Added another learning source for Nora. Also added the requirements.

train.py

@@ -37,6 +37,13 @@ def train(
 
     device = config.device
     model.to(device)
+
+    # ─── ensure optimizer state is on the same device ───
+    # (this moves any loaded CPU buffers for Adam/AdamW into CUDA)
+    for state in optimizer.state.values():
+        for k, v in state.items():
+            if isinstance(v, torch.Tensor):
+                state[k] = v.to(device)
     criterion = nn.CrossEntropyLoss(ignore_index=tokenizer.stoi["<pad>"])
     scaler = GradScaler()