[wwb] Add to vlm chat pipeline support of GenAI supported models

Author: sbalandi

tools/who_what_benchmark/whowhatbench/chat_visualtext_evaluator.py

@@ -2,6 +2,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import json
+import torch
 import numpy as np
 import pandas as pd
 from tqdm import tqdm
@@ -187,7 +188,14 @@ def default_gen_answer(
                     # The output tuple has format (<list of decoded outputs without question/prompt>, <GenerateDecoderOnlyOutput>)
                     answer_text = tokens[0][0]
                 else:
-                    answer_tokens = tokens[:, preprocess_inputs["input_ids"].shape[-1] :]
+                    # Some models includes the input_ids in the generated tokens, some - not, so we need to check and remove them if needed
+                    inputs_num = preprocess_inputs["input_ids"].shape[-1]
+                    if tokens.shape[-1] > inputs_num and torch.equal(
+                        tokens[:, :inputs_num], preprocess_inputs["input_ids"]
+                    ):
+                        answer_tokens = tokens[:, preprocess_inputs["input_ids"].shape[-1] :]
+                    else:
+                        answer_tokens = tokens
                     answer_text = tokenizer.batch_decode(answer_tokens, skip_special_tokens=True)[0]
 
                 inputs_processor.update_chat_history_with_answer(answer_text)