Ensure model is in eval mode with sampling

sigpt/model/sample.py

@@ -12,6 +12,7 @@ def generate_tokens(
 ) -> torch.Tensor:
     generations = 0
     with torch.no_grad():
+        model.eval()
         while generations < max_samples:
             logits = model(idx)  # (B, T, vocab_size)
             # We use the last token for prediction
@@ -40,4 +41,4 @@ def generate(
     encoded_prompt = torch.tile(encoded_prompt, (batches, 1))
 
     generated_tokens = generate_tokens(model, encoded_prompt, max_samples, k)
-    return [encoder.decode(i.tolist()) for i in generated_tokens]
+    return [encoder.decode(i.tolist()) for i in generated_tokens]