Fall back to CPU for backbone+SAE inference when no CUDA available

scripts/explorer/inference.py

@@ -47,16 +47,13 @@ def get_gpu_runner():
         return _gpu_runner
     if not args.sae_path:
         return None
-    if not torch.cuda.is_available():
-        print("[GPU runner] No CUDA device — on-the-fly inference disabled.")
-        return None
 
     src_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), '..', '..', 'src'))
     sys.path.insert(0, src_dir)
     from backbone_runners import load_batched_backbone
     from precompute_utils import load_sae, extract_tokens
 
-    dev = torch.device("cuda:0")
+    dev = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
     d_model = _all_datasets[0]['d_model']   # SAE output dim fixed to primary dataset
     print(f"[GPU runner] Loading {args.backbone} layer {args.layer} + SAE on {dev} ...")
     fwd, d_hidden, n_reg, tfm = load_batched_backbone(args.backbone, args.layer, dev)