fix: VL device detection, failed-capture error state, HF_HOME=/data/.cache

app.py

@@ -1,8 +1,12 @@
+import os
 import re
 import json
 import shutil
 import asyncio
 from pathlib import Path
+
+# Cache HF model downloads in persistent storage so cold starts skip re-downloading
+os.environ.setdefault("HF_HOME", "/data/.cache")
 from typing import Optional, List
 
 import httpx
@@ -408,7 +412,16 @@ async def _process(cid: int, type: str, **kwargs):
         recall_q = await loop.run_in_executor(None, generate_recall_question, summary, intent or "", claims)
         update_capture(cid, summary, result.get("tags", []), intent, emb, related, recall_q, claims, source_content_path, title)
     except Exception as e:
-        print(f"[process error] {e}")
+        print(f"[process error] cid={cid} {e}")
+        import sqlite3
+        try:
+            with sqlite3.connect(DB_PATH) as conn:
+                conn.execute(
+                    "UPDATE captures SET summary=?, intent='ephemeral' WHERE id=? AND summary IS NULL",
+                    ("⚠ Processing failed — delete and retry", cid)
+                )
+        except Exception:
+            pass
 
 
 async def _fetch_page(url: str) -> tuple[str, str]:

lm.py

@@ -380,7 +380,9 @@ else:
         img = _PILImage.open(file_path).convert("RGB")
         if max(img.width, img.height) > 768:
             img.thumbnail((768, 768), _PILImage.LANCZOS)
-        inputs = processor(text=text_prompt, images=img, return_tensors="pt").to("cuda")
+        # Match inputs to wherever the model actually is (ZeroGPU may or may not have moved it)
+        device = next(vl_model.parameters()).device
+        inputs = processor(text=text_prompt, images=img, return_tensors="pt").to(device)
         with torch.no_grad():
             out = vl_model.generate(**inputs, max_new_tokens=256)
         return processor.decode(out[0], skip_special_tokens=True)