addressed comments

megatron/text_generation_utils.py

@@ -207,13 +207,14 @@ def generate_samples_eval(model, context, max_gen_length, eos_token_id):
     with torch.no_grad():
         token_stream = get_token_stream(model, [context_tokens])
         for counter, decode_tokens in enumerate(token_stream):
-            decode_tokens, _ = decode_tokens
-            decode_tokens = decode_tokens[0].cpu().numpy().tolist()
-            trim_decode_tokens = tokenizer.detokenize(
-                decode_tokens)[raw_text_len:]
             if counter == args.out_seq_length:
                 break
 
+    decode_tokens, _ = decode_tokens
+    decode_tokens = decode_tokens[0].cpu().numpy().tolist()
+    trim_decode_tokens = tokenizer.detokenize(
+        decode_tokens)[raw_text_len:]
+ 
     return trim_decode_tokens