Force CPU device explicitly (ZeroGPU makes cuda.is_available() report True outside a real grant, breaking auto-detection)

app.py

@@ -27,6 +27,15 @@ Space's own runtime error, not guessed:
    on CPU exactly as the rest of this project does. It exists solely to
    satisfy that platform check without changing this project's actual
    CPU-only inference path.
+3. **`device="cpu"` passed explicitly to `build_interface` below**: its
+   default auto-detection (`drdetect.serve.demo`) checks
+   `torch.cuda.is_available()`, which ZeroGPU makes report `True` even
+   outside an actual `@spaces.GPU` grant -- auto-detection then picked
+   "cuda" and crashed the first real request with `RuntimeError: Low-level
+   CUDA init... reached` the moment a tensor was moved to it, since nothing
+   in the demo path is wrapped in `@spaces.GPU`. Forcing "cpu" here matches
+   this project's actual design (CPU-only inference throughout) rather than
+   trusting a detection heuristic that ZeroGPU's emulation invalidates.
 
 Both are wrapped so this file still runs unmodified outside a Space
 (`spaces` is not, and should not be, a dependency of the main `drdetect`
@@ -63,5 +72,7 @@ if __name__ == "__main__":
 
     # regression loss, not ce: it won the locked external evaluation decisively
     # (docs/22_PHASE8_VALIDATION_RESULTS.md, referable AUC 0.924 vs 0.888, DeLong p=6.1e-10)
-    demo = build_interface(checkpoint, backbone="efficientnet_b0", loss_name="regression", size=512)
+    demo = build_interface(
+        checkpoint, backbone="efficientnet_b0", loss_name="regression", size=512, device="cpu"
+    )
     demo.launch(ssr_mode=False)

src/drdetect/serve/demo.py

@@ -52,15 +52,23 @@ def build_interface(
     backbone: str = "efficientnet_b0",
     loss_name: str = "ce",
     size: int = 512,
+    device: str | None = None,
 ):
+    """`device=None` (default) auto-detects mps/cuda/cpu, the right choice for
+    local development. Pass an explicit `device` to override that -- needed
+    on HF Spaces' free ZeroGPU tier, where `torch.cuda.is_available()`
+    reports True outside an actual `@spaces.GPU` grant, so auto-detection
+    would pick "cuda" and crash the moment a tensor touches it (see app.py).
+    """
     import gradio as gr
     import torch
 
-    device = (
-        "mps"
-        if torch.backends.mps.is_available()
-        else ("cuda" if torch.cuda.is_available() else "cpu")
-    )
+    if device is None:
+        device = (
+            "mps"
+            if torch.backends.mps.is_available()
+            else ("cuda" if torch.cuda.is_available() else "cpu")
+        )
     model = load_grader(checkpoint, backbone=backbone, loss_name=loss_name, device=device)
     temperature = load_temperature(checkpoint)