F006/F008: serve Qwen models + model switcher (vanilla-first)

app.py

@@ -12,6 +12,15 @@ Run locally: ``uv run python app.py``.
 
 from __future__ import annotations
 
+# ZeroGPU: the `spaces` package MUST be imported before ANYTHING initializes CUDA
+# (i.e. before preload/torch below), or it raises "CUDA has been initialized before
+# importing the `spaces` package". So import it FIRST. Absent off-Space (local/tests),
+# where the `@spaces.GPU` decorator is a documented no-op anyway.
+try:
+    import spaces  # noqa: F401
+except ImportError:  # pragma: no cover - not on a ZeroGPU Space (local / tests)
+    pass
+
 # Prefer the ABSOLUTE packaged module so that `import app` (flat, via the test
 # pythonpath) and `import sql_env.app` BOTH delegate to the exact same
 # `sql_env.server.app_ui` — a single `Step`/class identity (R1). The flat

server/serving.py

@@ -141,16 +141,18 @@ def _load_model_and_tokenizer(source: str, revision: str | None) -> tuple[Any, A
 
 
 def _maybe_to_cuda(model: Any) -> Any:
-    """Move the model to CUDA when available (the only torch touch in the load path).
+    """Place the model on CUDA — the GPU-placement step of the load path.
 
-    Mirrors ``training/pipeline.py::run_eval`` and ``scripts/inference_smoke.py``. On a
-    CPU box this is a no-op; ``ModelPolicy._generate`` sends inputs to ``model.device``.
+    On a ZeroGPU Space, ``import spaces`` (done first in ``app.py``) enables a CUDA
+    emulation, so moving to ``cuda`` at load time is the REQUIRED pattern; the model
+    is materialized on the real GPU inside ``@spaces.GPU``. ``cuda.is_available()`` is
+    False on the startup container, so we must NOT guard on it. On a CPU-only box
+    ``.to("cuda")`` raises, so we fall back to CPU.
     """
-    import torch  # noqa: PLC0415
-
-    if torch.cuda.is_available():
+    try:
         return model.to("cuda")
-    return model
+    except Exception:  # CPU fallback when no CUDA is available (e.g. local dev)
+        return model
 
 
 def _ensure_loaded(key: str) -> tuple[Any, Any]: