Use 1-based indexing for position ids

README.md

@@ -42,7 +42,7 @@ past_key_values = {
 }
 input_ids = inputs['input_ids']
 attention_mask = inputs['attention_mask']
-position_ids = np.cumsum(inputs['attention_mask'], axis=-1) + 1
+position_ids = np.cumsum(inputs['attention_mask'], axis=-1)
 
 # 3. Generation loop
 max_new_tokens = 1024