fix : search engine

Dockerfile

@@ -3,21 +3,15 @@
 # Changes vs V3:
 #   • Removed deepface / GhostFaceNet / RetinaFace entirely
 #   • Added insightface + onnxruntime (SCRFD + ArcFace-R100)
-#   • Added huggingface_hub for AdaFace weight download
-#   • Pre-downloads AdaFace IR-50 WebFace4M weights at build time
-#   • Pre-downloads InsightFace buffalo_l pack at build time
-#   • Single worker (InsightFace ONNX is NOT thread-safe)
-#   • index dimensions: enterprise-faces=1024, enterprise-objects=1536
+#   • 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 ───────────────────────────────────────────────────
-# libGL + libGLib  : OpenCV headless
-# libgomp1         : OpenMP (used by ONNX runtime + numpy)
-# git              : needed by some HF hub downloads
-# curl             : useful for health checks / debug
 RUN apt-get update && apt-get install -y --no-install-recommends \
         libgl1 \
         libglib2.0-0 \
@@ -26,118 +20,119 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
         curl \
     && 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.
+ARG HF_TOKEN=""
+ENV HF_TOKEN=${HF_TOKEN}
+
 # ── Python deps ───────────────────────────────────────────────────
 COPY requirements.txt .
 RUN pip install --no-cache-dir --compile -r requirements.txt
 
-# ── Copy application code ────────────────────────────────────────
+# ── Copy application code ─────────────────────────────────────────
 COPY . .
 
 RUN mkdir -p temp_uploads saved_images && chmod -R 777 temp_uploads saved_images
 
-# ── Pre-download ALL AI models at BUILD time ─────────────────────
-# Bakes weights into image layer → cold start ~10s instead of ~5min
-#
-# Model sizes (approximate):
-#   SigLIP base    ~380 MB
-#   DINOv2 base    ~330 MB
-#   YOLO11n-seg    ~6 MB
-#   InsightFace buffalo_l (SCRFD-10GF + ArcFace-R100)  ~280 MB
-#   AdaFace IR-50 WebFace4M  ~170 MB
-#   Total image delta: ~1.2 GB
+# ── Pre-download public models at BUILD time ──────────────────────
+# SigLIP, DINOv2, YOLO, InsightFace buffalo_l — no token needed
 RUN python - <<'EOF'
 import os, sys
 
-# ── SigLIP ────────────────────────────────────────────────────────
-print("📦 Pre-downloading SigLIP...")
-from transformers import AutoProcessor, AutoModel
+print("Pre-downloading SigLIP...")
+from transformers import AutoProcessor, AutoModel, AutoImageProcessor
 AutoProcessor.from_pretrained("google/siglip-base-patch16-224", use_fast=True)
 AutoModel.from_pretrained("google/siglip-base-patch16-224")
-print("   ✅ SigLIP done")
+print("SigLIP done")
 
-# ── DINOv2 ───────────────────────────────────────────────────────
-print("📦 Pre-downloading DINOv2...")
-from transformers import AutoImageProcessor
+print("Pre-downloading DINOv2...")
 AutoImageProcessor.from_pretrained("facebook/dinov2-base")
 AutoModel.from_pretrained("facebook/dinov2-base")
-print("   ✅ DINOv2 done")
+print("DINOv2 done")
 
-# ── YOLO11n-seg ───────────────────────────────────────────────────
-print("📦 Pre-downloading YOLO11n-seg...")
+print("Pre-downloading YOLO11n-seg...")
 from ultralytics import YOLO
 YOLO("yolo11n-seg.pt")
-print("   ✅ YOLO done")
+print("YOLO done")
 
-# ── InsightFace buffalo_l ─────────────────────────────────────────
-# buffalo_l = SCRFD-10GF (detector) + ArcFace-R100 (encoder)
-# Handles small faces in group photos (det_size up to 1280x1280)
-print("📦 Pre-downloading InsightFace buffalo_l...")
+print("Pre-downloading InsightFace buffalo_l...")
 import numpy as np
 from insightface.app import FaceAnalysis
-face_app = FaceAnalysis(
-    name="buffalo_l",
-    providers=["CPUExecutionProvider"],
-)
+face_app = FaceAnalysis(name="buffalo_l", providers=["CPUExecutionProvider"])
 face_app.prepare(ctx_id=-1, det_size=(640, 640))
-# Warmup inference to confirm weights loaded
-test = np.zeros((112, 112, 3), dtype=np.uint8)
-face_app.get(test)
-print("   ✅ InsightFace buffalo_l done")
-
-# ── AdaFace IR-50 MS1MV2 ─────────────────────────────────────────
-# Repo: minchul/cvlface_adaface_ir50_ms1mv2
-# Loaded via AutoModel + trust_remote_code=True
-# Requires HF_TOKEN build arg (set in HF Space secrets)
-print("📦 Pre-downloading AdaFace IR-50 MS1MV2...")
-import os, sys
-from huggingface_hub import hf_hub_download
-from transformers import AutoModel
+face_app.get(np.zeros((112, 112, 3), dtype=np.uint8))
+print("InsightFace buffalo_l done")
 
-HF_TOKEN   = os.getenv("HF_TOKEN", None)
-REPO_ID    = "minchul/cvlface_adaface_ir50_ms1mv2"
-CACHE_PATH = os.path.expanduser("~/.cvlface_cache/minchul/cvlface_adaface_ir50_ms1mv2")
-os.makedirs(CACHE_PATH, exist_ok=True)
+print("All public models pre-downloaded successfully")
+EOF
 
-# Download files.txt manifest
-hf_hub_download(repo_id=REPO_ID, filename="files.txt",
-    token=HF_TOKEN, local_dir=CACHE_PATH, local_dir_use_symlinks=False)
+# ── 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).
+RUN python - <<'EOF'
+import os, sys, traceback
 
-with open(os.path.join(CACHE_PATH, "files.txt")) as f:
-    extra = [x.strip() for x in f.read().split("\n") if x.strip()]
+HF_TOKEN   = os.getenv("HF_TOKEN", "").strip()
+REPO_ID    = "minchul/cvlface_adaface_ir50_ms1mv2"
+CACHE_PATH = os.path.expanduser("~/.cvlface_cache/minchul/cvlface_adaface_ir50_ms1mv2")
 
-for fname in extra + ["config.json", "wrapper.py", "model.safetensors"]:
-    fpath = os.path.join(CACHE_PATH, fname)
-    if not os.path.exists(fpath):
-        hf_hub_download(repo_id=REPO_ID, filename=fname,
-            token=HF_TOKEN, local_dir=CACHE_PATH, local_dir_use_symlinks=False)
+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)")
+    sys.exit(0)
 
-# Load and verify
-cwd = os.getcwd(); os.chdir(CACHE_PATH); sys.path.insert(0, CACHE_PATH)
 try:
-    model = AutoModel.from_pretrained(CACHE_PATH, trust_remote_code=True, token=HF_TOKEN)
-finally:
-    os.chdir(cwd)
-    if CACHE_PATH in sys.path: sys.path.remove(CACHE_PATH)
-
-import torch
-with torch.no_grad():
-    out = model(torch.zeros(1, 3, 112, 112))
-emb = out if isinstance(out, torch.Tensor) else out.embedding
-print(f"   ✅ AdaFace loaded — output dim={emb.shape[-1]}")
-
-print("")
-print("✅ All V4 models pre-downloaded and verified!")
-print("   enterprise-faces  index dim : 1024  (ArcFace-512 + AdaFace-512)")
-print("   enterprise-objects index dim: 1536  (SigLIP-768 + DINOv2-768)")
+    from huggingface_hub import hf_hub_download
+    from transformers import AutoModel
+    import torch
+
+    print("Pre-downloading AdaFace IR-50 MS1MV2...")
+    os.makedirs(CACHE_PATH, exist_ok=True)
+
+    hf_hub_download(repo_id=REPO_ID, filename="files.txt",
+        token=HF_TOKEN, local_dir=CACHE_PATH, local_dir_use_symlinks=False)
+
+    with open(os.path.join(CACHE_PATH, "files.txt")) as f:
+        extra = [x.strip() for x in f.read().split("\n") if x.strip()]
+
+    for fname in extra + ["config.json", "wrapper.py", "model.safetensors"]:
+        fpath = os.path.join(CACHE_PATH, fname)
+        if not os.path.exists(fpath):
+            print("Downloading " + fname)
+            hf_hub_download(repo_id=REPO_ID, filename=fname,
+                token=HF_TOKEN, local_dir=CACHE_PATH, local_dir_use_symlinks=False)
+
+    cwd = os.getcwd()
+    os.chdir(CACHE_PATH)
+    sys.path.insert(0, CACHE_PATH)
+    try:
+        model = AutoModel.from_pretrained(
+            CACHE_PATH, trust_remote_code=True, token=HF_TOKEN)
+    finally:
+        os.chdir(cwd)
+        if CACHE_PATH in sys.path:
+            sys.path.remove(CACHE_PATH)
+
+    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")
+
+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")
+    sys.exit(0)
 EOF
 
 EXPOSE 7860
 
 # ── Single worker — InsightFace ONNX is NOT thread-safe ──────────
-# Each request acquires _face_lock before ONNX inference.
-# Multiple workers would each load their own model copy into RAM
-# (~1.5 GB each) which OOMs free HF Spaces (16 GB limit).
-# If you have a paid GPU Space with >32 GB RAM, set WEB_CONCURRENCY=2.
 ENV WEB_CONCURRENCY=1
 
 CMD uvicorn main:app \