Update app.py

app.py

@@ -1,10 +1,9 @@
 #!/usr/bin/env python3
-# app.py - Background Remover v32.2.3
-# - ONNX arena OFF (BFC fragmentation was OOMing BiRefNet-lite at 896px)
-# - VEHICLE_RES_CAP 640, FRAME_MAX_DIM 640 (fits in CPU memory)
-# - Video default: silueta (fast enough for 50s proxy budget)
-# - Frames-batch: non-blocking + temporal alpha smoothing
-# - Bulk: original filename preserved, 50-cap server-side
+# app.py - Background Remover v32.2.4
+# - RLIMIT_AS removed (was causing spurious bad_alloc at 12GB VA)
+# - BiRefNet-lite warmup skipped (was inflating RSS by ~6.8GB)
+# - Silueta default for video; proxy forces it
+# - ONNX arena OFF, VEHICLE_RES_CAP 640, FRAME_MAX_DIM 640
 
 import io
 import time
@@ -19,7 +18,6 @@ import gc
 import shutil
 import zipfile
 import json
-import resource
 import uuid
 import base64
 import urllib.parse
@@ -33,9 +31,6 @@ import psutil
 
 # ============================================================
 # ONNX RUNTIME
-# Arena OFF. BFC fragmented on BiRefNet-lite's deformable ASPP
-# and produced 822MB single-buffer allocation failures at 896px.
-# Per-inference memory is now predictable at the cost of ~10% speed.
 # ============================================================
 try:
     import onnxruntime as _ort
@@ -422,16 +417,8 @@ class ContentClassifier:
         return 'complex'
 
 
-def limit_memory():
-    try:
-        soft, hard = resource.getrlimit(resource.RLIMIT_AS)
-        resource.setrlimit(resource.RLIMIT_AS, (12 * 1024 * 1024 * 1024, hard))
-        logger.info("Memory limit set to 12GB")
-    except Exception as e:
-        logger.warning(f"Could not set memory limit: {e}")
-
-
-limit_memory()
+# RLIMIT_AS removed — see note above.
+logger.info("RLIMIT_AS not set — relying on cgroup memory.max")
 
 os.environ["U2NET_HOME"] = "/tmp/u2net_models"
 os.environ["OMP_NUM_THREADS"] = "2"
@@ -860,7 +847,14 @@ def _prewarm_models():
     try:
         test_img = Image.new('RGB', (64, 64), (255, 255, 255))
         ImageDraw.Draw(test_img).rectangle([20, 20, 44, 44], fill=(220, 40, 40))
+        # BiRefNet-lite warmup forces a 1024x1024 internal inference that
+        # reserves ~6 GB of transient allocator state which libc does not
+        # return to the OS on CPU-only boxes. Skip it.
+        WARMUP_SKIP = {'birefnet-general-lite'}
         for name, _ in PRELOAD_MODELS:
+            if name in WARMUP_SKIP:
+                logger.info(f"  {name} skipped (heavy warmup)")
+                continue
             session = model_manager.models.get(name)
             if session is None:
                 continue
@@ -879,12 +873,13 @@ def _prewarm_models():
                     logger.warning(f"  {name} WRONG (center={ca} corner={corner})")
             except Exception as e:
                 logger.warning(f"  {name}: {type(e).__name__}: {e}")
+        gc.collect()
     except Exception as e:
         logger.warning(f"Pre-warm skipped: {e}")
 
 
 logger.info("=" * 60)
-logger.info("Background Remover v32.2.3")
+logger.info("Background Remover v32.2.4")
 logger.info("=" * 60)
 pre_download_models()
 initialize_models()
@@ -1019,8 +1014,6 @@ def analyze_image_complexity(image):
             and cc_edge > 0.04
         )
 
-        # Vehicle path prefers BiRefNet-lite for quality on glossy surfaces.
-        # Falls back to Silueta if BiRefNet-lite isn't loaded.
         if is_vehicle_like:
             if has_birefnet:
                 return _route(
@@ -1259,7 +1252,7 @@ def process_image_sync(image_data, model_name=None, transparent=True,
 
 
 # ============================================================
-# VIDEO FRAME PIPELINE (BLOCKING WORKER + TEMPORAL SMOOTHING)
+# VIDEO FRAME PIPELINE
 # ============================================================
 def _process_frames_sync(raw_frames, resolved_model, transparent,
                          session_id: Optional[str] = None,
@@ -1362,7 +1355,7 @@ def _process_frames_sync(raw_frames, resolved_model, transparent,
 # ============================================================
 # FASTAPI
 # ============================================================
-app = FastAPI(title="Background Remover", version="32.2.3")
+app = FastAPI(title="Background Remover", version="32.2.4")
 app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_credentials=True,
                    allow_methods=["*"], allow_headers=["*"])
 
@@ -1570,9 +1563,6 @@ async def process_bulk_images(
         raise HTTPException(500, str(e)[:200])
 
 
-# ============================================================
-# FRAMES BATCH — non-blocking + hybrid temporal smoothing
-# ============================================================
 @app.post("/api/process/frames-batch")
 async def process_frames_batch(
     files: List[UploadFile] = File(...),
@@ -1734,6 +1724,7 @@ async def debug_env():
             'matting_models': sorted(MATTING_MODELS),
             'vehicle_res_cap': VEHICLE_RES_CAP,
             'frame_max_dim': FRAME_MAX_DIM,
+            'rlimit_as': None,
             'alpha_smoothing': _alpha_cache_stats()}
 
 
@@ -1744,7 +1735,7 @@ async def home():
     smoothing = _alpha_cache_stats()
     return HTMLResponse(f"""
     <!DOCTYPE html><html><head><meta charset="UTF-8">
-    <title>Background Remover v32.2.3</title>
+    <title>Background Remover v32.2.4</title>
     <style>body{{font-family:system-ui;background:#0A0A0A;
     color:#fff;min-height:100vh;display:flex;align-items:center;justify-content:center;
     margin:0;padding:20px}}.c{{max-width:900px;background:#111;
@@ -1752,13 +1743,13 @@ async def home():
     border:1px solid #222}}.s{{padding:15px;border-radius:12px;
     margin:20px 0;background:#1a1a1a}}a{{color:#7ED957}}</style>
     </head><body><div class="c">
-    <h1>Background Remover v32.2.3</h1>
+    <h1>Background Remover v32.2.4</h1>
     <div class="s">AI online. {len(stats['loaded_models'])} models.
     RAM {mem['percent']:.0f}% ({mem['used_mb']}/{mem['total_mb']}MB)</div>
     <p>Default: <strong>ISNet</strong> · Vehicles/F1: <strong>BiRefNet-lite @ {VEHICLE_RES_CAP}px</strong></p>
     <p>Video: <strong>{DEFAULT_VIDEO_MODEL}</strong> @ {FRAME_MAX_DIM}px · alpha smooth {smoothing['smooth_weight']}</p>
     <p>Sessions cached: {smoothing['sessions']}/{smoothing['max_sessions']}</p>
-    <p>ONNX: arena=OFF</p>
+    <p>ONNX: arena=OFF · RLIMIT_AS: unset</p>
     <p><a href="/docs">API Docs</a></p>
     </div></body></html>
     """)
@@ -1781,7 +1772,7 @@ async def health_check():
         'pymatting_available': PYMATTING_AVAILABLE,
         'image_formats': ['png', 'jpg', 'jpeg', 'webp', 'bmp'],
         'bulk_limit': 50,
-        'version': '32.2.3',
+        'version': '32.2.4',
         'default_model': get_default_model(),
         'default_video_model': DEFAULT_VIDEO_MODEL,
         'preload_models': [n for n, _ in PRELOAD_MODELS],
@@ -1791,6 +1782,7 @@ async def health_check():
         'vehicle_res_cap': VEHICLE_RES_CAP,
         'frame_max_dim': FRAME_MAX_DIM,
         'onnx_arena': False,
+        'rlimit_as': None,
         'alpha_smoothing': _alpha_cache_stats(),
         'frame_batch_endpoint': '/api/process/frames-batch',
         'frame_batch_size_max': 4,
@@ -1818,11 +1810,10 @@ async def gen_exc(request, exc):
 @app.on_event("startup")
 async def startup():
     logger.info("=" * 60)
-    logger.info("v32.2.3 ready")
+    logger.info("v32.2.4 ready")
     logger.info("ONNX: arena=OFF, mem_pattern=OFF")
+    logger.info("RLIMIT_AS: unset (relying on cgroup memory.max)")
     logger.info("Video: Silueta default + temporal alpha smoothing")
-    logger.info(f"Smoothing weight: {ALPHA_SMOOTH_WEIGHT}, "
-                f"max sessions: {ALPHA_CACHE_MAX_SESSIONS}")
     logger.info(f"Video default model: {DEFAULT_VIDEO_MODEL}")
     logger.info(f"Video max frame dim: {FRAME_MAX_DIM}")
     logger.info(f"Vehicle res cap: {VEHICLE_RES_CAP}")