Load on CPU, move to ZeroGPU

app.py

@@ -33,8 +33,12 @@ for name, repo in MODELS.items():
     opts = LoadOptions.from_env()
     if opts.dtype is None:
         opts = replace(opts, dtype=torch.bfloat16)
-    LOADED[name] = checkpoint.load("cuda", opts)
-    tok, model = LOADED[name]
+    # ZeroGPU has no GPU at start-up: load on the CPU, then move. ZeroGPU defers the move until a @spaces.GPU call, and
+    # Kev builds its input tensors on model.device.
+    tok, model = checkpoint.load("cpu", opts)
+    model.to("cuda")
+    model.device = torch.device("cuda")
+    LOADED[name] = tok, model
     print(f"[jevasio] {repo}: base={checkpoint.meta.base}@{checkpoint.meta.base_revision} dtype={model.dtype} "
           f"temperature={model.head.temperature:.2f} loaded in {time.perf_counter() - t0:.0f} s", flush=True)