diagnostics: surface gen_error + vram in /api/health

app.py

@@ -115,8 +115,17 @@ def _prep(req: "GenReq"):
 
 @app.get("/api/health")
 def health():
+    import torch
+    vram = None
+    try:
+        if torch.cuda.is_available():
+            free, total = torch.cuda.mem_get_info()
+            vram = {"free_gb": round(free / 1e9, 2), "total_gb": round(total / 1e9, 2),
+                    "allocated_gb": round(torch.cuda.memory_allocated() / 1e9, 2)}
+    except Exception:
+        pass
     return {"status": _state["status"], "device": _state["device"], "model": MODEL_ID,
-            "error": _state["error"]}
+            "error": _state["error"], "gen_error": _state.get("gen_error"), "vram": vram}
 
 
 def _not_ready_message():
@@ -162,12 +171,24 @@ def stream(req: GenReq):
 
 
 def _safe_generate(model, kwargs):
-    import torch
+    """Never raise into the request path — but DO record why, or failures are invisible."""
+    import torch, traceback
+    try:
+        torch.cuda.empty_cache()
+    except Exception:
+        pass
     try:
         with torch.no_grad():
             model.generate(**kwargs)
-    except Exception:
-        pass  # the streamer ends; the client sees whatever was produced
+        _state["gen_error"] = None
+    except Exception as e:
+        _state["gen_error"] = "%s: %s" % (type(e).__name__, str(e)[:400])
+        traceback.print_exc()
+    finally:
+        try:
+            torch.cuda.empty_cache()
+        except Exception:
+            pass
 
 
 @app.post("/api/generate")