\n+ style = MaterialTheme.typography.bodyMedium,\n+ )\n+ Spacer(modifier = Modifier.height(16.dp))\n+ }\n+ },\n+ confirmButton = {\n+ TextButton(onClick = onRetry) {\n+ Text(\"Retry\")\n+ }\n+ },\n+ dismissButton = {\n+ TextButton(onClick = onDismiss) {\n+ Text(\"Dismiss\")\n+ }\n+ },\n+ )\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePicker.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePicker.kt\nnew file mode 100644\nindex 00000000000..50588152f9f\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePicker.kt\n@@ -0,0 +1,83 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.component\n+\n+import androidx.compose.foundation.Image\n+import androidx.compose.foundation.clickable\n+import androidx.compose.foundation.layout.Arrangement\n+import androidx.compose.foundation.layout.Column\n+import androidx.compose.foundation.layout.Row\n+import androidx.compose.foundation.layout.Spacer\n+import androidx.compose.foundation.layout.fillMaxWidth\n+import androidx.compose.foundation.layout.height\n+import androidx.compose.foundation.layout.padding\n+import androidx.compose.foundation.layout.size\n+import androidx.compose.foundation.shape.RoundedCornerShape\n+import androidx.compose.material.icons.Icons\n+import androidx.compose.material.icons.filled.Add\n+import androidx.compose.material3.Button\n+import androidx.compose.material3.Icon\n+import androidx.compose.material3.MaterialTheme\n+import androidx.compose.material3.Text\n+import androidx.compose.runtime.Composable\n+import androidx.compose.ui.Alignment\n+import androidx.compose.ui.Modifier\n+import androidx.compose.ui.draw.clip\n+import androidx.compose.ui.layout.ContentScale\n+import androidx.compose.ui.res.painterResource\n+import androidx.compose.ui.unit.dp\n+\n+@Composable\n+fun ImagePicker(\n+ onGalleryClick: () -> Unit,\n+ onSampleClick: (Int) -> Unit,\n+ sampleImages: List<Int>,\n+ modifier: Modifier = Modifier,\n+) {\n+ Column(\n+ modifier = modifier\n+ .fillMaxWidth()\n+ .padding(16.dp),\n+ horizontalAlignment = Alignment.CenterHorizontally,\n+ ) {\n+ Button(onClick = onGalleryClick) {\n+ Icon(Icons.Default.Add, contentDescription = null)\n+ Text(\" Select from Gallery\", modifier = Modifier.padding(start = 8.dp))\n+ }\n+\n+ if (sampleImages.isNotEmpty()) {\n+ Spacer(modifier = Modifier.height(16.dp))\n+ Text(\"Or try a sample:\", style = MaterialTheme.typography.bodySmall)\n+ Spacer(modifier = Modifier.height(8.dp))\n+ Row(\n+ modifier = Modifier.fillMaxWidth(),\n+ horizontalArrangement = Arrangement.Center,\n+ ) {\n+ sampleImages.forEach { resId ->\n+ Image(\n+ painter = painterResource(resId),\n+ contentDescription = \"Sample image\",\n+ modifier = Modifier\n+ .size(64.dp)\n+ .padding(4.dp)\n+ .clip(RoundedCornerShape(8.dp))\n+ .clickable { onSampleClick(resId) },\n+ contentScale = ContentScale.Crop,\n+ )\n+ }\n+ }\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePreview.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePreview.kt\nnew file mode 100644\nindex 00000000000..327426c46c7\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePreview.kt\n@@ -0,0 +1,135 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.component\n+\n+import android.graphics.Bitmap\n+import android.graphics.Paint\n+import androidx.compose.foundation.Canvas\n+import androidx.compose.foundation.Image\n+import androidx.compose.foundation.layout.Box\n+import androidx.compose.foundation.layout.aspectRatio\n+import androidx.compose.foundation.layout.fillMaxWidth\n+import androidx.compose.runtime.Composable\n+import androidx.compose.runtime.remember\n+import androidx.compose.ui.Modifier\n+import androidx.compose.ui.geometry.Offset\n+import androidx.compose.ui.graphics.Color\n+import androidx.compose.ui.graphics.ImageBitmap\n+import androidx.compose.ui.graphics.Path\n+import androidx.compose.ui.graphics.asImageBitmap\n+import androidx.compose.ui.graphics.drawscope.Stroke\n+import androidx.compose.ui.graphics.nativeCanvas\n+import androidx.compose.ui.layout.ContentScale\n+import com.paddle.ocr.model.OCRResult\n+import kotlin.math.max\n+\n+private val BOX_COLORS = listOf(\n+ Color(0xFFE53935),\n+ Color(0xFF1E88E5),\n+ Color(0xFF43A047),\n+ Color(0xFFFB8C00),\n+ Color(0xFF8E24AA),\n+ Color(0xFF00ACC1),\n+ Color(0xFFD81B60),\n+ Color(0xFF6D4C41),\n+)\n+\n+@Composable\n+fun ImagePreview(\n+ bitmap: Bitmap,\n+ results: List<OCRResult>,\n+ modifier: Modifier = Modifier,\n+) {\n+ val imageBitmap: ImageBitmap = bitmap.asImageBitmap()\n+ val aspect = bitmap.width.toFloat() / bitmap.height.toFloat()\n+\n+ // Paints for index badges\n+ val numberPaint = remember {\n+ Paint().apply {\n+ color = android.graphics.Color.WHITE\n+ textSize = 36f\n+ isAntiAlias = true\n+ isFakeBoldText = true\n+ }\n+ }\n+ val badgePaint = remember {\n+ Paint().apply {\n+ color = 0xCC000000.toInt() // semi-transparent black\n+ isAntiAlias = true\n+ style = Paint.Style.FILL\n+ }\n+ }\n+\n+ Box(modifier = modifier) {\n+ Image(\n+ bitmap = imageBitmap,\n+ contentDescription = \"OCR Input\",\n+ contentScale = ContentScale.Fit,\n+ modifier = Modifier\n+ .fillMaxWidth()\n+ .aspectRatio(aspect),\n+ )\n+ if (results.isNotEmpty()) {\n+ Canvas(modifier = Modifier.matchParentSize()) {\n+ // Canvas matches Image exactly (both have same aspectRatio),\n+ // and ContentScale.Fit fills the Image composable with the bitmap\n+ // (same aspect ratio → no letterboxing), so coordinates map 1:1.\n+ val scaleX = size.width / bitmap.width\n+ val scaleY = size.height / bitmap.height\n+\n+ results.forEachIndexed { index, result ->\n+ val color = BOX_COLORS[index % BOX_COLORS.size]\n+ val path = Path().apply {\n+ result.box.points.forEachIndexed { i, pt ->\n+ val x = pt.x * scaleX\n+ val y = pt.y * scaleY\n+ if (i == 0) moveTo(x, y) else lineTo(x, y)\n+ }\n+ close()\n+ }\n+ // Box fill\n+ drawPath(path, color = color.copy(alpha = 0.15f))\n+ // Box outline\n+ drawPath(path, color = color, style = Stroke(width = 3f))\n+\n+ // Number badge: rounded rect background + white text\n+ val topLeft = result.box.points.minByOrNull { it.y }?.let { pt ->\n+ Offset(pt.x * scaleX, pt.y * scaleY)\n+ } ?: Offset.Zero\n+\n+ val label = \"${index + 1}\"\n+ val textWidth = numberPaint.measureText(label)\n+ val badgeH = 44f\n+ val badgeW = textWidth + 20f\n+ val badgeX = topLeft.x\n+ val badgeY = max(0f, topLeft.y - badgeH)\n+\n+ // Draw badge background\n+ drawContext.canvas.nativeCanvas.drawRoundRect(\n+ badgeX, badgeY, badgeX + badgeW, badgeY + badgeH,\n+ 8f, 8f, badgePaint,\n+ )\n+ // Draw number centered in badge\n+ drawContext.canvas.nativeCanvas.drawText(\n+ label,\n+ badgeX + (badgeW - textWidth) / 2f,\n+ badgeY + badgeH / 2f + 12f, // center vertically\n+ numberPaint,\n+ )\n+ }\n+ }\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/LoadingOverlay.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/LoadingOverlay.kt\nnew file mode 100644\nindex 00000000000..a02c3aaab75\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/LoadingOverlay.kt\n@@ -0,0 +1,44 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.component\n+\n+import androidx.compose.foundation.background\n+import androidx.compose.foundation.layout.Box\n+import androidx.compose.foundation.layout.Column\n+import androidx.compose.foundation.layout.Spacer\n+import androidx.compose.foundation.layout.fillMaxSize\n+import androidx.compose.foundation.layout.height\n+import androidx.compose.foundation.layout.size\n+import androidx.compose.material3.CircularProgressIndicator\n+import androidx.compose.material3.MaterialTheme\n+import androidx.compose.material3.Text\n+import androidx.compose.runtime.Composable\n+import androidx.compose.ui.Alignment\n+import androidx.compose.ui.Modifier\n+import androidx.compose.ui.unit.dp\n+\n+@Composable\n+fun LoadingOverlay(message: String = \"Loading models...\", modifier: Modifier = Modifier) {\n+ Box(\n+ modifier = modifier.fillMaxSize().background(MaterialTheme.colorScheme.background),\n+ contentAlignment = Alignment.Center,\n+ ) {\n+ Column(horizontalAlignment = Alignment.CenterHorizontally) {\n+ CircularProgressIndicator(modifier = Modifier.size(48.dp))\n+ Spacer(modifier = Modifier.height(16.dp))\n+ Text(text = message, style = MaterialTheme.typography.bodyLarge)\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ResultList.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ResultList.kt\nnew file mode 100644\nindex 00000000000..4d032bcfa66\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ResultList.kt\n@@ -0,0 +1,82 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.component\n+\n+import androidx.compose.foundation.layout.Column\n+import androidx.compose.foundation.layout.Row\n+import androidx.compose.foundation.layout.Spacer\n+import androidx.compose.foundation.layout.fillMaxWidth\n+import androidx.compose.foundation.layout.padding\n+import androidx.compose.foundation.layout.width\n+import androidx.compose.material3.HorizontalDivider\n+import androidx.compose.material3.MaterialTheme\n+import androidx.compose.material3.Text\n+import androidx.compose.material3.TextButton\n+import androidx.compose.runtime.Composable\n+import androidx.compose.ui.Alignment\n+import androidx.compose.ui.Modifier\n+import androidx.compose.ui.unit.dp\n+import com.paddle.ocr.model.OCRResult\n+\n+@Composable\n+fun ResultList(\n+ results: List<OCRResult>,\n+ onCopyAll: () -> Unit,\n+ modifier: Modifier = Modifier,\n+) {\n+ Column(modifier = modifier.padding(horizontal = 16.dp)) {\n+ Row(\n+ modifier = Modifier.fillMaxWidth(),\n+ verticalAlignment = Alignment.CenterVertically,\n+ ) {\n+ Text(\n+ text = \"Results (${results.size})\",\n+ style = MaterialTheme.typography.titleMedium,\n+ modifier = Modifier.weight(1f),\n+ )\n+ TextButton(onClick = onCopyAll) {\n+ Text(\"Copy All\")\n+ }\n+ }\n+ results.forEachIndexed { index, result ->\n+ Row(\n+ modifier = Modifier\n+ .fillMaxWidth()\n+ .padding(vertical = 8.dp),\n+ verticalAlignment = Alignment.CenterVertically,\n+ ) {\n+ Text(\n+ text = \"${index + 1}.\",\n+ style = MaterialTheme.typography.bodySmall,\n+ modifier = Modifier.width(24.dp),\n+ )\n+ Text(\n+ text = result.text,\n+ style = MaterialTheme.typography.bodyLarge,\n+ modifier = Modifier.weight(1f),\n+ )\n+ Spacer(modifier = Modifier.width(8.dp))\n+ Text(\n+ text = \"${\"%.2f\".format(result.confidence * 100)}%\",\n+ style = MaterialTheme.typography.labelSmall,\n+ color = MaterialTheme.colorScheme.secondary,\n+ )\n+ }\n+ if (index < results.lastIndex) {\n+ HorizontalDivider()\n+ }\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/TimingBar.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/TimingBar.kt\nnew file mode 100644\nindex 00000000000..09a14b60ec8\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/TimingBar.kt\n@@ -0,0 +1,60 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.component\n+\n+import androidx.compose.foundation.layout.Arrangement\n+import androidx.compose.foundation.layout.Row\n+import androidx.compose.foundation.layout.fillMaxWidth\n+import androidx.compose.foundation.layout.padding\n+import androidx.compose.material3.AssistChip\n+import androidx.compose.material3.MaterialTheme\n+import androidx.compose.material3.Text\n+import androidx.compose.runtime.Composable\n+import androidx.compose.ui.Modifier\n+import androidx.compose.ui.unit.dp\n+\n+@Composable\n+fun TimingBar(\n+ detectionMs: Long,\n+ recognitionMs: Long,\n+ totalMs: Long,\n+ modifier: Modifier = Modifier,\n+) {\n+ Row(\n+ modifier = modifier\n+ .fillMaxWidth()\n+ .padding(horizontal = 16.dp, vertical = 8.dp),\n+ horizontalArrangement = Arrangement.spacedBy(8.dp),\n+ ) {\n+ AssistChip(\n+ onClick = {},\n+ label = {\n+ Text(\"Det: ${detectionMs}ms\", style = MaterialTheme.typography.labelSmall)\n+ },\n+ )\n+ AssistChip(\n+ onClick = {},\n+ label = {\n+ Text(\"Rec: ${recognitionMs}ms\", style = MaterialTheme.typography.labelSmall)\n+ },\n+ )\n+ AssistChip(\n+ onClick = {},\n+ label = {\n+ Text(\"Total: ${totalMs}ms\", style = MaterialTheme.typography.labelSmall)\n+ },\n+ )\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/screen/HomeScreen.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/screen/HomeScreen.kt\nnew file mode 100644\nindex 00000000000..b1d0223d5cf\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/screen/HomeScreen.kt\n@@ -0,0 +1,106 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.screen\n+\n+import androidx.activity.compose.rememberLauncherForActivityResult\n+import androidx.activity.result.PickVisualMediaRequest\n+import androidx.activity.result.contract.ActivityResultContracts\n+import androidx.compose.foundation.layout.Column\n+import androidx.compose.foundation.layout.Spacer\n+import androidx.compose.foundation.layout.fillMaxSize\n+import androidx.compose.foundation.layout.fillMaxWidth\n+import androidx.compose.foundation.layout.height\n+import androidx.compose.foundation.rememberScrollState\n+import androidx.compose.foundation.verticalScroll\n+import androidx.compose.runtime.Composable\n+import androidx.compose.runtime.collectAsState\n+import androidx.compose.runtime.getValue\n+import androidx.compose.ui.Modifier\n+import androidx.compose.ui.unit.dp\n+import androidx.lifecycle.viewmodel.compose.viewModel\n+import com.paddle.ocr.demo.ui.component.*\n+import com.paddle.ocr.demo.ui.viewmodel.OCRViewModel\n+\n+@Composable\n+fun HomeScreen(viewModel: OCRViewModel = viewModel()) {\n+ val state by viewModel.uiState.collectAsState()\n+ val timing by viewModel.timing.collectAsState()\n+\n+ val galleryLauncher = rememberLauncherForActivityResult(\n+ contract = ActivityResultContracts.PickVisualMedia()\n+ ) { uri -> uri?.let { viewModel.onImageSelected(it) } }\n+\n+ Column(modifier = Modifier.fillMaxSize()) {\n+ when (val s = state) {\n+ is OCRViewModel.UIState.Loading -> {\n+ LoadingOverlay(\"Loading OCR models...\")\n+ }\n+ is OCRViewModel.UIState.Ready -> {\n+ Column(modifier = Modifier.verticalScroll(rememberScrollState())) {\n+ Spacer(modifier = Modifier.height(32.dp))\n+ ImagePicker(\n+ onGalleryClick = { galleryLauncher.launch(PickVisualMediaRequest(ActivityResultContracts.PickVisualMedia.ImageOnly)) },\n+ onSampleClick = { viewModel.onSampleImageClicked(it) },\n+ sampleImages = emptyList(),\n+ )\n+ }\n+ }\n+ is OCRViewModel.UIState.Processing -> {\n+ Column(modifier = Modifier.verticalScroll(rememberScrollState())) {\n+ Spacer(modifier = Modifier.height(16.dp))\n+ ImagePreview(bitmap = s.bitmap, results = emptyList())\n+ Spacer(modifier = Modifier.height(16.dp))\n+ LoadingOverlay(\"Processing...\")\n+ }\n+ }\n+ is OCRViewModel.UIState.Result -> {\n+ Column(modifier = Modifier.verticalScroll(rememberScrollState())) {\n+ ImagePreview(\n+ bitmap = s.bitmap,\n+ results = s.result.results,\n+ modifier = Modifier.fillMaxWidth(),\n+ )\n+ if (timing != null) {\n+ TimingBar(\n+ detectionMs = timing!!.detectionMs,\n+ recognitionMs = timing!!.recognitionMs,\n+ totalMs = timing!!.totalMs,\n+ )\n+ }\n+ Spacer(modifier = Modifier.height(4.dp))\n+ ResultList(\n+ results = s.result.results,\n+ onCopyAll = { viewModel.copyAllResults(s.result.results) },\n+ )\n+ Spacer(modifier = Modifier.height(8.dp))\n+ ImagePicker(\n+ onGalleryClick = { galleryLauncher.launch(PickVisualMediaRequest(ActivityResultContracts.PickVisualMedia.ImageOnly)) },\n+ onSampleClick = { viewModel.onSampleImageClicked(it) },\n+ sampleImages = emptyList(),\n+ )\n+ Spacer(modifier = Modifier.height(16.dp))\n+ }\n+ }\n+ is OCRViewModel.UIState.Error -> {\n+ LoadingOverlay(\"Error occurred\")\n+ ErrorDialog(\n+ message = s.message,\n+ onRetry = { viewModel.retry() },\n+ onDismiss = { viewModel.retry() },\n+ )\n+ }\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/theme/Theme.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/theme/Theme.kt\nnew file mode 100644\nindex 00000000000..84a4b19f0e9\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/theme/Theme.kt\n@@ -0,0 +1,45 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.theme\n+\n+import androidx.compose.material3.MaterialTheme\n+import androidx.compose.material3.lightColorScheme\n+import androidx.compose.runtime.Composable\n+import androidx.compose.ui.graphics.Color\n+\n+private val LightColorScheme = lightColorScheme(\n+ primary = Color(0xFF1565C0),\n+ onPrimary = Color.White,\n+ primaryContainer = Color(0xFFD1E4FF),\n+ onPrimaryContainer = Color(0xFF001D36),\n+ secondary = Color(0xFF006874),\n+ onSecondary = Color.White,\n+ secondaryContainer = Color(0xFF9EEFF7),\n+ onSecondaryContainer = Color(0xFF001F24),\n+ surface = Color(0xFFFDFBFF),\n+ onSurface = Color(0xFF1A1C1E),\n+ background = Color(0xFFFDFBFF),\n+ onBackground = Color(0xFF1A1C1E),\n+ error = Color(0xFFBA1A1A),\n+ onError = Color.White,\n+)\n+\n+@Composable\n+fun PPOCRTheme(content: @Composable () -> Unit) {\n+ MaterialTheme(\n+ colorScheme = LightColorScheme,\n+ content = content,\n+ )\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/viewmodel/OCRViewModel.kt b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/viewmodel/OCRViewModel.kt\nnew file mode 100644\nindex 00000000000..e3845cec292\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/viewmodel/OCRViewModel.kt\n@@ -0,0 +1,178 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.demo.ui.viewmodel\n+\n+import android.graphics.Bitmap\n+import android.graphics.BitmapFactory\n+import android.net.Uri\n+import androidx.lifecycle.ViewModel\n+import androidx.lifecycle.viewModelScope\n+import com.paddle.ocr.demo.OCRApplication\n+import com.paddle.ocr.model.OCRResult\n+import com.paddle.ocr.model.OCRRunResult\n+import kotlinx.coroutines.Dispatchers\n+import kotlinx.coroutines.flow.MutableStateFlow\n+import kotlinx.coroutines.flow.StateFlow\n+import kotlinx.coroutines.flow.asStateFlow\n+import kotlinx.coroutines.flow.collect\n+import kotlinx.coroutines.launch\n+import kotlinx.coroutines.withContext\n+\n+class OCRViewModel : ViewModel() {\n+\n+ sealed class UIState {\n+ data object Loading : UIState()\n+ data object Ready : UIState()\n+ data class Processing(val bitmap: Bitmap) : UIState()\n+ data class Result(val bitmap: Bitmap, val result: OCRRunResult) : UIState()\n+ data class Error(val message: String) : UIState()\n+ }\n+\n+ data class TimingInfo(\n+ val detectionMs: Long,\n+ val recognitionMs: Long,\n+ val totalMs: Long,\n+ )\n+\n+ private val _uiState = MutableStateFlow<UIState>(\n+ uiStateForModelState(OCRApplication.instance.modelState.value)\n+ )\n+ val uiState: StateFlow<UIState> = _uiState.asStateFlow()\n+\n+ private val _timing = MutableStateFlow<TimingInfo?>(null)\n+ val timing: StateFlow<TimingInfo?> = _timing.asStateFlow()\n+\n+ private fun uiStateForModelState(modelState: OCRApplication.ModelState): UIState {\n+ return when (modelState) {\n+ is OCRApplication.ModelState.Loading -> UIState.Loading\n+ is OCRApplication.ModelState.Ready -> UIState.Ready\n+ is OCRApplication.ModelState.Error -> UIState.Error(modelState.message)\n+ }\n+ }\n+\n+ init {\n+ viewModelScope.launch {\n+ OCRApplication.instance.modelState.collect { modelState ->\n+ when (modelState) {\n+ is OCRApplication.ModelState.Loading -> _uiState.value = UIState.Loading\n+ is OCRApplication.ModelState.Ready -> {\n+ if (_uiState.value is UIState.Loading || _uiState.value is UIState.Error) {\n+ _uiState.value = UIState.Ready\n+ }\n+ }\n+ is OCRApplication.ModelState.Error -> _uiState.value = UIState.Error(modelState.message)\n+ }\n+ }\n+ }\n+ }\n+\n+ fun onImageSelected(uri: Uri) {\n+ viewModelScope.launch {\n+ try {\n+ val bytes = withContext(Dispatchers.IO) {\n+ OCRApplication.instance.contentResolver.openInputStream(uri)?.use { it.readBytes() }\n+ }\n+ if (bytes != null) {\n+ val bitmap = withContext(Dispatchers.IO) {\n+ decodeSampledBitmap(bytes, maxWidth = 2048, maxHeight = 2048)\n+ }\n+ if (bitmap != null) {\n+ processImageBytes(bytes, bitmap)\n+ } else {\n+ _uiState.value = UIState.Error(\"Failed to load image\")\n+ }\n+ } else {\n+ _uiState.value = UIState.Error(\"Failed to load image\")\n+ }\n+ } catch (e: Exception) {\n+ _uiState.value = UIState.Error(e.message ?: \"Unknown error\")\n+ }\n+ }\n+ }\n+\n+ fun onSampleImageClicked(resId: Int) {\n+ viewModelScope.launch {\n+ val bytes = withContext(Dispatchers.IO) {\n+ OCRApplication.instance.resources.openRawResource(resId).use { it.readBytes() }\n+ }\n+ val bitmap = withContext(Dispatchers.IO) {\n+ BitmapFactory.decodeByteArray(bytes, 0, bytes.size)\n+ }\n+ if (bitmap != null) {\n+ processImageBytes(bytes, bitmap)\n+ } else {\n+ _uiState.value = UIState.Error(\"Failed to load image\")\n+ }\n+ }\n+ }\n+\n+ private suspend fun processImageBytes(bytes: ByteArray, bitmap: Bitmap) {\n+ _uiState.value = UIState.Processing(bitmap)\n+\n+ try {\n+ val ocr = OCRApplication.instance.ocr\n+ ?: throw IllegalStateException(\"OCR engine not initialized\")\n+\n+ val result = ocr.recognize(bytes)\n+ _uiState.value = UIState.Result(bitmap, result)\n+ _timing.value = TimingInfo(\n+ detectionMs = result.detectionTimeMs,\n+ recognitionMs = result.recognitionTimeMs,\n+ totalMs = result.totalTimeMs,\n+ )\n+ } catch (e: Exception) {\n+ _uiState.value = UIState.Error(e.message ?: \"OCR failed\")\n+ }\n+ }\n+\n+ fun retry() {\n+ val app = OCRApplication.instance\n+ if (app.isModelLoaded) {\n+ _uiState.value = UIState.Ready\n+ } else {\n+ app.retryLoadModels()\n+ }\n+ }\n+\n+ fun copyAllResults(results: List<OCRResult>) {\n+ val text = results.joinToString(\"\\n\") { it.text }\n+ val clipboard = OCRApplication.instance.getSystemService(android.content.Context.CLIPBOARD_SERVICE)\n+ as android.content.ClipboardManager\n+ clipboard.setPrimaryClip(android.content.ClipData.newPlainText(\"OCR Results\", text))\n+ }\n+\n+ private fun decodeSampledBitmap(\n+ bytes: ByteArray,\n+ maxWidth: Int,\n+ maxHeight: Int,\n+ ): Bitmap? {\n+ // First pass: decode bounds only\n+ val opts = BitmapFactory.Options().apply { inJustDecodeBounds = true }\n+ BitmapFactory.decodeByteArray(bytes, 0, bytes.size, opts)\n+ val (origW, origH) = opts.outWidth to opts.outHeight\n+ if (origW <= 0 || origH <= 0) return null\n+\n+ // Calculate sample size (power of 2)\n+ var sampleSize = 1\n+ while (sampleSize * 2 <= maxOf(origW / maxWidth, origH / maxHeight)) {\n+ sampleSize *= 2\n+ }\n+\n+ // Second pass: decode with sampling\n+ return BitmapFactory.Options().apply {\n+ inSampleSize = sampleSize\n+ }.let { BitmapFactory.decodeByteArray(bytes, 0, bytes.size, it) }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/app/src/main/res/drawable/ic_launcher_foreground.xml b/deploy/ppocr-android/app/src/main/res/drawable/ic_launcher_foreground.xml\nnew file mode 100644\nindex 00000000000..f9a506c939b\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/res/drawable/ic_launcher_foreground.xml\n@@ -0,0 +1,16 @@\n+<?xml version=\"1.0\" encoding=\"utf-8\"?>\n+<vector xmlns:android=\"http://schemas.android.com/apk/res/android\"\n+ android:width=\"108dp\"\n+ android:height=\"108dp\"\n+ android:viewportWidth=\"108\"\n+ android:viewportHeight=\"108\">\n+ <path\n+ android:fillColor=\"#FFFFFF\"\n+ android:pathData=\"M54,28 L54,28 C68.36,28 80,39.64 80,54 L80,54 C80,68.36 68.36,80 54,80 L54,80 C39.64,80 28,68.36 28,54 L28,54 C28,39.64 39.64,28 54,28 Z\" />\n+ <path\n+ android:fillColor=\"#1A73E8\"\n+ android:pathData=\"M46,42 L62,42 L62,46 L50,46 L50,52 L60,52 L60,56 L50,56 L50,62 L62,62 L62,66 L46,66 Z\" />\n+ <path\n+ android:fillColor=\"#1A73E8\"\n+ android:pathData=\"M42,50 L42,46 L38,46 L38,62 L42,62 L42,58 C43.5,60.5 46,62 49,62 C53.97,62 58,57.97 58,54 C58,50.03 53.97,46 49,46 C46,46 43.5,47.5 42,50 Z M49,58 C46.79,58 45,56.21 45,54 C45,51.79 46.79,50 49,50 C51.21,50 53,51.79 53,54 C53,56.21 51.21,58 49,58 Z\" />\n+</vector>\ndiff --git a/deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml b/deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml\nnew file mode 100644\nindex 00000000000..5ed0a2df704\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml\n@@ -0,0 +1,5 @@\n+<?xml version=\"1.0\" encoding=\"utf-8\"?>\n+<adaptive-icon xmlns:android=\"http://schemas.android.com/apk/res/android\">\n+ <background android:drawable=\"@color/ic_launcher_background\"/>\n+ <foreground android:drawable=\"@drawable/ic_launcher_foreground\"/>\n+</adaptive-icon>\ndiff --git a/deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml b/deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml\nnew file mode 100644\nindex 00000000000..5ed0a2df704\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml\n@@ -0,0 +1,5 @@\n+<?xml version=\"1.0\" encoding=\"utf-8\"?>\n+<adaptive-icon xmlns:android=\"http://schemas.android.com/apk/res/android\">\n+ <background android:drawable=\"@color/ic_launcher_background\"/>\n+ <foreground android:drawable=\"@drawable/ic_launcher_foreground\"/>\n+</adaptive-icon>\ndiff --git a/deploy/ppocr-android/app/src/main/res/values/colors.xml b/deploy/ppocr-android/app/src/main/res/values/colors.xml\nnew file mode 100644\nindex 00000000000..f67db2f66b9\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/res/values/colors.xml\n@@ -0,0 +1,4 @@\n+<?xml version=\"1.0\" encoding=\"utf-8\"?>\n+<resources>\n+ <color name=\"ic_launcher_background\">#1A73E8</color>\n+</resources>\ndiff --git a/deploy/ppocr-android/app/src/main/res/values/strings.xml b/deploy/ppocr-android/app/src/main/res/values/strings.xml\nnew file mode 100644\nindex 00000000000..d344644f808\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/res/values/strings.xml\n@@ -0,0 +1,4 @@\n+<?xml version=\"1.0\" encoding=\"utf-8\"?>\n+<resources>\n+ <string name=\"app_name\">PP-OCRv6 Demo</string>\n+</resources>\ndiff --git a/deploy/ppocr-android/app/src/main/res/values/themes.xml b/deploy/ppocr-android/app/src/main/res/values/themes.xml\nnew file mode 100644\nindex 00000000000..e6d6a9d03a1\n--- /dev/null\n+++ b/deploy/ppocr-android/app/src/main/res/values/themes.xml\n@@ -0,0 +1,4 @@\n+<?xml version=\"1.0\" encoding=\"utf-8\"?>\n+<resources>\n+ <style name=\"Theme.PaddleOCR\" parent=\"android:Theme.Material.Light.NoActionBar\" />\n+</resources>\ndiff --git a/deploy/ppocr-android/build.gradle.kts b/deploy/ppocr-android/build.gradle.kts\nnew file mode 100644\nindex 00000000000..81f6a880a24\n--- /dev/null\n+++ b/deploy/ppocr-android/build.gradle.kts\n@@ -0,0 +1,6 @@\n+plugins {\n+ alias(libs.plugins.android.application) apply false\n+ alias(libs.plugins.android.library) apply false\n+ alias(libs.plugins.kotlin.android) apply false\n+ alias(libs.plugins.kotlin.compose) apply false\n+}\ndiff --git a/deploy/ppocr-android/gradle.properties b/deploy/ppocr-android/gradle.properties\nnew file mode 100644\nindex 00000000000..f0a2e55f89e\n--- /dev/null\n+++ b/deploy/ppocr-android/gradle.properties\n@@ -0,0 +1,4 @@\n+org.gradle.jvmargs=-Xmx2048m -Dfile.encoding=UTF-8\n+android.useAndroidX=true\n+kotlin.code.style=official\n+android.nonTransitiveRClass=true\ndiff --git a/deploy/ppocr-android/gradle/libs.versions.toml b/deploy/ppocr-android/gradle/libs.versions.toml\nnew file mode 100644\nindex 00000000000..b13db7c021f\n--- /dev/null\n+++ b/deploy/ppocr-android/gradle/libs.versions.toml\n@@ -0,0 +1,41 @@\n+[versions]\n+agp = \"8.7.3\"\n+kotlin = \"2.1.0\"\n+coreKtx = \"1.15.0\"\n+lifecycleRuntime = \"2.8.7\"\n+activityCompose = \"1.9.3\"\n+composeBom = \"2024.12.01\"\n+onnxruntime = \"1.21.1\"\n+opencv = \"4.5.3\"\n+coroutines = \"1.9.0\"\n+coil = \"2.7.0\"\n+test-runner = \"1.2.0\"\n+test-ext-junit-ktx = \"1.2.1\"\n+test-monitor = \"1.7.2\"\n+junit = \"4.13.2\"\n+\n+[libraries]\n+androidx-core-ktx = { group = \"androidx.core\", name = \"core-ktx\", version.ref = \"coreKtx\" }\n+androidx-lifecycle-runtime-ktx = { group = \"androidx.lifecycle\", name = \"lifecycle-runtime-ktx\", version.ref = \"lifecycleRuntime\" }\n+androidx-lifecycle-viewmodel-compose = { group = \"androidx.lifecycle\", name = \"lifecycle-viewmodel-compose\", version.ref = \"lifecycleRuntime\" }\n+androidx-activity-compose = { group = \"androidx.activity\", name = \"activity-compose\", version.ref = \"activityCompose\" }\n+compose-bom = { group = \"androidx.compose\", name = \"compose-bom\", version.ref = \"composeBom\" }\n+compose-ui = { group = \"androidx.compose.ui\", name = \"ui\" }\n+compose-ui-graphics = { group = \"androidx.compose.ui\", name = \"ui-graphics\" }\n+compose-ui-tooling-preview = { group = \"androidx.compose.ui\", name = \"ui-tooling-preview\" }\n+compose-ui-tooling = { group = \"androidx.compose.ui\", name = \"ui-tooling\" }\n+compose-material3 = { group = \"androidx.compose.material3\", name = \"material3\" }\n+onnxruntime-android = { group = \"com.microsoft.onnxruntime\", name = \"onnxruntime-android\", version.ref = \"onnxruntime\" }\n+opencv-android = { group = \"com.quickbirdstudios\", name = \"opencv\", version.ref = \"opencv\" }\n+kotlinx-coroutines-android = { group = \"org.jetbrains.kotlinx\", name = \"kotlinx-coroutines-android\", version.ref = \"coroutines\" }\n+coil-compose = { group = \"io.coil-kt\", name = \"coil-compose\", version.ref = \"coil\" }\n+androidx-test-runner = { group = \"androidx.test\", name = \"runner\", version.ref = \"test-runner\" }\n+androidx-test-ext-junit-ktx = { group = \"androidx.test.ext\", name = \"junit-ktx\", version.ref = \"test-ext-junit-ktx\" }\n+androidx-test-monitor = { group = \"androidx.test\", name = \"monitor\", version.ref = \"test-monitor\" }\n+junit = { group = \"junit\", name = \"junit\", version.ref = \"junit\" }\n+\n+[plugins]\n+android-application = { id = \"com.android.application\", version.ref = \"agp\" }\n+android-library = { id = \"com.android.library\", version.ref = \"agp\" }\n+kotlin-android = { id = \"org.jetbrains.kotlin.android\", version.ref = \"kotlin\" }\n+kotlin-compose = { id = \"org.jetbrains.kotlin.plugin.compose\", version.ref = \"kotlin\" }\ndiff --git a/deploy/ppocr-android/gradle/wrapper/gradle-wrapper.jar b/deploy/ppocr-android/gradle/wrapper/gradle-wrapper.jar\nnew file mode 100644\nindex 00000000000..2c3521197d7\nBinary files /dev/null and b/deploy/ppocr-android/gradle/wrapper/gradle-wrapper.jar differ\ndiff --git a/deploy/ppocr-android/gradle/wrapper/gradle-wrapper.properties b/deploy/ppocr-android/gradle/wrapper/gradle-wrapper.properties\nnew file mode 100644\nindex 00000000000..09523c0e549\n--- /dev/null\n+++ b/deploy/ppocr-android/gradle/wrapper/gradle-wrapper.properties\n@@ -0,0 +1,7 @@\n+distributionBase=GRADLE_USER_HOME\n+distributionPath=wrapper/dists\n+distributionUrl=https\\://services.gradle.org/distributions/gradle-8.9-bin.zip\n+networkTimeout=10000\n+validateDistributionUrl=true\n+zipStoreBase=GRADLE_USER_HOME\n+zipStorePath=wrapper/dists\ndiff --git a/deploy/ppocr-android/gradlew b/deploy/ppocr-android/gradlew\nnew file mode 100644\nindex 00000000000..6696a4de892\n--- /dev/null\n+++ b/deploy/ppocr-android/gradlew\n@@ -0,0 +1,72 @@\n+#!/bin/sh\n+\n+#\n+# Gradle start up script for POSIX generated by Gradle.\n+#\n+\n+# Attempt to set APP_HOME\n+PRG=\"$0\"\n+while [ -h \"$PRG\" ] ; do\n+ ls=`ls -ld \"$PRG\"`\n+ link=`expr \"$ls\" : '.*-> \\(.*\\)$'`\n+ if expr \"$link\" : '/.*' > /dev/null; then\n+ PRG=\"$link\"\n+ else\n+ PRG=`dirname \"$PRG\"`\"/$link\"\n+ fi\n+done\n+SAVED=\"`pwd`\"\n+cd \"`dirname \\\"$PRG\\\"`/\" >/dev/null\n+APP_HOME=\"`pwd -P`\"\n+cd \"$SAVED\" >/dev/null\n+\n+APP_NAME=\"Gradle\"\n+APP_BASE_NAME=`basename \"$0\"`\n+\n+DEFAULT_JVM_OPTS='-Xmx64m -Xms64m'\n+\n+MAX_FD=\"maximum\"\n+\n+warn () {\n+ echo \"$*\"\n+} >&2\n+\n+die () {\n+ echo\n+ echo \"$*\"\n+ echo\n+ exit 1\n+} >&2\n+\n+OS_NAME=\"`uname`\"\n+case \"$OS_NAME\" in\n+ CYGWIN* | MINGW* | MSYS* )\n+ IS_WINDOWS=true\n+ ;;\n+esac\n+\n+CLASSPATH=$APP_HOME/gradle/wrapper/gradle-wrapper.jar\n+\n+if [ -n \"$JAVA_HOME\" ] ; then\n+ JAVACMD=\"$JAVA_HOME/bin/java\"\n+else\n+ JAVACMD=\"java\"\n+fi\n+\n+if [ ! -x \"$JAVACMD\" ] ; then\n+ die \"ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.\n+\n+Please set the JAVA_HOME variable in your environment to match the\n+location of your Java installation.\"\n+fi\n+\n+GRADLE_OPTS=\"${GRADLE_OPTS} -Dorg.gradle.appname=${APP_BASE_NAME}\"\n+\n+exec \"$JAVACMD\" \\\n+ $DEFAULT_JVM_OPTS \\\n+ $JAVA_OPTS \\\n+ $GRADLE_OPTS \\\n+ \"-Dorg.gradle.appname=$APP_BASE_NAME\" \\\n+ -classpath \"$CLASSPATH\" \\\n+ org.gradle.wrapper.GradleWrapperMain \\\n+ \"$@\"\ndiff --git a/deploy/ppocr-android/ppocr-sdk/build.gradle.kts b/deploy/ppocr-android/ppocr-sdk/build.gradle.kts\nnew file mode 100644\nindex 00000000000..1a220e392e1\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/build.gradle.kts\n@@ -0,0 +1,42 @@\n+plugins {\n+ alias(libs.plugins.android.library)\n+ alias(libs.plugins.kotlin.android)\n+}\n+\n+android {\n+ namespace = \"com.paddle.ocr\"\n+ compileSdk = 35\n+\n+ defaultConfig {\n+ minSdk = 26\n+ testInstrumentationRunner = \"androidx.test.runner.AndroidJUnitRunner\"\n+ }\n+\n+ buildTypes {\n+ release {\n+ isMinifyEnabled = false\n+ consumerProguardFiles(\"proguard-rules.pro\")\n+ }\n+ }\n+ compileOptions {\n+ sourceCompatibility = JavaVersion.VERSION_17\n+ targetCompatibility = JavaVersion.VERSION_17\n+ }\n+ kotlin {\n+ compilerOptions {\n+ jvmTarget.set(org.jetbrains.kotlin.gradle.dsl.JvmTarget.JVM_17)\n+ }\n+ }\n+}\n+\n+dependencies {\n+ implementation(libs.onnxruntime.android)\n+ implementation(libs.opencv.android)\n+ implementation(libs.kotlinx.coroutines.android)\n+ implementation(libs.androidx.core.ktx)\n+\n+ androidTestImplementation(libs.androidx.test.runner)\n+ androidTestImplementation(libs.androidx.test.ext.junit.ktx)\n+ androidTestImplementation(libs.androidx.test.monitor)\n+ androidTestImplementation(libs.junit)\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/proguard-rules.pro b/deploy/ppocr-android/ppocr-sdk/proguard-rules.pro\nnew file mode 100644\nindex 00000000000..5c912dc1b9f\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/proguard-rules.pro\n@@ -0,0 +1,2 @@\n+-keep class com.paddle.ocr.** { *; }\n+-keep class ai.onnxruntime.** { *; }\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/BenchmarkFixtures.kt b/deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/BenchmarkFixtures.kt\nnew file mode 100644\nindex 00000000000..c5fbdea05e2\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/BenchmarkFixtures.kt\n@@ -0,0 +1,19 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.benchmark\n+\n+object BenchmarkFixtures {\n+ const val defaultReferenceImageName = \"android_ocr_benchmark_reference.jpg\"\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/OCRBenchmarkTest.kt b/deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/OCRBenchmarkTest.kt\nnew file mode 100644\nindex 00000000000..b8153020762\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/OCRBenchmarkTest.kt\n@@ -0,0 +1,374 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.benchmark\n+\n+import android.os.Build\n+import android.os.Debug\n+import android.util.Log\n+import androidx.test.ext.junit.runners.AndroidJUnit4\n+import androidx.test.platform.app.InstrumentationRegistry\n+import com.paddle.ocr.EngineConfig\n+import com.paddle.ocr.PaddleOCR\n+import com.paddle.ocr.PaddleOCRConfig\n+import com.paddle.ocr.model.OCRRunResult\n+import com.paddle.ocr.util.OpenCVUtils\n+import kotlinx.coroutines.runBlocking\n+import org.json.JSONArray\n+import org.json.JSONObject\n+import org.junit.Assert.assertTrue\n+import org.junit.BeforeClass\n+import org.junit.Test\n+import org.junit.runner.RunWith\n+import java.io.File\n+import kotlin.math.pow\n+import kotlin.math.sqrt\n+\n+@RunWith(AndroidJUnit4::class)\n+class OCRBenchmarkTest {\n+\n+ companion object {\n+ private const val TAG = \"OCRBenchmark\"\n+\n+ // Instrumentation arguments (passed via -Pandroid.testInstrumentationRunnerArguments.*)\n+ private const val ARG_WARMUP = \"warmup\"\n+ private const val ARG_ITERATIONS = \"iterations\"\n+ private const val ARG_REC_BATCH_SIZE = \"rec_batch_size\"\n+\n+ private const val DEFAULT_WARMUP = 3\n+ private const val DEFAULT_ITERATIONS = 10\n+\n+ private lateinit var ocr: PaddleOCR\n+ private lateinit var outputDir: File\n+\n+ @JvmStatic\n+ @BeforeClass\n+ fun setUp() {\n+ val context = InstrumentationRegistry.getInstrumentation().targetContext\n+ OpenCVUtils.init(context)\n+\n+ outputDir = File(context.getExternalFilesDir(null), \"ocr_benchmark\")\n+ outputDir.mkdirs()\n+ println(\"[OCRBenchmark] Output dir: ${outputDir.absolutePath}\")\n+\n+ runBlocking {\n+ val batchSize = getInstrumentationArgument(ARG_REC_BATCH_SIZE)?.toIntOrNull() ?: 1\n+ val config = PaddleOCRConfig(recBatchSize = batchSize)\n+ ocr = PaddleOCR.create(context, config, EngineConfig())\n+ }\n+ println(\"[OCRBenchmark] Engine ready, coldLoadTimeMs=${ocr.coldLoadTimeMs}\")\n+ }\n+\n+ private fun getInstrumentationArgument(key: String): String? {\n+ val args = InstrumentationRegistry.getArguments()\n+ return args[key] as? String\n+ }\n+\n+ private fun saveAndPrint(filename: String, json: JSONObject) {\n+ val file = File(outputDir, filename)\n+ file.writeText(json.toString(2))\n+ println(\"[OCRBenchmark] Saved: ${file.absolutePath}\")\n+ println(\"[OCRBenchmark] ---- $filename ----\")\n+ json.toString(2).lines().forEach { println(\"[OCRBenchmark] $it\") }\n+ println(\"[OCRBenchmark] ---- end ----\")\n+ }\n+ }\n+\n+ // ========================================================================\n+ // Accuracy\n+ // ========================================================================\n+\n+ @Test\n+ fun testOCRExportJSON() {\n+ val imageBytes = loadBenchmarkImageBytes()\n+ val result = runBlocking { ocr.recognize(imageBytes) }\n+ val payload = buildExportPayload(result)\n+ saveAndPrint(\"accuracy_export.json\", payload)\n+\n+ assertTrue(\"Should have results\", payload.getJSONArray(\"items\").length() > 0)\n+ assertTrue(\"Should detect text\", result.lineCount > 0)\n+ }\n+\n+ // ========================================================================\n+ // Latency\n+ // ========================================================================\n+\n+ @Test\n+ fun testLatencyBenchmark() {\n+ val imageBytes = loadBenchmarkImageBytes()\n+ val warmup = getInstrumentationArgument(ARG_WARMUP)?.toIntOrNull() ?: DEFAULT_WARMUP\n+ val iterations = getInstrumentationArgument(ARG_ITERATIONS)?.toIntOrNull() ?: DEFAULT_ITERATIONS\n+\n+ println(\"[OCRBenchmark] Warming up $warmup iterations...\")\n+ runBlocking { repeat(warmup) { ocr.recognize(imageBytes) } }\n+\n+ println(\"[OCRBenchmark] Measuring $iterations iterations...\")\n+ val totals = mutableListOf<Long>()\n+ val dets = mutableListOf<Long>()\n+ val detPre = mutableListOf<Long>()\n+ val detInf = mutableListOf<Long>()\n+ val detPost = mutableListOf<Long>()\n+ val recs = mutableListOf<Long>()\n+ val recPre = mutableListOf<Long>()\n+ val recInf = mutableListOf<Long>()\n+ val recPost = mutableListOf<Long>()\n+ val overheads = mutableListOf<Long>()\n+ val perLineMs = mutableListOf<Long>()\n+ val detShapes = mutableListOf<List<Int>>()\n+ val recShapes = mutableListOf<List<Int>>()\n+ var firstLineCount = 0\n+\n+ runBlocking {\n+ repeat(iterations) { i ->\n+ val r = ocr.recognize(imageBytes)\n+ if (i == 0) firstLineCount = r.lineCount\n+ totals.add(r.totalTimeMs)\n+ dets.add(r.detectionTimeMs)\n+ detPre.add(r.detPreprocessMs)\n+ detInf.add(r.detInferenceMs)\n+ detPost.add(r.detPostprocessMs)\n+ recs.add(r.recognitionTimeMs)\n+ recPre.add(r.recPreprocessMs)\n+ recInf.add(r.recInferenceMs)\n+ recPost.add(r.recPostprocessMs)\n+ overheads.add(r.pipelineOverheadMs)\n+ detShapes.add(r.detInputShape)\n+ recShapes.addAll(r.recInputShapes)\n+ perLineMs.addAll(r.perLineRecMs)\n+ if ((i + 1) % 5 == 0) println(\"[OCRBenchmark] $i/${iterations}...\")\n+ }\n+ }\n+\n+ val memInfo = Debug.MemoryInfo()\n+ Debug.getMemoryInfo(memInfo)\n+\n+ val payload = JSONObject().apply {\n+ put(\"schemaVersion\", 1)\n+ put(\"deviceModel\", Build.MODEL)\n+ put(\"osVersion\", Build.VERSION.RELEASE)\n+ put(\"ortExecutionProvider\", \"CPU\")\n+ put(\"inputImage\", BenchmarkFixtures.defaultReferenceImageName)\n+ put(\"inputBytes\", imageBytes.size)\n+ put(\"warmupIterations\", warmup)\n+ put(\"measuredIterations\", iterations)\n+ put(\"firstMeasuredLineCount\", firstLineCount)\n+ put(\"coldLoadTimeMs\", ocr.coldLoadTimeMs)\n+ if (totals.isNotEmpty()) {\n+ put(\"firstMeasuredRun\", JSONObject().apply {\n+ put(\"totalTimeMs\", totals.first())\n+ put(\"detectionTimeMs\", dets.first())\n+ put(\"recognitionTimeMs\", recs.first())\n+ put(\"pipelineOverheadTimeMs\", overheads.first())\n+ })\n+ }\n+ put(\"totalTimeMs\", summarizeMs(totals).toJson())\n+ put(\"detectionTimeMs\", summarizeMs(dets).toJson())\n+ put(\"detectionPreprocessTimeMs\", summarizeMs(detPre).toJson())\n+ put(\"detectionInferenceTimeMs\", summarizeMs(detInf).toJson())\n+ put(\"detectionPostprocessTimeMs\", summarizeMs(detPost).toJson())\n+ put(\"recognitionTimeMs\", summarizeMs(recs).toJson())\n+ put(\"recognitionPreprocessTimeMs\", summarizeMs(recPre).toJson())\n+ put(\"recognitionInferenceTimeMs\", summarizeMs(recInf).toJson())\n+ put(\"recognitionPostprocessTimeMs\", summarizeMs(recPost).toJson())\n+ put(\"pipelineOverheadTimeMs\", summarizeMs(overheads).toJson())\n+ put(\"memoryPssKb\", memInfo.totalPss)\n+ put(\"inputShapeDistribution\", JSONObject().apply {\n+ put(\"detection\", summarizeShapes(detShapes))\n+ put(\"recognition\", summarizeShapes(recShapes))\n+ })\n+ if (perLineMs.isNotEmpty()) {\n+ put(\"recognitionPerLine\", JSONObject().apply {\n+ put(\"pooled\", JSONObject().apply {\n+ put(\"count\", perLineMs.size)\n+ put(\"totalMs\", summarizeMs(perLineMs).toJson())\n+ })\n+ })\n+ }\n+ }\n+\n+ saveAndPrint(\"latency_benchmark.json\", payload)\n+ printSpeedSummary(warmup, iterations, totals, dets, detPre, detInf, detPost, recs, recPre, recInf, recPost, overheads, firstLineCount)\n+\n+ assertTrue(\"Should complete all iterations\", totals.size == iterations)\n+ assertTrue(\"Should detect text\", firstLineCount > 0)\n+ }\n+\n+ private fun printSpeedSummary(\n+ warmup: Int,\n+ iterations: Int,\n+ totals: List<Long>,\n+ dets: List<Long>,\n+ detPre: List<Long>,\n+ detInf: List<Long>,\n+ detPost: List<Long>,\n+ recs: List<Long>,\n+ recPre: List<Long>,\n+ recInf: List<Long>,\n+ recPost: List<Long>,\n+ overheads: List<Long>,\n+ lineCount: Int,\n+ ) {\n+ val bar = \"+-----------------------------+----------+----------+----------+----------+\"\n+ val header = \"| Stage | Mean ms | Stdev | P90 | Min ms |\"\n+ fun row(label: String, samples: List<Long>): String {\n+ if (samples.isEmpty()) return \"| %-27s | %8s | %8s | %8s | %8s |\".format(label, \"-\", \"-\", \"-\", \"N/A\")\n+ val mean = samples.map { it.toDouble() }.average()\n+ val variance = if (samples.size > 1) samples.map { (it - mean).pow(2) }.average() else 0.0\n+ val stdev = sqrt(variance)\n+ val sorted = samples.sorted()\n+ val p90 = sorted[(0.9 * (sorted.size - 1)).toInt()]\n+ val min = sorted.first()\n+ return \"| %-27s | %8.2f | %8.2f | %8d | %8d |\".format(label, mean, stdev, p90, min)\n+ }\n+\n+ println(\"[OCRBenchmark] \")\n+ println(\"[OCRBenchmark] ╔════════════════════════════════════════════════════════════════════════════════╗\")\n+ println(\"[OCRBenchmark] ║ PP-OCRv6 Speed Benchmark Results ║\")\n+ println(\"[OCRBenchmark] ╠════════════════════════════════════════════════════════════════════════════════╣\")\n+ println(\"[OCRBenchmark] ║ Device: ${Build.MODEL} | OS: Android ${Build.VERSION.RELEASE} | Lines: $lineCount\")\n+ println(\"[OCRBenchmark] ║ Cold load: ${ocr.coldLoadTimeMs}ms | Warmup: $warmup | Measured: $iterations\")\n+ println(\"[OCRBenchmark] ╠════════════════════════════════════════════════════════════════════════════════╣\")\n+ println(\"[OCRBenchmark] $bar\")\n+ println(\"[OCRBenchmark] $header\")\n+ println(\"[OCRBenchmark] $bar\")\n+ println(\"[OCRBenchmark] ${row(\"Total pipeline\", totals)}\")\n+ println(\"[OCRBenchmark] $bar\")\n+ println(\"[OCRBenchmark] ${row(\" Detection (total)\", dets)}\")\n+ println(\"[OCRBenchmark] ${row(\" Preprocess\", detPre)}\")\n+ println(\"[OCRBenchmark] ${row(\" Inference\", detInf)}\")\n+ println(\"[OCRBenchmark] ${row(\" Postprocess\", detPost)}\")\n+ println(\"[OCRBenchmark] ${row(\" Recognition (total)\", recs)}\")\n+ println(\"[OCRBenchmark] ${row(\" Preprocess\", recPre)}\")\n+ println(\"[OCRBenchmark] ${row(\" Inference\", recInf)}\")\n+ println(\"[OCRBenchmark] ${row(\" Postprocess\", recPost)}\")\n+ println(\"[OCRBenchmark] ${row(\" Pipeline overhead\", overheads)}\")\n+ println(\"[OCRBenchmark] $bar\")\n+ println(\"[OCRBenchmark] ╚════════════════════════════════════════════════════════════════════════════════╝\")\n+ println(\"[OCRBenchmark] \")\n+ }\n+\n+ // ========================================================================\n+ // Memory\n+ // ========================================================================\n+\n+ @Test\n+ fun testMemoryBenchmark() {\n+ val imageBytes = loadBenchmarkImageBytes()\n+ val warmup = getInstrumentationArgument(ARG_WARMUP)?.toIntOrNull() ?: DEFAULT_WARMUP\n+ val iterations = getInstrumentationArgument(ARG_ITERATIONS)?.toIntOrNull() ?: DEFAULT_ITERATIONS\n+\n+ runBlocking { repeat(warmup) { ocr.recognize(imageBytes) } }\n+ Runtime.getRuntime().gc()\n+ Thread.sleep(500)\n+\n+ val before = Debug.MemoryInfo()\n+ Debug.getMemoryInfo(before)\n+\n+ runBlocking { repeat(iterations) { ocr.recognize(imageBytes) } }\n+\n+ val after = Debug.MemoryInfo()\n+ Debug.getMemoryInfo(after)\n+\n+ val payload = JSONObject().apply {\n+ put(\"schemaVersion\", 1)\n+ put(\"deviceModel\", Build.MODEL)\n+ put(\"iterations\", iterations)\n+ put(\"memoryBeforePssKb\", before.totalPss)\n+ put(\"memoryAfterPssKb\", after.totalPss)\n+ put(\"memoryDeltaPssKb\", after.totalPss - before.totalPss)\n+ put(\"nativeBeforeKb\", before.nativePss)\n+ put(\"nativeAfterKb\", after.nativePss)\n+ put(\"dalvikBeforeKb\", before.dalvikPss)\n+ put(\"dalvikAfterKb\", after.dalvikPss)\n+ }\n+\n+ saveAndPrint(\"memory_benchmark.json\", payload)\n+ assertTrue(\"Memory should be measurable\", after.totalPss > 0)\n+ }\n+\n+ // ========================================================================\n+ // Helpers\n+ // ========================================================================\n+\n+ private fun loadBenchmarkImageBytes(): ByteArray {\n+ val context = InstrumentationRegistry.getInstrumentation().context\n+ val stem = BenchmarkFixtures.defaultReferenceImageName.substringBeforeLast(\".\")\n+ val resId = context.resources.getIdentifier(stem, \"raw\", context.packageName)\n+ if (resId == 0) {\n+ throw IllegalStateException(\n+ \"Missing ${BenchmarkFixtures.defaultReferenceImageName} \" +\n+ \"(package=${context.packageName})\"\n+ )\n+ }\n+ return context.resources.openRawResource(resId).use { it.readBytes() }\n+ }\n+\n+ data class TimingSummary(val mean: Double, val stdev: Double, val p90: Double) {\n+ fun toJson() = JSONObject().apply {\n+ put(\"mean\", mean)\n+ put(\"stdev\", stdev)\n+ put(\"p90\", p90)\n+ }\n+ }\n+\n+ private fun summarizeMs(samples: List<Long>): TimingSummary {\n+ if (samples.isEmpty()) return TimingSummary(0.0, 0.0, 0.0)\n+ val doubles = samples.map { it.toDouble() }\n+ val mean = doubles.average()\n+ val variance = if (doubles.size > 1)\n+ doubles.map { (it - mean).pow(2) }.average()\n+ else 0.0\n+ val stdev = sqrt(variance)\n+ val sorted = doubles.sorted()\n+ val p90Idx = (0.9 * (sorted.size - 1)).toInt().coerceIn(0, sorted.size - 1)\n+ return TimingSummary(mean, stdev, sorted[p90Idx])\n+ }\n+\n+ private fun summarizeShapes(shapes: List<List<Int>>): JSONArray {\n+ val counts = mutableMapOf<String, Int>()\n+ shapes.forEach { shape ->\n+ val key = shape.joinToString(\",\")\n+ counts[key] = (counts[key] ?: 0) + 1\n+ }\n+ return JSONArray().apply {\n+ counts.entries.sortedByDescending { it.value }.forEach { (shape, count) ->\n+ put(JSONObject().apply {\n+ put(\"shape\", JSONArray(shape.split(\",\").map { it.toInt() }))\n+ put(\"count\", count)\n+ })\n+ }\n+ }\n+ }\n+\n+ private fun buildExportPayload(result: OCRRunResult): JSONObject {\n+ val items = JSONArray().apply {\n+ result.results.forEach { r ->\n+ put(JSONObject().apply {\n+ put(\"polygon\", JSONArray().apply {\n+ r.box.points.forEach { pt ->\n+ put(JSONArray(listOf(pt.x.toInt(), pt.y.toInt())))\n+ }\n+ })\n+ put(\"text\", r.text)\n+ put(\"score\", r.confidence.toDouble())\n+ })\n+ }\n+ }\n+ return JSONObject().apply {\n+ put(\"schemaVersion\", 1)\n+ put(\"source\", \"android_ocr_demo\")\n+ put(\"items\", items)\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/androidTest/res/raw/android_ocr_benchmark_reference.png b/deploy/ppocr-android/ppocr-sdk/src/androidTest/res/raw/android_ocr_benchmark_reference.png\nnew file mode 100644\nindex 00000000000..3cd2e791820\nBinary files /dev/null and b/deploy/ppocr-android/ppocr-sdk/src/androidTest/res/raw/android_ocr_benchmark_reference.png differ\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/AndroidManifest.xml b/deploy/ppocr-android/ppocr-sdk/src/main/AndroidManifest.xml\nnew file mode 100644\nindex 00000000000..9a40236b947\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/AndroidManifest.xml\n@@ -0,0 +1,3 @@\n+<?xml version=\"1.0\" encoding=\"utf-8\"?>\n+<manifest xmlns:android=\"http://schemas.android.com/apk/res/android\">\n+</manifest>\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/EngineConfig.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/EngineConfig.kt\nnew file mode 100644\nindex 00000000000..54a62d5728d\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/EngineConfig.kt\n@@ -0,0 +1,19 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr\n+\n+data class EngineConfig(\n+ val numThreads: Int = 4,\n+)\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCR.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCR.kt\nnew file mode 100644\nindex 00000000000..6d934906241\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCR.kt\n@@ -0,0 +1,118 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr\n+\n+import android.content.Context\n+import android.graphics.Bitmap\n+import com.paddle.ocr.engine.OCREngine\n+import com.paddle.ocr.engine.OCREngineResult\n+import com.paddle.ocr.model.OCRRunResult\n+import com.paddle.ocr.model.OCRError\n+import kotlinx.coroutines.Dispatchers\n+import kotlinx.coroutines.withContext\n+\n+class PaddleOCR private constructor(\n+ private val engine: OCREngine,\n+) {\n+ /** Time spent loading ONNX models (milliseconds). */\n+ val coldLoadTimeMs: Long get() = engine.coldLoadTimeMs\n+\n+ companion object {\n+\n+ suspend fun create(context: Context): PaddleOCR {\n+ val appContext = context.applicationContext\n+ return withContext(Dispatchers.IO) {\n+ val engine = OCREngine(appContext, PaddleOCRConfig(), EngineConfig())\n+ PaddleOCR(engine)\n+ }\n+ }\n+\n+ suspend fun create(\n+ context: Context,\n+ config: PaddleOCRConfig,\n+ engineConfig: EngineConfig = EngineConfig(),\n+ ): PaddleOCR {\n+ val appContext = context.applicationContext\n+ return withContext(Dispatchers.IO) {\n+ val engine = OCREngine(appContext, config, engineConfig)\n+ PaddleOCR(engine)\n+ }\n+ }\n+\n+ suspend fun create(\n+ context: Context,\n+ config: PaddleOCRConfig,\n+ engineConfig: EngineConfig,\n+ detModelAssetPath: String,\n+ recModelAssetPath: String,\n+ recConfigAssetPath: String,\n+ ): PaddleOCR {\n+ val appContext = context.applicationContext\n+ return withContext(Dispatchers.IO) {\n+ val engine = OCREngine(\n+ appContext, config, engineConfig,\n+ detModelAsset = detModelAssetPath,\n+ recModelAsset = recModelAssetPath,\n+ recConfigAsset = recConfigAssetPath,\n+ )\n+ PaddleOCR(engine)\n+ }\n+ }\n+ }\n+\n+ suspend fun recognize(bitmap: Bitmap): OCRRunResult {\n+ if (bitmap.width == 0 || bitmap.height == 0) {\n+ throw OCRError.InvalidImage()\n+ }\n+ return recognizeResult { engine.run(bitmap) }\n+ }\n+\n+ suspend fun recognize(imageBytes: ByteArray): OCRRunResult {\n+ if (imageBytes.isEmpty()) {\n+ throw OCRError.InvalidImage()\n+ }\n+ return recognizeResult { engine.run(imageBytes) }\n+ }\n+\n+ private suspend fun recognizeResult(runEngine: () -> OCREngineResult): OCRRunResult {\n+ return withContext(Dispatchers.IO) {\n+ val result = runEngine()\n+ OCRRunResult(\n+ results = result.results,\n+ detectionTimeMs = result.detectionTimeMs,\n+ recognitionTimeMs = result.recognitionTimeMs,\n+ totalTimeMs = result.totalTimeMs,\n+ lineCount = result.lineCount,\n+ detPreprocessMs = result.detPreprocessMs,\n+ detInferenceMs = result.detInferenceMs,\n+ detPostprocessMs = result.detPostprocessMs,\n+ recPreprocessMs = result.recPreprocessMs,\n+ recInferenceMs = result.recInferenceMs,\n+ recPostprocessMs = result.recPostprocessMs,\n+ pipelineOverheadMs = result.pipelineOverheadMs,\n+ coldLoadTimeMs = result.coldLoadTimeMs,\n+ detInputShape = result.detInputShape,\n+ recInputShapes = result.recInputShapes,\n+ perLineRecMs = result.perLineRecMs,\n+ )\n+ }\n+ }\n+\n+ suspend fun release() {\n+ withContext(Dispatchers.IO) {\n+ engine.release()\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCRConfig.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCRConfig.kt\nnew file mode 100644\nindex 00000000000..95a63aec2fa\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCRConfig.kt\n@@ -0,0 +1,31 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr\n+\n+data class PaddleOCRConfig(\n+ val detImgMode: String = \"BGR\",\n+ val detLimitSideLen: Int = 64,\n+ val detLimitType: String = \"min\",\n+ val detMaxSideLimit: Int = 4000,\n+ val detThresh: Float = 0.3f,\n+ val detBoxThresh: Float = 0.6f,\n+ val detUnclipRatio: Float = 1.5f,\n+ val detMaxCandidates: Int = 3000,\n+ val detUseDilation: Boolean = false,\n+ val detScoreMode: String = \"fast\",\n+ val detBoxType: String = \"quad\",\n+ val recScoreThresh: Float = 0.0f,\n+ val recBatchSize: Int = 1,\n+)\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/DetectionEngine.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/DetectionEngine.kt\nnew file mode 100644\nindex 00000000000..ab7f2ce4c60\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/DetectionEngine.kt\n@@ -0,0 +1,96 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.engine\n+\n+import android.graphics.Bitmap\n+import com.paddle.ocr.PaddleOCRConfig\n+import com.paddle.ocr.model.OCRBox\n+import com.paddle.ocr.postprocess.DBPostProcessor\n+import com.paddle.ocr.preprocess.DetPreprocessResult\n+import com.paddle.ocr.preprocess.DetPreprocessor\n+import org.opencv.core.Mat\n+\n+class DetectionEngine(\n+ private val ortManager: ORTSessionManager,\n+ private val config: PaddleOCRConfig,\n+) {\n+ data class DetectionResult(\n+ val boxes: List<OCRBox>,\n+ val preprocessMs: Long,\n+ val inferenceMs: Long,\n+ val postprocessMs: Long,\n+ val timeMs: Long,\n+ val inputShape: List<Int>,\n+ )\n+\n+ fun detect(bitmap: Bitmap): DetectionResult {\n+ return detect {\n+ DetPreprocessor.preprocess(\n+ bitmap,\n+ config.detLimitSideLen,\n+ config.detLimitType,\n+ config.detMaxSideLimit,\n+ config.detImgMode,\n+ )\n+ }\n+ }\n+\n+ fun detect(src: Mat): DetectionResult {\n+ return detect {\n+ DetPreprocessor.preprocess(\n+ src,\n+ config.detLimitSideLen,\n+ config.detLimitType,\n+ config.detMaxSideLimit,\n+ config.detImgMode,\n+ )\n+ }\n+ }\n+\n+ private fun detect(preprocessor: () -> DetPreprocessResult): DetectionResult {\n+ val preStart = System.currentTimeMillis()\n+ val preResult = preprocessor()\n+ val preprocessMs = System.currentTimeMillis() - preStart\n+\n+ val infStart = System.currentTimeMillis()\n+ val (outputData, outputShape) = ortManager.runDetection(preResult.tensorData, preResult.shape)\n+ val inferenceMs = System.currentTimeMillis() - infStart\n+\n+ val postStart = System.currentTimeMillis()\n+ val boxes = DBPostProcessor.process(\n+ pred = outputData,\n+ predShape = outputShape,\n+ thresh = config.detThresh,\n+ boxThresh = config.detBoxThresh,\n+ unclipRatio = config.detUnclipRatio,\n+ maxCandidates = config.detMaxCandidates,\n+ useDilation = config.detUseDilation,\n+ scoreMode = config.detScoreMode,\n+ boxType = config.detBoxType,\n+ originalH = preResult.originalH,\n+ originalW = preResult.originalW,\n+ )\n+ val postprocessMs = System.currentTimeMillis() - postStart\n+\n+ return DetectionResult(\n+ boxes = boxes,\n+ preprocessMs = preprocessMs,\n+ inferenceMs = inferenceMs,\n+ postprocessMs = postprocessMs,\n+ timeMs = preprocessMs + inferenceMs + postprocessMs,\n+ inputShape = listOf(1, 3, preResult.shape[2].toInt(), preResult.shape[3].toInt()),\n+ )\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngine.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngine.kt\nnew file mode 100644\nindex 00000000000..77fd8fa57e9\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngine.kt\n@@ -0,0 +1,184 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.engine\n+\n+import android.content.Context\n+import android.graphics.Bitmap\n+import com.paddle.ocr.EngineConfig\n+import com.paddle.ocr.PaddleOCRConfig\n+import com.paddle.ocr.model.ModelConfig\n+import com.paddle.ocr.model.OCRError\n+import com.paddle.ocr.model.OCRResult\n+import com.paddle.ocr.postprocess.BoxSorter\n+import com.paddle.ocr.postprocess.QuadTextCrop\n+import com.paddle.ocr.util.BitmapUtils\n+\n+class OCREngine(\n+ context: Context,\n+ private val config: PaddleOCRConfig,\n+ engineConfig: EngineConfig,\n+ detModelAsset: String = \"models/det/inference.onnx\",\n+ recModelAsset: String = \"models/rec/inference.onnx\",\n+ recConfigAsset: String = \"models/rec/inference.yml\",\n+) {\n+ private val ortManager = ORTSessionManager(context, engineConfig)\n+ private val detectionEngine: DetectionEngine\n+ private val recognitionEngine: RecognitionEngine\n+ val coldLoadTimeMs: Long get() = ortManager.coldLoadTimeMs\n+\n+ init {\n+ val configured = try {\n+ ortManager.loadModels(detModelAsset, recModelAsset)\n+ val recModelConfig = ModelConfig.parse(context, recConfigAsset)\n+ recModelConfig\n+ } catch (t: Throwable) {\n+ ortManager.release()\n+ throw t\n+ }\n+ detectionEngine = DetectionEngine(ortManager, config)\n+ recognitionEngine = RecognitionEngine(ortManager, configured.characterList)\n+ }\n+\n+ fun run(bitmap: Bitmap): OCREngineResult {\n+ val srcMat = BitmapUtils.bitmapToBGRMat(bitmap)\n+ return runWithOwnedMat(srcMat)\n+ }\n+\n+ fun run(imageBytes: ByteArray): OCREngineResult {\n+ val srcMat = BitmapUtils.imdecodeBGR(imageBytes)\n+ if (srcMat.empty()) {\n+ srcMat.release()\n+ throw OCRError.InvalidImage()\n+ }\n+ return runWithOwnedMat(srcMat)\n+ }\n+\n+ private fun runWithOwnedMat(srcMat: org.opencv.core.Mat): OCREngineResult {\n+ return try {\n+ run(srcMat)\n+ } finally {\n+ srcMat.release()\n+ }\n+ }\n+\n+ private fun run(srcMat: org.opencv.core.Mat): OCREngineResult {\n+ val totalStart = System.currentTimeMillis()\n+ val detResult = detectionEngine.detect(srcMat)\n+ val boxes = detResult.boxes\n+\n+ if (boxes.isEmpty()) {\n+ val elapsed = System.currentTimeMillis() - totalStart\n+ return OCREngineResult(\n+ results = emptyList(),\n+ detectionTimeMs = detResult.timeMs,\n+ recognitionTimeMs = 0,\n+ totalTimeMs = elapsed,\n+ lineCount = 0,\n+ detPreprocessMs = detResult.preprocessMs,\n+ detInferenceMs = detResult.inferenceMs,\n+ detPostprocessMs = detResult.postprocessMs,\n+ detInputShape = detResult.inputShape,\n+ coldLoadTimeMs = ortManager.coldLoadTimeMs,\n+ )\n+ }\n+\n+ // 2. Sort boxes\n+ val sortedBoxes = BoxSorter.sortInReadingOrder(boxes)\n+\n+ // 3. Crop and recognize text regions\n+ var totalRecPreMs = 0L\n+ var totalRecInfMs = 0L\n+ var totalRecPostMs = 0L\n+ var totalRecMs = 0L\n+ val allResults = mutableListOf<OCRResult>()\n+ val recInputShapes = mutableListOf<List<Int>>()\n+ val perLineRecMs = mutableListOf<Long>()\n+ val batchSize = config.recBatchSize.coerceAtLeast(1)\n+\n+ var i = 0\n+ while (i < sortedBoxes.size) {\n+ val batchCrops = mutableListOf<org.opencv.core.Mat>()\n+ val batchBoxIndices = mutableListOf<Int>()\n+ var next = i\n+ while (next < sortedBoxes.size && batchCrops.size < batchSize) {\n+ val crop = QuadTextCrop.crop(srcMat, sortedBoxes[next])\n+ if (crop.rows() > 0 && crop.cols() > 0) {\n+ batchCrops.add(crop)\n+ batchBoxIndices.add(next)\n+ } else {\n+ crop.release()\n+ }\n+ next++\n+ }\n+\n+ try {\n+ if (batchCrops.isNotEmpty()) {\n+ val batchResult = recognitionEngine.recognize(batchCrops)\n+ totalRecPreMs += batchResult.preprocessMs\n+ totalRecInfMs += batchResult.inferenceMs\n+ totalRecPostMs += batchResult.postprocessMs\n+ totalRecMs += batchResult.timeMs\n+ recInputShapes.add(batchResult.inputShape)\n+ if (batchSize == 1) {\n+ perLineRecMs.add(batchResult.timeMs)\n+ }\n+\n+ for (j in batchResult.texts.indices) {\n+ val boxIdx = batchBoxIndices[j]\n+ val (text, confidence) = batchResult.texts[j]\n+ if (confidence >= config.recScoreThresh) {\n+ allResults.add(\n+ OCRResult(\n+ box = sortedBoxes[boxIdx],\n+ text = text,\n+ confidence = confidence,\n+ )\n+ )\n+ }\n+ }\n+ }\n+ } finally {\n+ batchCrops.forEach { it.release() }\n+ }\n+ i = next\n+ }\n+\n+ val totalElapsed = System.currentTimeMillis() - totalStart\n+ val pipelineOverhead = totalElapsed - detResult.timeMs - totalRecMs\n+\n+ return OCREngineResult(\n+ results = allResults,\n+ detectionTimeMs = detResult.timeMs,\n+ recognitionTimeMs = totalRecMs,\n+ totalTimeMs = totalElapsed,\n+ lineCount = allResults.size,\n+ detPreprocessMs = detResult.preprocessMs,\n+ detInferenceMs = detResult.inferenceMs,\n+ detPostprocessMs = detResult.postprocessMs,\n+ recPreprocessMs = totalRecPreMs,\n+ recInferenceMs = totalRecInfMs,\n+ recPostprocessMs = totalRecPostMs,\n+ pipelineOverheadMs = pipelineOverhead,\n+ coldLoadTimeMs = ortManager.coldLoadTimeMs,\n+ detInputShape = detResult.inputShape,\n+ recInputShapes = recInputShapes,\n+ perLineRecMs = perLineRecMs,\n+ )\n+ }\n+\n+ fun release() {\n+ ortManager.release()\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngineResult.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngineResult.kt\nnew file mode 100644\nindex 00000000000..e8a240ed276\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngineResult.kt\n@@ -0,0 +1,39 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.engine\n+\n+import com.paddle.ocr.model.OCRResult\n+\n+data class OCREngineResult(\n+ val results: List<OCRResult>,\n+ val detectionTimeMs: Long,\n+ val recognitionTimeMs: Long,\n+ val totalTimeMs: Long,\n+ val lineCount: Int,\n+ // Detailed per-stage timing\n+ val detPreprocessMs: Long = 0,\n+ val detInferenceMs: Long = 0,\n+ val detPostprocessMs: Long = 0,\n+ val recPreprocessMs: Long = 0,\n+ val recInferenceMs: Long = 0,\n+ val recPostprocessMs: Long = 0,\n+ val pipelineOverheadMs: Long = 0,\n+ val coldLoadTimeMs: Long = 0,\n+ // Input tensor shapes\n+ val detInputShape: List<Int> = emptyList(),\n+ val recInputShapes: List<List<Int>> = emptyList(),\n+ // Per-line timing (only populated when recBatchSize == 1)\n+ val perLineRecMs: List<Long> = emptyList(),\n+)\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/ORTSessionManager.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/ORTSessionManager.kt\nnew file mode 100644\nindex 00000000000..f0531040aff\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/ORTSessionManager.kt\n@@ -0,0 +1,161 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.engine\n+\n+import ai.onnxruntime.OnnxTensor\n+import ai.onnxruntime.OrtEnvironment\n+import ai.onnxruntime.OrtSession\n+import android.content.Context\n+import com.paddle.ocr.EngineConfig\n+import com.paddle.ocr.model.OCRError\n+import java.nio.FloatBuffer\n+\n+class ORTSessionManager(\n+ private val context: Context,\n+ private val config: EngineConfig,\n+) {\n+ private var env: OrtEnvironment? = null\n+ private var detSession: OrtSession? = null\n+ private var recSession: OrtSession? = null\n+ private var detInputName: String = \"x\"\n+ private var recInputName: String = \"x\"\n+ var coldLoadTimeMs: Long = 0\n+ private set\n+\n+ fun loadModels(detAssetPath: String, recAssetPath: String) {\n+ val loadStart = System.currentTimeMillis()\n+ env = OrtEnvironment.getEnvironment()\n+ val opts = OrtSession.SessionOptions().apply {\n+ setOptimizationLevel(OrtSession.SessionOptions.OptLevel.ALL_OPT)\n+ setIntraOpNumThreads(config.numThreads)\n+ }\n+ try {\n+ val ortEnv = env ?: throw OCRError.ModelLoadFailed(\"OCR\", Exception(\"Environment not initialized\"))\n+ val detBytes = readModelAsset(detAssetPath)\n+ val recBytes = readModelAsset(recAssetPath)\n+ try {\n+ detSession = ortEnv.createSession(detBytes, opts)\n+ } catch (t: Throwable) {\n+ throw OCRError.ModelLoadFailed(\"detection\", t)\n+ }\n+ try {\n+ recSession = ortEnv.createSession(recBytes, opts)\n+ } catch (t: Throwable) {\n+ detSession?.close()\n+ detSession = null\n+ throw OCRError.ModelLoadFailed(\"recognition\", t)\n+ }\n+\n+ detInputName = try {\n+ detSession!!.inputNames.iterator().next()\n+ } catch (t: Throwable) {\n+ throw OCRError.ModelLoadFailed(\"detection\", t)\n+ }\n+ recInputName = try {\n+ recSession!!.inputNames.iterator().next()\n+ } catch (t: Throwable) {\n+ throw OCRError.ModelLoadFailed(\"recognition\", t)\n+ }\n+ coldLoadTimeMs = System.currentTimeMillis() - loadStart\n+ } finally {\n+ opts.close()\n+ }\n+ }\n+\n+ fun runDetection(input: FloatArray, shape: LongArray): Pair<FloatArray, LongArray> {\n+ val session = detSession\n+ ?: throw OCRError.ModelLoadFailed(\"detection\", Exception(\"Session not initialized\"))\n+ val ortEnv = env\n+ ?: throw OCRError.ModelLoadFailed(\"detection\", Exception(\"Environment not initialized\"))\n+ return runSession(ortEnv, session, detInputName, input, shape, \"detection\")\n+ }\n+\n+ fun runRecognition(input: FloatArray, shape: LongArray): Pair<FloatArray, LongArray> {\n+ val session = recSession\n+ ?: throw OCRError.ModelLoadFailed(\"recognition\", Exception(\"Session not initialized\"))\n+ val ortEnv = env\n+ ?: throw OCRError.ModelLoadFailed(\"recognition\", Exception(\"Environment not initialized\"))\n+ return runSession(ortEnv, session, recInputName, input, shape, \"recognition\")\n+ }\n+\n+ fun release() {\n+ try {\n+ detSession?.close()\n+ } finally {\n+ detSession = null\n+ try {\n+ recSession?.close()\n+ } finally {\n+ recSession = null\n+ env = null\n+ }\n+ }\n+ }\n+\n+ private fun readModelAsset(assetPath: String): ByteArray {\n+ return try {\n+ context.assets.open(assetPath).use { it.readBytes() }\n+ } catch (t: Throwable) {\n+ throw OCRError.ModelNotFound(assetPath, t)\n+ }\n+ }\n+\n+ private fun runSession(\n+ ortEnv: OrtEnvironment,\n+ session: OrtSession,\n+ inputName: String,\n+ input: FloatArray,\n+ shape: LongArray,\n+ modelName: String,\n+ ): Pair<FloatArray, LongArray> {\n+ val tensor = try {\n+ OnnxTensor.createTensor(ortEnv, FloatBuffer.wrap(input), shape)\n+ } catch (t: Throwable) {\n+ throw OCRError.InferenceFailed(modelName, t)\n+ }\n+ val result = try {\n+ try {\n+ session.run(mapOf(inputName to tensor))\n+ } catch (t: Throwable) {\n+ throw OCRError.InferenceFailed(modelName, t)\n+ }\n+ } finally {\n+ tensor.close()\n+ }\n+\n+ return try {\n+ try {\n+ val outputName = session.outputNames.iterator().next()\n+ val ortValue = result.get(outputName)\n+ .orElseThrow { Exception(\"No output tensor found\") }\n+ val outputTensor = ortValue as? OnnxTensor\n+ ?: throw Exception(\"Output is not an ONNX tensor\")\n+ Pair(copyFloatBuffer(outputTensor.floatBuffer), outputTensor.info.shape)\n+ } catch (t: Throwable) {\n+ throw OCRError.InferenceFailed(modelName, t)\n+ }\n+ } finally {\n+ result.close()\n+ }\n+ }\n+\n+ private fun copyFloatBuffer(buffer: FloatBuffer): FloatArray {\n+ val duplicate = buffer.duplicate()\n+ duplicate.rewind()\n+ val output = FloatArray(duplicate.remaining())\n+ duplicate.get(output)\n+ return output\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/RecognitionEngine.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/RecognitionEngine.kt\nnew file mode 100644\nindex 00000000000..4e6e0f3ded2\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/RecognitionEngine.kt\n@@ -0,0 +1,61 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.engine\n+\n+import com.paddle.ocr.postprocess.CTCDecoder\n+import com.paddle.ocr.preprocess.RecPreprocessor\n+import org.opencv.core.Mat\n+\n+class RecognitionEngine(\n+ private val ortManager: ORTSessionManager,\n+ private val characterList: List<String>,\n+) {\n+ data class RecognitionResult(\n+ val texts: List<Pair<String, Float>>,\n+ val preprocessMs: Long,\n+ val inferenceMs: Long,\n+ val postprocessMs: Long,\n+ val timeMs: Long,\n+ val inputShape: List<Int>,\n+ )\n+\n+ fun recognize(crops: List<Mat>): RecognitionResult {\n+ // Preprocess\n+ val preStart = System.currentTimeMillis()\n+ val preResult = RecPreprocessor.preprocessBatch(crops)\n+ val preprocessMs = System.currentTimeMillis() - preStart\n+\n+ // Inference\n+ val infStart = System.currentTimeMillis()\n+ val (outputData, outputShape) = ortManager.runRecognition(preResult.tensorData, preResult.shape)\n+ val inferenceMs = System.currentTimeMillis() - infStart\n+\n+ // Postprocess (CTC decode)\n+ val postStart = System.currentTimeMillis()\n+ val decoded = CTCDecoder.decode(outputData, outputShape, characterList)\n+ val postprocessMs = System.currentTimeMillis() - postStart\n+\n+ val inputShape = preResult.shape.map { it.toInt() }\n+ val timeMs = preprocessMs + inferenceMs + postprocessMs\n+ return RecognitionResult(\n+ texts = decoded,\n+ preprocessMs = preprocessMs,\n+ inferenceMs = inferenceMs,\n+ postprocessMs = postprocessMs,\n+ timeMs = timeMs,\n+ inputShape = inputShape,\n+ )\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/ModelConfig.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/ModelConfig.kt\nnew file mode 100644\nindex 00000000000..111770fe8f9\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/ModelConfig.kt\n@@ -0,0 +1,123 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.model\n+\n+import android.content.Context\n+import com.paddle.ocr.util.YamlUtils\n+\n+data class ModelConfig(\n+ val characterList: List<String>,\n+) {\n+ companion object {\n+ fun parse(context: Context, assetPath: String): ModelConfig {\n+ val content = try {\n+ context.assets.open(assetPath).bufferedReader().use { it.readText() }\n+ } catch (t: Throwable) {\n+ throw OCRError.ConfigParseFailed(assetPath, t)\n+ }\n+ val characterDict = try {\n+ extractCharacterDict(content, assetPath)\n+ } catch (e: OCRError.ConfigParseFailed) {\n+ throw e\n+ } catch (t: Throwable) {\n+ throw OCRError.ConfigParseFailed(assetPath, t)\n+ }\n+ val charListWithSpace = characterDict.toMutableList().apply {\n+ if (lastOrNull() != \" \") add(\" \")\n+ }\n+\n+ return ModelConfig(characterList = charListWithSpace)\n+ }\n+\n+ private fun extractCharacterDict(content: String, assetPath: String): List<String> {\n+ val lines = content.replace(\"\\r\\n\", \"\\n\").replace('\\r', '\\n').lines()\n+ val postProcessLine = lines.indexOfFirst { it.trim() == \"PostProcess:\" }\n+ if (postProcessLine < 0) {\n+ throw OCRError.ConfigParseFailed(\"Missing PostProcess in $assetPath\")\n+ }\n+\n+ val postProcessIndent = YamlUtils.leadingSpaces(lines[postProcessLine])\n+ val characterDictLine = findCharacterDictLine(lines, postProcessLine + 1, postProcessIndent)\n+ if (characterDictLine < 0) {\n+ throw OCRError.ConfigParseFailed(\"Missing character_dict in $assetPath\")\n+ }\n+\n+ val characterDictIndent = YamlUtils.leadingSpaces(lines[characterDictLine])\n+ val characters = mutableListOf<String>()\n+ var lineIndex = characterDictLine + 1\n+ while (lineIndex < lines.size) {\n+ val line = lines[lineIndex]\n+ val indent = YamlUtils.leadingSpaces(line)\n+ val content = line.substring(indent)\n+ val keyLikeContent = content.trim()\n+ if (keyLikeContent.isEmpty() || keyLikeContent.startsWith(\"#\")) {\n+ lineIndex++\n+ continue\n+ }\n+\n+ if (!content.startsWith(\"-\")) {\n+ if (indent <= characterDictIndent) break\n+ lineIndex++\n+ continue\n+ }\n+\n+ characters.add(parseYamlListScalar(content.substring(1)))\n+ lineIndex++\n+ }\n+\n+ if (characters.isEmpty()) {\n+ throw OCRError.ConfigParseFailed(\"Empty character_dict in $assetPath\")\n+ }\n+ return characters\n+ }\n+\n+ private fun findCharacterDictLine(\n+ lines: List<String>,\n+ startLine: Int,\n+ postProcessIndent: Int,\n+ ): Int {\n+ var lineIndex = startLine\n+ while (lineIndex < lines.size) {\n+ val line = lines[lineIndex]\n+ val trimmed = line.trim()\n+ if (trimmed.isEmpty() || trimmed.startsWith(\"#\")) {\n+ lineIndex++\n+ continue\n+ }\n+ val indent = YamlUtils.leadingSpaces(line)\n+ if (indent <= postProcessIndent) break\n+ if (trimmed == \"character_dict:\") return lineIndex\n+ lineIndex++\n+ }\n+ return -1\n+ }\n+\n+ private fun parseYamlListScalar(rawValue: String): String {\n+ val value = rawValue.dropWhile { it == ' ' }\n+ if (value.length >= 2 && value.first() == '\\'' && value.last() == '\\'') {\n+ return value.substring(1, value.length - 1).replace(\"''\", \"'\")\n+ }\n+ if (value.length >= 2 && value.first() == '\"' && value.last() == '\"') {\n+ return value.substring(1, value.length - 1)\n+ .replace(\"\\\\\\\"\", \"\\\"\")\n+ .replace(\"\\\\\\\\\", \"\\\\\")\n+ .replace(\"\\\\n\", \"\\n\")\n+ .replace(\"\\\\r\", \"\\r\")\n+ .replace(\"\\\\t\", \"\\t\")\n+ }\n+ return value\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRBox.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRBox.kt\nnew file mode 100644\nindex 00000000000..bf4885cc46b\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRBox.kt\n@@ -0,0 +1,25 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.model\n+\n+import android.graphics.PointF\n+\n+data class OCRBox(\n+ val points: List<PointF>,\n+) {\n+ init {\n+ require(points.size == 4) { \"OCRBox must have exactly 4 points, got ${points.size}\" }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRError.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRError.kt\nnew file mode 100644\nindex 00000000000..e505ca89946\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRError.kt\n@@ -0,0 +1,24 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.model\n+\n+sealed class OCRError(message: String, cause: Throwable? = null) : Exception(message, cause) {\n+ class ModelNotFound(modelPath: String, cause: Throwable? = null) : OCRError(\"Model not found: $modelPath\", cause)\n+ class ModelLoadFailed(modelName: String, cause: Throwable) : OCRError(\"Failed to load $modelName model\", cause)\n+ class ConfigParseFailed(path: String, cause: Throwable? = null) : OCRError(\"Failed to parse config: $path\", cause)\n+ class InvalidImage : OCRError(\"Input image is empty or invalid\")\n+ class InferenceFailed(stage: String, cause: Throwable) : OCRError(\"Inference failed at stage '$stage'\", cause)\n+ class DecodeError(message: String, cause: Throwable? = null) : OCRError(message, cause)\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRResult.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRResult.kt\nnew file mode 100644\nindex 00000000000..159204f63f1\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRResult.kt\n@@ -0,0 +1,22 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.model\n+\n+data class OCRResult(\n+ val box: OCRBox,\n+ val text: String,\n+ val confidence: Float,\n+ val wordBoxes: List<OCRBox>? = null,\n+)\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRRunResult.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRRunResult.kt\nnew file mode 100644\nindex 00000000000..7b19596b548\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRRunResult.kt\n@@ -0,0 +1,37 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.model\n+\n+data class OCRRunResult(\n+ val results: List<OCRResult>,\n+ val detectionTimeMs: Long,\n+ val recognitionTimeMs: Long,\n+ val totalTimeMs: Long,\n+ val lineCount: Int,\n+ // Detailed per-stage timing\n+ val detPreprocessMs: Long = 0,\n+ val detInferenceMs: Long = 0,\n+ val detPostprocessMs: Long = 0,\n+ val recPreprocessMs: Long = 0,\n+ val recInferenceMs: Long = 0,\n+ val recPostprocessMs: Long = 0,\n+ val pipelineOverheadMs: Long = 0,\n+ val coldLoadTimeMs: Long = 0,\n+ // Input tensor shapes\n+ val detInputShape: List<Int> = emptyList(),\n+ val recInputShapes: List<List<Int>> = emptyList(),\n+ // Per-line timing (only populated when recBatchSize == 1)\n+ val perLineRecMs: List<Long> = emptyList(),\n+)\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/BoxSorter.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/BoxSorter.kt\nnew file mode 100644\nindex 00000000000..72d2bd36d43\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/BoxSorter.kt\n@@ -0,0 +1,47 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.postprocess\n+\n+import com.paddle.ocr.model.OCRBox\n+\n+object BoxSorter {\n+ private const val ROW_THRESHOLD_Y = 10f\n+\n+ fun sortInReadingOrder(boxes: List<OCRBox>): List<OCRBox> {\n+ if (boxes.size <= 1) return boxes\n+\n+ val list = boxes.toMutableList()\n+ list.sortWith(compareBy({ it.points[0].y }, { it.points[0].x }))\n+\n+ // Align with PaddleX SortQuadBoxes: bubble left-to-right inside a 10px row band.\n+ for (i in 0 until list.size - 1) {\n+ var j = i\n+ while (j >= 0) {\n+ val next = list[j + 1]\n+ val curr = list[j]\n+ if (kotlin.math.abs(next.points[0].y - curr.points[0].y) < ROW_THRESHOLD_Y &&\n+ next.points[0].x < curr.points[0].x\n+ ) {\n+ list[j] = next\n+ list[j + 1] = curr\n+ j--\n+ } else {\n+ break\n+ }\n+ }\n+ }\n+ return list\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/CTCDecoder.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/CTCDecoder.kt\nnew file mode 100644\nindex 00000000000..feb56d005c2\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/CTCDecoder.kt\n@@ -0,0 +1,66 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.postprocess\n+\n+object CTCDecoder {\n+ private const val BLANK_IDX = 0\n+\n+ fun decode(output: FloatArray, shape: LongArray, characterList: List<String>): List<Pair<String, Float>> {\n+ val batchSize = shape[0].toInt()\n+ val timeSteps = shape[1].toInt()\n+ val numClasses = shape[2].toInt()\n+\n+ val results = mutableListOf<Pair<String, Float>>()\n+ for (b in 0 until batchSize) {\n+ val baseOffset = b * timeSteps * numClasses\n+\n+ val indices = IntArray(timeSteps)\n+ val probs = FloatArray(timeSteps)\n+ for (t in 0 until timeSteps) {\n+ val offset = baseOffset + t * numClasses\n+ var maxIdx = 0\n+ var maxVal = output[offset]\n+ for (c in 1 until numClasses) {\n+ val v = output[offset + c]\n+ if (v > maxVal) {\n+ maxVal = v\n+ maxIdx = c\n+ }\n+ }\n+ indices[t] = maxIdx\n+ probs[t] = maxVal\n+ }\n+\n+ val keptProbs = mutableListOf<Float>()\n+ val sb = StringBuilder()\n+ var prevIdx = -1\n+ for (t in 0 until timeSteps) {\n+ val idx = indices[t]\n+ if (idx != BLANK_IDX && idx != prevIdx) {\n+ val charIdx = idx - 1\n+ if (charIdx >= 0 && charIdx < characterList.size) {\n+ sb.append(characterList[charIdx])\n+ keptProbs.add(probs[t])\n+ }\n+ }\n+ prevIdx = idx\n+ }\n+\n+ val confidence = if (keptProbs.isNotEmpty()) keptProbs.average().toFloat() else 0f\n+ results.add(Pair(sb.toString(), confidence))\n+ }\n+ return results\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/DBPostProcessor.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/DBPostProcessor.kt\nnew file mode 100644\nindex 00000000000..78fbb2e919a\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/DBPostProcessor.kt\n@@ -0,0 +1,189 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.postprocess\n+\n+import android.graphics.PointF\n+import com.paddle.ocr.model.OCRBox\n+import com.paddle.ocr.util.MathUtils\n+import org.opencv.core.Core\n+import org.opencv.core.CvType\n+import org.opencv.core.Mat\n+import org.opencv.core.MatOfPoint\n+import org.opencv.core.MatOfPoint2f\n+import org.opencv.core.Point\n+import org.opencv.core.Scalar\n+import org.opencv.imgproc.Imgproc\n+\n+object DBPostProcessor {\n+ private const val MIN_SIZE_BEFORE_UNCLIP = 3f\n+ private const val MIN_SIZE_AFTER_UNCLIP = 5f\n+\n+ fun process(\n+ pred: FloatArray,\n+ predShape: LongArray,\n+ thresh: Float,\n+ boxThresh: Float,\n+ unclipRatio: Float,\n+ maxCandidates: Int,\n+ useDilation: Boolean,\n+ scoreMode: String,\n+ boxType: String,\n+ originalH: Int,\n+ originalW: Int,\n+ ): List<OCRBox> {\n+ require(boxType == \"quad\") { \"Only DBPostProcess box_type=quad is supported\" }\n+\n+ val pH = predShape[2].toInt()\n+ val pW = predShape[3].toInt()\n+ val scaleX = originalW.toDouble() / pW\n+ val scaleY = originalH.toDouble() / pH\n+ val normalizedScoreMode = scoreMode.lowercase()\n+\n+ val rawProb = Mat(pH, pW, CvType.CV_32FC1)\n+ val mask = Mat(pH, pW, CvType.CV_8UC1)\n+ val contours = mutableListOf<MatOfPoint>()\n+ val hierarchy = Mat()\n+ var contourMask = mask\n+\n+ return try {\n+ rawProb.put(0, 0, pred)\n+ val threshMat = Mat()\n+ try {\n+ Imgproc.threshold(rawProb, threshMat, thresh.toDouble(), 255.0, Imgproc.THRESH_BINARY)\n+ threshMat.convertTo(mask, CvType.CV_8UC1)\n+ } finally {\n+ threshMat.release()\n+ }\n+\n+ contourMask = if (useDilation) {\n+ val kernel = Mat.ones(2, 2, CvType.CV_8UC1)\n+ try {\n+ Mat().also { Imgproc.dilate(mask, it, kernel) }\n+ } finally {\n+ kernel.release()\n+ }\n+ } else {\n+ mask\n+ }\n+\n+ Imgproc.findContours(contourMask, contours, hierarchy, Imgproc.RETR_LIST, Imgproc.CHAIN_APPROX_SIMPLE)\n+ val boxes = mutableListOf<OCRBox>()\n+ val numContours = minOf(contours.size, maxCandidates)\n+ for (index in 0 until numContours) {\n+ val contour = contours[index]\n+ val contour2f = MatOfPoint2f(*contour.toArray())\n+ val rect = try {\n+ Imgproc.minAreaRect(contour2f)\n+ } finally {\n+ contour2f.release()\n+ }\n+ val sside = minOf(rect.size.width, rect.size.height)\n+ if (sside < MIN_SIZE_BEFORE_UNCLIP) continue\n+\n+ val rectPoints = Array(4) { Point() }\n+ rect.points(rectPoints)\n+ val orderedPoints = QuadGeometry.orderMinAreaRectPoints(rectPoints)\n+ val score = if (normalizedScoreMode == \"slow\") {\n+ computeBoxScore(rawProb, contour.toList())\n+ } else {\n+ computeBoxScore(rawProb, orderedPoints)\n+ }\n+ if (score < boxThresh) continue\n+\n+ val expandedPts = PolygonUnclip.unclip(orderedPoints, unclipRatio)\n+ val expandedRect = MatOfPoint2f().let { expanded2f ->\n+ try {\n+ expanded2f.fromList(expandedPts)\n+ Imgproc.minAreaRect(expanded2f)\n+ } finally {\n+ expanded2f.release()\n+ }\n+ }\n+ val sside2 = minOf(expandedRect.size.width, expandedRect.size.height)\n+ if (sside2 < MIN_SIZE_AFTER_UNCLIP) continue\n+\n+ val expandedBox = Array(4) { Point() }\n+ expandedRect.points(expandedBox)\n+ val ePts = QuadGeometry.orderMinAreaRectPoints(expandedBox)\n+ val scaled = listOf(\n+ scalePoint(ePts[0], scaleX, scaleY, originalW, originalH),\n+ scalePoint(ePts[1], scaleX, scaleY, originalW, originalH),\n+ scalePoint(ePts[2], scaleX, scaleY, originalW, originalH),\n+ scalePoint(ePts[3], scaleX, scaleY, originalW, originalH),\n+ )\n+\n+ val boxW = kotlin.math.hypot(\n+ scaled[1].x - scaled[0].x,\n+ scaled[1].y - scaled[0].y,\n+ )\n+ val boxH = kotlin.math.hypot(\n+ scaled[3].x - scaled[0].x,\n+ scaled[3].y - scaled[0].y,\n+ )\n+ if (boxW <= 3 || boxH <= 3) continue\n+\n+ boxes.add(OCRBox(points = scaled))\n+ }\n+ boxes\n+ } finally {\n+ hierarchy.release()\n+ if (contourMask !== mask) {\n+ contourMask.release()\n+ }\n+ contours.forEach { it.release() }\n+ mask.release()\n+ rawProb.release()\n+ }\n+ }\n+\n+ private fun scalePoint(\n+ point: Point,\n+ scaleX: Double,\n+ scaleY: Double,\n+ originalW: Int,\n+ originalH: Int,\n+ ): PointF {\n+ return PointF(\n+ MathUtils.roundHalfToEven(point.x * scaleX).coerceIn(0, originalW).toFloat(),\n+ MathUtils.roundHalfToEven(point.y * scaleY).coerceIn(0, originalH).toFloat(),\n+ )\n+ }\n+\n+ private fun computeBoxScore(probMap: Mat, points: List<Point>): Float {\n+ if (points.isEmpty()) return 0f\n+\n+ val h = probMap.rows()\n+ val w = probMap.cols()\n+ val xmin = points.minOf { kotlin.math.floor(it.x).toInt() }.coerceIn(0, w - 1)\n+ val xmax = points.maxOf { kotlin.math.ceil(it.x).toInt() }.coerceIn(0, w - 1)\n+ val ymin = points.minOf { kotlin.math.floor(it.y).toInt() }.coerceIn(0, h - 1)\n+ val ymax = points.maxOf { kotlin.math.ceil(it.y).toInt() }.coerceIn(0, h - 1)\n+\n+ val mask = Mat(ymax - ymin + 1, xmax - xmin + 1, CvType.CV_8UC1, Scalar(0.0))\n+ val pts = MatOfPoint().apply {\n+ fromList(\n+ points.map { Point((it.x - xmin).toInt().toDouble(), (it.y - ymin).toInt().toDouble()) }\n+ )\n+ }\n+ Imgproc.fillPoly(mask, mutableListOf(pts), Scalar(1.0))\n+\n+ val roi = probMap.submat(ymin, ymax + 1, xmin, xmax + 1)\n+ val mean = Core.mean(roi, mask)\n+ roi.release()\n+ mask.release()\n+ pts.release()\n+ return mean.`val`[0].toFloat()\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/PolygonUnclip.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/PolygonUnclip.kt\nnew file mode 100644\nindex 00000000000..d4d0b4bf775\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/PolygonUnclip.kt\n@@ -0,0 +1,138 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.postprocess\n+\n+import org.opencv.core.Point\n+import kotlin.math.PI\n+import kotlin.math.abs\n+import kotlin.math.acos\n+import kotlin.math.atan2\n+import kotlin.math.ceil\n+import kotlin.math.cos\n+import kotlin.math.hypot\n+import kotlin.math.sin\n+\n+internal object PolygonUnclip {\n+ private const val EPS = 1e-6\n+ private const val ARC_TOLERANCE = 0.25\n+\n+ fun unclip(points: List<Point>, unclipRatio: Float): List<Point> {\n+ if (points.size < 3) return points\n+\n+ val signedArea = signedArea(points)\n+ val area = abs(signedArea)\n+ val perimeter = perimeter(points)\n+ if (!area.isFinite() || !perimeter.isFinite() || area <= EPS || perimeter <= EPS) {\n+ return points\n+ }\n+\n+ val distance = area * unclipRatio / perimeter\n+ if (!distance.isFinite() || distance <= EPS) return points\n+\n+ val clockwiseInImageCoords = signedArea > 0.0\n+ val normals = mutableListOf<Point>()\n+ for (i in points.indices) {\n+ val start = points[i]\n+ val end = points[(i + 1) % points.size]\n+ val dx = end.x - start.x\n+ val dy = end.y - start.y\n+ val len = hypot(dx, dy)\n+ if (!len.isFinite() || len <= EPS) return points\n+\n+ val normal = if (clockwiseInImageCoords) {\n+ Point(dy / len, -dx / len)\n+ } else {\n+ Point(-dy / len, dx / len)\n+ }\n+ normals.add(normal)\n+ }\n+\n+ val expanded = mutableListOf<Point>()\n+ for (i in points.indices) {\n+ appendRoundJoin(\n+ output = expanded,\n+ center = points[i],\n+ fromNormal = normals[(i - 1 + normals.size) % normals.size],\n+ toNormal = normals[i],\n+ distance = distance,\n+ clockwiseInImageCoords = clockwiseInImageCoords,\n+ )\n+ }\n+ return if (expanded.size >= 3) expanded else points\n+ }\n+\n+ private fun appendRoundJoin(\n+ output: MutableList<Point>,\n+ center: Point,\n+ fromNormal: Point,\n+ toNormal: Point,\n+ distance: Double,\n+ clockwiseInImageCoords: Boolean,\n+ ) {\n+ var startAngle = atan2(fromNormal.y, fromNormal.x)\n+ var endAngle = atan2(toNormal.y, toNormal.x)\n+ if (clockwiseInImageCoords) {\n+ while (endAngle < startAngle) endAngle += 2.0 * PI\n+ } else {\n+ while (endAngle > startAngle) endAngle -= 2.0 * PI\n+ }\n+\n+ val sweep = endAngle - startAngle\n+ val steps = ceil(abs(sweep) / arcStepAngle(distance)).toInt().coerceAtLeast(1)\n+ for (step in 0..steps) {\n+ val angle = startAngle + sweep * step.toDouble() / steps.toDouble()\n+ appendPoint(\n+ output,\n+ Point(\n+ center.x + cos(angle) * distance,\n+ center.y + sin(angle) * distance,\n+ )\n+ )\n+ }\n+ }\n+\n+ private fun arcStepAngle(distance: Double): Double {\n+ val ratio = (1.0 - ARC_TOLERANCE / distance).coerceIn(-1.0, 1.0)\n+ val step = 2.0 * acos(ratio)\n+ return if (step.isFinite() && step > EPS) step else PI / 8.0\n+ }\n+\n+ private fun appendPoint(output: MutableList<Point>, point: Point) {\n+ val last = output.lastOrNull()\n+ if (last == null || hypot(point.x - last.x, point.y - last.y) > EPS) {\n+ output.add(point)\n+ }\n+ }\n+\n+ private fun signedArea(points: List<Point>): Double {\n+ var sum = 0.0\n+ for (i in points.indices) {\n+ val p1 = points[i]\n+ val p2 = points[(i + 1) % points.size]\n+ sum += p1.x * p2.y - p2.x * p1.y\n+ }\n+ return sum / 2.0\n+ }\n+\n+ private fun perimeter(points: List<Point>): Double {\n+ var length = 0.0\n+ for (i in points.indices) {\n+ val p1 = points[i]\n+ val p2 = points[(i + 1) % points.size]\n+ length += hypot(p2.x - p1.x, p2.y - p1.y)\n+ }\n+ return length\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadGeometry.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadGeometry.kt\nnew file mode 100644\nindex 00000000000..1dee6cab32a\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadGeometry.kt\n@@ -0,0 +1,43 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.postprocess\n+\n+import org.opencv.core.Point\n+\n+internal object QuadGeometry {\n+ fun orderMinAreaRectPoints(points: Array<Point>): List<Point> {\n+ val sorted = points.sortedBy { it.x }\n+ val topLeft: Point\n+ val bottomLeft: Point\n+ if (sorted[1].y > sorted[0].y) {\n+ topLeft = sorted[0]\n+ bottomLeft = sorted[1]\n+ } else {\n+ topLeft = sorted[1]\n+ bottomLeft = sorted[0]\n+ }\n+\n+ val topRight: Point\n+ val bottomRight: Point\n+ if (sorted[3].y > sorted[2].y) {\n+ topRight = sorted[2]\n+ bottomRight = sorted[3]\n+ } else {\n+ topRight = sorted[3]\n+ bottomRight = sorted[2]\n+ }\n+ return listOf(topLeft, topRight, bottomRight, bottomLeft)\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadTextCrop.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadTextCrop.kt\nnew file mode 100644\nindex 00000000000..e8b5bb2be06\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadTextCrop.kt\n@@ -0,0 +1,86 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.postprocess\n+\n+import com.paddle.ocr.model.OCRBox\n+import org.opencv.core.Core\n+import org.opencv.core.CvType\n+import org.opencv.core.Mat\n+import org.opencv.core.MatOfPoint2f\n+import org.opencv.core.Point\n+import org.opencv.core.Size\n+import org.opencv.imgproc.Imgproc\n+import kotlin.math.hypot\n+import kotlin.math.max\n+\n+object QuadTextCrop {\n+ private const val VERTICAL_CROP_RATIO = 1.5\n+\n+ fun crop(src: Mat, box: OCRBox): Mat {\n+ // Align with PaddleX CropByPolys.get_minarea_rect_crop: recompute minAreaRect\n+ // from the detected quad before perspective transform.\n+ val rectPoints = box.points.map { Point(it.x.toDouble(), it.y.toDouble()) }\n+ val rectInput = MatOfPoint2f()\n+ rectInput.fromList(rectPoints)\n+ val boundingBox = Imgproc.minAreaRect(rectInput)\n+ rectInput.release()\n+\n+ val boxPoints = Array(4) { Point() }\n+ boundingBox.points(boxPoints)\n+ val ordered = QuadGeometry.orderMinAreaRectPoints(boxPoints)\n+\n+ val widthTop = hypot(ordered[0].x - ordered[1].x, ordered[0].y - ordered[1].y)\n+ val widthBottom = hypot(ordered[2].x - ordered[3].x, ordered[2].y - ordered[3].y)\n+ val heightLeft = hypot(ordered[0].x - ordered[3].x, ordered[0].y - ordered[3].y)\n+ val heightRight = hypot(ordered[1].x - ordered[2].x, ordered[1].y - ordered[2].y)\n+\n+ val dstW = max(widthTop, widthBottom).toInt().coerceAtLeast(1)\n+ val dstH = max(heightLeft, heightRight).toInt().coerceAtLeast(1)\n+\n+ val srcPts = MatOfPoint2f()\n+ srcPts.fromList(ordered)\n+ val dstPts = MatOfPoint2f()\n+ dstPts.fromList(\n+ listOf(\n+ Point(0.0, 0.0),\n+ Point(dstW.toDouble(), 0.0),\n+ Point(dstW.toDouble(), dstH.toDouble()),\n+ Point(0.0, dstH.toDouble()),\n+ )\n+ )\n+ val m = Imgproc.getPerspectiveTransform(srcPts, dstPts)\n+ srcPts.release()\n+ dstPts.release()\n+\n+ val dst = Mat(dstH, dstW, CvType.CV_8UC3)\n+ Imgproc.warpPerspective(\n+ src,\n+ dst,\n+ m,\n+ Size(dstW.toDouble(), dstH.toDouble()),\n+ Imgproc.INTER_CUBIC,\n+ Core.BORDER_REPLICATE,\n+ )\n+ m.release()\n+\n+ if (dst.rows().toDouble() / dst.cols() >= VERTICAL_CROP_RATIO) {\n+ val rotated = Mat()\n+ Core.rotate(dst, rotated, Core.ROTATE_90_COUNTERCLOCKWISE)\n+ dst.release()\n+ return rotated\n+ }\n+ return dst\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/DetPreprocessor.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/DetPreprocessor.kt\nnew file mode 100644\nindex 00000000000..7fd9d240c9e\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/DetPreprocessor.kt\n@@ -0,0 +1,103 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.preprocess\n+\n+import android.graphics.Bitmap\n+import com.paddle.ocr.util.BitmapUtils\n+import com.paddle.ocr.util.ImageUtils\n+import org.opencv.core.Core\n+import org.opencv.core.CvType\n+import org.opencv.core.Mat\n+import org.opencv.core.Scalar\n+import org.opencv.imgproc.Imgproc\n+\n+data class DetPreprocessResult(\n+ val tensorData: FloatArray,\n+ val shape: LongArray,\n+ val originalH: Int,\n+ val originalW: Int,\n+)\n+\n+object DetPreprocessor {\n+ private val mean = doubleArrayOf(0.485, 0.456, 0.406)\n+ private val std = doubleArrayOf(0.229, 0.224, 0.225)\n+ private const val scale = 1.0 / 255.0\n+\n+ fun preprocess(\n+ bitmap: Bitmap,\n+ limitSideLen: Int,\n+ limitType: String,\n+ maxSideLimit: Int,\n+ imgMode: String,\n+ ): DetPreprocessResult {\n+ val src = BitmapUtils.bitmapToBGRMat(bitmap)\n+ return try {\n+ preprocess(src, limitSideLen, limitType, maxSideLimit, imgMode)\n+ } finally {\n+ src.release()\n+ }\n+ }\n+\n+ fun preprocess(\n+ src: Mat,\n+ limitSideLen: Int,\n+ limitType: String,\n+ maxSideLimit: Int,\n+ imgMode: String,\n+ ): DetPreprocessResult {\n+ val originalH = src.rows()\n+ val originalW = src.cols()\n+ val input = if (imgMode.uppercase() == \"RGB\") {\n+ Mat().also { Imgproc.cvtColor(src, it, Imgproc.COLOR_BGR2RGB) }\n+ } else {\n+ src\n+ }\n+\n+ val resized = ImageUtils.resizeToMultipleOf32(input, limitSideLen, limitType, maxSideLimit)\n+ if (input !== src) input.release()\n+\n+ val h = resized.rows()\n+ val w = resized.cols()\n+ val floatMat = Mat(h, w, CvType.CV_32FC3)\n+ resized.convertTo(floatMat, CvType.CV_32F)\n+\n+ val channels = mutableListOf<Mat>()\n+ Core.split(floatMat, channels)\n+ for (c in 0..2) {\n+ Core.multiply(channels[c], Scalar(scale), channels[c])\n+ Core.subtract(channels[c], Scalar(mean[c]), channels[c])\n+ Core.divide(channels[c], Scalar(std[c]), channels[c])\n+ }\n+ resized.release()\n+ floatMat.release()\n+\n+ val channelSize = h * w\n+ val tensorData = FloatArray(3 * channelSize)\n+ for (c in 0..2) {\n+ val chan = channels[c]\n+ val buf = FloatArray(channelSize)\n+ chan.get(0, 0, buf)\n+ System.arraycopy(buf, 0, tensorData, c * channelSize, channelSize)\n+ chan.release()\n+ }\n+\n+ return DetPreprocessResult(\n+ tensorData = tensorData,\n+ shape = longArrayOf(1, 3, h.toLong(), w.toLong()),\n+ originalH = originalH,\n+ originalW = originalW,\n+ )\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/RecPreprocessor.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/RecPreprocessor.kt\nnew file mode 100644\nindex 00000000000..648be60a4d7\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/RecPreprocessor.kt\n@@ -0,0 +1,105 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.preprocess\n+\n+import org.opencv.core.Core\n+import org.opencv.core.CvType\n+import org.opencv.core.Mat\n+import org.opencv.core.Size\n+import org.opencv.imgproc.Imgproc\n+import kotlin.math.ceil\n+\n+data class RecPreprocessResult(\n+ val tensorData: FloatArray,\n+ val shape: LongArray,\n+)\n+\n+object RecPreprocessor {\n+ private const val FIXED_HEIGHT = 48\n+ private const val MAX_IMG_W = 3200\n+\n+ fun preprocessBatch(crops: List<Mat>): RecPreprocessResult {\n+ // Convert BGR to RGB and resize to fixed height while preserving aspect ratio\n+ val resizedMats = mutableListOf<Mat>()\n+ for (crop in crops) {\n+ // Convert BGR to RGB (model expects RGB input)\n+ val rgb = Mat()\n+ Imgproc.cvtColor(crop, rgb, Imgproc.COLOR_BGR2RGB)\n+ val h = rgb.rows()\n+ val w = rgb.cols()\n+ val aspectRatio = if (h > 0) w.toDouble() / h else 1.0\n+ val newW = ceil(FIXED_HEIGHT * aspectRatio).toInt().coerceAtMost(MAX_IMG_W)\n+ val dst = Mat()\n+ Imgproc.resize(rgb, dst, Size(newW.toDouble(), FIXED_HEIGHT.toDouble()), 0.0, 0.0, Imgproc.INTER_LINEAR)\n+ rgb.release()\n+ resizedMats.add(dst)\n+ }\n+\n+ // Convert to float and normalize: (x / 255 - 0.5) / 0.5 = x / 127.5 - 1\n+ val floatMats = mutableListOf<Mat>()\n+ for (mat in resizedMats) {\n+ val floatMat = Mat(mat.rows(), mat.cols(), CvType.CV_32FC3)\n+ mat.convertTo(floatMat, CvType.CV_32F)\n+ // Use Scalar with all 3 channels set — single-value Scalar only sets val[0]!\n+ Core.divide(floatMat, org.opencv.core.Scalar(127.5, 127.5, 127.5), floatMat)\n+ Core.subtract(floatMat, org.opencv.core.Scalar(1.0, 1.0, 1.0), floatMat)\n+\n+ floatMats.add(floatMat)\n+ mat.release() // Release resized mat\n+ }\n+ resizedMats.clear()\n+\n+ val maxW = floatMats.maxOf { it.cols() }\n+ val n = floatMats.size\n+\n+ // Pad to max width\n+ val paddedMats = mutableListOf<Mat>()\n+ for (mat in floatMats) {\n+ if (mat.cols() == maxW) {\n+ paddedMats.add(mat)\n+ } else {\n+ val padded = Mat(FIXED_HEIGHT, maxW, CvType.CV_32FC3, org.opencv.core.Scalar(0.0))\n+ val roi = padded.submat(0, FIXED_HEIGHT, 0, mat.cols())\n+ mat.copyTo(roi)\n+ roi.release()\n+ mat.release()\n+ paddedMats.add(padded)\n+ }\n+ }\n+ floatMats.clear()\n+\n+ // Build tensor data\n+ val channelSize = FIXED_HEIGHT * maxW\n+ val tensorData = FloatArray(n * 3 * channelSize)\n+ for (b in 0 until n) {\n+ val mat = paddedMats[b]\n+ val channels = mutableListOf<Mat>()\n+ Core.split(mat, channels)\n+ for (c in 0..2) {\n+ val buf = FloatArray(channelSize)\n+ channels[c].get(0, 0, buf)\n+ System.arraycopy(buf, 0, tensorData, (b * 3 + c) * channelSize, channelSize)\n+ channels[c].release()\n+ }\n+ mat.release()\n+ }\n+ paddedMats.clear()\n+\n+ return RecPreprocessResult(\n+ tensorData = tensorData,\n+ shape = longArrayOf(n.toLong(), 3, FIXED_HEIGHT.toLong(), maxW.toLong()),\n+ )\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/BitmapUtils.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/BitmapUtils.kt\nnew file mode 100644\nindex 00000000000..a504e7058f5\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/BitmapUtils.kt\n@@ -0,0 +1,72 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.util\n+\n+import android.graphics.Bitmap\n+import org.opencv.android.Utils\n+import org.opencv.core.CvType\n+import org.opencv.core.Mat\n+import org.opencv.core.MatOfByte\n+import org.opencv.imgcodecs.Imgcodecs\n+import org.opencv.imgproc.Imgproc\n+\n+object BitmapUtils {\n+\n+ fun imdecodeBGR(imageBytes: ByteArray): Mat {\n+ val encoded = MatOfByte(*imageBytes)\n+ return try {\n+ Imgcodecs.imdecode(encoded, Imgcodecs.IMREAD_COLOR)\n+ } finally {\n+ encoded.release()\n+ }\n+ }\n+\n+ fun bitmapToBGRMat(bitmap: Bitmap): Mat {\n+ return bitmapToMat(bitmap, Imgproc.COLOR_RGBA2BGR)\n+ }\n+\n+ fun bitmapToRGBMat(bitmap: Bitmap): Mat {\n+ return bitmapToMat(bitmap, Imgproc.COLOR_RGBA2RGB)\n+ }\n+\n+ fun bgrMatToBitmap(mat: Mat): Bitmap {\n+ val rgba = Mat()\n+ return try {\n+ Imgproc.cvtColor(mat, rgba, Imgproc.COLOR_BGR2RGBA)\n+ Bitmap.createBitmap(rgba.cols(), rgba.rows(), Bitmap.Config.ARGB_8888).also { bitmap ->\n+ Utils.matToBitmap(rgba, bitmap)\n+ }\n+ } finally {\n+ rgba.release()\n+ }\n+ }\n+\n+ private fun bitmapToMat(bitmap: Bitmap, colorConversionCode: Int): Mat {\n+ val bmp = bitmap.copy(Bitmap.Config.ARGB_8888, false)\n+ val rgba = Mat(bmp.height, bmp.width, CvType.CV_8UC4)\n+ val dst = Mat()\n+ return try {\n+ Utils.bitmapToMat(bmp, rgba)\n+ Imgproc.cvtColor(rgba, dst, colorConversionCode)\n+ dst\n+ } catch (t: Throwable) {\n+ dst.release()\n+ throw t\n+ } finally {\n+ bmp.recycle()\n+ rgba.release()\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/ImageUtils.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/ImageUtils.kt\nnew file mode 100644\nindex 00000000000..c672a641006\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/ImageUtils.kt\n@@ -0,0 +1,52 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.util\n+\n+import org.opencv.core.Mat\n+import org.opencv.core.Size\n+import org.opencv.imgproc.Imgproc\n+\n+object ImageUtils {\n+\n+ fun resizeToMultipleOf32(\n+ src: Mat,\n+ limitSideLen: Int,\n+ limitType: String,\n+ maxSideLimit: Int,\n+ ): Mat {\n+ val h = src.rows()\n+ val w = src.cols()\n+ var ratio = when (limitType.lowercase()) {\n+ \"max\" -> if (maxOf(h, w) > limitSideLen) limitSideLen.toDouble() / maxOf(h, w) else 1.0\n+ \"min\" -> if (minOf(h, w) < limitSideLen) limitSideLen.toDouble() / minOf(h, w) else 1.0\n+ \"resize_long\" -> limitSideLen.toDouble() / maxOf(h, w)\n+ else -> throw IllegalArgumentException(\"Unsupported det limit type: $limitType\")\n+ }\n+\n+ var newH = (h * ratio).toInt()\n+ var newW = (w * ratio).toInt()\n+ if (maxOf(newH, newW) > maxSideLimit) {\n+ ratio = maxSideLimit.toDouble() / maxOf(newH, newW)\n+ newH = (newH * ratio).toInt()\n+ newW = (newW * ratio).toInt()\n+ }\n+\n+ newH = maxOf(MathUtils.roundHalfToEven(newH / 32.0) * 32, 32)\n+ newW = maxOf(MathUtils.roundHalfToEven(newW / 32.0) * 32, 32)\n+ val dst = Mat()\n+ Imgproc.resize(src, dst, Size(newW.toDouble(), newH.toDouble()), 0.0, 0.0, Imgproc.INTER_LINEAR)\n+ return dst\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/MathUtils.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/MathUtils.kt\nnew file mode 100644\nindex 00000000000..bbee6ab15e2\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/MathUtils.kt\n@@ -0,0 +1,31 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.util\n+\n+import kotlin.math.floor\n+\n+object MathUtils {\n+\n+ fun roundHalfToEven(value: Double): Int {\n+ val floored = floor(value)\n+ val diff = value - floored\n+ return when {\n+ diff < 0.5 -> floored.toInt()\n+ diff > 0.5 -> floored.toInt() + 1\n+ floored.toInt() % 2 == 0 -> floored.toInt()\n+ else -> floored.toInt() + 1\n+ }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/OpenCVUtils.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/OpenCVUtils.kt\nnew file mode 100644\nindex 00000000000..77aa455d964\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/OpenCVUtils.kt\n@@ -0,0 +1,34 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.util\n+\n+import android.content.Context\n+import android.util.Log\n+\n+object OpenCVUtils {\n+\n+ private var initialized = false\n+\n+ fun init(context: Context): Boolean {\n+ if (initialized) return true\n+ try {\n+ System.loadLibrary(\"opencv_java4\")\n+ initialized = true\n+ } catch (e: UnsatisfiedLinkError) {\n+ Log.e(\"OpenCVUtils\", \"Failed to initialize OpenCV: ${e.message}\")\n+ }\n+ return initialized\n+ }\n+}\ndiff --git a/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/YamlUtils.kt b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/YamlUtils.kt\nnew file mode 100644\nindex 00000000000..8f066e6f186\n--- /dev/null\n+++ b/deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/YamlUtils.kt\n@@ -0,0 +1,22 @@\n+// Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+//\n+// Licensed under the Apache License, Version 2.0 (the \"License\");\n+// you may not use this file except in compliance with the License.\n+// You may obtain a copy of the License at\n+//\n+// http://www.apache.org/licenses/LICENSE-2.0\n+//\n+// Unless required by applicable law or agreed to in writing, software\n+// distributed under the License is distributed on an \"AS IS\" BASIS,\n+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n+// See the License for the specific language governing permissions and\n+// limitations under the License.\n+\n+package com.paddle.ocr.util\n+\n+object YamlUtils {\n+\n+ fun leadingSpaces(line: String): Int {\n+ return line.indexOfFirst { it != ' ' }.let { if (it < 0) line.length else it }\n+ }\n+}\ndiff --git a/deploy/ppocr-android/run_benchmark.sh b/deploy/ppocr-android/run_benchmark.sh\nnew file mode 100644\nindex 00000000000..0b543935576\n--- /dev/null\n+++ b/deploy/ppocr-android/run_benchmark.sh\n@@ -0,0 +1,39 @@\n+#!/bin/bash\n+# PP-OCRv6 Android Speed Benchmark\n+# Usage: JAVA_HOME=/opt/android-jdk ./run_benchmark.sh [iterations] [warmup]\n+# e.g. JAVA_HOME=/opt/android-jdk ./run_benchmark.sh 20 5\n+# defaults: 10 iterations, 3 warmup\n+\n+set -e\n+SCRIPT_DIR=\"$(cd \"$(dirname \"$0\")\" && pwd)\"\n+cd \"$SCRIPT_DIR\"\n+\n+ITERATIONS=\"${1:-50}\"\n+WARMUP=\"${2:-30}\"\n+\n+echo \"========================================\"\n+echo \" PP-OCRv6 Android Speed Benchmark\"\n+echo \"========================================\"\n+echo \" Warmup: ${WARMUP}\"\n+echo \" Iterations: ${ITERATIONS}\"\n+echo \"========================================\"\n+echo \"\"\n+\n+adb logcat -c 2>/dev/null || true\n+\n+echo \"[1/2] Running speed benchmark (warmup + ${ITERATIONS} measured iterations)...\"\n+echo \"----------------------------------------\"\n+./gradlew :ppocr-sdk:connectedAndroidTest \\\n+ -Pandroid.testInstrumentationRunnerArguments.class=com.paddle.ocr.benchmark.OCRBenchmarkTest#testLatencyBenchmark \\\n+ -Pandroid.testInstrumentationRunnerArguments.warmup=\"${WARMUP}\" \\\n+ -Pandroid.testInstrumentationRunnerArguments.iterations=\"${ITERATIONS}\" \\\n+ 2>&1\n+echo \"----------------------------------------\"\n+echo \"\"\n+\n+echo \"[2/2] Speed benchmark results:\"\n+echo \"========================================\"\n+adb logcat -d -s System.out:I 2>/dev/null | grep \"OCRBenchmark\" | sed 's/^.*System.out/I/'\n+echo \"========================================\"\n+echo \"\"\n+echo \" Done\"\ndiff --git a/deploy/ppocr-android/settings.gradle.kts b/deploy/ppocr-android/settings.gradle.kts\nnew file mode 100644\nindex 00000000000..7210b454a5b\n--- /dev/null\n+++ b/deploy/ppocr-android/settings.gradle.kts\n@@ -0,0 +1,29 @@\n+pluginManagement {\n+ repositories {\n+ maven { url = uri(\"https://maven.aliyun.com/repository/google\") }\n+ maven { url = uri(\"https://maven.aliyun.com/repository/central\") }\n+ maven { url = uri(\"https://maven.aliyun.com/repository/gradle-plugin\") }\n+ google {\n+ content {\n+ includeGroupByRegex(\"com\\\\.android.*\")\n+ includeGroupByRegex(\"com\\\\.google.*\")\n+ includeGroupByRegex(\"androidx.*\")\n+ }\n+ }\n+ mavenCentral()\n+ gradlePluginPortal()\n+ }\n+}\n+dependencyResolutionManagement {\n+ repositoriesMode.set(RepositoriesMode.FAIL_ON_PROJECT_REPOS)\n+ repositories {\n+ maven { url = uri(\"https://maven.aliyun.com/repository/google\") }\n+ maven { url = uri(\"https://maven.aliyun.com/repository/central\") }\n+ google()\n+ mavenCentral()\n+ }\n+}\n+\n+rootProject.name = \"ppocr-android\"\n+include(\":app\")\n+include(\":ppocr-sdk\")\ndiff --git a/docs/version3.x/inference_deployment/cross_platform/android_deployment.en.md b/docs/version3.x/inference_deployment/cross_platform/android_deployment.en.md\nnew file mode 100644\nindex 00000000000..2618d0524c2\n--- /dev/null\n+++ b/docs/version3.x/inference_deployment/cross_platform/android_deployment.en.md\n@@ -0,0 +1,251 @@\n+# PP-OCRv6 Android Demo\n+\n+## Introduction\n+\n+This project is an Android deployment example for PaddleOCR v6, implementing mobile OCR inference using ONNX Runtime. The project adopts a **SDK and Demo separation** architecture, where the SDK module can be independently integrated into third-party applications.\n+\n+## Features\n+\n+- End-to-end text detection and recognition pipeline\n+- Supports PP-OCRv6 series ONNX models\n+- Detailed performance timing (detection/recognition stage breakdown)\n+- MVVM + Jetpack Compose Demo application\n+- AAR integration support\n+\n+## Project Structure\n+\n+```\n+ppocr-android/\n+├── ppocr-sdk/ # OCR SDK (Android Library)\n+│ ├── src/main/\n+│ │ ├── assets/models/ # Model files directory\n+│ │ │ ├── det/ # Detection model: inference.onnx\n+│ │ │ └── rec/ # Recognition model: inference.onnx, inference.yml\n+│ │ └── java/com/paddle/ocr/\n+│ │ ├── PaddleOCR.kt # [Public API] SDK entry point\n+│ │ ├── PaddleOCRConfig.kt # [Public API] Inference configuration\n+│ │ └── ...\n+│ └── build.gradle.kts\n+├── app/ # Demo App\n+│ ├── src/main/java/com/paddle/ocr/demo/\n+│ │ ├── OCRApplication.kt # Initialize SDK\n+│ │ └── ui/ # Compose UI\n+│ └── build.gradle.kts\n+├── run_benchmark.sh # Performance test script\n+└── README.md\n+```\n+\n+## Requirements\n+\n+| Dependency | Version |\n+|------------|---------|\n+| Android Studio | Ladybug (2024.2+) |\n+| JDK | 17 |\n+| Kotlin | 2.1.0 |\n+| minSdk | 26 (Android 8.0) |\n+| ONNX Runtime | 1.21.1 |\n+| OpenCV | 4.5.3 |\n+\n+## Quick Start\n+\n+### 1. Clone the Project\n+\n+```bash\n+git clone https://github.com/PaddlePaddle/PaddleOCR.git\n+cd PaddleOCR/deploy/ppocr-android\n+```\n+\n+### 2. Prepare Models\n+\n+This project supports the following models:\n+\n+| Model | HuggingFace | BOS |\n+|------|-------------|-----|\n+| **PP-OCRv6_small** | [Detection model](https://huggingface.co/PaddlePaddle/PP-OCRv6_small_det_onnx) / [Recognition model](https://huggingface.co/PaddlePaddle/PP-OCRv6_small_rec_onnx) | [Detection model](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv6_small_det_onnx_infer.tar) / [Recognition model](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv6_small_rec_onnx_infer.tar) |\n+| **PP-OCRv6_tiny** | [Detection model](https://huggingface.co/PaddlePaddle/PP-OCRv6_tiny_det_onnx) / [Recognition model](https://huggingface.co/PaddlePaddle/PP-OCRv6_tiny_rec_onnx) | [Detection model](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv6_tiny_det_onnx_infer.tar) / [Recognition model](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv6_tiny_rec_onnx_infer.tar) |\n+| **PP-OCRv5_mobile** | [Detection model](https://huggingface.co/PaddlePaddle/PP-OCRv5_mobile_det_onnx) / [Recognition model](https://huggingface.co/PaddlePaddle/PP-OCRv5_mobile_rec_onnx) | [Detection model](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv5_mobile_det_onnx_infer.tar) / [Recognition model](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv5_mobile_rec_onnx_infer.tar) |\n+\n+After downloading and extracting, place the files in `ppocr-sdk/src/main/assets/models/`:\n+\n+- Detection model: place `inference.onnx` in `models/det/`\n+- Recognition model: place `inference.onnx` and `inference.yml` in `models/rec/`\n+\n+### 3. Build and Run\n+\n+```bash\n+# Build Debug APK\n+./gradlew :app:assembleDebug\n+\n+# Install to device\n+./gradlew :app:installDebug\n+```\n+\n+Or run directly from Android Studio.\n+\n+### 4. Try the Demo\n+\n+1. Open \"PP-OCRv6 Demo\" application\n+2. Wait for model loading to complete\n+3. Tap \"Select from Gallery\" to choose an image\n+4. View recognition results and timing statistics\n+\n+## SDK Integration\n+\n+### Option 1: Source Code Dependency\n+\n+1. Copy `ppocr-sdk/` to your project root\n+2. Add to `settings.gradle.kts`:\n+ ```kotlin\n+ include(\":ppocr-sdk\")\n+ ```\n+3. Add to your app module's `build.gradle.kts`:\n+ ```kotlin\n+ implementation(project(\":ppocr-sdk\"))\n+ ```\n+\n+### Option 2: AAR Dependency\n+\n+```bash\n+# Build AAR\n+./gradlew :ppocr-sdk:assembleRelease\n+```\n+\n+AAR output: `ppocr-sdk/build/outputs/aar/ppocr-sdk-release.aar`\n+\n+Add to your app module's `build.gradle.kts`:\n+\n+```kotlin\n+dependencies {\n+ implementation(files(\"libs/ppocr-sdk-release.aar\"))\n+ // AAR doesn't transit dependencies, add manually\n+ implementation(\"com.microsoft.onnxruntime:onnxruntime-android:1.21.1\")\n+ implementation(\"com.quickbirdstudios:opencv:4.5.3\")\n+ implementation(\"org.jetbrains.kotlinx:kotlinx-coroutines-android:1.9.0\")\n+}\n+```\n+\n+## API Reference\n+\n+### Create Instance\n+\n+```kotlin\n+// Default configuration\n+val ocr = PaddleOCR.create(context)\n+\n+// Custom configuration\n+val ocr = PaddleOCR.create(\n+ context = context,\n+ config = PaddleOCRConfig(\n+ detThresh = 0.3f,\n+ detBoxThresh = 0.6f,\n+ recScoreThresh = 0.0f,\n+ recBatchSize = 1,\n+ ),\n+ engineConfig = EngineConfig(numThreads = 4),\n+ detModelAssetPath = \"models/det/inference.onnx\",\n+ recModelAssetPath = \"models/rec/inference.onnx\",\n+ recConfigAssetPath = \"models/rec/inference.yml\",\n+)\n+```\n+\n+### Perform OCR\n+\n+```kotlin\n+// Pass Bitmap\n+val result = ocr.recognize(bitmap)\n+\n+// Pass image bytes (recommended, consistent with Python pipeline)\n+val result = ocr.recognize(imageBytes)\n+\n+// Read results\n+result.results.forEach { item ->\n+ println(\"Text: ${item.text}, Confidence: ${item.confidence}\")\n+ println(\"Box: ${item.box.points}\")\n+}\n+println(\"Detection: ${result.detectionTimeMs}ms, Recognition: ${result.recognitionTimeMs}ms\")\n+```\n+\n+### Release Resources\n+\n+```kotlin\n+ocr.release()\n+```\n+\n+### Configuration Parameters\n+\n+```kotlin\n+data class PaddleOCRConfig(\n+ val detImgMode: String = \"BGR\", // Input color mode\n+ val detLimitSideLen: Int = 64, // Detection side length limit\n+ val detLimitType: String = \"min\", // Limit strategy\n+ val detMaxSideLimit: Int = 4000, // Maximum side length\n+ val detThresh: Float = 0.3f, // Binarization threshold\n+ val detBoxThresh: Float = 0.6f, // Detection box confidence threshold\n+ val detUnclipRatio: Float = 1.5f, // Detection box expansion ratio\n+ val detMaxCandidates: Int = 3000, // Maximum candidate boxes\n+ val detUseDilation: Boolean = false, // Whether to dilate\n+ val detScoreMode: String = \"fast\", // Scoring mode\n+ val detBoxType: String = \"quad\", // Detection box type\n+ val recScoreThresh: Float = 0.0f, // Recognition confidence threshold\n+ val recBatchSize: Int = 1, // Recognition batch size\n+)\n+```\n+\n+### Result Models\n+\n+```kotlin\n+data class OCRRunResult(\n+ val results: List<OCRResult>, // Recognition result list\n+ val detectionTimeMs: Long, // Detection time\n+ val recognitionTimeMs: Long, // Recognition time\n+ val totalTimeMs: Long, // Total time\n+ val lineCount: Int, // Number of lines\n+ // Detailed timing...\n+)\n+\n+data class OCRResult(\n+ val box: OCRBox, // Detection box coordinates\n+ val text: String, // Recognized text\n+ val confidence: Float, // Confidence score\n+)\n+```\n+\n+## Performance Testing\n+\n+The project provides an automated performance testing script:\n+\n+```bash\n+# Run benchmark (10 tests, 3 warmup)\n+./run_benchmark.sh 10 3\n+\n+# Sample output\n+╔═════════════════════════════════════════════════════════════════════════╗\n+║ PP-OCRv6 Speed Benchmark Results ║\n+╠═════════════════════════════════════════════════════════════════════════╣\n+║ Device: GM1900 | OS: Android 9 | Lines: 5 ║\n+║ Cold load: 158ms | Warmup: 3 | Measured: 10 ║\n+╠═════════════════════════════════════════════════════════════════════════╣\n++-----------------------------+----------+----------+----------+----------+\n+| Stage | Mean ms | Stdev | P90 | Min ms|\n++-----------------------------+----------+----------+----------+----------+\n+| Total pipeline | 420.40 | 6.37 | 427 | 413 |\n++-----------------------------+----------+----------+----------+----------+\n+| Detection (total) | 348.70 | 4.67 | 356 | 343 |\n+| Preprocess | 33.30 | 2.90 | 36 | 28 |\n+| Inference | 311.00 | 2.93 | 315 | 304 |\n+| Postprocess | 4.40 | 0.49 | 5 | 4 |\n+| Recognition (total) | 66.20 | 3.16 | 68 | 64 |\n+| Preprocess | 3.00 | 0.89 | 4 | 2 |\n+| Inference | 60.60 | 3.14 | 63 | 58 |\n+| Postprocess | 2.60 | 0.92 | 4 | 1 |\n+| Pipeline overhead | 5.50 | 0.50 | 6 | 5 |\n++-----------------------------+----------+----------+----------+----------+\n+╚═════════════════════════════════════════════════════════════════════════╝\n+```\n+\n+## Notes\n+\n+1. **OpenCV Initialization**: Call `OpenCVUtils.init(context)` before `PaddleOCR.create()`\n+2. **Coroutine Usage**: `create()` and `recognize()` are suspend functions, call them in coroutines\n+3. **Memory Management**: Call `release()` when no longer needed\n+4. **ProGuard Rules**: Refer to `ppocr-sdk/proguard-rules.pro`\ndiff --git a/docs/version3.x/inference_deployment/cross_platform/android_deployment.md b/docs/version3.x/inference_deployment/cross_platform/android_deployment.md\nnew file mode 100644\nindex 00000000000..69008d3b33d\n--- /dev/null\n+++ b/docs/version3.x/inference_deployment/cross_platform/android_deployment.md\n@@ -0,0 +1,251 @@\n+# PP-OCRv6 Android Demo\n+\n+## 简介\n+\n+本项目是 PaddleOCR v6 在 Android 平台的部署示例，基于 ONNX Runtime 实现移动端 OCR 推理。项目采用 **SDK 与 Demo 分离** 架构，SDK 模块可独立集成到第三方应用。\n+\n+## 功能特性\n+\n+- 文本检测 + 文本识别端到端流程\n+- 支持 PP-OCRv6 系列 ONNX 模型\n+- 详细的性能计时（检测/识别各阶段耗时）\n+- MVVM + Jetpack Compose Demo 应用\n+- 支持 AAR 方式集成\n+\n+## 项目结构\n+\n+```\n+ppocr-android/\n+├── ppocr-sdk/ # OCR SDK（Android Library）\n+│ ├── src/main/\n+│ │ ├── assets/models/ # 模型文件目录\n+│ │ │ ├── det/ # 检测模型：inference.onnx\n+│ │ │ └── rec/ # 识别模型：inference.onnx, inference.yml\n+│ │ └── java/com/paddle/ocr/\n+│ │ ├── PaddleOCR.kt # [公开 API] SDK 入口\n+│ │ ├── PaddleOCRConfig.kt # [公开 API] 推理参数配置\n+│ │ └── ...\n+│ └── build.gradle.kts\n+├── app/ # Demo App\n+│ ├── src/main/java/com/paddle/ocr/demo/\n+│ │ ├── OCRApplication.kt # 初始化 SDK\n+│ │ └── ui/ # Compose UI\n+│ └── build.gradle.kts\n+├── run_benchmark.sh # 性能测试脚本\n+└── README.md\n+```\n+\n+## 环境要求\n+\n+| 依赖 | 版本 |\n+|------|------|\n+| Android Studio | Ladybug (2024.2+) |\n+| JDK | 17 |\n+| Kotlin | 2.1.0 |\n+| minSdk | 26 (Android 8.0) |\n+| ONNX Runtime | 1.21.1 |\n+| OpenCV | 4.5.3 |\n+\n+## 快速开始\n+\n+### 1. 克隆项目\n+\n+```bash\n+git clone https://github.com/PaddlePaddle/PaddleOCR.git\n+cd PaddleOCR/deploy/ppocr-android\n+```\n+\n+### 2. 准备模型\n+\n+本项目支持以下模型：\n+\n+| 模型 | HuggingFace | BOS |\n+|------|-------------|-----|\n+| **PP-OCRv6_small** | [检测模型](https://huggingface.co/PaddlePaddle/PP-OCRv6_small_det_onnx) / [识别模型](https://huggingface.co/PaddlePaddle/PP-OCRv6_small_rec_onnx) | [检测模型](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv6_small_det_onnx_infer.tar) / [识别模型](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv6_small_rec_onnx_infer.tar) |\n+| **PP-OCRv6_tiny** | [检测模型](https://huggingface.co/PaddlePaddle/PP-OCRv6_tiny_det_onnx) / [识别模型](https://huggingface.co/PaddlePaddle/PP-OCRv6_tiny_rec_onnx) | [检测模型](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv6_tiny_det_onnx_infer.tar) / [识别模型](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv6_tiny_rec_onnx_infer.tar) |\n+| **PP-OCRv5_mobile** | [检测模型](https://huggingface.co/PaddlePaddle/PP-OCRv5_mobile_det_onnx) / [识别模型](https://huggingface.co/PaddlePaddle/PP-OCRv5_mobile_rec_onnx) | [检测模型](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv5_mobile_det_onnx_infer.tar) / [识别模型](https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv5_mobile_rec_onnx_infer.tar) |\n+\n+下载并解压后，将文件放入 `ppocr-sdk/src/main/assets/models/` 目录：\n+\n+- 检测模型：将 `inference.onnx` 放入 `models/det/`\n+- 识别模型：将 `inference.onnx` 和 `inference.yml` 放入 `models/rec/`\n+\n+### 3. 编译运行\n+\n+```bash\n+# 编译 Debug APK\n+./gradlew :app:assembleDebug\n+\n+# 安装到设备\n+./gradlew :app:installDebug\n+```\n+\n+或使用 Android Studio 直接运行。\n+\n+### 4. 体验 Demo\n+\n+1. 打开 \"PP-OCRv6 Demo\" 应用\n+2. 等待模型加载完成\n+3. 点击 \"Select from Gallery\" 选择图片\n+4. 查看识别结果和耗时统计\n+\n+## SDK 集成\n+\n+### 方式一：源码依赖\n+\n+1. 将 `ppocr-sdk/` 复制到项目根目录\n+2. 在 `settings.gradle.kts` 添加：\n+ ```kotlin\n+ include(\":ppocr-sdk\")\n+ ```\n+3. 在 App 模块 `build.gradle.kts` 添加：\n+ ```kotlin\n+ implementation(project(\":ppocr-sdk\"))\n+ ```\n+\n+### 方式二：AAR 依赖\n+\n+```bash\n+# 构建 AAR\n+./gradlew :ppocr-sdk:assembleRelease\n+```\n+\n+AAR 输出：`ppocr-sdk/build/outputs/aar/ppocr-sdk-release.aar`\n+\n+在 App 模块 `build.gradle.kts` 添加：\n+\n+```kotlin\n+dependencies {\n+ implementation(files(\"libs/ppocr-sdk-release.aar\"))\n+ // AAR 不传递依赖，需手动添加\n+ implementation(\"com.microsoft.onnxruntime:onnxruntime-android:1.21.1\")\n+ implementation(\"com.quickbirdstudios:opencv:4.5.3\")\n+ implementation(\"org.jetbrains.kotlinx:kotlinx-coroutines-android:1.9.0\")\n+}\n+```\n+\n+## API 参考\n+\n+### 创建实例\n+\n+```kotlin\n+// 默认配置\n+val ocr = PaddleOCR.create(context)\n+\n+// 自定义配置\n+val ocr = PaddleOCR.create(\n+ context = context,\n+ config = PaddleOCRConfig(\n+ detThresh = 0.3f,\n+ detBoxThresh = 0.6f,\n+ recScoreThresh = 0.0f,\n+ recBatchSize = 1,\n+ ),\n+ engineConfig = EngineConfig(numThreads = 4),\n+ detModelAssetPath = \"models/det/inference.onnx\",\n+ recModelAssetPath = \"models/rec/inference.onnx\",\n+ recConfigAssetPath = \"models/rec/inference.yml\",\n+)\n+```\n+\n+### 执行 OCR\n+\n+```kotlin\n+// 传入 Bitmap\n+val result = ocr.recognize(bitmap)\n+\n+// 传入图片字节数据（推荐，与 Python 流程一致）\n+val result = ocr.recognize(imageBytes)\n+\n+// 读取结果\n+result.results.forEach { item ->\n+ println(\"文本: ${item.text}, 置信度: ${item.confidence}\")\n+ println(\"坐标: ${item.box.points}\")\n+}\n+println(\"检测: ${result.detectionTimeMs}ms, 识别: ${result.recognitionTimeMs}ms\")\n+```\n+\n+### 释放资源\n+\n+```kotlin\n+ocr.release()\n+```\n+\n+### 配置参数\n+\n+```kotlin\n+data class PaddleOCRConfig(\n+ val detImgMode: String = \"BGR\", // 输入色彩模式\n+ val detLimitSideLen: Int = 64, // 检测侧边长限制\n+ val detLimitType: String = \"min\", // 限制策略\n+ val detMaxSideLimit: Int = 4000, // 最长边上限\n+ val detThresh: Float = 0.3f, // 二值化阈值\n+ val detBoxThresh: Float = 0.6f, // 检测框置信度阈值\n+ val detUnclipRatio: Float = 1.5f, // 检测框扩展比例\n+ val detMaxCandidates: Int = 3000, // 最大候选框数\n+ val detUseDilation: Boolean = false, // 是否膨胀\n+ val detScoreMode: String = \"fast\", // 打分模式\n+ val detBoxType: String = \"quad\", // 检测框类型\n+ val recScoreThresh: Float = 0.0f, // 识别置信度阈值\n+ val recBatchSize: Int = 1, // 识别批大小\n+)\n+```\n+\n+### 结果模型\n+\n+```kotlin\n+data class OCRRunResult(\n+ val results: List<OCRResult>, // 识别结果列表\n+ val detectionTimeMs: Long, // 检测耗时\n+ val recognitionTimeMs: Long, // 识别耗时\n+ val totalTimeMs: Long, // 总耗时\n+ val lineCount: Int, // 识别行数\n+ // 详细计时...\n+)\n+\n+data class OCRResult(\n+ val box: OCRBox, // 检测框坐标\n+ val text: String, // 识别文本\n+ val confidence: Float, // 置信度\n+)\n+```\n+\n+## 性能测试\n+\n+项目提供自动化性能测试脚本：\n+\n+```bash\n+# 运行 benchmark（10次测试，3次预热）\n+./run_benchmark.sh 10 3\n+\n+# 输出示例\n+╔═════════════════════════════════════════════════════════════════════════╗\n+║ PP-OCRv6 Speed Benchmark Results ║\n+╠═════════════════════════════════════════════════════════════════════════╣\n+║ Device: GM1900 | OS: Android 9 | Lines: 5 ║\n+║ Cold load: 158ms | Warmup: 3 | Measured: 10 ║\n+╠═════════════════════════════════════════════════════════════════════════╣\n++-----------------------------+----------+----------+----------+----------+\n+| Stage | Mean ms | Stdev | P90 | Min ms|\n++-----------------------------+----------+----------+----------+----------+\n+| Total pipeline | 420.40 | 6.37 | 427 | 413 |\n++-----------------------------+----------+----------+----------+----------+\n+| Detection (total) | 348.70 | 4.67 | 356 | 343 |\n+| Preprocess | 33.30 | 2.90 | 36 | 28 |\n+| Inference | 311.00 | 2.93 | 315 | 304 |\n+| Postprocess | 4.40 | 0.49 | 5 | 4 |\n+| Recognition (total) | 66.20 | 3.16 | 68 | 64 |\n+| Preprocess | 3.00 | 0.89 | 4 | 2 |\n+| Inference | 60.60 | 3.14 | 63 | 58 |\n+| Postprocess | 2.60 | 0.92 | 4 | 1 |\n+| Pipeline overhead | 5.50 | 0.50 | 6 | 5 |\n++-----------------------------+----------+----------+----------+----------+\n+╚═════════════════════════════════════════════════════════════════════════╝\n+```\n+\n+## 注意事项\n+\n+1. **OpenCV 初始化**：调用 `PaddleOCR.create()` 前需先调用 `OpenCVUtils.init(context)`\n+2. **协程调用**：`create()` 和 `recognize()` 都是 suspend 函数，需在协程中调用\n+3. **内存管理**：不再使用时调用 `release()` 释放资源\n+4. **混淆规则**：参考 `ppocr-sdk/proguard-rules.pro`\ndiff --git a/mkdocs.yml b/mkdocs.yml\nindex 6a64d1ed041..3e9ae264e4c 100644\n--- a/mkdocs.yml\n+++ b/mkdocs.yml\n@@ -139,6 +139,7 @@ plugins:\n 打包 PaddleOCR 项目: Package PaddleOCR Projects\n 获取ONNX模型: Obtaining ONNX Models\n 端侧部署: On-Device Deployment\n+ Android 部署: Android Deployment\n 服务化部署: Serving\n 浏览器端部署: Browser Deployment\n 模块列表: Module List\n@@ -362,6 +363,7 @@ nav:\n - CLI: version3.x/inference_deployment/serving/paddleocr_official_api/cli.md\n - 跨端部署:\n - 端侧部署: version3.x/inference_deployment/cross_platform/on_device_deployment.md\n+ - Android 部署: version3.x/inference_deployment/cross_platform/android_deployment.md\n - 浏览器端部署: version3.x/inference_deployment/cross_platform/browser.md\n - 其它:\n - 获取ONNX模型: version3.x/inference_deployment/others/obtaining_onnx_models.md\n"},"changed_files":{"kind":"list like","value":["deploy/ppocr-android/.gitignore","deploy/ppocr-android/README.md","deploy/ppocr-android/README_en.md","deploy/ppocr-android/app/build.gradle.kts","deploy/ppocr-android/app/src/main/AndroidManifest.xml","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/MainActivity.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/OCRApplication.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ErrorDialog.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePicker.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePreview.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/LoadingOverlay.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ResultList.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/TimingBar.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/screen/HomeScreen.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/theme/Theme.kt","deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/viewmodel/OCRViewModel.kt","deploy/ppocr-android/app/src/main/res/drawable/ic_launcher_foreground.xml","deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml","deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml","deploy/ppocr-android/app/src/main/res/values/colors.xml","deploy/ppocr-android/app/src/main/res/values/strings.xml","deploy/ppocr-android/app/src/main/res/values/themes.xml","deploy/ppocr-android/build.gradle.kts","deploy/ppocr-android/gradle.properties","deploy/ppocr-android/gradle/libs.versions.toml","deploy/ppocr-android/gradle/wrapper/gradle-wrapper.jar","deploy/ppocr-android/gradle/wrapper/gradle-wrapper.properties","deploy/ppocr-android/gradlew","deploy/ppocr-android/ppocr-sdk/build.gradle.kts","deploy/ppocr-android/ppocr-sdk/proguard-rules.pro","deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/BenchmarkFixtures.kt","deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/OCRBenchmarkTest.kt","deploy/ppocr-android/ppocr-sdk/src/androidTest/res/raw/android_ocr_benchmark_reference.png","deploy/ppocr-android/ppocr-sdk/src/main/AndroidManifest.xml","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/EngineConfig.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCR.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCRConfig.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/DetectionEngine.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngine.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngineResult.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/ORTSessionManager.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/RecognitionEngine.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/ModelConfig.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRBox.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRError.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRResult.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRRunResult.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/BoxSorter.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/CTCDecoder.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/DBPostProcessor.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/PolygonUnclip.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadGeometry.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadTextCrop.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/DetPreprocessor.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/RecPreprocessor.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/BitmapUtils.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/ImageUtils.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/MathUtils.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/OpenCVUtils.kt","deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/YamlUtils.kt","deploy/ppocr-android/run_benchmark.sh","deploy/ppocr-android/settings.gradle.kts","docs/version3.x/inference_deployment/cross_platform/android_deployment.en.md","docs/version3.x/inference_deployment/cross_platform/android_deployment.md","mkdocs.yml"],"string":"[\n \"deploy/ppocr-android/.gitignore\",\n \"deploy/ppocr-android/README.md\",\n \"deploy/ppocr-android/README_en.md\",\n \"deploy/ppocr-android/app/build.gradle.kts\",\n \"deploy/ppocr-android/app/src/main/AndroidManifest.xml\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/MainActivity.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/OCRApplication.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ErrorDialog.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePicker.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ImagePreview.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/LoadingOverlay.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/ResultList.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/component/TimingBar.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/screen/HomeScreen.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/theme/Theme.kt\",\n \"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/ui/viewmodel/OCRViewModel.kt\",\n \"deploy/ppocr-android/app/src/main/res/drawable/ic_launcher_foreground.xml\",\n \"deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher.xml\",\n \"deploy/ppocr-android/app/src/main/res/mipmap-anydpi-v26/ic_launcher_round.xml\",\n \"deploy/ppocr-android/app/src/main/res/values/colors.xml\",\n \"deploy/ppocr-android/app/src/main/res/values/strings.xml\",\n \"deploy/ppocr-android/app/src/main/res/values/themes.xml\",\n \"deploy/ppocr-android/build.gradle.kts\",\n \"deploy/ppocr-android/gradle.properties\",\n \"deploy/ppocr-android/gradle/libs.versions.toml\",\n \"deploy/ppocr-android/gradle/wrapper/gradle-wrapper.jar\",\n \"deploy/ppocr-android/gradle/wrapper/gradle-wrapper.properties\",\n \"deploy/ppocr-android/gradlew\",\n \"deploy/ppocr-android/ppocr-sdk/build.gradle.kts\",\n \"deploy/ppocr-android/ppocr-sdk/proguard-rules.pro\",\n \"deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/BenchmarkFixtures.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/androidTest/java/com/paddle/ocr/benchmark/OCRBenchmarkTest.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/androidTest/res/raw/android_ocr_benchmark_reference.png\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/AndroidManifest.xml\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/EngineConfig.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCR.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/PaddleOCRConfig.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/DetectionEngine.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngine.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/OCREngineResult.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/ORTSessionManager.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/engine/RecognitionEngine.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/ModelConfig.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRBox.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRError.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRResult.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/model/OCRRunResult.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/BoxSorter.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/CTCDecoder.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/DBPostProcessor.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/PolygonUnclip.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadGeometry.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/postprocess/QuadTextCrop.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/DetPreprocessor.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/preprocess/RecPreprocessor.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/BitmapUtils.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/ImageUtils.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/MathUtils.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/OpenCVUtils.kt\",\n \"deploy/ppocr-android/ppocr-sdk/src/main/java/com/paddle/ocr/util/YamlUtils.kt\",\n \"deploy/ppocr-android/run_benchmark.sh\",\n \"deploy/ppocr-android/settings.gradle.kts\",\n \"docs/version3.x/inference_deployment/cross_platform/android_deployment.en.md\",\n \"docs/version3.x/inference_deployment/cross_platform/android_deployment.md\",\n \"mkdocs.yml\"\n]","isConversational":false},"review_threads":{"kind":"list like","value":[],"string":"[]","isConversational":false},"test_patch":{"kind":"string","value":""},"merged":{"kind":"bool","value":true,"string":"true"}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":89,"numItemsPerPage":100,"numTotalItems":87912,"offset":8900,"length":100}},"jwt":"eyJhbGciOiJFZERTQSIsImtpZCI6IjVHZDBvd0g5MTM2eDZjc1FvbE1zcktNWUZoRVFoUm5PVVVybEpjOUhaUEEifQ.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4NjczMDMwNCwianRpIjoiYTc3N2EwZWQtMGM5NS00MDk3LWI3ZjQtZTkyNGQ4ZDg3MDExIiwic3ViIjoiL2RhdGFzZXRzL1BvdGF0b0hEL2dpdGh1Yi1wci1yZXNvbHZlIiwiZXhwIjoxNzg2NzMzOTA0LCJpc3MiOiJodHRwczovL2h1Z2dpbmdmYWNlLmNvIn0.BBIaIx4-kgjuFOcVwLioTgVIiTV-sU35ks3HvoGlB0CpXSuBITiHLdu-zKbBKSbKHW4GYy6dVh4vy_dgssVvAg","displayUrls":true,"splitSizeSummaries":[{"config":"default","split":"train","numRows":87912,"numBytesParquet":1412740926}]},"discussionsStats":{"closed":0,"open":0,"total":0},"fullWidth":true,"hasGatedAccess":true,"hasFullAccess":true,"isEmbedded":false,"savedQueries":{"community":[],"user":[]}}">
repo stringlengths 5 53 | pr_number int32 1 321k | task_type stringclasses 2
values | issue_text stringlengths 0 81.2k | pr_title stringlengths 1 319 | pr_body stringlengths 0 105k | base_sha stringlengths 40 40 | head_sha stringlengths 40 40 | gold_diff stringlengths 0 202M | changed_files listlengths 0 100 | review_threads listlengths 0 100 | test_patch stringlengths 0 23.4M | merged bool 1
class |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
PaddlePaddle/Paddle | 67,371 | issue_to_patch | 【Mthreads】support features for llama 13 b | 1. Support/optimize operators for LLaMA 13B
- [x] mudnn matmul
- [x] mudnn flashattn
- [x] some unary/binary ops
- [x] optimze cross entropy/rope
2. redefine macro, PADDLE_ENFORCE and PADDLE_ENFORCE_NOT_NULL for improving performance.
3. extend cpp_extension module for building rmsnorm of paddlenlp repo.
4. skip... | cb5568b738f291724573c9b03690a5e075bbfb36 | 65d9c5886090bab5dafa12e086cfa83824c96264 | diff --git a/build.sh b/build.sh
new file mode 100644
index 00000000000000..08d3542ecbccff
--- /dev/null
+++ b/build.sh
@@ -0,0 +1,7 @@
+#!/bin/bash
+mkdir -p build
+pushd build
+cmake .. -DWITH_MKL=ON -DWITH_MUSA=ON -DWITH_MCCL=ON -DWITH_GPU=OFF -DPY_VERSION=3.9 -DWITH_CINN=OFF -DCMAKE_EXPORT_COMPILE_COMMANDS=on
+mak... | [
"build.sh",
"cmake/configure.cmake",
"cmake/musa.cmake",
"cmake/phi.cmake",
"paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc",
"paddle/fluid/operators/ops_signature/generated_fused_sig.cc",
"paddle/fluid/operators/ops_signature/generated_sig.cc",
"paddle/fluid/operators/ops_signature/gene... | [] | diff --git a/test/collective/fleet/dygraph_group_sharded_stage2.py b/test/collective/fleet/dygraph_group_sharded_stage2.py
index 81f6df163f1db5..b7213e819e2f4c 100644
--- a/test/collective/fleet/dygraph_group_sharded_stage2.py
+++ b/test/collective/fleet/dygraph_group_sharded_stage2.py
@@ -208,7 +208,7 @@ def test_dp_s... | true | |
PaddlePaddle/Paddle | 79,277 | issue_to_patch | [API Compatibility] align smooth_l1_loss by adding it to compat module -part |
### PR Category
User Experience
### PR Types
New features
### Description
对 smooth_l1_loss 的修复
相关 PR：
- https://github.com/PaddlePaddle/Paddle/pull/78811#discussion_r3256346398
### 是否引起精度变化
否
| e3133879709c6bf869b8952d62109f9296c64cd5 | 6974db011f184153077b113d4d95f8c5cf343ea3 | diff --git a/python/paddle/compat/nn/__init__.py b/python/paddle/compat/nn/__init__.py
index a567b396d8815b..7f0c1f3cb39bef 100644
--- a/python/paddle/compat/nn/__init__.py
+++ b/python/paddle/compat/nn/__init__.py
@@ -15,6 +15,7 @@
from __future__ import annotations
import collections
+import warnings
from iterto... | [
"python/paddle/compat/nn/__init__.py",
"python/paddle/compat/nn/functional/__init__.py",
"test/legacy_test/test_compat_smooth_l1_loss.py"
] | [
{
"comment": "把这些注释移除一下",
"path": "python/paddle/compat/nn/functional/__init__.py",
"hunk": "@@ -380,3 +383,96 @@ def to_list_if_necessary(x):\n paddings=to_list_if_necessary(padding),\n dilations=to_list_if_necessary(dilation),\n )\n+\n+\n+@ForbidKeywordsDecorator(\n+ illegal_key... | diff --git a/test/legacy_test/test_compat_smooth_l1_loss.py b/test/legacy_test/test_compat_smooth_l1_loss.py
new file mode 100644
index 00000000000000..9dfafeacac257d
--- /dev/null
+++ b/test/legacy_test/test_compat_smooth_l1_loss.py
@@ -0,0 +1,229 @@
+# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+#
... | true | |
PaddlePaddle/Paddle | 79,307 | issue_to_patch | [Allocator] Add generic all block info API (#79272) | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | d3c4748dfe0f71ada11774cf2518f3a3ad2327d8 | e37974522fe115a975a41fb932e484ae37727ded | diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 70b6c6263bbb33..5df439fe6ebc6a 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -3753,8 +3753,11 @@ All parameter, weight, gradient are variables in Paddle.
m.def("vmm_free_block_info", [](int device_id)... | [
"paddle/fluid/pybind/pybind.cc",
"paddle/phi/core/memory/allocation/auto_growth_best_fit_allocator.h",
"paddle/phi/core/memory/mem_utils.cc",
"paddle/phi/core/memory/mem_utils.h",
"paddle/phi/core/memory/mem_visitor.cc",
"paddle/phi/core/memory/mem_visitor.h",
"python/paddle/device/cuda/memory_analyzer.... | [
{
"comment": "**🔴 兼容性** 这里把已导出的 `AllBlockInfoOfVmmAllocator` 直接改名成 `AllBlockInfoOfAllocator`，但没有保留旧 C++ 符号。\n\n`mem_utils.h` 中旧函数带 `PADDLE_API extern`，下游 C++ 代码或已编译扩展如果继续包含/链接 `paddle::memory::AllBlockInfoOfVmmAllocator`，在 release/3.4 分支升级后会出现源码或链接兼容性破坏。Python 的 `vmm_all_block_info` 旧入口已经保留，C++ 层也应保持同样的兼容 alia... | diff --git a/test/cpp/fluid/memory/allocator_visitor_test.cc b/test/cpp/fluid/memory/allocator_visitor_test.cc
index 1df632e3b1eac4..91c9dce4f59800 100644
--- a/test/cpp/fluid/memory/allocator_visitor_test.cc
+++ b/test/cpp/fluid/memory/allocator_visitor_test.cc
@@ -12,6 +12,8 @@
// See the License for the specific la... | true | |
PaddlePaddle/Paddle | 79,263 | issue_to_patch | [CI] Fix empty C++ coverage reports | ### PR Category
Execute Infrastructure
### PR Types
Bug fixes
### Description
修复 Coverage 任务中 C++ 覆盖率数据为空仍可能通过的问题。
背景是部分 Coverage job 中构建使用的 GCC 与采集阶段默认 gcov 主版本不一致，lcov/gcov 读取 gcno/gcda 失败后会生成空的 coverage-diff.info。后续 coverage_judge.sh 对空数据只输出 no data found，没有把这种采集失败识别为失败，导致 C++ diff coverage 没有实际数据也可能被判定通过。
本 PR ... | 4d12baa6a2c7ea4dbb0e799e6475bc1d0a9e6794 | cd6fd34f1b1a8d50cdbc477f0fc765347f4f8977 | diff --git a/ci/coverage_info.sh b/ci/coverage_info.sh
index 703a3d886151d..d785719f99987 100644
--- a/ci/coverage_info.sh
+++ b/ci/coverage_info.sh
@@ -32,6 +32,92 @@ echo "::endgroup::"
cd ${PADDLE_ROOT}/build
+function init_gcov_tool() {
+ local cxx_compiler=""
+ local compiler_version=""
+ local compi... | [
"ci/coverage_info.sh"
] | [
{
"comment": "init_gcov_tool() does not reset the global GCOV_TOOL before searching. If GCOV_TOOL is already set in the environment and no candidate gcov binary is found, the function will silently reuse the stale value and skip the intended \"no gcov\" failure path (and version match check), which can undermin... | true | ||
PaddlePaddle/Paddle | 79,233 | issue_to_patch | 【Hackathon 10th Spring No.51】Apply yaml-cpp GCC 15 compatibility patch - part 5 | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 8d925a3e4842bc41c4c11af8d37265d2be635d4f | fb301d72c2725f4fcabf78c513b91a1b5b108364 | diff --git a/cmake/external/yaml.cmake b/cmake/external/yaml.cmake
index 9c094ea4e573f2..93ae4ba9204371 100644
--- a/cmake/external/yaml.cmake
+++ b/cmake/external/yaml.cmake
@@ -23,6 +23,12 @@ set(SOURCE_INCLUDE_DIR ${SOURCE_DIR}/include)
include_directories(${YAML_INCLUDE_DIR})
+file(TO_NATIVE_PATH
+ ${PADDL... | [
"cmake/external/yaml.cmake",
"patches/yaml-cpp/emitterutils.cpp.patch"
] | [] | true | ||
PaddlePaddle/Paddle | 79,203 | issue_to_patch | [CI] Upgrade H-Coverage to Python 3.12 | ### PR Category
Environment Adaptation
### PR Types
Improvements
### Description
本 PR 仅将 H-Coverage (`CI-H`) 流水线从 Python 3.10 升级到 Python 3.12，不修改 Linux-CPU / NPU / 普通 Coverage 等其它流水线。
主要改动：
- `H-Coverage.yml`：build / test / Fleet 相关 job 的 `PY_VERSION` 改为 `"3.12"`，并同步更新显式 `cmake -DPY_VERSION=3.12`。
- H-Coverage 相关容器统... | 8da1ec306684956ad70cc6f06bbc6f626d51243a | 175d43b72937821819d4467b64e02057c5fd0ff4 | diff --git a/.github/workflows/H-Coverage.yml b/.github/workflows/H-Coverage.yml
index 49f9efa3ba31fc..766fd8092fb9fa 100644
--- a/.github/workflows/H-Coverage.yml
+++ b/.github/workflows/H-Coverage.yml
@@ -104,7 +104,7 @@ jobs:
LITE_GIT_TAG: develop
WITH_UNITY_BUILD: "ON"
WITH_FA_BUILD... | [
".github/workflows/H-Coverage.yml",
"test/test_flashmask_ci/flashmask_gt.json",
"test/test_flashmask_ci/run.sh"
] | [
{
"comment": " **优先级：P1**\n\n这里直接假设 `cuda129-coverage-test` 容器内已经有 `python3.12`/`pip3.12`，但当前 H-Coverage `Coverage build` 已在这组新增检查处失败：日志在 flashattn URL 检查后没有进入 `pip install`/`cmake`，随后退出 1（见 https://github.com/PaddlePaddle/Paddle/actions/runs/26721401586/job/787... | diff --git a/test/test_flashmask_ci/flashmask_gt.json b/test/test_flashmask_ci/flashmask_gt.json
index e644d6e4e2faff..768a3c2f1df60c 100644
--- a/test/test_flashmask_ci/flashmask_gt.json
+++ b/test/test_flashmask_ci/flashmask_gt.json
@@ -19,16 +19,16 @@
"gen_startend_row_indices7-1-8192-33792-2-1-1-80-80-3-dtype0":... | true | |
PaddlePaddle/Paddle | 79,119 | issue_to_patch | [CI] Upgrade Linux-CPU and NPU CI to Python 3.10 | ### PR Category
Environment Adaptation
### PR Types
Deprecations
### Description
按 @SigureMo 的建议，单独把 Linux-CPU CI 从 Python 3.9 升级，并同步更新下游 Linux-NPU 消费 CPU wheel 的路径。
NPU 现有 CI 镜像只显式准备了 Python 3.9 / 3.10 相关依赖；本 PR 的目标是退掉 3.9，不强制升到 3.12。因此当前改为 Linux-CPU 与 Linux-NPU 统一使用 Python 3.10，继续使用现有 NPU 镜像。
主要改动：
- `_Linux-CPU.... | 7ac7614c37f9cd21a8f6bb009ec857afa253aee7 | 50ad302f5da18f8ec0debb8e9bc7dfff6e6a9c90 | diff --git a/.github/workflows/_Linux-CPU.yml b/.github/workflows/_Linux-CPU.yml
index 91e2c6fec445ff..d54486e1430ee5 100644
--- a/.github/workflows/_Linux-CPU.yml
+++ b/.github/workflows/_Linux-CPU.yml
@@ -63,7 +63,7 @@ jobs:
WITH_DISTRIBUTE: "ON"
PREC_SUFFIX: .py3
WITH_UNITY_BUILD: "O... | [
".github/workflows/_Linux-CPU.yml",
".github/workflows/_Linux-NPU.yml"
] | [
{
"comment": "这个文件的改动是不是只应该是 3.9 -> 3.10 的改动，现在是否有一些没有必要的改动？",
"path": ".github/workflows/_Linux-NPU.yml",
"hunk": "",
"resolving_sha": "50ad302f5da18f8ec0debb8e9bc7dfff6e6a9c90",
"resolving_diff": "diff --git a/.github/workflows/_Linux-NPU.yml b/.github/workflows/_Linux-NPU.yml\nindex 6e143c106... | true | ||
PaddlePaddle/Paddle | 79,107 | issue_to_patch | 【Hackathon 10th Spring No.52】[CI] Update `Coverage` Python version to 3.12 in CI configuration and scripts - part 2 | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 2665e9e6d263e7b022ab54cacc0cbafd7ae5b2d5 | cbff14314b9c05c75ad02ee82be3767b945eb9b1 | diff --git a/.github/workflows/Coverage.yml b/.github/workflows/Coverage.yml
index d4e2786fa15a50..687e1d39be5421 100644
--- a/.github/workflows/Coverage.yml
+++ b/.github/workflows/Coverage.yml
@@ -109,7 +109,7 @@ jobs:
WITH_FLAGCX: "ON"
LITE_GIT_TAG: develop
WITH_UNITY_BUILD: "ON"
- ... | [
".github/workflows/Coverage.yml",
"ci/coverage_test.sh",
"paddle/phi/kernels/fusion/gpu/block_multi_head_attention_kernel.cu",
"paddle/phi/kernels/fusion/gpu/quant_dequant_kernel.h",
"python/paddle/distributed/auto_parallel/process_mesh.py",
"test/ai_edited_test/test_ai_amp_advanced.py",
"test/ai_edited... | [
{
"comment": "This change bumps Paddle's global CMake `PY_VERSION` default/minimum from 3.9 to 3.10, which is broader than the PR description (Coverage CI to py3.12) and conflicts with the current Python support contract elsewhere (e.g., `setup.py` enforces >=3.9 and docs list 3.9 support). As-is, builds config... | diff --git a/ci/coverage_test.sh b/ci/coverage_test.sh
index dfd9abca4fac67..fed2ed0b3048c6 100644
--- a/ci/coverage_test.sh
+++ b/ci/coverage_test.sh
@@ -52,8 +52,9 @@ unset GREP_OPTIONS
source $(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/utils.sh
init
-ln -sf $(which python3.9) /usr/local/bin/python
-ln -sf $(wh... | true | |
PaddlePaddle/Paddle | 79,275 | issue_to_patch | [API Compatibility] Align torch.nn.attention.flex_attention.or_masks/and_masks | ### PR Category
User Experience
### PR Types
New features
### Description
- 新增 python/paddle/nn/attention/flex_attention.py
- or_masks(*mask_mods)：返回一个新 mask，结果是多个 mask 的逻辑 OR。
- and_masks(*mask_mods)：返回一个新 mask，结果是多个 mask 的逻辑 AND。
- 入参非 callable 时按 PyTorch 抛 RuntimeError。
### 是否引起精度变化
否
... | 8cacdfd15bc89296682c784df5b1685a7ca6e408 | 48788851973ac09bea02bdfa12a231a9846b8538 | diff --git a/python/paddle/nn/attention/__init__.py b/python/paddle/nn/attention/__init__.py
index ba0ae208316b33..dbc795783c277a 100644
--- a/python/paddle/nn/attention/__init__.py
+++ b/python/paddle/nn/attention/__init__.py
@@ -12,6 +12,7 @@
# See the License for the specific language governing permissions and
# l... | [
"python/paddle/nn/attention/__init__.py",
"python/paddle/nn/attention/flex_attention.py",
"test/legacy_test/test_api_compatibility_part2.py"
] | [
{
"comment": "pir下也有new_zeros，这个是动静统一的",
"path": "python/paddle/nn/attention/flex_attention.py",
"hunk": "@@ -0,0 +1,130 @@\n+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+#\n+# Licensed under the Apache License, Version 2.0 (the \"License\");\n+# you may not use this file except i... | diff --git a/test/legacy_test/test_api_compatibility_part2.py b/test/legacy_test/test_api_compatibility_part2.py
index 4ce6d5c0f789cb..85b498f44de23d 100644
--- a/test/legacy_test/test_api_compatibility_part2.py
+++ b/test/legacy_test/test_api_compatibility_part2.py
@@ -19,6 +19,125 @@
import paddle
+# Test flex_a... | true | |
PaddlePaddle/Paddle | 79,267 | issue_to_patch | [API Compatibility] Support `create_graph` in `paddle.Tensor.backward()` and `paddle.autograd.backward()` | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
User Experience
### PR Types
Improvements
### Description
<!-- Describe what you’ve done -->
- Support param `create... | c3a820dbcb6b100fcd1cde032753521ebbf88831 | 709ce80565ad23ad50a36518a77ea73ab954f971 | diff --git a/paddle/fluid/eager/backward.cc b/paddle/fluid/eager/backward.cc
index f0ace4b0fe828..7ea5bdb003f3e 100644
--- a/paddle/fluid/eager/backward.cc
+++ b/paddle/fluid/eager/backward.cc
@@ -746,6 +746,7 @@ std::vector<paddle::Tensor> RunBackward(
void Backward(const std::vector<paddle::Tensor>& tensors, // out... | [
"paddle/fluid/eager/backward.cc",
"paddle/fluid/eager/backward.h",
"paddle/fluid/pybind/eager_functions.cc",
"python/paddle/autograd/backward_mode.py",
"python/paddle/base/dygraph/tensor_patch_methods.py",
"test/legacy_test/test_backward.py"
] | [
{
"comment": "**🟡 建议** 新增 `create_graph` 公开参数后，docstring 的 `Args` 段仍只说明到 `retain_graph`，没有描述 `create_graph` 的语义、默认值以及和 `retain_graph` 的关系。\n\n`paddle.grad` 已经在文档中说明 `create_graph` 用于构建梯度计算图、支持高阶导数；这里是同类公开入口，缺文档会让生成的 API 文档和用户可见签名不一致。\n\n建议修复方式：\n在 `Args` 中补充 `create_graph(bool, optional)`，说明 `True` 时会构建反向计算过程的... | diff --git a/test/legacy_test/test_backward.py b/test/legacy_test/test_backward.py
index 05fdb572c79de..9f18de07f9270 100644
--- a/test/legacy_test/test_backward.py
+++ b/test/legacy_test/test_backward.py
@@ -101,6 +101,175 @@ def test_strip_grad_suffix(self):
self.assertEqual(backward._strip_grad_suffix_(... | true | |
PaddlePaddle/Paddle | 79,259 | issue_to_patch | acc paddle.sqrt | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 9d69b2f2501224aff7c69fb7204cd4e8281ec740 | 14247c71999fbbcff6c9b18a143e36ecb50d9b00 | diff --git a/paddle/phi/kernels/funcs/activation_functor.h b/paddle/phi/kernels/funcs/activation_functor.h
index 78bec8faa5185a..fd03b6a9e6483a 100644
--- a/paddle/phi/kernels/funcs/activation_functor.h
+++ b/paddle/phi/kernels/funcs/activation_functor.h
@@ -4608,7 +4608,7 @@ struct CudaRsqrtFunctor<ComplexType<T>>
... | [
"paddle/phi/kernels/funcs/activation_functor.h",
"paddle/phi/kernels/gpu/activation_grad_kernel.cu"
] | [
{
"comment": "用FLAGS_use_accuracy_compatible_kernel控制一下吧~",
"path": "paddle/phi/kernels/funcs/activation_functor.h",
"hunk": "@@ -4610,15 +4610,12 @@ struct CudaRsqrtFunctor<ComplexType<T>>\n \n template <typename T>\n struct CudaRsqrtGradFunctor : public BaseActivationFunctor<T> {\n- using MT = typena... | true | ||
PaddlePaddle/Paddle | 79,305 | issue_to_patch | [Operator Mechanism] Add uint32 max enforce helper | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 4ce8ac4a365da4acdff66a9004fc9bc1596b739e | 9b122fc85fc4f9131b22536f5156affcef5aa820 | diff --git a/paddle/common/enforce.h b/paddle/common/enforce.h
index 430f9ab575df08..fa23e810733413 100644
--- a/paddle/common/enforce.h
+++ b/paddle/common/enforce.h
@@ -17,6 +17,7 @@
#ifdef __GNUC__
#include <cxxabi.h> // for __cxa_demangle
#endif // __GNUC__
+#include <cinttypes>
#include <excepti... | [
"paddle/common/enforce.h"
] | [
{
"comment": "**🟡 建议** 这个公共 helper 当前没有限制整数类型，后续误传 `double` / `float` 时会先做整数转换。\n\n`std::is_signed<double>` 为 true，负数错误分支会在格式化时执行 `static_cast<int64_t>(var)`；上界分支也会先执行 `static_cast<uint64_t>(__var)` 再比较。对 NaN 或超出目标整数表示范围的浮点值，这个转换本身就可能触发未定义行为；对 `1.5` 这类值还会先截断后通过检查，和 `fit in uint32_t` 的语义不一致。\n\n建议把宏限定为整数输入，例如对 ... | true | ||
PaddlePaddle/Paddle | 79,305 | comment_to_fix | [Operator Mechanism] Add uint32 max enforce helper | **🟡 建议** 这个公共 helper 当前没有限制整数类型，后续误传 `double` / `float` 时会先做整数转换。
`std::is_signed<double>` 为 true，负数错误分支会在格式化时执行 `static_cast<int64_t>(var)`；上界分支也会先执行 `static_cast<uint64_t>(__var)` 再比较。对 NaN 或超出目标整数表示范围的浮点值，这个转换本身就可能触发未定义行为；对 `1.5` 这类值还会先截断后通过检查，和 `fit in uint32_t` 的语义不一致。
建议把宏限定为整数输入，例如对 `std::decay_t<decltype(__v... | 4ce8ac4a365da4acdff66a9004fc9bc1596b739e | 9b122fc85fc4f9131b22536f5156affcef5aa820 | diff --git a/paddle/common/enforce.h b/paddle/common/enforce.h
index 430f9ab575df08..fa23e810733413 100644
--- a/paddle/common/enforce.h
+++ b/paddle/common/enforce.h
@@ -17,6 +17,7 @@
#ifdef __GNUC__
#include <cxxabi.h> // for __cxa_demangle
#endif // __GNUC__
+#include <cinttypes>
#include <excepti... | [
"paddle/common/enforce.h"
] | [
{
"comment": "**🟡 建议** 这个公共 helper 当前没有限制整数类型，后续误传 `double` / `float` 时会先做整数转换。\n\n`std::is_signed<double>` 为 true，负数错误分支会在格式化时执行 `static_cast<int64_t>(var)`；上界分支也会先执行 `static_cast<uint64_t>(__var)` 再比较。对 NaN 或超出目标整数表示范围的浮点值，这个转换本身就可能触发未定义行为；对 `1.5` 这类值还会先截断后通过检查，和 `fit in uint32_t` 的语义不一致。\n\n建议把宏限定为整数输入，例如对 ... | true | ||
PaddlePaddle/Paddle | 79,305 | comment_to_fix | [Operator Mechanism] Add uint32 max enforce helper | **🟡 建议** 这里在 inline helper 函数里直接 `PADDLE_THROW`，负数输入的报错位置会固定到 `paddle/common/enforce.h`，而不是 `PADDLE_ENFORCE_LE_UINT32_MAX` 的实际调用点。
`PADDLE_THROW` 展开时使用的是当前函数定义处的 `__FILE__` / `__LINE__`；我用预处理展开验证，`PADDLE_ENFORCE_LE_UINT32_MAX(x, "x")` 的负数分支会生成 `EnforceNotMet(..., "paddle/common/enforce.h", 359)`，而上界检查分支才会保留调用点。这样后续算子... | 4ce8ac4a365da4acdff66a9004fc9bc1596b739e | 9b122fc85fc4f9131b22536f5156affcef5aa820 | diff --git a/paddle/common/enforce.h b/paddle/common/enforce.h
index 430f9ab575df08..fa23e810733413 100644
--- a/paddle/common/enforce.h
+++ b/paddle/common/enforce.h
@@ -17,6 +17,7 @@
#ifdef __GNUC__
#include <cxxabi.h> // for __cxa_demangle
#endif // __GNUC__
+#include <cinttypes>
#include <excepti... | [
"paddle/common/enforce.h"
] | [
{
"comment": "**🟡 建议** 这里在 inline helper 函数里直接 `PADDLE_THROW`，负数输入的报错位置会固定到 `paddle/common/enforce.h`，而不是 `PADDLE_ENFORCE_LE_UINT32_MAX` 的实际调用点。\n\n`PADDLE_THROW` 展开时使用的是当前函数定义处的 `__FILE__` / `__LINE__`；我用预处理展开验证，`PADDLE_ENFORCE_LE_UINT32_MAX(x, \"x\")` 的负数分支会生成 `EnforceNotMet(..., \"paddle/common/enforce.h\",... | true | ||
PaddlePaddle/Paddle | 79,270 | issue_to_patch | Assert headdim == headdim_v in fa2, fm1, fm3 kernels | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 4ce8ac4a365da4acdff66a9004fc9bc1596b739e | 5a4875511d708f52fea3ad00f283fcd5475801b8 | diff --git a/paddle/phi/kernels/gpu/flash_attn_grad_kernel.cu b/paddle/phi/kernels/gpu/flash_attn_grad_kernel.cu
index 31f56972ff0ba8..206768dfef2f77 100644
--- a/paddle/phi/kernels/gpu/flash_attn_grad_kernel.cu
+++ b/paddle/phi/kernels/gpu/flash_attn_grad_kernel.cu
@@ -243,7 +243,7 @@ void FlashAttnUnpaddedGradBaseKer... | [
"paddle/phi/kernels/gpu/flash_attn_grad_kernel.cu",
"paddle/phi/kernels/gpu/flash_attn_kernel.cu",
"paddle/phi/kernels/gpu/flash_attn_v3_grad_kernel.cu",
"test/test_flashmask_ci/test_flashmask_ci.py"
] | [
{
"comment": "**🔴 Bug** 这里只在 padded `FlashAttnBaseKernel` 中校验了 `q` 和 `v` 的 head dim，`flash_attn_unpadded` 仍会漏过同样的非法输入。\n\n`flash_attn_unpadded` 在同文件的 `FlashAttnUnpaddedBaseKernel` 直接接收独立的 `q/k/v`，并把 `q.dims()[2]` 作为 `params.head_size` 传给 `flash_attn_varlen_fwd`；Python wrapper 也允许用户分别传 `query/key/value`。因此 `v.d... | diff --git a/test/test_flashmask_ci/test_flashmask_ci.py b/test/test_flashmask_ci/test_flashmask_ci.py
index 5704c330837ac1..9b0bc0fb2812df 100644
--- a/test/test_flashmask_ci/test_flashmask_ci.py
+++ b/test/test_flashmask_ci/test_flashmask_ci.py
@@ -261,3 +261,192 @@ def test_flashmask(
assert (v.grad - v_ref.gra... | true | |
PaddlePaddle/Paddle | 79,272 | issue_to_patch | [Allocator] Add generic all block info API | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 83c656a73c4cede65b9107826d4fc3c153e688be | ae76453b0bd2a4b39adefdbf04a17eb24975a9c6 | diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index ef3d512cf882ff..eeea571f8cebc9 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -3787,8 +3787,11 @@ All parameter, weight, gradient are variables in Paddle.
m.def("vmm_free_block_info", [](int device_id)... | [
"paddle/fluid/pybind/pybind.cc",
"paddle/phi/core/memory/allocation/auto_growth_best_fit_allocator.h",
"paddle/phi/core/memory/mem_utils.cc",
"paddle/phi/core/memory/mem_utils.h",
"paddle/phi/core/memory/mem_visitor.cc",
"paddle/phi/core/memory/mem_visitor.h",
"python/paddle/device/cuda/memory_analyzer.... | [
{
"comment": "**🔴 兼容性** 这里直接把已导出的 `AllBlockInfoOfVmmAllocator` 声明替换成新名字，会删除旧 C++ 符号。\n\n`AllBlockInfoOfVmmAllocator` 原来是 `PADDLE_API` 声明，`pybind` 之外的 C++ 调用方或外部工具如果包含 `mem_utils.h`，升级后会编译或链接失败；Python 侧虽然保留了 `vmm_all_block_info` 别名，但 C++ 侧没有同等兼容层。\n\n建议保留旧函数作为 wrapper：在头文件继续声明 `AllBlockInfoOfVmmAllocator`，在实现中让... | diff --git a/test/cpp/fluid/memory/allocator_visitor_test.cc b/test/cpp/fluid/memory/allocator_visitor_test.cc
index 1df632e3b1eac4..91c9dce4f59800 100644
--- a/test/cpp/fluid/memory/allocator_visitor_test.cc
+++ b/test/cpp/fluid/memory/allocator_visitor_test.cc
@@ -12,6 +12,8 @@
// See the License for the specific la... | true | |
PaddlePaddle/Paddle | 79,292 | issue_to_patch | align kaiming uniform and gaussian with torch | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 9d69b2f2501224aff7c69fb7204cd4e8281ec740 | 61f4381eace4b5c22b14ce25515b301c67787947 | diff --git a/paddle/fluid/ir_adaptor/translator/op_translator.cc b/paddle/fluid/ir_adaptor/translator/op_translator.cc
index 64faa56652095..1d8d519717fda 100644
--- a/paddle/fluid/ir_adaptor/translator/op_translator.cc
+++ b/paddle/fluid/ir_adaptor/translator/op_translator.cc
@@ -4196,6 +4196,43 @@ struct Pad3dOpTransc... | [
"paddle/fluid/ir_adaptor/translator/op_translator.cc",
"paddle/fluid/pir/serialize_deserialize/patch/4.yaml",
"paddle/phi/infermeta/nullary.cc",
"paddle/phi/infermeta/nullary.h",
"paddle/phi/kernels/cpu/gaussian_kernel.cc",
"paddle/phi/kernels/gaussian_kernel.h",
"paddle/phi/kernels/gpu/gaussian_kernel.... | [
{
"comment": "**🔴 Bug** CPU `gaussian` 仍会把 double 参数截断成 float。\n\n本 PR 将 `GaussianKernel` 签名改成 `double mean/std`，但这里马上 `static_cast<float>` 传给 `NormalDistribution`。该 helper 的通用版本和 `complex128` 特化参数也都是 `const float&`，所以 `dtype=float64`/`complex128` 在 CPU 上仍按 float32 的 `mean/std` 构造分布，PR 目标里的 double 精度不会生效。\n\n建... | true | ||
PaddlePaddle/Paddle | 79,301 | issue_to_patch | [CodeStyle] Clean up Python 3.9 SOT C++ compatibility code | ### PR Category
Execute Infrastructure
### PR Types
Improvements
### Description
Paddle now requires Python >= 3.10, so the SOT C++ pybind code no longer needs Python 3.9 compatibility helpers.
This PR cleans up `paddle/fluid/pybind/sot` by:
- removing the unused `PY_3_9_0_HEX` and `PY_3_9_PLUS` macros;
- simplifyin... | 38248622c5716a6900aaea7322667bf0ab233f26 | cd48e6a1388b3fce263d7e8c31c1da7d27aec1bb | diff --git a/paddle/fluid/pybind/sot/guards.cc b/paddle/fluid/pybind/sot/guards.cc
index 93fe4ad424d2e7..c7c5167dbfcdb9 100644
--- a/paddle/fluid/pybind/sot/guards.cc
+++ b/paddle/fluid/pybind/sot/guards.cc
@@ -27,10 +27,6 @@ limitations under the License. */
#include <internal/pycore_interpframe.h>
#endif
-#if !PY... | [
"paddle/fluid/pybind/sot/guards.cc",
"paddle/fluid/pybind/sot/macros.h"
] | [
{
"comment": "这是为啥？因为触发了检查？#79300 已经移除了这个检查，恢复这里的改动，并 merge 最新的 develop",
"path": "ci/coverage_info.sh",
"hunk": "@@ -159,6 +159,7 @@ function gen_full_report() {\n \"${PADDLE_ROOT}/paddle/fluid/inference/*\" \\\n \"${PADDLE_ROOT}/paddle/fluid/memory/*\" \\\n \"${PADDLE_ROOT}/pad... | true | ||
PaddlePaddle/Paddle | 79,285 | issue_to_patch | [Release/3.4][Operator] cherry-pick mm out_dtype dynamic path | ### PR Category
Operator Mechanism
### PR Types
Improvements
### Description
该 PR 基于 release/3.4 手动完整 cherry-pick develop PR #79252 中的 `paddle.mm` dynamic-only `out_dtype` 改动，用于替代自动 cherry-pick PR #79282 中 CI 暴露的问题。
主要内容：
- 为动态图 `paddle.mm` 增加临时最小 `out_dtype` 支持，仅支持 CUDA 2-D BF16 x BF16 -> FP32。
- 保持静态图/P... | 6627d52e102c596b610068b15a8c319a28e99205 | 2f4528c34436aebc65850ef2742fbf9cf048ef4f | diff --git a/paddle/phi/infermeta/binary.cc b/paddle/phi/infermeta/binary.cc
index d06657173568df..9ee8d920eb2d58 100644
--- a/paddle/phi/infermeta/binary.cc
+++ b/paddle/phi/infermeta/binary.cc
@@ -3170,6 +3170,60 @@ void MatmulInferMeta(const MetaTensor& x,
out->share_lod(x);
}
+void MmOutDtypeInferMeta(const M... | [
"paddle/phi/infermeta/binary.cc",
"paddle/phi/infermeta/binary.h",
"paddle/phi/kernels/funcs/blas/blas.h",
"paddle/phi/kernels/funcs/blas/blas_impl.cu.h",
"paddle/phi/kernels/gpu/matmul_kernel.cu",
"paddle/phi/kernels/impl/matmul_kernel_impl.h",
"paddle/phi/kernels/matmul_kernel.h",
"paddle/phi/ops/ya... | [] | diff --git a/test/legacy_test/test_mm_out.py b/test/legacy_test/test_mm_out.py
index 3e9f2cceff5e7b..6677f04eb65b4d 100644
--- a/test/legacy_test/test_mm_out.py
+++ b/test/legacy_test/test_mm_out.py
@@ -174,5 +174,92 @@ def test_2d_1d(self):
)
+class TestMmOutDtypeDynamicOnly(unittest.TestCase):
+ def s... | true | |
PaddlePaddle/Paddle | 79,283 | issue_to_patch | [release/3.4] fix add_n 0-size bug | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 6627d52e102c596b610068b15a8c319a28e99205 | 4ed3607a07f013e0122bedde3482aeaf54104438 | diff --git a/paddle/phi/infermeta/multiary.cc b/paddle/phi/infermeta/multiary.cc
index ac98369145405..a2d8336ef397c 100644
--- a/paddle/phi/infermeta/multiary.cc
+++ b/paddle/phi/infermeta/multiary.cc
@@ -406,6 +406,7 @@ void AddNInferMeta(const std::vector<const MetaTensor*>& x,
}
bool is_all_0d_tensor = true;
... | [
"paddle/phi/infermeta/multiary.cc",
"test/legacy_test/test_add_n_op.py"
] | [] | diff --git a/test/legacy_test/test_add_n_op.py b/test/legacy_test/test_add_n_op.py
index ceb89af53dc55..74cbcf5bad07a 100644
--- a/test/legacy_test/test_add_n_op.py
+++ b/test/legacy_test/test_add_n_op.py
@@ -119,5 +119,17 @@ def test_add_n_zerosize(self):
np.testing.assert_allclose(x_g_np_32[i].shape, [0,... | true | |
PaddlePaddle/Paddle | 79,300 | issue_to_patch | [CI] Keep gcov tool selection only | ### PR Category
Execute Infrastructure
### PR Types
Bug fixes
### Description
本 PR 调整 Coverage 脚本，保留 #79263 中用于解决 GCC/gcov 主版本不一致的 `GCOV_TOOL` 搜索和 `lcov --gcov-tool` 显式选择逻辑，但撤回基于 `.c/.cc/.cpp/.cxx/.cu` 文件变更强制要求 C++ `coverage-diff.info`/`DA:` 的检查，以及 #79274 为该检查补充的 diff filtering wrapper。
现阶段不适合继续加这个校验：Coverage CI 跑的是... | 6e6b2ee3ccb4ca96a5b1802cdd7eb2c30d47e8ab | c2109500acaffbb9e863d2e7a8bdce7fcf5968b6 | diff --git a/ci/coverage_info.sh b/ci/coverage_info.sh
index 8c80c63557398e..19ab52f9579088 100644
--- a/ci/coverage_info.sh
+++ b/ci/coverage_info.sh
@@ -90,54 +90,6 @@ function init_gcov_tool() {
fi
}
-function pr_has_cpp_source_changes() {
- local diff_file="$1"
-
- awk '
- index($0, "+++ ") == ... | [
"ci/coverage_info.sh"
] | [
{
"comment": "`coverage_info.sh` runs with `set +e`, so failures from `coverage_pull_request.py` here will be silently ignored and the script may continue with an empty `COVERAGE_DIFF_PATTERN` / missing `git-diff.out`, producing incorrect (often empty) diff coverage outputs. Please restore explicit error handli... | true | ||
PaddlePaddle/Paddle | 79,291 | issue_to_patch | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | ### PR Category
Environment Adaptation
### PR Types
Deprecations
### Description
Related to #79289 and #79290.
This PR cleans Python 3.10 type-hint modernization diagnostics and enables Ruff's Python 3.10 target:
- merge latest `develop` after #79289/#79290 and resolve the resulting conflict without force-push
- ... | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/pyproject.toml b/pyproject.toml
index 29ce1c5ca4b28c..7137ed71bc29a4 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -5,7 +5,7 @@ exclude = [
"./python/paddle/utils/gast/**",
]
line-length = 80
-target-version = "py39"
+target-version = "py310"
[tool.ruff.format]
# Prevent change to double q... | [
"pyproject.toml",
"python/paddle/_ops.py",
"python/paddle/_typing/basic.py",
"python/paddle/_typing/device_like.py",
"python/paddle/_typing/dtype_like.py",
"python/paddle/_typing/layout.py",
"python/paddle/_typing/shape.py",
"python/paddle/amp/auto_cast.py",
"python/paddle/amp/debugging.py",
"pyth... | [
{
"comment": "这里还得测 `Optional[torch.device]`，至少在我们支持 Python 3.10 的时候还需要支持这种 case，可以记一个 TODO，将来不支持 3.10 的时候可以清理掉这块代码",
"path": "test/compat/test_torch_proxy.py",
"hunk": "@@ -185,15 +185,13 @@ def test_fake_interface(self):\n class TestDeviceAsTypeHints(unittest.TestCase):\n @paddle.use_compat_guard(... | diff --git a/test/compat/test_torch_proxy.py b/test/compat/test_torch_proxy.py
index 82f8e6fb155035..57e607c945db79 100644
--- a/test/compat/test_torch_proxy.py
+++ b/test/compat/test_torch_proxy.py
@@ -15,6 +15,7 @@
import pathlib
import sys
import unittest
+from typing import Optional
from unittest.mock import Ma... | true | |
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 这里还得测 `Optional[torch.device]`，至少在我们支持 Python 3.10 的时候还需要支持这种 case，可以记一个 TODO，将来不支持 3.10 的时候可以清理掉这块代码 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/test/compat/test_torch_proxy.py b/test/compat/test_torch_proxy.py
index 82f8e6fb155035..57e607c945db79 100644
--- a/test/compat/test_torch_proxy.py
+++ b/test/compat/test_torch_proxy.py
@@ -15,6 +15,7 @@
import pathlib
import sys
import unittest
+from typing import Optional
from unittest.mock import Ma... | [
"test/compat/test_torch_proxy.py"
] | [
{
"comment": "这里还得测 `Optional[torch.device]`，至少在我们支持 Python 3.10 的时候还需要支持这种 case，可以记一个 TODO，将来不支持 3.10 的时候可以清理掉这块代码",
"path": "test/compat/test_torch_proxy.py",
"hunk": "@@ -185,15 +185,13 @@ def test_fake_interface(self):\n class TestDeviceAsTypeHints(unittest.TestCase):\n @paddle.use_compat_guard(... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | ```suggestion
[Any, FunctionGraph, Tracker], "VariableBase" | None
```
只是一个疑问，是否可以 / 应该改成这样？ | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/jit/sot/opcode_translator/executor/variables/base.py b/python/paddle/jit/sot/opcode_translator/executor/variables/base.py
index f712bada5e88b4..9ffb81037f0867 100644
--- a/python/paddle/jit/sot/opcode_translator/executor/variables/base.py
+++ b/python/paddle/jit/sot/opcode_translator/executor... | [
"python/paddle/jit/sot/opcode_translator/executor/variables/base.py"
] | [
{
"comment": "```suggestion\n [Any, FunctionGraph, Tracker], \"VariableBase\" | None\n```\n\n只是一个疑问，是否可以 / 应该改成这样？",
"path": "python/paddle/jit/sot/opcode_translator/executor/variables/base.py",
"hunk": "@@ -64,7 +64,7 @@\n # Each variable object should implement a method called `from_value`,... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 原来的注释可以看看怎么补充一下 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/device/cpu.py b/python/paddle/device/cpu.py
index 955dbf2bd8f7f6..d47f96aa6850aa 100644
--- a/python/paddle/device/cpu.py
+++ b/python/paddle/device/cpu.py
@@ -14,9 +14,7 @@
from __future__ import annotations
-from typing import TYPE_CHECKING, Union
-
-from typing_extensions import TypeA... | [
"python/paddle/device/cpu.py"
] | [
{
"comment": "原来的注释可以看看怎么补充一下",
"path": "python/paddle/device/cpu.py",
"hunk": "@@ -30,11 +28,7 @@\n if TYPE_CHECKING:\n from paddle import CPUPlace\n \n- _CPUPlaceLike: TypeAlias = Union[\n- CPUPlace,\n- str, # some string like \"iluvatar_gpu\" \"metax_gpu:0\", etc.\n- int,... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 现在 RUF100 应该可以不用了吧 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/test/compat/test_torch_proxy.py b/test/compat/test_torch_proxy.py
index 82f8e6fb155035..57e607c945db79 100644
--- a/test/compat/test_torch_proxy.py
+++ b/test/compat/test_torch_proxy.py
@@ -15,6 +15,7 @@
import pathlib
import sys
import unittest
+from typing import Optional
from unittest.mock import Ma... | [
"test/compat/test_torch_proxy.py"
] | [
{
"comment": "现在 RUF100 应该可以不用了吧",
"path": "test/compat/test_torch_proxy.py",
"hunk": "@@ -185,15 +186,17 @@ def test_fake_interface(self):\n class TestDeviceAsTypeHints(unittest.TestCase):\n @paddle.use_compat_guard()\n def test_device_as_type_hints(self):\n- from typing import Optional\... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 和上面的一样，你保留下注释吧 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/device/cuda/__init__.py b/python/paddle/device/cuda/__init__.py
index e61d537a71daff..4abda00c0dd28e 100644
--- a/python/paddle/device/cuda/__init__.py
+++ b/python/paddle/device/cuda/__init__.py
@@ -14,9 +14,7 @@
from __future__ import annotations
import os
-from typing import TYPE_CHECK... | [
"python/paddle/device/cuda/__init__.py"
] | [
{
"comment": "和上面的一样，你保留下注释吧",
"path": "python/paddle/device/cuda/__init__.py",
"hunk": "@@ -29,12 +27,7 @@\n from paddle import CUDAPlace, CustomPlace\n from paddle.base.libpaddle import _gpuDeviceProperties\n \n- _CudaPlaceLike: TypeAlias = Union[\n- CUDAPlace,\n- CustomPlace,... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 这里也保留下注释 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/device/xpu/__init__.py b/python/paddle/device/xpu/__init__.py
index 9a47dcb07cd8b5..6c915473bcee54 100644
--- a/python/paddle/device/xpu/__init__.py
+++ b/python/paddle/device/xpu/__init__.py
@@ -13,9 +13,7 @@
# limitations under the License.
from __future__ import annotations
-from typin... | [
"python/paddle/device/xpu/__init__.py"
] | [
{
"comment": "这里也保留下注释",
"path": "python/paddle/device/xpu/__init__.py",
"hunk": "@@ -27,11 +25,7 @@\n from paddle import XPUPlace\n from paddle.base.libpaddle import _gpuDeviceProperties\n \n- _XPUPlaceLike: TypeAlias = Union[\n- XPUPlace,\n- str, # some str like 'xpu:0', 'xpu... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 这里也 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/device/__init__.py b/python/paddle/device/__init__.py
index 1a845df1fdb1b7..067e4fa030f44a 100644
--- a/python/paddle/device/__init__.py
+++ b/python/paddle/device/__init__.py
@@ -22,9 +22,7 @@
import sys
import types
from contextlib import contextmanager
-from typing import TYPE_CHECKING,... | [
"python/paddle/device/__init__.py"
] | [
{
"comment": "这里也",
"path": "python/paddle/device/__init__.py",
"hunk": "@@ -63,22 +61,13 @@\n from paddle._typing.device_like import PlaceLike\n from paddle.base.core import Place\n \n- _InitStreamBase = Union[\n- core.CUDAStream, core.CustomDeviceStream, core.XPUStream\n- ]\n- ... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 对对对，前面的，就像这样保留下 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/device/cpu.py b/python/paddle/device/cpu.py
index 955dbf2bd8f7f6..d47f96aa6850aa 100644
--- a/python/paddle/device/cpu.py
+++ b/python/paddle/device/cpu.py
@@ -14,9 +14,7 @@
from __future__ import annotations
-from typing import TYPE_CHECKING, Union
-
-from typing_extensions import TypeA... | [
"python/paddle/device/cpu.py"
] | [
{
"comment": "对对对，前面的，就像这样保留下",
"path": "python/paddle/device/cpu.py",
"hunk": "@@ -30,11 +28,11 @@\n if TYPE_CHECKING:\n from paddle import CPUPlace\n \n- _CPUPlaceLike: TypeAlias = Union[\n- CPUPlace,\n- str, # some string like \"iluvatar_gpu\" \"metax_gpu:0\", etc.\n- int... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 这里也 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/device/custom_device.py b/python/paddle/device/custom_device.py
index 161c6c2646138f..3b9cb73a22cfc8 100644
--- a/python/paddle/device/custom_device.py
+++ b/python/paddle/device/custom_device.py
@@ -13,9 +13,7 @@
# limitations under the License.
from __future__ import annotations
-from t... | [
"python/paddle/device/custom_device.py"
] | [
{
"comment": "这里也",
"path": "python/paddle/device/custom_device.py",
"hunk": "@@ -30,11 +28,7 @@\n if TYPE_CHECKING:\n from paddle import CustomPlace\n \n- _CustomPlaceLike: TypeAlias = Union[\n- CustomPlace,\n- str, # some string like \"iluvatar_gpu\" \"metax_gpu:0\", etc.\n- ... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 这里也别改人家代码语义啊，这不合理啊，如果是我们类型上哪里有问题，就补全，如果实在不行，就 ignore | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/distribution/uniform.py b/python/paddle/distribution/uniform.py
index 474ab7b0521f40..7ab12661b75bcb 100644
--- a/python/paddle/distribution/uniform.py
+++ b/python/paddle/distribution/uniform.py
@@ -28,20 +28,18 @@
if TYPE_CHECKING:
from collections.abc import Sequence
- from typi... | [
"python/paddle/distribution/uniform.py"
] | [
{
"comment": "这里也别改人家代码语义啊，这不合理啊，如果是我们类型上哪里有问题，就补全，如果实在不行，就 ignore",
"path": "python/paddle/distribution/uniform.py",
"hunk": "@@ -91,8 +91,8 @@ class Uniform(distribution.Distribution):\n >>> u2 = Uniform(low=[1.0, 2.0], high=[3.0, 4.0])\n >>> # 4 distributions\n >>>... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 现在是不会认为 `dtype` 是 literal 了吗？不建议改用法哈 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/amp/debugging.py b/python/paddle/amp/debugging.py
index 5c158e1906dddb..a24cb6a9b822d0 100644
--- a/python/paddle/amp/debugging.py
+++ b/python/paddle/amp/debugging.py
@@ -19,7 +19,6 @@
from typing import (
TYPE_CHECKING,
Any,
- Callable,
TypeVar,
)
@@ -33,7 +32,7 @@
fr... | [
"python/paddle/amp/debugging.py"
] | [
{
"comment": "现在是不会认为 `dtype` 是 literal 了吗？不建议改用法哈",
"path": "python/paddle/amp/debugging.py",
"hunk": "@@ -103,7 +103,7 @@ def check_layer_numerics(\n ... def forward(self, x):\n ... # return 1/x * self._w + self._b open it you will see the error log\n ... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 这里可以改 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/vision/datasets/mnist.py b/python/paddle/vision/datasets/mnist.py
index a8fba2c7b0fe7f..4c672ad0c42106 100644
--- a/python/paddle/vision/datasets/mnist.py
+++ b/python/paddle/vision/datasets/mnist.py
@@ -64,6 +64,7 @@ class MNIST(Dataset[tuple["_ImageDataType", "npt.NDArray[np.int64]"]]):
... | [
"python/paddle/vision/datasets/mnist.py"
] | [
{
"comment": "这里可以改",
"path": "python/paddle/vision/datasets/mnist.py",
"hunk": "",
"resolving_sha": "72e139da04d615705fb44b93d329af59b1ffaed8",
"resolving_diff": "diff --git a/python/paddle/vision/datasets/mnist.py b/python/paddle/vision/datasets/mnist.py\nindex a8fba2c7b0fe7f..4c672ad0c42106 1... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 这个也可以改 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/vision/datasets/voc2012.py b/python/paddle/vision/datasets/voc2012.py
index be7b8636e46080..3f84c8f3741f01 100644
--- a/python/paddle/vision/datasets/voc2012.py
+++ b/python/paddle/vision/datasets/voc2012.py
@@ -75,6 +75,7 @@ class VOC2012(Dataset[tuple["_ImageDataType", "npt.NDArray[Any]"]])... | [
"python/paddle/vision/datasets/voc2012.py"
] | [
{
"comment": "这个也可以改",
"path": "python/paddle/vision/datasets/voc2012.py",
"hunk": "",
"resolving_sha": "72e139da04d615705fb44b93d329af59b1ffaed8",
"resolving_diff": "diff --git a/python/paddle/vision/datasets/voc2012.py b/python/paddle/vision/datasets/voc2012.py\nindex be7b8636e46080..3f84c8f37... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 这个可以改 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/vision/datasets/mnist.py b/python/paddle/vision/datasets/mnist.py
index a8fba2c7b0fe7f..4c672ad0c42106 100644
--- a/python/paddle/vision/datasets/mnist.py
+++ b/python/paddle/vision/datasets/mnist.py
@@ -64,6 +64,7 @@ class MNIST(Dataset[tuple["_ImageDataType", "npt.NDArray[np.int64]"]]):
... | [
"python/paddle/vision/datasets/mnist.py"
] | [
{
"comment": "这个可以改",
"path": "python/paddle/vision/datasets/mnist.py",
"hunk": "",
"resolving_sha": "72e139da04d615705fb44b93d329af59b1ffaed8",
"resolving_diff": "diff --git a/python/paddle/vision/datasets/mnist.py b/python/paddle/vision/datasets/mnist.py\nindex a8fba2c7b0fe7f..4c672ad0c42106 1... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | datasets 的改动我觉得没问题 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/vision/datasets/cifar.py b/python/paddle/vision/datasets/cifar.py
index 700efdfd3f6851..a9513c3901ab73 100644
--- a/python/paddle/vision/datasets/cifar.py
+++ b/python/paddle/vision/datasets/cifar.py
@@ -94,6 +94,7 @@ class Cifar10(Dataset[tuple["_ImageDataType", "npt.NDArray[Any]"]]):
... | [
"python/paddle/vision/datasets/cifar.py"
] | [
{
"comment": "datasets 的改动我觉得没问题",
"path": "python/paddle/vision/datasets/cifar.py",
"hunk": "",
"resolving_sha": "72e139da04d615705fb44b93d329af59b1ffaed8",
"resolving_diff": "diff --git a/python/paddle/vision/datasets/cifar.py b/python/paddle/vision/datasets/cifar.py\nindex 700efdfd3f6851..a95... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | 可以考虑 ignore，这种写法对用户来说并不友好 | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/amp/debugging.py b/python/paddle/amp/debugging.py
index 5c158e1906dddb..a24cb6a9b822d0 100644
--- a/python/paddle/amp/debugging.py
+++ b/python/paddle/amp/debugging.py
@@ -19,7 +19,6 @@
from typing import (
TYPE_CHECKING,
Any,
- Callable,
TypeVar,
)
@@ -33,7 +32,7 @@
fr... | [
"python/paddle/amp/debugging.py"
] | [
{
"comment": "可以考虑 ignore，这种写法对用户来说并不友好",
"path": "python/paddle/amp/debugging.py",
"hunk": "@@ -103,7 +104,7 @@ def check_layer_numerics(\n ... def forward(self, x):\n ... # return 1/x * self._w + self._b open it you will see the error log\n ... ... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | ignore | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/incubate/jit/inference_decorator.py b/python/paddle/incubate/jit/inference_decorator.py
index de4cb2833c3f43..ca6803f9b111bd 100644
--- a/python/paddle/incubate/jit/inference_decorator.py
+++ b/python/paddle/incubate/jit/inference_decorator.py
@@ -20,7 +20,7 @@
import textwrap
import warnin... | [
"python/paddle/incubate/jit/inference_decorator.py"
] | [
{
"comment": "ignore",
"path": "python/paddle/incubate/jit/inference_decorator.py",
"hunk": "@@ -610,7 +611,7 @@ def inference(\n \n >>> batch = 4096\n >>> hidd = 1024\n- >>> dtype = paddle.bfloat16\n+ >>> dtype: Literal[\"bfloat16\"] = \"bfloat16\"",
"r... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | ```suggestion
>>> label = paddle.randint(low=0, high=int(num_classes.item()), size=[batch_size], dtype='int64') # type: ignore[call-overload, arg-type]
``` | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/nn/functional/common.py b/python/paddle/nn/functional/common.py
index 2b2e4d31c37ea0..3a8ad6c07caa18 100644
--- a/python/paddle/nn/functional/common.py
+++ b/python/paddle/nn/functional/common.py
@@ -50,8 +50,7 @@
if TYPE_CHECKING:
from collections.abc import Sequence
-
- from typi... | [
"python/paddle/nn/functional/common.py"
] | [
{
"comment": "```suggestion\n >>> label = paddle.randint(low=0, high=int(num_classes.item()), size=[batch_size], dtype='int64') # type: ignore[call-overload, arg-type]\n```",
"path": "python/paddle/nn/functional/common.py",
"hunk": "@@ -2767,7 +2766,7 @@ class centers and the shape of sampled_cl... | true | ||
PaddlePaddle/Paddle | 79,291 | comment_to_fix | [CodeStyle] Clean up Python 3.10 type-hint diagnostics | ```suggestion
>>> label = paddle.randint(low=0, high=int(num_classes.item()), size=[batch_size], dtype='int64') # type: ignore[call-overload, arg-type]
``` | 83c656a73c4cede65b9107826d4fc3c153e688be | 72e139da04d615705fb44b93d329af59b1ffaed8 | diff --git a/python/paddle/nn/functional/loss.py b/python/paddle/nn/functional/loss.py
index f7de9a16718576..2de571f8587169 100644
--- a/python/paddle/nn/functional/loss.py
+++ b/python/paddle/nn/functional/loss.py
@@ -35,8 +35,8 @@
from ...tensor.manipulation import reshape
if TYPE_CHECKING:
- from collections.... | [
"python/paddle/nn/functional/loss.py"
] | [
{
"comment": "```suggestion\n >>> label = paddle.randint(low=0, high=int(num_classes.item()), size=[batch_size], dtype='int64') # type: ignore[call-overload, arg-type]\n```",
"path": "python/paddle/nn/functional/loss.py",
"hunk": "@@ -2530,7 +2530,7 @@ def margin_cross_entropy(\n ... | true | ||
PaddlePaddle/Paddle | 79,289 | issue_to_patch | [Typing] Upgrade mypy to 2.1.0 and target Python 3.10 | ### PR Category
User Experience
### PR Types
Devs
### Description
本 PR 升级 Paddle 的 mypy 工具链配置，并修复 mypy 2.1.0 / Python 3.10 下暴露出的示例类型检查问题。
#### 1. mypy 版本与配置升级
原因：mypy 2.0 起不再支持 `--python-version 3.9`，因此升级 mypy 时需要同步提升类型检查目标版本。
- `python/unittest_py/requirements.txt`: `mypy==1.19.1` -> `mypy==2.1.0`
-... | 9d69b2f2501224aff7c69fb7204cd4e8281ec740 | d690a2dfff35716e768e925a763e5898db9ae735 | diff --git a/pyproject.toml b/pyproject.toml
index f9abad63991ae..29ce1c5ca4b28 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -172,8 +172,9 @@ known-first-party = ["paddle"]
"test/dygraph_to_static/seq2seq_dygraph_model.py" = ["RUF005"]
[tool.mypy]
-python_version = "3.9"
+python_version = "3.10"
cache_dir =... | [
"pyproject.toml",
"python/paddle/base/framework.py",
"python/paddle/distributed/communication/recv.py",
"python/paddle/distributed/communication/send.py",
"python/paddle/distributed/fleet/base/topology.py",
"python/paddle/nn/functional/loss.py",
"python/paddle/nn/initializer/assign.py",
"python/paddle... | [
{
"comment": "这注释没啥用啊，清理掉吧，另外你看看有没有啥选项自动匹配核心数的，如果没有就用 8 吧\n\n以及我们的 CI type_checking 脚本里加一下 worker 数自动匹配核心数的逻辑吧",
"path": "pyproject.toml",
"hunk": "@@ -172,8 +172,10 @@ known-first-party = [\"paddle\"]\n \"test/dygraph_to_static/seq2seq_dygraph_model.py\" = [\"RUF005\"]\n \n [tool.mypy]\n-python_version... | diff --git a/python/unittest_py/requirements.txt b/python/unittest_py/requirements.txt
index 2b8b2010c65a2..051ecd93892a5 100644
--- a/python/unittest_py/requirements.txt
+++ b/python/unittest_py/requirements.txt
@@ -17,7 +17,7 @@ wandb>=0.23.1
xlsxwriter==3.0.9
xdoctest==1.3.0
ubelt==1.4.0 # just for xdoctest
-mypy... | true | |
PaddlePaddle/Paddle | 79,274 | issue_to_patch | [CI] Check C++ coverage before diff filtering | ### PR Category
Execute Infrastructure
### PR Types
Bug fixes
### Description
修复 Coverage C++ 空数据 guard 在 diff 行过滤后误伤 PR 的问题，并补上 diff filtering 失败时的临时文件保护。
背景：#79263 合入后，`ci/coverage_info.sh` 会在 `coverage_diff.py` 过滤 PR diff 行之后检查 `coverage-diff.info` 是否仍有 `DA:` 行。这个检查可以发现 gcov/lcov 采集失败导致的空报告，但也会误伤一种正常场景：C++ covera... | 002722386dbe898fb76b3d77f9cd1f20c175180b | e76b8bf3297bbfc9c2a8e54effe64e1dd867f932 | diff --git a/ci/coverage_info.sh b/ci/coverage_info.sh
index d785719f999872..8c80c63557398e 100644
--- a/ci/coverage_info.sh
+++ b/ci/coverage_info.sh
@@ -107,17 +107,37 @@ function pr_has_cpp_source_changes() {
' "${diff_file}"
}
-function ensure_cpp_diff_coverage_data() {
+function ensure_cpp_extracted_covera... | [
"ci/coverage_info.sh"
] | [
{
"comment": "After moving the C++-coverage DA guard before diff filtering, a failure in `ci/coverage_diff.py` (non-zero exit) can silently produce an empty `coverage-diff.tmp` (due to shell redirection), then `mv` overwrites `coverage-diff.info`. Since `tools/coverage/coverage_lines.py` exits 0 on \"no data fo... | true | ||
PaddlePaddle/Paddle | 79,266 | issue_to_patch | 【Hackathon 10th Spring No.52】[ThirdParty][Crypto] Upgrade Crypto++ to 8.6.0 - part 5 | ### PR Category
Environment Adaptation
### PR Types
Improvements
### Description
本 PR 将 Paddle 使用的 Crypto++ 从 `CRYPTOPP_8_2_0` 升级到 `CRYPTOPP_8_6_0`，并同步更新 `cryptopp-cmake` 到对应版本。
升级原因：Crypto++ 8.6.0 已包含上游 PR [weidai11/cryptopp#970](https://github.com/weidai11/cryptopp/pull/970)，该修复解决了 GCC 11 在 C++20 模式下编译 `SourceTemp... | 4d12baa6a2c7ea4dbb0e799e6475bc1d0a9e6794 | b76a1baa7dc4788e4e02f26d367324ff69f9dc91 | diff --git a/cmake/external/cryptopp.cmake b/cmake/external/cryptopp.cmake
index 84112fe6b7228a..2bb6f450140be8 100644
--- a/cmake/external/cryptopp.cmake
+++ b/cmake/external/cryptopp.cmake
@@ -21,21 +21,20 @@ set(CRYPTOPP_INSTALL_DIR ${THIRD_PARTY_PATH}/install/cryptopp)
set(CRYPTOPP_INCLUDE_DIR
"${CRYPTOPP_INS... | [
"cmake/external/cryptopp.cmake",
"patches/cryptopp/CMakeLists.txt",
"third_party/cryptopp",
"third_party/cryptopp-cmake"
] | [
{
"comment": "The Aarch64 option block defines `DISABLE_ARM_AES` twice (for both ASIMD and AES), which means the ASIMD toggle is not actually exposed as an option and can also confuse later checks that reference `DISABLE_ARM_ASIMD`. Rename the first option to `DISABLE_ARM_ASIMD`.",
"path": "patches/cryptopp... | true | ||
PaddlePaddle/Paddle | 79,290 | issue_to_patch | [CodeStyle] Clean up Python 3.9 compatibility code | ### PR Category
Environment Adaptation
### PR Types
Improvements
### Description
Related to #59685.
Python 3.9 has retired, so this PR removes focused Python <3.10 compatibility code:
- delete the private `_typing.backport` shim and use `types.EllipsisType` directly
- remove SOT `co_lnotab` generation and Python ... | 9d69b2f2501224aff7c69fb7204cd4e8281ec740 | 01a1740aaed3f9a1c8730e98e82ea00741ec5dd8 | diff --git a/.github/workflows/H-Coverage.yml b/.github/workflows/H-Coverage.yml
index bf5c34b2393cb..2f8e317c26656 100644
--- a/.github/workflows/H-Coverage.yml
+++ b/.github/workflows/H-Coverage.yml
@@ -505,7 +505,7 @@ jobs:
if: needs.build.outputs.can-skip != 'true'
runs-on:
group: Fleet-H-single-ca... | [
".github/workflows/H-Coverage.yml",
"python/paddle/_typing/__init__.py",
"python/paddle/_typing/backport.py",
"python/paddle/_typing/basic.py",
"python/paddle/jit/dy2static/convert_call_func.py",
"python/paddle/jit/sot/opcode_translator/executor/pycode_generator.py",
"python/paddle/jit/sot/opcode_transl... | [
{
"comment": "backport 文件可以直接清理了，相关地方直接用对应类型就好了",
"path": "python/paddle/_typing/backport.py",
"hunk": "",
"resolving_sha": "01a1740aaed3f9a1c8730e98e82ea00741ec5dd8",
"resolving_diff": "diff --git a/python/paddle/_typing/backport.py b/python/paddle/_typing/backport.py\ndeleted file mode 100644\... | diff --git a/test/sot/test_dataclass.py b/test/sot/test_dataclass.py
index ea4a78c8e2457..5120ace1c42fc 100644
--- a/test/sot/test_dataclass.py
+++ b/test/sot/test_dataclass.py
@@ -14,7 +14,6 @@
from __future__ import annotations
-import sys
import unittest
from dataclasses import dataclass, field
@@ -161,11 +... | true | |
PaddlePaddle/Paddle | 79,276 | issue_to_patch | fix add_n 0-size bug | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 8cacdfd15bc89296682c784df5b1685a7ca6e408 | fa6bfed3dde252b97c9db8e32ce4d8bdd813b8a4 | diff --git a/paddle/phi/infermeta/multiary.cc b/paddle/phi/infermeta/multiary.cc
index 426a04daf3ed43..76b58fda898562 100644
--- a/paddle/phi/infermeta/multiary.cc
+++ b/paddle/phi/infermeta/multiary.cc
@@ -406,6 +406,7 @@ void AddNInferMeta(const std::vector<const MetaTensor*>& x,
}
bool is_all_0d_tensor = true;... | [
"paddle/phi/infermeta/multiary.cc",
"test/legacy_test/test_add_n_op.py"
] | [
{
"comment": "**🔴 Bug** 这里不能直接把 `in_dim` 初始化为 `x[0]->dims()`。\n\n下面的循环仍会跳过 `x[i]->is_selected_rows() && x_dim.size() == 1` 的输入，但 `in_dim` 已经固定成这个被跳过的 shape。动态图 `add_n_impl` 只要输入里有 DenseTensor 就会走 dense `add_n` kernel，并在 kernel 前用 `AddNInferMeta` 设置 Dense 输出；`sum` 也有 `add_n_sr` / `add_n` 的 SelectedRows 路径。因此当首个... | diff --git a/test/legacy_test/test_add_n_op.py b/test/legacy_test/test_add_n_op.py
index ceb89af53dc55c..74cbcf5bad07ad 100644
--- a/test/legacy_test/test_add_n_op.py
+++ b/test/legacy_test/test_add_n_op.py
@@ -119,5 +119,17 @@ def test_add_n_zerosize(self):
np.testing.assert_allclose(x_g_np_32[i].shape, [... | true | |
PaddlePaddle/Paddle | 77,430 | issue_to_patch | fixing 77340 | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
Inference
### PR Types
Bug fixes
### Description
Fixing #77340
### 是否引起精度变化
否 | c8061a30ad66b0a1e3c47f10fb685bc230758c61 | 745e94e4d8744e6a9579f1d256757525cb78ca40 | diff --git a/paddle/fluid/framework/new_executor/instruction/onednn/onednn_instruction.cc b/paddle/fluid/framework/new_executor/instruction/onednn/onednn_instruction.cc
index 175efaab654803..d5685e9986b60c 100644
--- a/paddle/fluid/framework/new_executor/instruction/onednn/onednn_instruction.cc
+++ b/paddle/fluid/frame... | [
"paddle/fluid/framework/new_executor/instruction/onednn/onednn_instruction.cc"
] | [] | true | ||
PaddlePaddle/Paddle | 79,252 | issue_to_patch | [Operator Mechanism] Support mm out_dtype for BF16 CUDA |
<!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization... | d9242d558766b5d24fa3231a798864c37cdd5cda | 11579f81f3f9f8b5483c7e4d7294bc8f4aa39b38 | diff --git a/paddle/phi/infermeta/binary.cc b/paddle/phi/infermeta/binary.cc
index a6be6c5cb89aeb..367230ee6769b2 100644
--- a/paddle/phi/infermeta/binary.cc
+++ b/paddle/phi/infermeta/binary.cc
@@ -3171,6 +3171,60 @@ void MatmulInferMeta(const MetaTensor& x,
out->share_lod(x);
}
+void MmOutDtypeInferMeta(const M... | [
"paddle/phi/infermeta/binary.cc",
"paddle/phi/infermeta/binary.h",
"paddle/phi/kernels/funcs/blas/blas.h",
"paddle/phi/kernels/funcs/blas/blas_impl.cu.h",
"paddle/phi/kernels/gpu/matmul_kernel.cu",
"paddle/phi/kernels/impl/matmul_kernel_impl.h",
"paddle/phi/kernels/matmul_kernel.h",
"paddle/phi/ops/ya... | [
{
"comment": "**🔴 Bug** 反向 kernel 递归调用自身，在非 CUDA 后端（XPU/OneDNN）若 bf16 out_grad 以 fp32 形式传入会触发此分支，而 XPU 上没有 Cast<float->bf16> kernel，导致运行时崩溃。\n\n建议改为提前在前向输出阶段就记录是否用了 out_dtype，或在反向 kernel 顶层通过 `PADDLE_ENFORCE` 限制仅 CUDA 后端进入此分支：\n```cpp\n#if defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)\n if constexpr ... | diff --git a/test/legacy_test/test_mm_out.py b/test/legacy_test/test_mm_out.py
index 3e9f2cceff5e7b..6677f04eb65b4d 100644
--- a/test/legacy_test/test_mm_out.py
+++ b/test/legacy_test/test_mm_out.py
@@ -174,5 +174,92 @@ def test_2d_1d(self):
)
+class TestMmOutDtypeDynamicOnly(unittest.TestCase):
+ def s... | true | |
PaddlePaddle/Paddle | 79,252 | comment_to_fix | [Operator Mechanism] Support mm out_dtype for BF16 CUDA | **🟡 建议** `test_out_dtype_rejects_unsupported_cases` 测试开头没有调用 `self._skip_if_no_bf16_cuda()`。在 CPU-only CI 环境中创建 `paddle.randn([3,4], dtype='bfloat16')` 会抛出异常，测试以 error 而非 skip 结束。建议在方法开头添加 CUDA 检查，或使用 float32 tensor 测试异常路径（不依赖 bf16 设备支持）。
| d9242d558766b5d24fa3231a798864c37cdd5cda | 11579f81f3f9f8b5483c7e4d7294bc8f4aa39b38 | diff --git a/test/legacy_test/test_mm_out.py b/test/legacy_test/test_mm_out.py
index 3e9f2cceff5e7b..6677f04eb65b4d 100644
--- a/test/legacy_test/test_mm_out.py
+++ b/test/legacy_test/test_mm_out.py
@@ -174,5 +174,92 @@ def test_2d_1d(self):
)
+class TestMmOutDtypeDynamicOnly(unittest.TestCase):
+ def s... | [
"test/legacy_test/test_mm_out.py"
] | [
{
"comment": "**🟡 建议** `test_out_dtype_rejects_unsupported_cases` 测试开头没有调用 `self._skip_if_no_bf16_cuda()`。在 CPU-only CI 环境中创建 `paddle.randn([3,4], dtype='bfloat16')` 会抛出异常，测试以 error 而非 skip 结束。建议在方法开头添加 CUDA 检查，或使用 float32 tensor 测试异常路径（不依赖 bf16 设备支持）。\n",
"path": "test/legacy_test/test_mm_out.py",
"hu... | true | ||
PaddlePaddle/Paddle | 79,252 | comment_to_fix | [Operator Mechanism] Support mm out_dtype for BF16 CUDA | **🟡 建议** `TestMmOutDtype` 类目前仅在 `setUp` 中调用 `paddle.disable_static()`，所有测试只覆盖动态图路径。
由于本 PR 同时修改了 `dygraph_ops.yaml` / `static_ops.yaml`，并在静态图分支切换了 `matmul_v2` → `matmul` op，建议补充静态图 / PIR 模式测试，例如：
```python
class TestMmOutDtypeStatic(unittest.TestCase):
def test_bf16_to_fp32_static(self):
if not paddle.is_... | d9242d558766b5d24fa3231a798864c37cdd5cda | 11579f81f3f9f8b5483c7e4d7294bc8f4aa39b38 | diff --git a/test/legacy_test/test_mm_out.py b/test/legacy_test/test_mm_out.py
index 3e9f2cceff5e7b..6677f04eb65b4d 100644
--- a/test/legacy_test/test_mm_out.py
+++ b/test/legacy_test/test_mm_out.py
@@ -174,5 +174,92 @@ def test_2d_1d(self):
)
+class TestMmOutDtypeDynamicOnly(unittest.TestCase):
+ def s... | [
"test/legacy_test/test_mm_out.py"
] | [
{
"comment": "**🟡 建议** `TestMmOutDtype` 类目前仅在 `setUp` 中调用 `paddle.disable_static()`，所有测试只覆盖动态图路径。\n\n由于本 PR 同时修改了 `dygraph_ops.yaml` / `static_ops.yaml`，并在静态图分支切换了 `matmul_v2` → `matmul` op，建议补充静态图 / PIR 模式测试，例如：\n```python\nclass TestMmOutDtypeStatic(unittest.TestCase):\n def test_bf16_to_fp32_static(self)... | true | ||
PaddlePaddle/Paddle | 79,252 | comment_to_fix | [Operator Mechanism] Support mm out_dtype for BF16 CUDA | **🟡 建议** `TestMmOutDtype` 所有测试方法只覆盖动态图路径，缺少静态图和 PIR 路径测试。
`mm()` 静态图路径已改为使用 `matmul` op（原为 `matmul_v2`）并新增 `out_dtype` 属性传递。建议至少补充一个静态图测试：
```python
def test_bf16_to_fp32_static(self):
self._skip_if_no_bf16_cuda()
with paddle.static.program_guard(
paddle.static.Program(), paddle.static.Program()
)... | d9242d558766b5d24fa3231a798864c37cdd5cda | 11579f81f3f9f8b5483c7e4d7294bc8f4aa39b38 | diff --git a/test/legacy_test/test_mm_out.py b/test/legacy_test/test_mm_out.py
index 3e9f2cceff5e7b..6677f04eb65b4d 100644
--- a/test/legacy_test/test_mm_out.py
+++ b/test/legacy_test/test_mm_out.py
@@ -174,5 +174,92 @@ def test_2d_1d(self):
)
+class TestMmOutDtypeDynamicOnly(unittest.TestCase):
+ def s... | [
"test/legacy_test/test_mm_out.py"
] | [
{
"comment": "**🟡 建议** `TestMmOutDtype` 所有测试方法只覆盖动态图路径，缺少静态图和 PIR 路径测试。\n\n`mm()` 静态图路径已改为使用 `matmul` op（原为 `matmul_v2`）并新增 `out_dtype` 属性传递。建议至少补充一个静态图测试：\n```python\ndef test_bf16_to_fp32_static(self):\n self._skip_if_no_bf16_cuda()\n with paddle.static.program_guard(\n paddle.static.Program(), ... | true | ||
PaddlePaddle/Paddle | 79,252 | comment_to_fix | [Operator Mechanism] Support mm out_dtype for BF16 CUDA | **🔴 Bug** `mm_out_dtype` 没有配置 `backward`，但 Python 入口已把 `paddle.mm(..., out_dtype=paddle.float32)` 直接路由到这个 op。
现有 `matmul` 在同文件上方配置了 `backward : matmul_grad`；eager codegen 也会在 forward YAML 没有 `backward` 时把该 op 当成 forward-only 并跳过 GradNode 生成。因此用户把 BF16 输入设为 `stop_gradient=False` 后，`out_dtype` 分支不会像普通 `paddle.mm` 一样向 `... | d9242d558766b5d24fa3231a798864c37cdd5cda | 11579f81f3f9f8b5483c7e4d7294bc8f4aa39b38 | diff --git a/paddle/phi/ops/yaml/inconsistent/dygraph_ops.yaml b/paddle/phi/ops/yaml/inconsistent/dygraph_ops.yaml
index 6f45a6c46a8d68..67da0dbaae7578 100755
--- a/paddle/phi/ops/yaml/inconsistent/dygraph_ops.yaml
+++ b/paddle/phi/ops/yaml/inconsistent/dygraph_ops.yaml
@@ -286,6 +286,16 @@
backward : minimum_grad
... | [
"paddle/phi/ops/yaml/inconsistent/dygraph_ops.yaml"
] | [
{
"comment": "**🔴 Bug** `mm_out_dtype` 没有配置 `backward`，但 Python 入口已把 `paddle.mm(..., out_dtype=paddle.float32)` 直接路由到这个 op。\n\n现有 `matmul` 在同文件上方配置了 `backward : matmul_grad`；eager codegen 也会在 forward YAML 没有 `backward` 时把该 op 当成 forward-only 并跳过 GradNode 生成。因此用户把 BF16 输入设为 `stop_gradient=False` 后，`out_dtype` 分... | true | ||
PaddlePaddle/Paddle | 79,271 | issue_to_patch | [Cherry-pick][Precision Depth Alignment] UniformKernel in GPU | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | f3f5f3fd150c07b242d63f37999050f2b0e43d73 | 19474a269a3362404583e59443fe3fca3819e1f3 | diff --git a/paddle/phi/kernels/funcs/distribution_helper.h b/paddle/phi/kernels/funcs/distribution_helper.h
index d0c6e21d0148e6..dbf9817f06da06 100644
--- a/paddle/phi/kernels/funcs/distribution_helper.h
+++ b/paddle/phi/kernels/funcs/distribution_helper.h
@@ -69,22 +69,30 @@ struct exponential_transform {
T lambd... | [
"paddle/phi/kernels/funcs/distribution_helper.h",
"paddle/phi/kernels/gpu/uniform_kernel.cu",
"test/legacy_test/test_sparse_unary_op.py"
] | [] | diff --git a/test/legacy_test/test_sparse_unary_op.py b/test/legacy_test/test_sparse_unary_op.py
index 63062a92d87095..a9ae7d1cce4f16 100644
--- a/test/legacy_test/test_sparse_unary_op.py
+++ b/test/legacy_test/test_sparse_unary_op.py
@@ -62,6 +62,12 @@ def check_result(
mask = paddle.randint(0, 2, [8, 16,... | true | |
PaddlePaddle/Paddle | 79,130 | issue_to_patch | [Precision Depth Alignment] UniformKernel in GPU | ### PR Category
Operator Mechanism
### PR Types
Improvements
### Description
修复 GPU `UniformKernel` 在 float16/bfloat16 低精度类型下的精度问题，与 PyTorch 对齐。
**问题根因**：旧实现在 float（MT）精度中完成 `val * range + min` 后将结果隐式 cast 到 float16/bfloat16 输出类型。当某个 `val < 1.0` 的 curand 随机数计算结果非常接近 max 时，cast 会因精度截断舍入到 max，违反 `[min, max)` ... | de9bb946fdc161956631e363a83256a804a69256 | a303f1da89224ef5e1dcfbe80bd064310cfae482 | diff --git a/paddle/phi/kernels/funcs/distribution_helper.h b/paddle/phi/kernels/funcs/distribution_helper.h
index 774deb55766887..ace594160235b0 100644
--- a/paddle/phi/kernels/funcs/distribution_helper.h
+++ b/paddle/phi/kernels/funcs/distribution_helper.h
@@ -69,22 +69,30 @@ struct exponential_transform {
T lambd... | [
"paddle/phi/kernels/funcs/distribution_helper.h",
"paddle/phi/kernels/gpu/uniform_kernel.cu",
"test/legacy_test/test_sparse_unary_op.py"
] | [
{
"comment": "**🟡 建议** `seed != 0` 分支未同步精度修复，与 `seed == 0` 分支行为不一致\n\n`seed == 0` 分支已通过 `uniform_real_transform<MT, T>` 修复了 float16/bfloat16 cast 后可能等于 max 的边界问题，但 `seed != 0` 分支仍使用 `UniformGenerator<T>`，其内部以旧方式处理边界，未经过同等保护，导致两条路径的精度行为不一致。\n\n建议修复方式：检查 `UniformGenerator<T>` 的边界处理逻辑，或在 `seed != 0` 分支中同样先在 MT 精度... | diff --git a/test/legacy_test/test_sparse_unary_op.py b/test/legacy_test/test_sparse_unary_op.py
index 63062a92d87095..a9ae7d1cce4f16 100644
--- a/test/legacy_test/test_sparse_unary_op.py
+++ b/test/legacy_test/test_sparse_unary_op.py
@@ -62,6 +62,12 @@ def check_result(
mask = paddle.randint(0, 2, [8, 16,... | true | |
PaddlePaddle/Paddle | 79,228 | issue_to_patch | [API Compatibility] Align torch.Tensor.transpose_/ reshape_as | ### PR Category
User Experience
### PR Types
New features
### Description
torch.Tensor.transpose_：
Paddle 原来只支持 x.transpose_(perm=[...])，PyTorch 是 x.transpose_(dim0, dim1)。现在给 transpose_ 加了现有 transpose_decorator()，所以同时支持：
- Paddle 原用法：x.transpose_([1, 0, 2])、x.transpose_(perm=[1, 0, 2])
- PyT... | 002722386dbe898fb76b3d77f9cd1f20c175180b | a999c720b4df55e13f00122af9554aa2cb48d59f | diff --git a/python/paddle/tensor/__init__.py b/python/paddle/tensor/__init__.py
index aa0aa3c0c2d0c..79fc2ca567637 100644
--- a/python/paddle/tensor/__init__.py
+++ b/python/paddle/tensor/__init__.py
@@ -203,6 +203,7 @@
repeat_interleave,
reshape,
reshape_,
+ reshape_as,
roll,
rot90,
r... | [
"python/paddle/tensor/__init__.py",
"python/paddle/tensor/linalg.py",
"python/paddle/tensor/manipulation.py",
"test/legacy_test/test_api_compatibility_part3.py"
] | [
{
"comment": "按规范写API，英文文档",
"path": "python/paddle/tensor/manipulation.py",
"hunk": "@@ -5654,6 +5654,10 @@ def get_attr_shape(list_shape):\n return out\n \n \n+def reshape_as(x: Tensor, other: Tensor, name: str | None = None) -> Tensor:",
"resolving_sha": "a999c720b4df55e13f00122af9554aa2c... | diff --git a/test/legacy_test/test_api_compatibility_part3.py b/test/legacy_test/test_api_compatibility_part3.py
index 5ef1a7ac724d8..7b79a8098953a 100644
--- a/test/legacy_test/test_api_compatibility_part3.py
+++ b/test/legacy_test/test_api_compatibility_part3.py
@@ -4883,5 +4883,107 @@ def test_static_Compatibility(s... | true | |
PaddlePaddle/Paddle | 79,215 | issue_to_patch | [API Compatibility] select_scatter/sgn/signbit/slice_scatter/take/tensordot/tril_indices/triu_indices/vander/logaddexp/logspace/moveaxis/nan_to_num/nanmean/nansum/masked_fill/addmv/addr/fix/histc/trunc Edit By AI Agent | ### PR Category
User Experience
### PR Types
Improvements
### Description
**API Compatibility Edit By AI Agent：**
- `torch.select_scatter`
- `torch.Tensor.select_scatter`
- `torch.sgn`
- `torch.Tensor.sgn`
- `torch.signbit`
- `torch.Tensor.signbit`
- `torch.slice_scatter`
- `torch.Tensor.slice_scatter`
- `torch.take`... | 181bff1bf262b7dd2cb2b83502f9102fc550d3c0 | 1a418cb466d11d9f955268a6137ff9de7d170138 | diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py
index 387e50b2e85dfc..295cf12e8986de 100644
--- a/python/paddle/__init__.py
+++ b/python/paddle/__init__.py
@@ -582,6 +582,10 @@ def new_init(self, *args, **kwargs):
add_n,
addmm,
addmm_,
+ addmv,
+ addmv_,
+ addr,
+ addr_,
... | [
"python/paddle/__init__.py",
"python/paddle/base/dygraph/tensor_patch_methods.py",
"python/paddle/special.py",
"python/paddle/tensor/__init__.py",
"python/paddle/tensor/creation.py",
"python/paddle/tensor/manipulation.py",
"python/paddle/tensor/math.py",
"python/paddle/utils/decorator_utils.py",
"te... | [
{
"comment": "**🟡 建议** `out` 参数未设为 keyword-only\n\n在本 PR 的其他 API（`logaddexp`、`sgn`、`signbit`）中，`out` 参数均正确地放在 `*` 之后作为 keyword-only 参数。但 `tensordot` 的 `out` 是普通位置参数，放在 `name` 之后。这导致用户可以通过位置传参意外传入 `out`，且与 PyTorch 的 `out` 语义（始终为 keyword-only）不一致。\n\n建议修复：将 `out` 移到 `*` 之后，使其成为 keyword-only 参数：\n```python\ndef t... | diff --git a/test/dygraph_to_static/test_tensor_attr_consistency.py b/test/dygraph_to_static/test_tensor_attr_consistency.py
index 1d520355d40d8a..57968728c3eecc 100644
--- a/test/dygraph_to_static/test_tensor_attr_consistency.py
+++ b/test/dygraph_to_static/test_tensor_attr_consistency.py
@@ -30,8 +30,10 @@
'... | true | |
PaddlePaddle/Paddle | 79,278 | issue_to_patch | add fleet time | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 641cde71b3dcad673d60d915be85a563c4026a48 | 00d2d7b78e7f3171f12e8277bed3564c67c5b34c | diff --git a/.github/workflows/H-Coverage.yml b/.github/workflows/H-Coverage.yml
index 22ca839444cbc..bf5c34b2393cb 100644
--- a/.github/workflows/H-Coverage.yml
+++ b/.github/workflows/H-Coverage.yml
@@ -587,7 +587,7 @@ jobs:
'
- name: Single card test
- timeout-minutes: 50
+ timeout-... | [
".github/workflows/H-Coverage.yml"
] | [] | true | ||
PaddlePaddle/Paddle | 79,268 | issue_to_patch | [API Compatibility] Add alias `paddle.utils.data.DistributedSampler` for `paddle.io.DistributedBatchSampler` | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
User Experience
### PR Types
Improvements
### Description
<!-- Describe what you’ve done -->
- Add API alias and arg... | c3a820dbcb6b100fcd1cde032753521ebbf88831 | 6bcb52be4603674088a39231b1f67d97d1022c1e | diff --git a/python/paddle/io/dataloader/batch_sampler.py b/python/paddle/io/dataloader/batch_sampler.py
index 0b3ab75d06e24..571ce330c8f8b 100644
--- a/python/paddle/io/dataloader/batch_sampler.py
+++ b/python/paddle/io/dataloader/batch_sampler.py
@@ -233,6 +233,9 @@ class DistributedBatchSampler(BatchSampler):
... | [
"python/paddle/io/dataloader/batch_sampler.py",
"python/paddle/utils/data/__init__.py",
"python/paddle/utils/data/distributed.py",
"test/legacy_test/test_api_compatibility_part1.py",
"test/legacy_test/test_batch_sampler.py"
] | [
{
"comment": "看看seed是不是用在这里\n\nhttps://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/io/dataloader/batch_sampler.py#L339\n\n然后原来的DistributedBatchSampler可以把新增参数放到末尾，避免后向不兼容",
"path": "python/paddle/utils/data/distributed.py",
"hunk": "@@ -0,0 +1,41 @@\n+# Copyright (c) 2026 PaddlePaddle Autho... | diff --git a/test/legacy_test/test_api_compatibility_part1.py b/test/legacy_test/test_api_compatibility_part1.py
index 05b5df17d4324..25f31d9115c49 100644
--- a/test/legacy_test/test_api_compatibility_part1.py
+++ b/test/legacy_test/test_api_compatibility_part1.py
@@ -2396,5 +2396,60 @@ def test_Compatibility(self):
... | true | |
PaddlePaddle/Paddle | 79,164 | issue_to_patch | [API compatibility] Align Module state_dict_post_hook | ### PR Category
User Experience
### PR Types
New features, BC Breaking
### Description
state_dict_post_hook同时支持1个参数与4个参数的重载
### 是否引起精度变化
否 | 44b28a74657074dc7b37e1cf3f9c5335dcb27a1c | 2fbd87bbc58f014a8ac881a0e62cc6f9a61bbe69 | diff --git a/python/paddle/nn/layer/layers.py b/python/paddle/nn/layer/layers.py
index 2b274ddb3d92b5..f9b8c2bfbcb3d3 100644
--- a/python/paddle/nn/layer/layers.py
+++ b/python/paddle/nn/layer/layers.py
@@ -2755,8 +2755,17 @@ def _state_dict_impl(
)
if use_hook:
+ local_metada... | [
"python/paddle/nn/layer/layers.py",
"test/legacy_test/test_base_layer.py"
] | [
{
"comment": "**🔴 Bug** 字符串拼接缺少空格，错误信息拼接后存在粘连\n\n三段字符串相邻拼接时均缺少空格分隔符：\n- `\"are not\" + \"expected\"` → 输出 `\"are notexpected\"`\n- `\"modified,\" + \"it\"` → 输出 `\"modified,it should\"`\n\n建议修复：\n```python\nraise AssertionError(\n \"Hooks registered with ``register_load_state_dict_post_hook`` are not \"\n ... | diff --git a/test/legacy_test/test_base_layer.py b/test/legacy_test/test_base_layer.py
index 09dbf17c8e4d4c..099578e310e187 100644
--- a/test/legacy_test/test_base_layer.py
+++ b/test/legacy_test/test_base_layer.py
@@ -347,6 +347,88 @@ def state_dict_post_hook(destination):
hook_remove_helper._hook_id,... | true | |
PaddlePaddle/Paddle | 79,197 | issue_to_patch | [API Compatibility] Support param `optimizer` for `lr_scheduler` | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
User Experience
### PR Types
Improvements
### Description
<!-- Describe what you’ve done -->
- Add alias `optimizer`... | 44b28a74657074dc7b37e1cf3f9c5335dcb27a1c | 0fe31e929c0f371de856052c79fb65dc0579876c | diff --git a/python/paddle/optimizer/lr.py b/python/paddle/optimizer/lr.py
index 385bf9e4dd6a4..09f161a2169b5 100644
--- a/python/paddle/optimizer/lr.py
+++ b/python/paddle/optimizer/lr.py
@@ -20,7 +20,7 @@
import numpy
import numpy.typing as npt
-from typing_extensions import NotRequired
+from typing_extensions im... | [
"python/paddle/optimizer/lr.py",
"python/paddle/utils/decorator_utils.py",
"test/legacy_test/test_lr_scheduler.py"
] | [
{
"comment": " **优先级：P2**\n\n这里直接调用 `optimizer.set_lr_scheduler(self)`，但该 API 被 `@framework.dygraph_only` 限制；因此在 static/PIR 模式下，新引入的 `optimizer` 参数会在构造 scheduler 时抛 `AssertionError`。这些 lr scheduler 现有文档和测试都覆盖 static 图用法，PR 也没有声明 `optimizer` 只支持动态图，会形成不一致的兼容行为。... | diff --git a/test/legacy_test/test_lr_scheduler.py b/test/legacy_test/test_lr_scheduler.py
index 9b1e48e97e8e3..2721b25a63f66 100644
--- a/test/legacy_test/test_lr_scheduler.py
+++ b/test/legacy_test/test_lr_scheduler.py
@@ -1332,6 +1332,138 @@ def test_pir_linear_warmup_lr(self):
scheduler.step()
... | true | |
PaddlePaddle/Paddle | 79,210 | issue_to_patch | [Security] Harden CIFAR and Flowers dataset loading | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
User Experience
### PR Types
Security
### Description
本 PR 对 CIFAR 和 Flowers 数据集加载流程进行安全加固，降低非官方或被篡改数据文件带来的反序列化和解压风... | 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/python/paddle/dataset/cifar.py b/python/paddle/dataset/cifar.py
index f04acdfcc4b2b..bf7b12f94033e 100644
--- a/python/paddle/dataset/cifar.py
+++ b/python/paddle/dataset/cifar.py
@@ -27,6 +27,7 @@
"""
+import os
import pickle
import tarfile
@@ -44,7 +45,7 @@
CIFAR100_MD5 = 'eb9058c3a382ffc7106e4... | [
"python/paddle/dataset/cifar.py",
"python/paddle/utils/download.py",
"python/paddle/vision/datasets/cifar.py",
"python/paddle/vision/datasets/flowers.py",
"test/legacy_test/test_dataset_security.py"
] | [
{
"comment": "**🟡 建议** 此处新建了一个 `_safe_extract_tar` 实现，但 `python/paddle/utils/download.py` 中已有功能相近的 `_safe_extract_tar`（line 311），同样进行路径穿越检查和特殊文件过滤。两套实现行为略有不同（本处 hard-fail，`download.py` 版本 skip+warn），长期维护存在分歧修复风险。\n\n建议修复方式：\n考虑复用 `paddle.utils.download._safe_extract_tar`，或将其提取为公共工具函数并通过参数（如 `on_unsafe='raise'|... | diff --git a/test/legacy_test/test_dataset_security.py b/test/legacy_test/test_dataset_security.py
new file mode 100644
index 0000000000000..0b566a6308197
--- /dev/null
+++ b/test/legacy_test/test_dataset_security.py
@@ -0,0 +1,276 @@
+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under... | true | |
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | **🟡 建议** 此处新建了一个 `_safe_extract_tar` 实现，但 `python/paddle/utils/download.py` 中已有功能相近的 `_safe_extract_tar`（line 311），同样进行路径穿越检查和特殊文件过滤。两套实现行为略有不同（本处 hard-fail，`download.py` 版本 skip+warn），长期维护存在分歧修复风险。
建议修复方式：
考虑复用 `paddle.utils.download._safe_extract_tar`，或将其提取为公共工具函数并通过参数（如 `on_unsafe='raise'|'skip'`）控制遇到不安全成员时的行为，避免两... | 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/python/paddle/vision/datasets/flowers.py b/python/paddle/vision/datasets/flowers.py
index 792f399087b5b..012a500749bf8 100644
--- a/python/paddle/vision/datasets/flowers.py
+++ b/python/paddle/vision/datasets/flowers.py
@@ -35,6 +35,7 @@
from paddle.dataset.common import _check_exists_and_download
from p... | [
"python/paddle/vision/datasets/flowers.py"
] | [
{
"comment": "**🟡 建议** 此处新建了一个 `_safe_extract_tar` 实现，但 `python/paddle/utils/download.py` 中已有功能相近的 `_safe_extract_tar`（line 311），同样进行路径穿越检查和特殊文件过滤。两套实现行为略有不同（本处 hard-fail，`download.py` 版本 skip+warn），长期维护存在分歧修复风险。\n\n建议修复方式：\n考虑复用 `paddle.utils.download._safe_extract_tar`，或将其提取为公共工具函数并通过参数（如 `on_unsafe='raise'|... | true | ||
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | **🟡 建议** `test_reject_invalid_on_unsafe` 在同一 `with` 块内以写入模式（`'w:gz'`）开启 `data_tar` 并传给 `_safe_extract_tar`。
测试目前能通过，因为 `on_unsafe='bad'` 的参数校验在访问 `tar.getmembers()` 之前就抛出 ValueError，但写入模式的 tarfile 对象被用于解压函数，代码语义混乱，未来可能因 tarfile 或测试框架行为变化而产生误报。
建议将 tar 文件的创建与传参分开：
```python
def test_reject_invalid_on_unsafe(self):
... | 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/test/legacy_test/test_dataset_security.py b/test/legacy_test/test_dataset_security.py
new file mode 100644
index 0000000000000..0b566a6308197
--- /dev/null
+++ b/test/legacy_test/test_dataset_security.py
@@ -0,0 +1,276 @@
+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under... | [
"test/legacy_test/test_dataset_security.py"
] | [
{
"comment": "**🟡 建议** `test_reject_invalid_on_unsafe` 在同一 `with` 块内以写入模式（`'w:gz'`）开启 `data_tar` 并传给 `_safe_extract_tar`。\n\n测试目前能通过，因为 `on_unsafe='bad'` 的参数校验在访问 `tar.getmembers()` 之前就抛出 ValueError，但写入模式的 tarfile 对象被用于解压函数，代码语义混乱，未来可能因 tarfile 或测试框架行为变化而产生误报。\n\n建议将 tar 文件的创建与传参分开：\n```python\ndef test_reject... | true | ||
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | **❓ 疑问** `md5file(filename)` 在每次调用 `reader()` 时都会全量读取整个 CIFAR archive（CIFAR-10 约 170 MB），以计算 MD5。
对于 `cycle=True` 场景（`reader()` 只调用一次），开销仅在首次迭代时产生；但若用户在多个 epoch 中反复调用 `reader()`，每个 epoch 均会触发一次全量 I/O 校验。
PR 描述说明此设计是为了「避免缓存文件在下载校验后被替换」（TOCTOU 防护），若确认此为预期行为，建议在代码注释中显式说明，以免后续维护者以为这是性能 bug 而移除。
| 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/python/paddle/dataset/cifar.py b/python/paddle/dataset/cifar.py
index f04acdfcc4b2b..bf7b12f94033e 100644
--- a/python/paddle/dataset/cifar.py
+++ b/python/paddle/dataset/cifar.py
@@ -27,6 +27,7 @@
"""
+import os
import pickle
import tarfile
@@ -44,7 +45,7 @@
CIFAR100_MD5 = 'eb9058c3a382ffc7106e4... | [
"python/paddle/dataset/cifar.py"
] | [
{
"comment": "**❓ 疑问** `md5file(filename)` 在每次调用 `reader()` 时都会全量读取整个 CIFAR archive（CIFAR-10 约 170 MB），以计算 MD5。\n\n对于 `cycle=True` 场景（`reader()` 只调用一次），开销仅在首次迭代时产生；但若用户在多个 epoch 中反复调用 `reader()`，每个 epoch 均会触发一次全量 I/O 校验。\n\nPR 描述说明此设计是为了「避免缓存文件在下载校验后被替换」（TOCTOU 防护），若确认此为预期行为，建议在代码注释中显式说明，以免后续维护者以为这是性能 bug 而移除。\... | true | ||
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | **🟡 建议** 三处错误消息以 `Skipping` 开头，但在 `on_unsafe='raise'` 模式下作为 `ValueError` 抛出时语义不准确。
用户遇到 symlink 时会收到：`ValueError: Skipping symbolic link in tar for security: jpg/link.jpg`。"Skipping" 暗示成员被跳过继续执行，而实际上已抛出异常停止。hardlink 和 special file 同理。
建议区分两种场景的消息：
```python
if on_unsafe == 'raise':
raise ValueError(
f"Un... | 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/python/paddle/utils/download.py b/python/paddle/utils/download.py
index 4ba4a3938fbad..26261056fcf1c 100644
--- a/python/paddle/utils/download.py
+++ b/python/paddle/utils/download.py
@@ -308,15 +308,20 @@ def _uncompress_file_zip(filepath):
return uncompressed_path
-def _safe_extract_tar(tar,... | [
"python/paddle/utils/download.py"
] | [
{
"comment": "**🟡 建议** 三处错误消息以 `Skipping` 开头，但在 `on_unsafe='raise'` 模式下作为 `ValueError` 抛出时语义不准确。\n\n用户遇到 symlink 时会收到：`ValueError: Skipping symbolic link in tar for security: jpg/link.jpg`。\"Skipping\" 暗示成员被跳过继续执行，而实际上已抛出异常停止。hardlink 和 special file 同理。\n\n建议区分两种场景的消息：\n```python\nif on_unsafe == 'raise':\n ... | true | ||
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | ❓ **疑问（F3 仍存在）** `md5file(filename)` 在每次调用 `reader()` 时对整个 CIFAR archive（约 170 MB）全量计算 MD5。
对于多 epoch 训练，每轮重新调用 `reader()` 会额外读取 170 MB 文件。可考虑两种优化：
1. 用闭包标志位在首次调用时校验，后续 epoch 跳过；
2. 以文件 mtime + size 做快速预检，仅 mtime 变化时才做全量 MD5。
| 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/python/paddle/dataset/cifar.py b/python/paddle/dataset/cifar.py
index f04acdfcc4b2b..bf7b12f94033e 100644
--- a/python/paddle/dataset/cifar.py
+++ b/python/paddle/dataset/cifar.py
@@ -27,6 +27,7 @@
"""
+import os
import pickle
import tarfile
@@ -44,7 +45,7 @@
CIFAR100_MD5 = 'eb9058c3a382ffc7106e4... | [
"python/paddle/dataset/cifar.py"
] | [
{
"comment": "❓ **疑问（F3 仍存在）** `md5file(filename)` 在每次调用 `reader()` 时对整个 CIFAR archive（约 170 MB）全量计算 MD5。\n\n对于多 epoch 训练，每轮重新调用 `reader()` 会额外读取 170 MB 文件。可考虑两种优化：\n1. 用闭包标志位在首次调用时校验，后续 epoch 跳过；\n2. 以文件 mtime + size 做快速预检，仅 mtime 变化时才做全量 MD5。\n",
"path": "python/paddle/dataset/cifar.py",
"hunk": "@@ -... | true | ||
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | **🟡 建议** `md5file()` 在每次实例化（`data_file` 非 None 路径）时全量扫描 ~170MB 归档，缺少类似 `reader_creator` 中基于 `(mtime_ns, size)` 的缓存机制。
当用户每个 epoch 重建 `Cifar10(data_file=my_file)` 实例，或在超参数搜索中多次实例化时，每次都会触发一次完整的 MD5 计算，导致额外 ~0.5–1s 延迟。
建议在模块级添加一个简单的 mtime+size 缓存：
```python
_cifar_md5_cache: dict[str, tuple[tuple, str]] = {} # path ->... | 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/python/paddle/vision/datasets/cifar.py b/python/paddle/vision/datasets/cifar.py
index 9020da84cb147..700efdfd3f685 100644
--- a/python/paddle/vision/datasets/cifar.py
+++ b/python/paddle/vision/datasets/cifar.py
@@ -13,8 +13,10 @@
# limitations under the License.
from __future__ import annotations
+imp... | [
"python/paddle/vision/datasets/cifar.py"
] | [
{
"comment": "**🟡 建议** `md5file()` 在每次实例化（`data_file` 非 None 路径）时全量扫描 ~170MB 归档，缺少类似 `reader_creator` 中基于 `(mtime_ns, size)` 的缓存机制。\n\n当用户每个 epoch 重建 `Cifar10(data_file=my_file)` 实例，或在超参数搜索中多次实例化时，每次都会触发一次完整的 MD5 计算，导致额外 ~0.5–1s 延迟。\n\n建议在模块级添加一个简单的 mtime+size 缓存：\n```python\n_cifar_md5_cache: dict[str, tuple[... | true | ||
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | **🟡 建议** 测试文件中出现 `if __name__ == '__main__'`
按 Paddle CI checklist §C，`if __name__ == '__main__'` 出现在 `test/**` 下属于环境依赖信号，Paddle CI 通常以 `pytest` 模块方式运行测试，而非直接执行脚本，该块不会被 CI 触达。
建议删除这两行，统一由 CI 框架发现和运行测试用例。
| 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/test/legacy_test/test_dataset_security.py b/test/legacy_test/test_dataset_security.py
new file mode 100644
index 0000000000000..0b566a6308197
--- /dev/null
+++ b/test/legacy_test/test_dataset_security.py
@@ -0,0 +1,276 @@
+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under... | [
"test/legacy_test/test_dataset_security.py"
] | [
{
"comment": "**🟡 建议** 测试文件中出现 `if __name__ == '__main__'`\n\n按 Paddle CI checklist §C，`if __name__ == '__main__'` 出现在 `test/**` 下属于环境依赖信号，Paddle CI 通常以 `pytest` 模块方式运行测试，而非直接执行脚本，该块不会被 CI 触达。\n\n建议删除这两行，统一由 CI 框架发现和运行测试用例。\n",
"path": "test/legacy_test/test_dataset_security.py",
"hunk": "@@ -0,0 +1,26... | true | ||
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | **❓ 疑问** `Cifar100` 是否也继承或覆盖了此处的 `__init__` 安全校验逻辑？
PR 描述称 `Cifar10/Cifar100` 均已加固，但当前 diff 只可见一处 `__init__` 修改（hunk `@@ -154,6 +172,12 @@`）。如果 `Cifar100` 有自己的 `__init__` 且没有类似的 `_check_local_cifar_md5` 调用，则 `Cifar100(data_file=...)` 仍缺少 MD5 校验。请确认 `Cifar100` 是否共用此 `__init__` 或有独立的安全修改。
| 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/python/paddle/vision/datasets/cifar.py b/python/paddle/vision/datasets/cifar.py
index 9020da84cb147..700efdfd3f685 100644
--- a/python/paddle/vision/datasets/cifar.py
+++ b/python/paddle/vision/datasets/cifar.py
@@ -13,8 +13,10 @@
# limitations under the License.
from __future__ import annotations
+imp... | [
"python/paddle/vision/datasets/cifar.py"
] | [
{
"comment": "**❓ 疑问** `Cifar100` 是否也继承或覆盖了此处的 `__init__` 安全校验逻辑？\n\nPR 描述称 `Cifar10/Cifar100` 均已加固，但当前 diff 只可见一处 `__init__` 修改（hunk `@@ -154,6 +172,12 @@`）。如果 `Cifar100` 有自己的 `__init__` 且没有类似的 `_check_local_cifar_md5` 调用，则 `Cifar100(data_file=...)` 仍缺少 MD5 校验。请确认 `Cifar100` 是否共用此 `__init__` 或有独立的安全修改。\n",
... | true | ||
PaddlePaddle/Paddle | 79,210 | comment_to_fix | [Security] Harden CIFAR and Flowers dataset loading | **🟡 建议** 缓存测试（`_count_local_cifar_md5_calls`）仅覆盖 `Cifar10`，如果 `Cifar100` 也有独立的 MD5 校验路径，建议同样补充缓存命中和缓存刷新测试，确保安全加固行为的完整性。
| 82c1f9003cab94bbbbe35379699b053cd8811279 | 1a9ae4c8ae2f52d34326aa71603aad7597fbc178 | diff --git a/test/legacy_test/test_dataset_security.py b/test/legacy_test/test_dataset_security.py
new file mode 100644
index 0000000000000..0b566a6308197
--- /dev/null
+++ b/test/legacy_test/test_dataset_security.py
@@ -0,0 +1,276 @@
+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under... | [
"test/legacy_test/test_dataset_security.py"
] | [
{
"comment": "**🟡 建议** 缓存测试（`_count_local_cifar_md5_calls`）仅覆盖 `Cifar10`，如果 `Cifar100` 也有独立的 MD5 校验路径，建议同样补充缓存命中和缓存刷新测试，确保安全加固行为的完整性。\n",
"path": "test/legacy_test/test_dataset_security.py",
"hunk": "@@ -0,0 +1,256 @@\n+# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.\n+#\n+# Licensed un... | true | ||
PaddlePaddle/Paddle | 79,124 | issue_to_patch | [API compatibility] Align torch.distributions.multivariate_normal.MultivariateNormal、torch.distributions.distribution.Distribution、torch.distributions.distribution.Distribution.icdf / entropy、torch.distributions.normal.Normal | ### PR Category
User Experience
### PR Types
New features
### Description
- 支持 PyTorch 构造参数 `validate_args`，补充 `arg_constraints`、`support`、`mode`、`expand`，并对齐 `covariance_matrix`、
`precision_matrix`、`scale_tril` 三种参数入口及 lazy property 行为。
- 同时补充了分布参数校验能力：`loc` 必须是一维及以上，协方差矩阵和精度矩阵必须正定，`scale_tril` 必须是下三角... | 44b28a74657074dc7b37e1cf3f9c5335dcb27a1c | f514240249fbd852d84eb534f059479a1dc15c6b | diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py
index 387e50b2e85df..6f9f49449179a 100644
--- a/python/paddle/__init__.py
+++ b/python/paddle/__init__.py
@@ -19,6 +19,7 @@
# implemented as independent modules with no runtime dependency on PyTorch.
import math
+import sys as _sys
import typing
... | [
"python/paddle/__init__.py",
"python/paddle/distribution/constraint.py",
"python/paddle/distribution/distribution.py",
"python/paddle/distribution/multivariate_normal.py",
"python/paddle/distribution/normal.py",
"python/paddle/optim/__init__.py",
"python/paddle/optim/adadelta.py",
"python/paddle/optim... | [
{
"comment": "**🔴 兼容性** `paddle.optim` 子模块物理文件已全部删除，仅注册顶层 `sys.modules['paddle.optim']` 别名不足以支持子模块导入。\n\n**影响**：用户代码 `from paddle.optim.adam import Adam`、`import paddle.optim.adadelta` 等将在运行时抛出 `ModuleNotFoundError`，因为 `paddle/optim/adam.py` 等物理文件已被删除，而 `sys.modules` 没有为 `paddle.optim.adam`、`paddle.optim.adade... | diff --git a/test/distribution/test_distribution_multivariate_normal.py b/test/distribution/test_distribution_multivariate_normal.py
index fd54db855f7e8..f9c55f09393ac 100644
--- a/test/distribution/test_distribution_multivariate_normal.py
+++ b/test/distribution/test_distribution_multivariate_normal.py
@@ -24,6 +24,7 ... | true | |
PaddlePaddle/Paddle | 79,257 | issue_to_patch | [cherry-pick] Optimize the compilation of sleef (#79242) | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 47eac8a2fa862a738353430f74c248935f7c913f | fd3803fb418efecfac55409580c745ace883d975 | diff --git a/.gitmodules b/.gitmodules
index 9cc275e82c7c1..8acd3a38a7a0d 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -125,3 +125,4 @@
[submodule "third_party/sleef"]
path = third_party/sleef
url = https://github.com/shibatch/sleef.git
+ ignore = dirty
diff --git a/cmake/external/sleef.cmake b/cmake/external/sle... | [
".gitmodules",
"cmake/external/sleef.cmake",
"cmake/sleef.cmake"
] | [] | true | ||
PaddlePaddle/Paddle | 79,249 | issue_to_patch | [Revert] Revert PR #79240 and #79170 | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 44b28a74657074dc7b37e1cf3f9c5335dcb27a1c | a145a2998f93ba983a347e399f8410d27ce3b339 | diff --git a/cmake/third_party.cmake b/cmake/third_party.cmake
index 1bad4938a95e2..290d1fd198e3a 100755
--- a/cmake/third_party.cmake
+++ b/cmake/third_party.cmake
@@ -575,9 +575,29 @@ if(WITH_CUSPARSELT)
list(APPEND third_party_deps extern_cusparselt)
endif()
-set(FLASHATTN_ENABLED OFF)
+string(FIND "${CUDA_ARC... | [
"cmake/third_party.cmake",
"paddle/fluid/CMakeLists.txt",
"paddle/fluid/distributed/collective/CMakeLists.txt",
"paddle/fluid/pybind/CMakeLists.txt"
] | [
{
"comment": "这块是不是可以清理掉了？",
"path": "python/paddle/distributed/fleet/meta_parallel/dualpipev.py",
"hunk": "@@ -30,6 +30,56 @@\n except ImportError:\n deep_ep = None",
"resolving_sha": "a145a2998f93ba983a347e399f8410d27ce3b339",
"resolving_diff": ""
},
{
"comment": "**🟡 建议** `get_ev... | true | ||
PaddlePaddle/Paddle | 79,039 | issue_to_patch | [API compatibility] Align torch.nn.Module | ### PR Category
User Experience
### PR Types
New features, BC Breaking
### Description
对齐 `torch.nn.Module` 的以下 API 语义：
- 补齐 `get_extra_state/set_extra_state` 与 `state_dict/load_state_dict` 联动，支持 `_extra_state` 保存、加载、`strict` 错误处理、`metadata`、`load hooks` 顺序、`dict-like` 校验和 `assign` 基本语义。
- 对齐 `state/load h... | 44b28a74657074dc7b37e1cf3f9c5335dcb27a1c | 824ba22084ecb5a15d51e0fb5e98d6cef1ff62b8 | diff --git a/paddle/fluid/pybind/eager_method.cc b/paddle/fluid/pybind/eager_method.cc
index 0564e61297e43..adeab579a40f8 100644
--- a/paddle/fluid/pybind/eager_method.cc
+++ b/paddle/fluid/pybind/eager_method.cc
@@ -1453,7 +1453,7 @@ static PyObject* tensor_method_detach(TensorObject* self,
auto v = reinterpret_c... | [
"paddle/fluid/pybind/eager_method.cc",
"paddle/phi/api/include/compat/ATen/ops/detach.h",
"python/paddle/amp/auto_cast.py",
"python/paddle/nn/layer/layers.py",
"test/legacy_test/test_base_layer.py",
"test/legacy_test/test_base_module.py",
"test/legacy_test/test_detach.py",
"test/legacy_test/test_imper... | [
{
"comment": "这两个self._state_dict_hooks、self._state_dict_post_hooks\n\n是否只需要一个字典，本质是同一个东西吧。如果只是别名吗，那只用保留一个字典",
"path": "python/paddle/nn/layer/layers.py",
"hunk": "@@ -2645,9 +2664,14 @@ def register_state_dict_hook(\n return hook_remove_helper\n \n def register_state_dict_post_hook(\n- ... | diff --git a/test/legacy_test/test_base_layer.py b/test/legacy_test/test_base_layer.py
index 09dbf17c8e4d4..f12e00da38ed6 100644
--- a/test/legacy_test/test_base_layer.py
+++ b/test/legacy_test/test_base_layer.py
@@ -467,6 +467,192 @@ def load_state_dict_post_hook(layer, incompatible_keys):
{"weigh... | true | |
PaddlePaddle/Paddle | 79,219 | issue_to_patch | Fix VMM allocator hot paths and IPC slices | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 28e83f2ade8f4f850fe6dbc599d8ffbb6b4d1fbc | fc5b0a279004160a05e894129d73c0c9d1a25573 | diff --git a/paddle/fluid/pybind/tensor.cc b/paddle/fluid/pybind/tensor.cc
index cd795997f5243..bb6610d58f0de 100644
--- a/paddle/fluid/pybind/tensor.cc
+++ b/paddle/fluid/pybind/tensor.cc
@@ -200,7 +200,11 @@ namespace {
void ShareTensorViaVmm(const DenseTensor &self, py::tuple *out) {
auto *holder =
dynami... | [
"paddle/fluid/pybind/tensor.cc",
"paddle/phi/core/memory/allocation/allocator.h",
"paddle/phi/core/memory/allocation/virtual_memory_auto_growth_best_fit_allocator.cc",
"paddle/phi/core/memory/allocation/virtual_memory_auto_growth_best_fit_allocator.h",
"paddle/phi/core/memory/allocation/vmm_ipc_allocation.h... | [] | diff --git a/test/cpp/fluid/memory/vmm_auto_growth_best_fit_allocator_test.cu b/test/cpp/fluid/memory/vmm_auto_growth_best_fit_allocator_test.cu
index 70dfa14b5eb63..164cf745ec7aa 100644
--- a/test/cpp/fluid/memory/vmm_auto_growth_best_fit_allocator_test.cu
+++ b/test/cpp/fluid/memory/vmm_auto_growth_best_fit_allocator... | true | |
PaddlePaddle/Paddle | 79,261 | issue_to_patch | Fix VMM allocator hot paths and IPC slices | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 47eac8a2fa862a738353430f74c248935f7c913f | 0301193ffde88422226789f02b39e73f73669a88 | diff --git a/paddle/fluid/pybind/tensor.cc b/paddle/fluid/pybind/tensor.cc
index 12ab9d1b58c044..de795e2b4b277b 100644
--- a/paddle/fluid/pybind/tensor.cc
+++ b/paddle/fluid/pybind/tensor.cc
@@ -200,7 +200,11 @@ namespace {
void ShareTensorViaVmm(const DenseTensor &self, py::tuple *out) {
auto *holder =
dyna... | [
"paddle/fluid/pybind/tensor.cc",
"paddle/phi/core/memory/allocation/allocator.h",
"paddle/phi/core/memory/allocation/virtual_memory_auto_growth_best_fit_allocator.cc",
"paddle/phi/core/memory/allocation/virtual_memory_auto_growth_best_fit_allocator.h",
"paddle/phi/core/memory/allocation/vmm_ipc_allocation.h... | [] | diff --git a/test/cpp/fluid/memory/vmm_auto_growth_best_fit_allocator_test.cu b/test/cpp/fluid/memory/vmm_auto_growth_best_fit_allocator_test.cu
index 70dfa14b5eb637..164cf745ec7aa7 100644
--- a/test/cpp/fluid/memory/vmm_auto_growth_best_fit_allocator_test.cu
+++ b/test/cpp/fluid/memory/vmm_auto_growth_best_fit_allocat... | true | |
PaddlePaddle/Paddle | 79,242 | issue_to_patch | Optimize the compilation of sleef | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 82c1f9003cab94bbbbe35379699b053cd8811279 | b5e4a52a1bdc40122e82796cdbec367f54d47f7b | diff --git a/.gitmodules b/.gitmodules
index 6cf4b6b5c06426..8e090e15727b02 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -125,3 +125,4 @@
[submodule "third_party/sleef"]
path = third_party/sleef
url = https://github.com/shibatch/sleef.git
+ ignore = dirty
diff --git a/cmake/external/sleef.cmake b/cmake/external/s... | [
".gitmodules",
"cmake/external/sleef.cmake",
"cmake/sleef.cmake"
] | [
{
"comment": "**🟡 建议** 版本注释未同步更新\n\n当前注释仍写 `# Sleef version: 3.6.1`，但原来是通过 `set(SLEEF_TAG \"3.6.1\")` 明确锁定版本，现已改为跟随 submodule commit `6ee14bcae5fe92c2ff8b000d5a01102dab08d774`，两者对应关系不清晰。\n\n建议将注释更新为：\n```cmake\n# Sleef submodule: third_party/sleef @ 6ee14bc\n```\n或在 Description 中说明新 commit 对应的版本/标签。\n",
"p... | true | ||
PaddlePaddle/Paddle | 79,241 | issue_to_patch | [cherry-pick] fix repeat_interleave and normalize and swiglu_grad | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | e23968c8d7ceaeb1882328ae023ba045263d205f | 3ee24ad0fb409af7bdc4a76ad734948826297b52 | diff --git a/paddle/phi/kernels/cpu/p_norm_grad_kernel.cc b/paddle/phi/kernels/cpu/p_norm_grad_kernel.cc
index 03347aad33beae..d8a93cc31d2ad5 100644
--- a/paddle/phi/kernels/cpu/p_norm_grad_kernel.cc
+++ b/paddle/phi/kernels/cpu/p_norm_grad_kernel.cc
@@ -55,6 +55,9 @@ void PNormGradKernel(const Context& dev_ctx,
aut... | [
"paddle/phi/kernels/cpu/p_norm_grad_kernel.cc",
"paddle/phi/kernels/funcs/repeat_tensor2index_tensor.cc",
"paddle/phi/kernels/funcs/repeat_tensor2index_tensor.cu",
"paddle/phi/kernels/gpu/p_norm_grad_kernel.cu",
"paddle/phi/kernels/swiglu_grad_kernel.h",
"test/legacy_test/test_norm_all.py",
"test/legacy... | [] | diff --git a/test/legacy_test/test_norm_all.py b/test/legacy_test/test_norm_all.py
index 9ece64f2f4487c..7f118a963aab89 100644
--- a/test/legacy_test/test_norm_all.py
+++ b/test/legacy_test/test_norm_all.py
@@ -1806,6 +1806,38 @@ def test_dygraph(self):
)
+class API_PnormGradTest_ZeroSize(unittest.Test... | true | |
PaddlePaddle/Paddle | 79,253 | issue_to_patch | [CI] Fix docker image build failure handling | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | d9242d558766b5d24fa3231a798864c37cdd5cda | b40db187a8187cb20b909bd9b86f21263ae5a130 | diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index 37a2ca40079b1..f967e3cf938a6 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -120,14 +120,15 @@ jobs:
echo "${name}_image=${docker_image}" >> $GITHUB_OUTPUT
set +e
docker... | [
".github/workflows/docker.yml"
] | [
{
"comment": "You can simplify this logic and avoid toggling `set +e` / capturing `$?` by relying on Bash's `set -e` exception for `if` conditionals: `if docker pull ...; then` will not abort the script on pull failure, but `docker build` / `docker push` will still fail-fast. This also avoids potential quoting/... | true | ||
PaddlePaddle/Paddle | 79,250 | issue_to_patch | [API Compatibility] Add cuda_stream to Paddle stream wrapper | ### PR Category
User Experience
### PR Types
Improvements
### Description
参考 https://github.com/NVIDIA/cudnn-frontend/compare/develop...PFCCLab:cudnn-frontend:paddle 中 Paddle 适配对当前 CUDA stream 的处理：外部库原本使用 PyTorch 形态的 `torch.cuda.current_stream().cuda_stream` 获取 `cudaStream_t`，Paddle 适配需要改成 `stream_base.raw_str... | d9242d558766b5d24fa3231a798864c37cdd5cda | 775654c85ba3c4c82b477745c26e46d282896b87 | diff --git a/python/paddle/device/__init__.py b/python/paddle/device/__init__.py
index 9fc1bcb5cb747..4347c99078ef6 100644
--- a/python/paddle/device/__init__.py
+++ b/python/paddle/device/__init__.py
@@ -1497,6 +1497,13 @@ def _as_parameter_(self):
else:
return ctypes.c_void_p(self.stream_base.ra... | [
"python/paddle/device/__init__.py",
"test/legacy_test/test_cuda_stream_event.py"
] | [
{
"comment": "这行是不是没测？覆盖率过不了吧？",
"path": "python/paddle/device/__init__.py",
"hunk": "@@ -1497,6 +1497,12 @@ def _as_parameter_(self):\n else:\n return ctypes.c_void_p(self.stream_base.raw_stream)\n \n+ @property\n+ def cuda_stream(self) -> int:\n+ if isinstance(self.str... | diff --git a/test/legacy_test/test_cuda_stream_event.py b/test/legacy_test/test_cuda_stream_event.py
index d57965cf5ab85..21a375bba3217 100644
--- a/test/legacy_test/test_cuda_stream_event.py
+++ b/test/legacy_test/test_cuda_stream_event.py
@@ -193,6 +193,17 @@ def test_cuda_stream(self):
self.assertTrue(t... | true | |
PaddlePaddle/Paddle | 79,251 | issue_to_patch | [API Compatibility] Add torch-compatible NVTX range context manager | ### PR Category
User Experience
### PR Types
Improvements
### Description
This PR adds a torch-compatible NVTX range context manager to Paddle's existing NVTX namespace:
- `paddle.cuda.nvtx.range(msg, *args, **kwargs)`
- `paddle.device.nvtx.range(msg, *args, **kwargs)`
The wrapper follows `torch.cuda.nvtx.range` be... | d9242d558766b5d24fa3231a798864c37cdd5cda | bedbde59a5daedb2e00463161c8fc9e44db9bb5b | diff --git a/python/paddle/cuda/__init__.py b/python/paddle/cuda/__init__.py
index a5599bbf517c9b..065b116c571869 100644
--- a/python/paddle/cuda/__init__.py
+++ b/python/paddle/cuda/__init__.py
@@ -16,7 +16,8 @@
from __future__ import annotations
-from typing import TYPE_CHECKING, Union
+from contextlib import co... | [
"python/paddle/cuda/__init__.py",
"python/paddle/device/__init__.py",
"test/legacy_test/test_cuda_unittest.py"
] | [
{
"comment": "```suggestion\n def range(msg: str, *args: Any, **kwargs: Any) -> generator 的返回值咋标来着太久了我忘了，你自己查去:\n```",
"path": "python/paddle/device/__init__.py",
"hunk": "@@ -2012,6 +2013,23 @@ class amp:\n class nvtx:\n \"\"\"Namespace for NVTX marker operations.\"\"\"\n \n+ @staticmethod\n+... | diff --git a/test/legacy_test/test_cuda_unittest.py b/test/legacy_test/test_cuda_unittest.py
index 6e2bdf44d4f803..1f3d11cf7ce12d 100644
--- a/test/legacy_test/test_cuda_unittest.py
+++ b/test/legacy_test/test_cuda_unittest.py
@@ -17,6 +17,7 @@
import types
import unittest
import warnings
+from unittest import mock
... | true | |
PaddlePaddle/Paddle | 79,247 | issue_to_patch | 【Hackathon 10th Spring No.51】Deprecated Python 3.9 - part 6 | <!-- TemplateReference: https://github.com/PaddlePaddle/Paddle/wiki/PULL-REQUEST-TEMPLATE--REFERENCE -->
<!-- Demo: https://github.com/PaddlePaddle/Paddle/pull/24810 -->
### PR Category
<!-- One of [ User Experience | Execute Infrastructure | Operator Mechanism | CINN | Custom Device | Performance Optimization | D... | 7ef7b9c04aa379adf21738e3f9a76af66dd2c409 | ff62e665e78f1e7a8ef3337360711073d671e457 | diff --git a/.github/workflows/H-Coverage.yml b/.github/workflows/H-Coverage.yml
index 766fd8092fb9fa..22ca839444cbc6 100644
--- a/.github/workflows/H-Coverage.yml
+++ b/.github/workflows/H-Coverage.yml
@@ -209,11 +209,12 @@ jobs:
docker exec -t ${{ env.container_name }} /bin/bash -c '
source ~/.b... | [
".github/workflows/H-Coverage.yml",
".github/workflows/Night_ALL_Coverage.yml",
"CMakeLists.txt",
"ci/inference_build.sh",
"ci/run_mac_test.sh",
"ci/run_setup.sh",
"ci/utils.sh",
"cmake/python_module.cmake",
"paddle/common/flags.cc",
"paddle/scripts/conda_build.py",
"paddle/scripts/paddle_build.... | [
{
"comment": "The RuntimeError message concatenates two f-strings without a separating space, which will produce `now,you are using...` in the final message. Add a trailing space after the comma (or merge into one string) to keep the message readable.",
"path": "setup.py",
"hunk": "@@ -43,9 +43,9 @@\n v... | diff --git a/ci/run_mac_test.sh b/ci/run_mac_test.sh
index 3eb384bd1b1abc..bfcee12eac4188 100644
--- a/ci/run_mac_test.sh
+++ b/ci/run_mac_test.sh
@@ -27,11 +27,7 @@ function run_mac_test() {
========================================
EOF
- if [ "$1" == "cp38-cp38" ]; then
- pip3.8 uninstall -y ... | true | |
PaddlePaddle/Paddle | 79,240 | issue_to_patch | [Build] Keep NVSHMEM available for FlashMask | ### PR Category
Environment Adaptation
### PR Types
Bug fixes
### Description
修复 #79170 合入后 H-Coverage / FlashMask V3 仍需要 NVSHMEM 的场景。
#79170 已经在 DeepEP 相关 CMake 入口禁用了 GPU DeepEP；但它额外在 `cmake/third_party.cmake` 中对所有非 XPU 构建强制 `WITH_NVSHMEM=OFF`。这会覆盖 H-Coverage 显式传入的 `-DWITH_NVSHMEM=ON`，导致 `extern_nvshmem` 不再进入 thir... | d9242d558766b5d24fa3231a798864c37cdd5cda | 6d8c56681e0e2d6dcfe12782b562cebfe65d28e8 | diff --git a/cmake/third_party.cmake b/cmake/third_party.cmake
index d9776f1cc54ad..1bad4938a95e2 100755
--- a/cmake/third_party.cmake
+++ b/cmake/third_party.cmake
@@ -575,33 +575,9 @@ if(WITH_CUSPARSELT)
list(APPEND third_party_deps extern_cusparselt)
endif()
-string(FIND "${CUDA_ARCH_BIN}" "90" ARCH_BIN_CONTAI... | [
"cmake/third_party.cmake"
] | [
{
"comment": "是否可以，在 XPU OR FA 的场景下开启，感觉不用 FA 的场景也不用 nvshmem",
"path": "cmake/third_party.cmake",
"hunk": "@@ -581,10 +581,6 @@ if(NOT WITH_GPU\n OR (ARCH_BIN_CONTAINS_90 EQUAL -1))\n set(WITH_NVSHMEM OFF)\n endif()\n-# DeepEP GPU retired, force disable NVSHMEM for non-XPU builds\n-if(NOT WITH_XPU)... | true | ||
PaddlePaddle/Paddle | 79,240 | comment_to_fix | [Build] Keep NVSHMEM available for FlashMask | NVSHMEM gate is checking `CUDA_ARCH_BIN` via `string(FIND ...)`, but `CUDA_ARCH_BIN` is only defined when `CUDA_ARCH_NAME=Manual` (see `cmake/cuda.cmake` where it is `unset(CUDA_ARCH_BIN CACHE)` for all other modes). In common configurations like `CUDA_ARCH_NAME=Hopper/Blackwell/Auto`, this makes `ARCH_BIN_CONTAINS_90`... | d9242d558766b5d24fa3231a798864c37cdd5cda | 6d8c56681e0e2d6dcfe12782b562cebfe65d28e8 | diff --git a/cmake/third_party.cmake b/cmake/third_party.cmake
index d9776f1cc54ad..1bad4938a95e2 100755
--- a/cmake/third_party.cmake
+++ b/cmake/third_party.cmake
@@ -575,33 +575,9 @@ if(WITH_CUSPARSELT)
list(APPEND third_party_deps extern_cusparselt)
endif()
-string(FIND "${CUDA_ARCH_BIN}" "90" ARCH_BIN_CONTAI... | [
"cmake/third_party.cmake"
] | [
{
"comment": "NVSHMEM gate is checking `CUDA_ARCH_BIN` via `string(FIND ...)`, but `CUDA_ARCH_BIN` is only defined when `CUDA_ARCH_NAME=Manual` (see `cmake/cuda.cmake` where it is `unset(CUDA_ARCH_BIN CACHE)` for all other modes). In common configurations like `CUDA_ARCH_NAME=Hopper/Blackwell/Auto`, this makes ... | true | ||
PaddlePaddle/PaddleOCR | 18,158 | issue_to_patch | [cherry-pick] add PP-OCRv6 model download links (HuggingFace & ModelScope) to all README | 57d1cb6ab797f8676da7ab82070ef873a60b07dc | e50e1137e1b6a39c4bec6bea17f2fd87ea84e99b | diff --git a/README.md b/README.md
index 668ab730355..6458242fc13 100644
--- a/README.md
+++ b/README.md
@@ -70,6 +70,7 @@ English | [简体中文](./readme/README_cn.md) | [繁體中文](./readme/README
- **Specialized scenarios**: Major improvements in digital displays, dot-matrix characters, tire prints, and industrial text re... | [
"README.md",
"readme/README_ar.md",
"readme/README_cn.md",
"readme/README_es.md",
"readme/README_fr.md",
"readme/README_ja.md",
"readme/README_ko.md",
"readme/README_ru.md",
"readme/README_tcn.md"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,156 | issue_to_patch | add PP-OCRv6 model download links (HuggingFace & ModelScope) to README | f0bef31f676ad73b950469ebfcef1904cf07eefb | 40d9e917edfd72b92d66d2dbf278bfd87399aca0 | diff --git a/README.md b/README.md
index 668ab730355..6458242fc13 100644
--- a/README.md
+++ b/README.md
@@ -70,6 +70,7 @@ English | [简体中文](./readme/README_cn.md) | [繁體中文](./readme/README
- **Specialized scenarios**: Major improvements in digital displays, dot-matrix characters, tire prints, and industrial text re... | [
"README.md",
"readme/README_ar.md",
"readme/README_cn.md",
"readme/README_es.md",
"readme/README_fr.md",
"readme/README_ja.md",
"readme/README_ko.md",
"readme/README_ru.md",
"readme/README_tcn.md"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,153 | issue_to_patch | fix inference_engine doc | 83c5b7ba2f62bcc308bb9108bef7baf952b8b7a5 | e93158c647aa06db30f0ff29287f9c322aea6226 | diff --git a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md b/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md
index cfd0821b5b8..2a04434bf59 100644
--- a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md
+++ b/docs/version3.x/inference_d... | [
"docs/version3.x/inference_deployment/local_inference/inference_engine.en.md",
"docs/version3.x/inference_deployment/local_inference/inference_engine.md"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,150 | issue_to_patch | 【cherry-pick】fix v6 docs | 0bc13de323482dba0bd610f54693963dc8a0167d | 03973038f889c842f267bfb767bedd0a4354bc9d | diff --git a/docs/version3.x/module_usage/text_detection.en.md b/docs/version3.x/module_usage/text_detection.en.md
index 0e9f0359573..3cd8ac91269 100644
--- a/docs/version3.x/module_usage/text_detection.en.md
+++ b/docs/version3.x/module_usage/text_detection.en.md
@@ -25,7 +25,7 @@ The text detection module is a critic... | [
"docs/version3.x/module_usage/text_detection.en.md",
"docs/version3.x/module_usage/text_detection.md",
"docs/version3.x/module_usage/text_recognition.en.md",
"docs/version3.x/module_usage/text_recognition.md",
"docs/version3.x/pipeline_usage/OCR.en.md",
"docs/version3.x/pipeline_usage/OCR.md"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,149 | issue_to_patch | 【docs】fix v6 docs | 0345142187dd23375f3c9ceaf74b27c51250be4e | 92d3e4cf516b11a1e37e7fd1dca1ae133a6d4bd0 | diff --git a/docs/version3.x/module_usage/text_detection.en.md b/docs/version3.x/module_usage/text_detection.en.md
index 0e9f0359573..3cd8ac91269 100644
--- a/docs/version3.x/module_usage/text_detection.en.md
+++ b/docs/version3.x/module_usage/text_detection.en.md
@@ -25,7 +25,7 @@ The text detection module is a critic... | [
"docs/version3.x/module_usage/text_detection.en.md",
"docs/version3.x/module_usage/text_detection.md",
"docs/version3.x/module_usage/text_recognition.en.md",
"docs/version3.x/module_usage/text_recognition.md",
"docs/version3.x/pipeline_usage/OCR.en.md",
"docs/version3.x/pipeline_usage/OCR.md"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,143 | issue_to_patch | 【cp】fix readme | b03f46425e8ff4442b268ce449e3eef758146cd4 | 62044c0c63ad5ec12bc0831bdea7acffa6f3f031 | diff --git a/README.md b/README.md
index 3c976b2ba8f..668ab730355 100644
--- a/README.md
+++ b/README.md
@@ -70,7 +70,6 @@ English | [简体中文](./readme/README_cn.md) | [繁體中文](./readme/README
- **Specialized scenarios**: Major improvements in digital displays, dot-matrix characters, tire prints, and industrial text re... | [
"README.md",
"mkdocs.yml",
"readme/README_ar.md",
"readme/README_cn.md",
"readme/README_es.md",
"readme/README_fr.md",
"readme/README_ja.md",
"readme/README_ko.md",
"readme/README_ru.md",
"readme/README_tcn.md"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,142 | issue_to_patch | fix readme | 7265049b05adf69b7982b5e7c894ee088335cbc5 | 5adc73e690ef181d7668a78b421440f238d6edbe | diff --git a/README.md b/README.md
index 3c976b2ba8f..668ab730355 100644
--- a/README.md
+++ b/README.md
@@ -70,7 +70,6 @@ English | [简体中文](./readme/README_cn.md) | [繁體中文](./readme/README
- **Specialized scenarios**: Major improvements in digital displays, dot-matrix characters, tire prints, and industrial text re... | [
"README.md",
"readme/README_ar.md",
"readme/README_cn.md",
"readme/README_es.md",
"readme/README_fr.md",
"readme/README_ja.md",
"readme/README_ko.md",
"readme/README_ru.md",
"readme/README_tcn.md"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,140 | issue_to_patch | update new mkdocs | b03f46425e8ff4442b268ce449e3eef758146cd4 | 823e1e2cf9f1d870d1f5cac2774389bc8c1f471a | diff --git a/mkdocs.yml b/mkdocs.yml
index 6e98aa6db30..f9a6072031f 100644
--- a/mkdocs.yml
+++ b/mkdocs.yml
@@ -104,8 +104,6 @@ plugins:
PP-OCRv6: PP-OCRv6
使用教程: Usage Tutorial
PP-OCRv6简介: PP-OCRv6 Introduction
- PP-OCRv5简介: PP-OCRv5 Introduction
- PP-OCRv5多... | [
"mkdocs.yml"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,104 | issue_to_patch | [Feat] Support PP-OCRv6 | b53e29e858f47140bcfb8dc9d14846efa4edad22 | 0b7d2b2ac41f18dd989a2778a4732f830b1a2ef8 | diff --git a/README.md b/README.md
index aeb7f7e9a34..3c976b2ba8f 100644
--- a/README.md
+++ b/README.md
@@ -44,9 +44,9 @@ English | [简体中文](./readme/README_cn.md) | [繁體中文](./readme/README
### 🔍 Universal Text Recognition (Scene OCR)
> *The global gold standard for high-speed, multilingual text spotting.*
-* **100+... | [
"README.md",
"api_sdk/README.md",
"api_sdk/README_cn.md",
"api_sdk/go/README.md",
"api_sdk/go/README_cn.md",
"api_sdk/go/client_platform_test.go",
"api_sdk/go/client_test.go",
"api_sdk/go/doc.go",
"api_sdk/go/models.go",
"api_sdk/go/ocr.go",
"api_sdk/go/operation.go",
"api_sdk/go/results.go",
... | [] | diff --git a/api_sdk/go/client_platform_test.go b/api_sdk/go/client_platform_test.go
deleted file mode 100644
index f959cdb5a06..00000000000
--- a/api_sdk/go/client_platform_test.go
+++ /dev/null
@@ -1,85 +0,0 @@
-package paddleocr
-
-import (
- "context"
- "encoding/json"
- "net/http"
- "net/http/httptest"
- "testing"... | true | ||
PaddlePaddle/PaddleOCR | 18,129 | issue_to_patch | Fix PaddleOCR-VL HPS pipeline not using the vLLM server | The PaddleOCR-VL-1.6 HPS SDK ships a pipeline_config.yaml with `genai_config.backend: native`, which makes the pipeline run the VLM inside the Triton container instead of calling the dedicated vLLM server deployed by compose.yaml. The vLLM server stays completely idle while inference runs unbatched in the pipeline cont... | 71c7fefbf5e8478bc7f14b2d9b095bb80580178b | e6e9c61332c5e86e185aeaf8df2fbffc8a0231d8 | diff --git a/deploy/paddleocr_vl_docker/hps/prepare.sh b/deploy/paddleocr_vl_docker/hps/prepare.sh
index 6cb724dd318..929e5c86b93 100755
--- a/deploy/paddleocr_vl_docker/hps/prepare.sh
+++ b/deploy/paddleocr_vl_docker/hps/prepare.sh
@@ -53,6 +53,14 @@ sed -i.bak \
"${PIPELINE_CONFIG}"
rm -f "${PIPELINE_CONFIG}.ba... | [
"deploy/paddleocr_vl_docker/hps/prepare.sh"
] | [
{
"comment": "```suggestion\n```",
"path": "deploy/paddleocr_vl_docker/hps/prepare.sh",
"hunk": "@@ -53,6 +53,19 @@ sed -i.bak \\\n \"${PIPELINE_CONFIG}\"\n rm -f \"${PIPELINE_CONFIG}.bak\"\n \n+# Some SDKs (e.g. PaddleOCR-VL-1.6) ship a pipeline config with\n+# `backend: native`, which makes the pi... | true | ||
PaddlePaddle/PaddleOCR | 18,129 | comment_to_fix | Fix PaddleOCR-VL HPS pipeline not using the vLLM server | ```suggestion
``` | 71c7fefbf5e8478bc7f14b2d9b095bb80580178b | e6e9c61332c5e86e185aeaf8df2fbffc8a0231d8 | diff --git a/deploy/paddleocr_vl_docker/hps/prepare.sh b/deploy/paddleocr_vl_docker/hps/prepare.sh
index 6cb724dd318..929e5c86b93 100755
--- a/deploy/paddleocr_vl_docker/hps/prepare.sh
+++ b/deploy/paddleocr_vl_docker/hps/prepare.sh
@@ -53,6 +53,14 @@ sed -i.bak \
"${PIPELINE_CONFIG}"
rm -f "${PIPELINE_CONFIG}.ba... | [
"deploy/paddleocr_vl_docker/hps/prepare.sh"
] | [
{
"comment": "```suggestion\n```",
"path": "deploy/paddleocr_vl_docker/hps/prepare.sh",
"hunk": "@@ -53,6 +53,19 @@ sed -i.bak \\\n \"${PIPELINE_CONFIG}\"\n rm -f \"${PIPELINE_CONFIG}.bak\"\n \n+# Some SDKs (e.g. PaddleOCR-VL-1.6) ship a pipeline config with\n+# `backend: native`, which makes the pi... | true | ||
PaddlePaddle/PaddleOCR | 18,130 | issue_to_patch | refactor(langchain-paddleocr): migrate from HTTP API to PaddleOCR SDK | ## Summary
Migrate `langchain-paddleocr` from direct synchronous HTTP API calls to the official PaddleOCR SDK (`PaddleOCRClient`), aligning with the async Job API architecture in PaddleOCR 3.6.0+.
## Changes
- **`pyproject.toml`**: Upgrade to `paddleocr>=3.6.0`, remove `paddlex[serving]`, `requests`, `numpy`, `proto... | 71c7fefbf5e8478bc7f14b2d9b095bb80580178b | edd7f353bb50e897152efa33df0b3eebdca4579f | diff --git a/langchain-paddleocr/langchain_paddleocr/document_loaders/paddleocr_vl.py b/langchain-paddleocr/langchain_paddleocr/document_loaders/paddleocr_vl.py
index cff731d3775..9e2fb2f377a 100644
--- a/langchain-paddleocr/langchain_paddleocr/document_loaders/paddleocr_vl.py
+++ b/langchain-paddleocr/langchain_paddle... | [
"langchain-paddleocr/langchain_paddleocr/document_loaders/paddleocr_vl.py",
"langchain-paddleocr/pyproject.toml",
"langchain-paddleocr/tests/integration_tests/document_loaders/test_paddleocr_vl_loader.py",
"langchain-paddleocr/tests/unit_tests/document_loaders/test_paddleocr_vl_loader.py"
] | [] | diff --git a/langchain-paddleocr/tests/integration_tests/document_loaders/test_paddleocr_vl_loader.py b/langchain-paddleocr/tests/integration_tests/document_loaders/test_paddleocr_vl_loader.py
index c511fba5c80..0a7b14c3a94 100644
--- a/langchain-paddleocr/tests/integration_tests/document_loaders/test_paddleocr_vl_load... | true | |
PaddlePaddle/PaddleOCR | 17,933 | issue_to_patch | [Feat] Add PP-OCRv6 iOS Demo | 77ddbafccdee5020368c16c6960e6ff51e40b971 | 5834b8e7cd627f006307fb8d48ddcd5c84f3e3f0 | diff --git a/RELEASING.md b/RELEASING.md
new file mode 100644
index 00000000000..fe260a2c98f
--- /dev/null
+++ b/RELEASING.md
@@ -0,0 +1,194 @@
+# PaddleOCR Release SOP
+
+[中文](./RELEASING_cn.md)
+
+## Scope
+
+This document describes the standard release process for PaddleOCR.
+
+## Release Types
+
+The current proces... | [
"RELEASING.md",
"RELEASING_cn.md",
"deploy/ios_demo/.gitignore",
"deploy/ios_demo/NOTICE",
"deploy/ios_demo/PaddleOCRDemo.xcodeproj/project.pbxproj",
"deploy/ios_demo/PaddleOCRDemo.xcodeproj/xcshareddata/xcschemes/PaddleOCRDemo.xcscheme",
"deploy/ios_demo/PaddleOCRDemo.xcworkspace/contents.xcworkspaceda... | [] | diff --git a/deploy/ios_demo/scripts/extract_xctest_metrics.py b/deploy/ios_demo/scripts/extract_xctest_metrics.py
new file mode 100755
index 00000000000..24b9e375513
--- /dev/null
+++ b/deploy/ios_demo/scripts/extract_xctest_metrics.py
@@ -0,0 +1,129 @@
+#!/usr/bin/env python3
+# Copyright (c) 2026 PaddlePaddle Author... | true | ||
PaddlePaddle/PaddleOCR | 18,133 | issue_to_patch | add onnxruntime installation desc | 71c7fefbf5e8478bc7f14b2d9b095bb80580178b | 062b5f9b208d471bf5e9850c461b01fa5e9f2442 | diff --git a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md b/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md
index 45c741a1a4d..cfd0821b5b8 100644
--- a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md
+++ b/docs/version3.x/inference_d... | [
"docs/version3.x/inference_deployment/local_inference/inference_engine.en.md",
"docs/version3.x/inference_deployment/local_inference/inference_engine.md"
] | [] | true | |||
PaddlePaddle/PaddleOCR | 18,124 | issue_to_patch | add speed data | 6d2e6fb9c7df029f9f5d626952f179d4c9e49344 | e02cc271977066ac7974fd2f6ae8e567b9559204 | diff --git a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md b/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md
index 9501d5ed48b..45c741a1a4d 100644
--- a/docs/version3.x/inference_deployment/local_inference/inference_engine.en.md
+++ b/docs/version3.x/inference_d... | [
"docs/version3.x/inference_deployment/local_inference/inference_engine.en.md",
"docs/version3.x/inference_deployment/local_inference/inference_engine.md",
"docs/version3.x/module_usage/chart_parsing.en.md",
"docs/version3.x/module_usage/chart_parsing.md",
"docs/version3.x/module_usage/doc_img_orientation_cl... | [
{
"comment": "ONNXRuntime 官方名字两个单词是分开写的：ONNX Runtime\n\nhttps://onnxruntime.ai/docs/",
"path": "docs/version3.x/module_usage/chart_parsing.en.md",
"hunk": "@@ -52,6 +52,14 @@ paddleocr chart_parsing -i \"{'image': 'https://paddle-model-ecology.bj.bcebos.co\n --engine transformers\n ```\n \n+If you c... | true | |||
PaddlePaddle/PaddleOCR | 18,121 | issue_to_patch | [Feat] Add PP-OCRv6 Android Demo | 1e5aa0ad31bc8a82cd8e1daef7adc24e577d2534 | 35b1339167286d0a5e4872e620daf06b115c3ec9 | diff --git a/deploy/ppocr-android/.gitignore b/deploy/ppocr-android/.gitignore
new file mode 100644
index 00000000000..62d8c741289
--- /dev/null
+++ b/deploy/ppocr-android/.gitignore
@@ -0,0 +1,93 @@
+# OS generated files
+.DS_Store
+.DS_Store?
+._*
+.Spotlight-V100
+.Trashes
+ehthumbs.db
+Thumbs.db
+
+# Built applicat... | [
"deploy/ppocr-android/.gitignore",
"deploy/ppocr-android/README.md",
"deploy/ppocr-android/README_en.md",
"deploy/ppocr-android/app/build.gradle.kts",
"deploy/ppocr-android/app/src/main/AndroidManifest.xml",
"deploy/ppocr-android/app/src/main/java/com/paddle/ocr/demo/MainActivity.kt",
"deploy/ppocr-andr... | [] | true |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.