Upload app.py with huggingface_hub

app.py

@@ -134,7 +134,7 @@ print(f"Sage 1B loaded - {sum(p.numel() for p in model.parameters()):,} params")
 
 def generate_text(prompt, max_length, temperature):
     tokens = tok.encode(prompt).ids[:50]
-    inp = torch.tensor([[1] + tokens], dtype=torch.long)
+    inp = torch.tensor([[2] + tokens], dtype=torch.long)
     out = model.generate(inp, max_new=int(max_length), temp=temperature, top_k=40)
     return tok.decode(out[0].tolist(), skip_special_tokens=True)