fix: build issues

Dockerfile

@@ -1,18 +1,17 @@
 # Dockerfile — Enterprise Lens V4
-#
-# Changes vs V3:
-#   • Removed deepface / GhostFaceNet / RetinaFace entirely
-#   • Added insightface + onnxruntime (SCRFD + ArcFace-R100)
-#   • AdaFace IR-50 MS1MV2 pre-downloaded at build time (needs HF_TOKEN)
-#   • Single worker — InsightFace ONNX is NOT thread-safe
-#   • Index dims: enterprise-faces=1024, enterprise-objects=1536
 
 FROM python:3.10-slim
 
 WORKDIR /app
 
 # ── System deps ───────────────────────────────────────────────────
+# g++ / build-essential : required to compile insightface C++ extensions
+# libgl1 + libglib2.0-0 : OpenCV headless
+# libgomp1              : OpenMP for ONNX runtime
+# git + curl            : HF hub downloads
 RUN apt-get update && apt-get install -y --no-install-recommends \
+        build-essential \
+        g++ \
         libgl1 \
         libglib2.0-0 \
         libgomp1 \
@@ -21,9 +20,8 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     && rm -rf /var/lib/apt/lists/*
 
 # ── HF_TOKEN build arg ────────────────────────────────────────────
-# Set in HF Space -> Settings -> Repository Secrets as HF_TOKEN.
-# HF Spaces passes Repository Secrets as both runtime env vars
-# AND Docker build ARGs with the same name automatically.
+# Add HF_TOKEN to HF Space -> Settings -> Repository Secrets.
+# HF Spaces passes Repository Secrets as Docker build ARGs automatically.
 ARG HF_TOKEN=""
 ENV HF_TOKEN=${HF_TOKEN}
 
@@ -37,7 +35,6 @@ COPY . .
 RUN mkdir -p temp_uploads saved_images && chmod -R 777 temp_uploads saved_images
 
 # ── Pre-download public models at BUILD time ──────────────────────
-# SigLIP, DINOv2, YOLO, InsightFace buffalo_l — no token needed
 RUN python - <<'EOF'
 import os, sys
 
@@ -68,10 +65,7 @@ print("InsightFace buffalo_l done")
 print("All public models pre-downloaded successfully")
 EOF
 
-# ── Pre-download AdaFace (separate step — graceful failure) ───────
-# Split from above so a failure here does NOT fail the whole build.
-# If HF_TOKEN is missing or wrong, build still succeeds and the app
-# runs in ArcFace-only fallback mode (zero-padded to 1024-D).
+# ── Pre-download AdaFace (graceful — never fails the build) ───────
 RUN python - <<'EOF'
 import os, sys, traceback
 
@@ -81,8 +75,7 @@ CACHE_PATH = os.path.expanduser("~/.cvlface_cache/minchul/cvlface_adaface_ir50_m
 
 if not HF_TOKEN:
     print("HF_TOKEN not set - skipping AdaFace pre-download")
-    print("AdaFace will retry at runtime if HF_TOKEN is set then")
-    print("Fallback: ArcFace-only zero-padded to 1024-D (build succeeds)")
+    print("Fallback: ArcFace zero-padded to 1024-D (build succeeds)")
     sys.exit(0)
 
 try:
@@ -120,19 +113,17 @@ try:
     with torch.no_grad():
         out = model(torch.zeros(1, 3, 112, 112))
     emb = out if isinstance(out, torch.Tensor) else out.embedding
-    print("AdaFace pre-download complete - output dim=" + str(emb.shape[-1]))
-    print("1024-D FULL FUSION will be active at runtime")
+    print("AdaFace done - output dim=" + str(emb.shape[-1]))
 
 except Exception as e:
     print("AdaFace pre-download failed: " + str(e))
     print(traceback.format_exc()[-400:])
-    print("Build continues - fallback to ArcFace-only zero-padded 1024-D")
+    print("Build continues - ArcFace-only fallback active")
     sys.exit(0)
 EOF
 
 EXPOSE 7860
 
-# ── Single worker — InsightFace ONNX is NOT thread-safe ──────────
 ENV WEB_CONCURRENCY=1
 
 CMD uvicorn main:app \

requirements.txt

@@ -12,6 +12,7 @@ torch==2.4.1+cpu
 torchvision==0.19.1+cpu
 
 # ── HuggingFace stack ─────────────────────────────────────────────
+# transformers unpinned — let pip resolve to avoid conflicts with ultralytics
 transformers
 huggingface_hub>=0.26.0
 safetensors>=0.4.0
@@ -19,6 +20,9 @@ tokenizers>=0.20.0
 accelerate>=1.1.0
 
 # ── InsightFace — SCRFD + ArcFace-R100 ───────────────────────────
+# insightface 0.7.3 has no pre-built wheel → pip tries to compile C++
+# and fails without g++. Using 0.7.3 with g++ now added to Dockerfile.
+# onnxruntime provides CPU ONNX inference for InsightFace models.
 insightface==0.7.3
 onnxruntime==1.19.2