From f2c5997847012a349e9d5bcf6f1463cb623c82c5 Mon Sep 17 00:00:00 2001
From: Your Name <your.email@example.com>
Date: Sat, 24 Jan 2026 23:46:44 +0800
Subject: [PATCH] fix: use huggingface_hub for ONNX download and add CACHE_BUST

- Use huggingface_hub.hf_hub_download() instead of curl (supports xet storage)
- Add CACHE_BUST ARG to force fresh downloads after cache delete
- Add font file size validation after download
- Re-enable registry cache (buildcache tag deleted manually)
- ARM64 graceful degradation for MinerU
---
 .github/workflows/release.yml |  7 ++++---
 Dockerfile                    | 26 ++++++++++++++++++++++++--
 2 files changed, 28 insertions(+), 5 deletions(-)

diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index e7663c3..328d20e 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -155,6 +155,7 @@ jobs:
       #   - 啟用 zstd compression 減少傳輸量和儲存空間
       #   - Multi-arch build 需要更多磁碟空間，已在上方清理
       #   - 明確指定 file: Dockerfile 避免與 Lite 版混淆
+      #   - 使用 CACHE_BUST 強制重新下載模型
       # ========================================
       - name: Build and push Docker image
         uses: docker/build-push-action@v6
@@ -166,15 +167,15 @@ jobs:
             ${{ env.DOCKER_IMAGE }}:${{ steps.tag.outputs.version }}
             ${{ env.DOCKER_IMAGE }}:latest
           platforms: linux/amd64,linux/arm64
-          # 🔥 使用獨立的 cache key 避免與 Lite 版交叉污染
-          # registry cache 沒有 10GB 限制，更適合大型模型
+          # 使用 registry cache（已手動刪除 Docker Hub 上的 buildcache tag）
           cache-from: type=registry,ref=${{ env.DOCKER_IMAGE }}:buildcache-full
           cache-to: type=registry,ref=${{ env.DOCKER_IMAGE }}:buildcache-full,mode=max,compression=zstd
           # 啟用壓縮減少 image 大小
           outputs: type=image,compression=zstd,compression-level=3
-          # 設定 build 參數
+          # 設定 build 參數（使用時間戳強制重新下載模型層）
           build-args: |
             BUILDKIT_INLINE_CACHE=1
+            CACHE_BUST=${{ github.run_id }}
 
       - name: Docker Build Summary
         run: |
diff --git a/Dockerfile b/Dockerfile
index 012392c..0bbc857 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -392,7 +392,15 @@ ENV PDF_SIGN_REASON="ConvertX-CN PDF Packager"
 ENV PDF_SIGN_LOCATION="Taiwan"
 ENV PDF_SIGN_CONTACT="convertx-cn@localhost"
 
-RUN set -eu && \
+# ==============================================================================
+# 🔥 Cache Busting（強制重新執行模型下載層）
+# ==============================================================================
+# 當 CACHE_BUST 改變時，後續所有層都會重新執行
+# 這確保模型下載不會被損壞的 cache 跳過
+# ==============================================================================
+ARG CACHE_BUST=1
+RUN echo "Cache bust: ${CACHE_BUST}" && \
+  set -eu && \
   echo "===========================================================" && \
   echo "🚀 階段 12-UNIFIED：Python 工具 + 模型統一安裝" && \
   echo "===========================================================" && \
@@ -453,7 +461,8 @@ RUN set -eu && \
   echo "" && \
   echo "📥 [6/8] 下載 PDFMathTranslate/BabelDOC DocLayout-YOLO ONNX 模型..." && \
   mkdir -p /root/.cache/babeldoc/models && \
-  python3 -c "from huggingface_hub import hf_hub_download; import shutil; import os; p=hf_hub_download(repo_id='wybxc/DocLayout-YOLO-DocStructBench-onnx', filename='doclayout_yolo_docstructbench_imgsz1024.onnx'); t='/root/.cache/babeldoc/models/doclayout_yolo_docstructbench_imgsz1024.onnx'; shutil.copy2(p, t); print(f'Downloaded and copied to {t}, size: {os.path.getsize(t)/1024/1024:.2f} MB')" && \
+  echo "   正在下載 ONNX 模型（約 75MB）..." && \
+  python3 -c "from huggingface_hub import hf_hub_download; import shutil, os, sys; print('   Downloading from HuggingFace...'); p=hf_hub_download(repo_id='wybxc/DocLayout-YOLO-DocStructBench-onnx', filename='doclayout_yolo_docstructbench_imgsz1024.onnx'); print(f'   Downloaded to cache: {p}'); t='/root/.cache/babeldoc/models/doclayout_yolo_docstructbench_imgsz1024.onnx'; shutil.copy2(p, t); size=os.path.getsize(t); print(f'   Copied to: {t}'); print(f'   File size: {size} bytes ({size/1024/1024:.2f} MB)'); sys.exit(1) if size < 10000000 else print('   SUCCESS: ONNX model downloaded and verified')" && \
   echo "✅ ONNX 模型下載完成" && \
   ls -lh /root/.cache/babeldoc/models/*.onnx && \
   \
@@ -483,6 +492,19 @@ RUN set -eu && \
   "https://github.com/timelic/source-han-serif/releases/download/main/SourceHanSerifKR-Regular.ttf" && \
   echo "✅ 字型下載完成" && \
   ls -lh /app/*.ttf && \
+  # 驗證字型檔案大小（確保下載完整）
+  echo "   驗證字型檔案..." && \
+  for font in GoNotoKurrent-Regular.ttf SourceHanSerifCN-Regular.ttf SourceHanSerifTW-Regular.ttf SourceHanSerifJP-Regular.ttf SourceHanSerifKR-Regular.ttf; do \
+    if [ ! -f "/app/$font" ]; then \
+      echo "   ERROR: Font not found: /app/$font" && exit 1; \
+    fi; \
+    size=$(stat -c%s "/app/$font" 2>/dev/null || stat -f%z "/app/$font" 2>/dev/null || echo 0); \
+    if [ "$size" -lt 1000000 ]; then \
+      echo "   ERROR: Font too small: /app/$font ($size bytes)" && exit 1; \
+    fi; \
+    echo "   ✓ $font ($size bytes)"; \
+  done && \
+  echo "✅ 所有字型驗證通過" && \
   \
   # ========================================
   # [7/8] 下載 BabelDOC 完整資源