Increment start_pos by encoded size in generate

torchchat/generate.py

@@ -1192,6 +1192,8 @@ def callback(x, *, done_generating=False):
                     max_seq_length=max_seq_length,
                     attention_backend=self.builder_args.attention_backend,
                 )
+                if generator_args.chat_mode:
+                    start_pos += encoded.size(0)
                 for token_tensor, metrics in generator_func:
                     if token_tensor is not None:
                         start_pos += token_tensor.size(0)