fix: pre-cache RMBG-2.0 and DINOv2 in preload worker to prevent GPU lease timeout

briaai/RMBG-2.0 and DINOv2 (facebookresearch/dinov2) were being downloaded
inside the 240s @GPU callback on first use, exhausting the ZeroGPU lease before
inference could run. Preload worker now warms both caches at Space startup.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -161,6 +161,30 @@ def _preload_worker() -> None:
     except Exception as exc:
         print(f"[NeAR] preload: NeAR disk cache failed: {exc}", flush=True)
 
+    # Step 3: warm rembg model cache (briaai/RMBG-2.0, referenced in pipeline.yaml).
+    # Without this, the download happens inside the 240s GPU callback and times out.
+    try:
+        from huggingface_hub import snapshot_download
+        snapshot_download(repo_id="briaai/RMBG-2.0", token=os.environ.get("HF_TOKEN"))
+        print("[NeAR] preload: RMBG-2.0 disk cache ready.", flush=True)
+    except Exception as exc:
+        print(f"[NeAR] preload: RMBG-2.0 disk cache failed: {exc}", flush=True)
+
+    # Step 4: warm DINOv2 torch.hub cache.
+    # If NEAR_AUX_REPO is set, snapshot_download handles it inside load_dinov2_model.
+    # Otherwise we must pre-fetch facebookresearch/dinov2 from GitHub now (CPU-only).
+    if not (os.environ.get("NEAR_DINO_LOCAL_REPO") or os.environ.get("NEAR_AUX_REPO")):
+        try:
+            import torch
+            _dino_tmp = torch.hub.load(
+                "facebookresearch/dinov2", "dinov2_vitl14_reg",
+                pretrained=True, verbose=False,
+            )
+            del _dino_tmp
+            print("[NeAR] preload: DINOv2 torch.hub cache ready.", flush=True)
+        except Exception as exc:
+            print(f"[NeAR] preload: DINOv2 torch.hub cache failed: {exc}", flush=True)
+
 
 # ── GPU ensure helpers ────────────────────────────────────────────────────────
 # Called at the top of EVERY @GPU callback.  Always re-creates renderer and