Add files using upload-large-folder tool

LLaVA/docs/Data.md

@@ -0,0 +1,29 @@
+## Data
+
+| Data file name | Size |
+| --- | ---: |
+| [llava_instruct_150k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_instruct_150k.json) | 229 MB |
+| [llava_instruct_80k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/llava_instruct_80k.json) | 229 MB |
+| [conversation_58k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/conversation_58k.json) | 126 MB |
+| [detail_23k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/detail_23k.json) | 20.5 MB |
+| [complex_reasoning_77k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Instruct-150K/blob/main/complex_reasoning_77k.json) | 79.6 MB |
+
+### Pretraining Dataset
+The pretraining dataset used in this release is a subset of CC-3M dataset, filtered with a more balanced concept coverage distribution.  Please see [here](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K) for a detailed description of the dataset structure and how to download the images.
+
+If you already have CC-3M dataset on your disk, the image names follow this format: `GCC_train_000000000.jpg`.  You may edit the `image` field correspondingly if necessary.
+
+| Data | Chat File | Meta Data | Size |
+| --- |  --- |  --- | ---: |
+| CC-3M Concept-balanced 595K | [chat.json](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K/blob/main/chat.json) | [metadata.json](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K/blob/main/metadata.json) | 211 MB
+| LAION/CC/SBU BLIP-Caption Concept-balanced 558K | [blip_laion_cc_sbu_558k.json](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain/blob/main/blip_laion_cc_sbu_558k.json) | [metadata.json](#) | 181 MB
+
+**Important notice**: Upon the request from the community, as ~15% images of the original CC-3M dataset are no longer accessible, we upload [`images.zip`](https://huggingface.co/datasets/liuhaotian/LLaVA-CC3M-Pretrain-595K/blob/main/images.zip) for better reproducing our work in research community. It must not be used for any other purposes. The use of these images must comply with the CC-3M license. This may be taken down at any time when requested by the original CC-3M dataset owner or owners of the referenced images.
+
+### GPT-4 Prompts
+
+We provide our prompts and few-shot samples for GPT-4 queries, to better facilitate research in this domain.  Please check out the [`prompts`](https://github.com/haotian-liu/LLaVA/tree/main/playground/data/prompts) folder for three kinds of questions: conversation, detail description, and complex reasoning.
+
+They are organized in a format of `system_message.txt` for system message, pairs of `abc_caps.txt` for few-shot sample user input, and `abc_conv.txt` for few-shot sample reference output.
+
+Note that you may find them in different format. For example, `conversation` is in `jsonl`, and detail description is answer-only.  The selected format in our preliminary experiments works slightly better than a limited set of alternatives that we tried: `jsonl`, more natural format, answer-only.  If interested, you may try other variants or conduct more careful study in this.  Contributions are welcomed!

LLaVA/llava/eval/eval_gpt_review_bench.py

@@ -0,0 +1,121 @@
+import argparse
+import json
+import os
+
+import openai
+import time
+
+NUM_SECONDS_TO_SLEEP = 0.5
+
+
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            response = openai.ChatCompletion.create(
+                model='gpt-4-0314',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except openai.error.RateLimitError:
+            pass
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    return response['choices'][0]['message']['content']
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-c', '--context')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    if os.path.isfile(os.path.expanduser(args.output)):
+        cur_reviews = [json.loads(line) for line in open(os.path.expanduser(args.output))]
+    else:
+        cur_reviews = []
+
+    review_file = open(f'{args.output}', 'a')
+
+    context_list = [json.loads(line) for line in open(os.path.expanduser(args.context))]
+    image_to_context = {context['image']: context for context in context_list}
+
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+
+        inst = image_to_context[ques['image']]
+
+        if isinstance(inst['caption'], list):
+            cap_str = '\n'.join(inst['caption'])
+        else:
+            cap_str = inst['caption']
+
+        category = 'llava_bench_' + json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            assert False, f"Visual QA category not found in rule file: {category}."
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Context]\n{cap_str}\n\n'
+                   f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        cur_js = {
+            'id': idx+1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1.get('answer_id', ans1['question_id']),
+            'answer2_id': ans2.get('answer_id', ans2['answer_id']),
+            'category': category
+        }
+        if idx >= len(cur_reviews):
+            review = get_eval(content, args.max_tokens)
+            scores = parse_score(review)
+            cur_js['content'] = review
+            cur_js['tuple'] = scores
+            review_file.write(json.dumps(cur_js) + '\n')
+            review_file.flush()
+        else:
+            print(f'Skipping {idx} as we already have it.')
+        idx += 1
+        print(idx)
+    review_file.close()

LLaVA/llava/eval/eval_gpt_review_visual.py

@@ -0,0 +1,118 @@
+import argparse
+import json
+import os
+
+import openai
+import time
+
+NUM_SECONDS_TO_SLEEP = 0.5
+
+
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            response = openai.ChatCompletion.create(
+                model='gpt-4-0314',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except openai.error.RateLimitError:
+            pass
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    return response['choices'][0]['message']['content']
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-c', '--context')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    if os.path.isfile(os.path.expanduser(args.output)):
+        cur_reviews = [json.loads(line) for line in open(os.path.expanduser(args.output))]
+    else:
+        cur_reviews = []
+
+    review_file = open(f'{args.output}', 'a')
+
+    context_list = [json.loads(line) for line in open(os.path.expanduser(args.context))]
+    image_to_context = {context['image']: context for context in context_list}
+
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+
+        inst = image_to_context[ques['image']]
+        cap_str = '\n'.join(inst['captions'])
+        box_str = '\n'.join([f'{instance["category"]}: {instance["bbox"]}' for instance in inst['instances']])
+
+        category = json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            assert False, f"Visual QA category not found in rule file: {category}."
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Context]\n{cap_str}\n\n{box_str}\n\n'
+                   f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        cur_js = {
+            'id': idx+1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1.get('answer_id', ans1['question_id']),
+            'answer2_id': ans2.get('answer_id', ans2['answer_id']),
+            'category': category
+        }
+        if idx >= len(cur_reviews):
+            review = get_eval(content, args.max_tokens)
+            scores = parse_score(review)
+            cur_js['content'] = review
+            cur_js['tuple'] = scores
+            review_file.write(json.dumps(cur_js) + '\n')
+            review_file.flush()
+        else:
+            print(f'Skipping {idx} as we already have it.')
+        idx += 1
+        print(idx)
+    review_file.close()

LLaVA/llava/eval/eval_pope.py

@@ -0,0 +1,81 @@
+import os
+import json
+import argparse
+
+def eval_pope(answers, label_file):
+    label_list = [json.loads(q)['label'] for q in open(label_file, 'r')]
+
+    for answer in answers:
+        text = answer['text']
+
+        # Only keep the first sentence
+        if text.find('.') != -1:
+            text = text.split('.')[0]
+
+        text = text.replace(',', '')
+        words = text.split(' ')
+        if 'No' in words or 'not' in words or 'no' in words:
+            answer['text'] = 'no'
+        else:
+            answer['text'] = 'yes'
+
+    for i in range(len(label_list)):
+        if label_list[i] == 'no':
+            label_list[i] = 0
+        else:
+            label_list[i] = 1
+
+    pred_list = []
+    for answer in answers:
+        if answer['text'] == 'no':
+            pred_list.append(0)
+        else:
+            pred_list.append(1)
+
+    pos = 1
+    neg = 0
+    yes_ratio = pred_list.count(1) / len(pred_list)
+
+    TP, TN, FP, FN = 0, 0, 0, 0
+    for pred, label in zip(pred_list, label_list):
+        if pred == pos and label == pos:
+            TP += 1
+        elif pred == pos and label == neg:
+            FP += 1
+        elif pred == neg and label == neg:
+            TN += 1
+        elif pred == neg and label == pos:
+            FN += 1
+
+    print('TP\tFP\tTN\tFN\t')
+    print('{}\t{}\t{}\t{}'.format(TP, FP, TN, FN))
+
+    precision = float(TP) / float(TP + FP)
+    recall = float(TP) / float(TP + FN)
+    f1 = 2*precision*recall / (precision + recall)
+    acc = (TP + TN) / (TP + TN + FP + FN)
+    print('Accuracy: {}'.format(acc))
+    print('Precision: {}'.format(precision))
+    print('Recall: {}'.format(recall))
+    print('F1 score: {}'.format(f1))
+    print('Yes ratio: {}'.format(yes_ratio))
+    print('%.3f, %.3f, %.3f, %.3f, %.3f' % (f1, acc, precision, recall, yes_ratio) )
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--annotation-dir", type=str)
+    parser.add_argument("--question-file", type=str)
+    parser.add_argument("--result-file", type=str)
+    args = parser.parse_args()
+
+    questions = [json.loads(line) for line in open(args.question_file)]
+    questions = {question['question_id']: question for question in questions}
+    answers = [json.loads(q) for q in open(args.result_file)]
+    for file in os.listdir(args.annotation_dir):
+        assert file.startswith('coco_pope_')
+        assert file.endswith('.json')
+        category = file[10:-5]
+        cur_answers = [x for x in answers if questions[x['question_id']]['category'] == category]
+        print('Category: {}, # samples: {}'.format(category, len(cur_answers)))
+        eval_pope(cur_answers, os.path.join(args.annotation_dir, file))
+        print("====================================")

LLaVA/llava/eval/eval_science_qa_gpt4.py

@@ -0,0 +1,104 @@
+import argparse
+import json
+import os
+import re
+import random
+from collections import defaultdict
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--base-dir', type=str)
+    parser.add_argument('--gpt4-result', type=str)
+    parser.add_argument('--our-result', type=str)
+    parser.add_argument('--split', type=str, default='test')
+    parser.add_argument('--options', type=list, default=["A", "B", "C", "D", "E"])
+    return parser.parse_args()
+
+
+def convert_caps(results):
+    fakecaps = []
+    for result in results:
+        image_id = result['question_id']
+        caption = result['text']
+        fakecaps.append({"image_id": int(image_id), "caption": caption})
+    return fakecaps
+
+
+def get_pred_idx(prediction, choices, options):
+    """
+    Get the index (e.g. 2) from the prediction (e.g. 'C')
+    """
+    if prediction in options[:len(choices)]:
+        return options.index(prediction)
+    else:
+        return random.choice(range(len(choices)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    base_dir = args.base_dir
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[args.split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+    our_predictions = [json.loads(line) for line in open(args.our_result)]
+    our_predictions = {pred['question_id']: pred for pred in our_predictions}
+    split_problems = {idx: problems[idx] for idx in split_indices}
+
+    gpt4_predictions = json.load(open(args.gpt4_result))['outputs']
+
+    results = defaultdict(lambda: 0)
+
+    for prob_id, prob in split_problems.items():
+        if prob_id not in our_predictions:
+            continue
+        if prob_id not in gpt4_predictions:
+            continue
+        our_pred = our_predictions[prob_id]['text']
+        gpt4_pred = gpt4_predictions[prob_id]
+
+        pattern = re.compile(r'The answer is ([A-Z]).')
+        our_res = pattern.findall(our_pred)
+        if len(our_res) == 1:
+            our_answer = our_res[0]  # 'A', 'B', ...
+        else:
+            our_answer = "FAILED"
+        gpt4_res = pattern.findall(gpt4_pred)
+        if len(gpt4_res) == 1:
+            gpt4_answer = gpt4_res[0]  # 'A', 'B', ...
+        else:
+            gpt4_answer = "FAILED"
+
+        our_pred_idx = get_pred_idx(our_answer, prob['choices'], args.options)
+        gpt4_pred_idx = get_pred_idx(gpt4_answer, prob['choices'], args.options)
+
+        if gpt4_answer == 'FAILED':
+            results['gpt4_failed'] += 1
+            # continue
+            gpt4_pred_idx = our_pred_idx
+            # if our_pred_idx != prob['answer']:
+            #     print(our_predictions[prob_id]['prompt'])
+            #     print('-----------------')
+            #     print(f'LECTURE: {prob["lecture"]}')
+            #     print(f'SOLUTION: {prob["solution"]}')
+            #     print('=====================')
+        else:
+            # continue
+            pass
+        # gpt4_pred_idx = our_pred_idx
+
+        if gpt4_pred_idx == prob['answer']:
+            results['correct'] += 1
+        else:
+            results['incorrect'] += 1
+
+
+        if gpt4_pred_idx == prob['answer'] or our_pred_idx == prob['answer']:
+            results['correct_upperbound'] += 1
+
+    correct = results['correct']
+    total = results['correct'] + results['incorrect']
+    print(f'Total: {total}, Correct: {correct}, Accuracy: {correct / total * 100:.2f}%')
+    print(f'Total: {total}, Correct (upper): {results["correct_upperbound"]}, Accuracy: {results["correct_upperbound"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4 NO-ANS (RANDOM): {results["gpt4_failed"]}, Percentage: {results["gpt4_failed"] / total * 100:.2f}%')
+

LLaVA/llava/eval/eval_textvqa.py

@@ -0,0 +1,65 @@
+import os
+import argparse
+import json
+import re
+
+from llava.eval.m4c_evaluator import TextVQAAccuracyEvaluator
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--annotation-file', type=str)
+    parser.add_argument('--result-file', type=str)
+    parser.add_argument('--result-dir', type=str)
+    return parser.parse_args()
+
+
+def prompt_processor(prompt):
+    if prompt.startswith('OCR tokens: '):
+        pattern = r"Question: (.*?) Short answer:"
+        match = re.search(pattern, prompt, re.DOTALL)
+        question = match.group(1)
+    elif 'Reference OCR token: ' in prompt and len(prompt.split('\n')) == 3:
+        if prompt.startswith('Reference OCR token:'):
+            question = prompt.split('\n')[1]
+        else:
+            question = prompt.split('\n')[0]
+    elif len(prompt.split('\n')) == 2:
+        question = prompt.split('\n')[0]
+    else:
+        assert False
+
+    return question.lower()
+
+
+def eval_single(annotation_file, result_file):
+    experiment_name = os.path.splitext(os.path.basename(result_file))[0]
+    print(experiment_name)
+    annotations = json.load(open(annotation_file))['data']
+    annotations = {(annotation['image_id'], annotation['question'].lower()): annotation for annotation in annotations}
+    results = [json.loads(line) for line in open(result_file)]
+
+    pred_list = []
+    for result in results:
+        annotation = annotations[(result['question_id'], prompt_processor(result['prompt']))]
+        pred_list.append({
+            "pred_answer": result['text'],
+            "gt_answers": annotation['answers'],
+        })
+
+    evaluator = TextVQAAccuracyEvaluator()
+    print('Samples: {}\nAccuracy: {:.2f}%\n'.format(len(pred_list), 100. * evaluator.eval_pred_list(pred_list)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    if args.result_file is not None:
+        eval_single(args.annotation_file, args.result_file)
+
+    if args.result_dir is not None:
+        for result_file in sorted(os.listdir(args.result_dir)):
+            if not result_file.endswith('.jsonl'):
+                print(f'Skipping {result_file}')
+                continue
+            eval_single(args.annotation_file, os.path.join(args.result_dir, result_file))

LLaVA/llava/eval/generate_webpage_data_from_table.py

@@ -0,0 +1,111 @@
+"""Generate json file for webpage."""
+import json
+import os
+import re
+
+# models = ['llama', 'alpaca', 'gpt35', 'bard']
+models = ['vicuna']
+
+
+def read_jsonl(path: str, key: str=None):
+    data = []
+    with open(os.path.expanduser(path)) as f:
+        for line in f:
+            if not line:
+                continue
+            data.append(json.loads(line))
+    if key is not None:
+        data.sort(key=lambda x: x[key])
+        data = {item[key]: item for item in data}
+    return data
+
+
+def trim_hanging_lines(s: str, n: int) -> str:
+    s = s.strip()
+    for _ in range(n):
+        s = s.split('\n', 1)[1].strip()
+    return s
+
+
+if __name__ == '__main__':
+    questions = read_jsonl('table/question.jsonl', key='question_id')
+
+    # alpaca_answers = read_jsonl('table/answer/answer_alpaca-13b.jsonl', key='question_id')
+    # bard_answers = read_jsonl('table/answer/answer_bard.jsonl', key='question_id')
+    # gpt35_answers = read_jsonl('table/answer/answer_gpt35.jsonl', key='question_id')
+    # llama_answers = read_jsonl('table/answer/answer_llama-13b.jsonl', key='question_id')
+    vicuna_answers = read_jsonl('table/answer/answer_vicuna-13b.jsonl', key='question_id')
+    ours_answers = read_jsonl('table/results/llama-13b-hf-alpaca.jsonl', key='question_id')
+
+    review_vicuna = read_jsonl('table/review/review_vicuna-13b_llama-13b-hf-alpaca.jsonl', key='question_id')
+    # review_alpaca = read_jsonl('table/review/review_alpaca-13b_vicuna-13b.jsonl', key='question_id')
+    # review_bard = read_jsonl('table/review/review_bard_vicuna-13b.jsonl', key='question_id')
+    # review_gpt35 = read_jsonl('table/review/review_gpt35_vicuna-13b.jsonl', key='question_id')
+    # review_llama = read_jsonl('table/review/review_llama-13b_vicuna-13b.jsonl', key='question_id')
+
+    records = []
+    for qid in questions.keys():
+        r = {
+            'id': qid,
+            'category': questions[qid]['category'],
+            'question': questions[qid]['text'],
+            'answers': {
+                # 'alpaca': alpaca_answers[qid]['text'],
+                # 'llama': llama_answers[qid]['text'],
+                # 'bard': bard_answers[qid]['text'],
+                # 'gpt35': gpt35_answers[qid]['text'],
+                'vicuna': vicuna_answers[qid]['text'],
+                'ours': ours_answers[qid]['text'],
+            },
+            'evaluations': {
+                # 'alpaca': review_alpaca[qid]['text'],
+                # 'llama': review_llama[qid]['text'],
+                # 'bard': review_bard[qid]['text'],
+                'vicuna': review_vicuna[qid]['content'],
+                # 'gpt35': review_gpt35[qid]['text'],
+            },
+            'scores': {
+                'vicuna': review_vicuna[qid]['tuple'],
+                # 'alpaca': review_alpaca[qid]['score'],
+                # 'llama': review_llama[qid]['score'],
+                # 'bard': review_bard[qid]['score'],
+                # 'gpt35': review_gpt35[qid]['score'],
+            },
+        }
+
+        # cleanup data
+        cleaned_evals = {}
+        for k, v in r['evaluations'].items():
+            v = v.strip()
+            lines = v.split('\n')
+            # trim the first line if it's a pair of numbers
+            if re.match(r'\d+[, ]+\d+', lines[0]):
+                lines = lines[1:]
+            v = '\n'.join(lines)
+            cleaned_evals[k] = v.replace('Assistant 1', "**Assistant 1**").replace('Assistant 2', '**Assistant 2**')
+
+        r['evaluations'] = cleaned_evals
+        records.append(r)
+
+    # Reorder the records, this is optional
+    for r in records:
+        if r['id'] <= 20:
+            r['id'] += 60
+        else:
+            r['id'] -= 20
+    for r in records:
+        if r['id'] <= 50:
+            r['id'] += 10
+        elif 50 < r['id'] <= 60:
+            r['id'] -= 50
+    for r in records:
+        if r['id'] == 7:
+            r['id'] = 1
+        elif r['id'] < 7:
+            r['id'] += 1 
+
+    records.sort(key=lambda x: x['id'])
+
+    # Write to file
+    with open('webpage/data.json', 'w') as f:
+        json.dump({'questions': records, 'models': models}, f, indent=2)

LLaVA/llava/eval/m4c_evaluator.py

@@ -0,0 +1,334 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import re
+
+from tqdm import tqdm
+
+
+class EvalAIAnswerProcessor:
+    """
+    Processes an answer similar to Eval AI
+        copied from
+        https://github.com/facebookresearch/mmf/blob/c46b3b3391275b4181567db80943473a89ab98ab/pythia/tasks/processors.py#L897
+    """
+
+    CONTRACTIONS = {
+        "aint": "ain't",
+        "arent": "aren't",
+        "cant": "can't",
+        "couldve": "could've",
+        "couldnt": "couldn't",
+        "couldn'tve": "couldn't've",
+        "couldnt've": "couldn't've",
+        "didnt": "didn't",
+        "doesnt": "doesn't",
+        "dont": "don't",
+        "hadnt": "hadn't",
+        "hadnt've": "hadn't've",
+        "hadn'tve": "hadn't've",
+        "hasnt": "hasn't",
+        "havent": "haven't",
+        "hed": "he'd",
+        "hed've": "he'd've",
+        "he'dve": "he'd've",
+        "hes": "he's",
+        "howd": "how'd",
+        "howll": "how'll",
+        "hows": "how's",
+        "Id've": "I'd've",
+        "I'dve": "I'd've",
+        "Im": "I'm",
+        "Ive": "I've",
+        "isnt": "isn't",
+        "itd": "it'd",
+        "itd've": "it'd've",
+        "it'dve": "it'd've",
+        "itll": "it'll",
+        "let's": "let's",
+        "maam": "ma'am",
+        "mightnt": "mightn't",
+        "mightnt've": "mightn't've",
+        "mightn'tve": "mightn't've",
+        "mightve": "might've",
+        "mustnt": "mustn't",
+        "mustve": "must've",
+        "neednt": "needn't",
+        "notve": "not've",
+        "oclock": "o'clock",
+        "oughtnt": "oughtn't",
+        "ow's'at": "'ow's'at",
+        "'ows'at": "'ow's'at",
+        "'ow'sat": "'ow's'at",
+        "shant": "shan't",
+        "shed've": "she'd've",
+        "she'dve": "she'd've",
+        "she's": "she's",
+        "shouldve": "should've",
+        "shouldnt": "shouldn't",
+        "shouldnt've": "shouldn't've",
+        "shouldn'tve": "shouldn't've",
+        "somebody'd": "somebodyd",
+        "somebodyd've": "somebody'd've",
+        "somebody'dve": "somebody'd've",
+        "somebodyll": "somebody'll",
+        "somebodys": "somebody's",
+        "someoned": "someone'd",
+        "someoned've": "someone'd've",
+        "someone'dve": "someone'd've",
+        "someonell": "someone'll",
+        "someones": "someone's",
+        "somethingd": "something'd",
+        "somethingd've": "something'd've",
+        "something'dve": "something'd've",
+        "somethingll": "something'll",
+        "thats": "that's",
+        "thered": "there'd",
+        "thered've": "there'd've",
+        "there'dve": "there'd've",
+        "therere": "there're",
+        "theres": "there's",
+        "theyd": "they'd",
+        "theyd've": "they'd've",
+        "they'dve": "they'd've",
+        "theyll": "they'll",
+        "theyre": "they're",
+        "theyve": "they've",
+        "twas": "'twas",
+        "wasnt": "wasn't",
+        "wed've": "we'd've",
+        "we'dve": "we'd've",
+        "weve": "we've",
+        "werent": "weren't",
+        "whatll": "what'll",
+        "whatre": "what're",
+        "whats": "what's",
+        "whatve": "what've",
+        "whens": "when's",
+        "whered": "where'd",
+        "wheres": "where's",
+        "whereve": "where've",
+        "whod": "who'd",
+        "whod've": "who'd've",
+        "who'dve": "who'd've",
+        "wholl": "who'll",
+        "whos": "who's",
+        "whove": "who've",
+        "whyll": "why'll",
+        "whyre": "why're",
+        "whys": "why's",
+        "wont": "won't",
+        "wouldve": "would've",
+        "wouldnt": "wouldn't",
+        "wouldnt've": "wouldn't've",
+        "wouldn'tve": "wouldn't've",
+        "yall": "y'all",
+        "yall'll": "y'all'll",
+        "y'allll": "y'all'll",
+        "yall'd've": "y'all'd've",
+        "y'alld've": "y'all'd've",
+        "y'all'dve": "y'all'd've",
+        "youd": "you'd",
+        "youd've": "you'd've",
+        "you'dve": "you'd've",
+        "youll": "you'll",
+        "youre": "you're",
+        "youve": "you've",
+    }
+
+    NUMBER_MAP = {
+        "none": "0",
+        "zero": "0",
+        "one": "1",
+        "two": "2",
+        "three": "3",
+        "four": "4",
+        "five": "5",
+        "six": "6",
+        "seven": "7",
+        "eight": "8",
+        "nine": "9",
+        "ten": "10",
+    }
+    ARTICLES = ["a", "an", "the"]
+    PERIOD_STRIP = re.compile(r"(?!<=\d)(\.)(?!\d)")
+    COMMA_STRIP = re.compile(r"(?<=\d)(\,)+(?=\d)")
+    PUNCTUATIONS = [
+        ";",
+        r"/",
+        "[",
+        "]",
+        '"',
+        "{",
+        "}",
+        "(",
+        ")",
+        "=",
+        "+",
+        "\\",
+        "_",
+        "-",
+        ">",
+        "<",
+        "@",
+        "`",
+        ",",
+        "?",
+        "!",
+    ]
+
+    def __init__(self, *args, **kwargs):
+        pass
+
+    def word_tokenize(self, word):
+        word = word.lower()
+        word = word.replace(",", "").replace("?", "").replace("'s", " 's")
+        return word.strip()
+
+    def process_punctuation(self, in_text):
+        out_text = in_text
+        for p in self.PUNCTUATIONS:
+            if (p + " " in in_text or " " + p in in_text) or (
+                re.search(self.COMMA_STRIP, in_text) is not None
+            ):
+                out_text = out_text.replace(p, "")
+            else:
+                out_text = out_text.replace(p, " ")
+        out_text = self.PERIOD_STRIP.sub("", out_text, re.UNICODE)
+        return out_text
+
+    def process_digit_article(self, in_text):
+        out_text = []
+        temp_text = in_text.lower().split()
+        for word in temp_text:
+            word = self.NUMBER_MAP.setdefault(word, word)
+            if word not in self.ARTICLES:
+                out_text.append(word)
+            else:
+                pass
+        for word_id, word in enumerate(out_text):
+            if word in self.CONTRACTIONS:
+                out_text[word_id] = self.CONTRACTIONS[word]
+        out_text = " ".join(out_text)
+        return out_text
+
+    def __call__(self, item):
+        item = self.word_tokenize(item)
+        item = item.replace("\n", " ").replace("\t", " ").strip()
+        item = self.process_punctuation(item)
+        item = self.process_digit_article(item)
+        return item
+
+
+class TextVQAAccuracyEvaluator:
+    def __init__(self):
+        self.answer_processor = EvalAIAnswerProcessor()
+
+    def _compute_answer_scores(self, raw_answers):
+        """
+        compute the accuracy (soft score) of human answers
+        """
+        answers = [self.answer_processor(a) for a in raw_answers]
+        assert len(answers) == 10
+        gt_answers = list(enumerate(answers))
+        unique_answers = set(answers)
+        unique_answer_scores = {}
+
+        for unique_answer in unique_answers:
+            accs = []
+            for gt_answer in gt_answers:
+                other_answers = [item for item in gt_answers if item != gt_answer]
+                matching_answers = [
+                    item for item in other_answers if item[1] == unique_answer
+                ]
+                acc = min(1, float(len(matching_answers)) / 3)
+                accs.append(acc)
+            unique_answer_scores[unique_answer] = sum(accs) / len(accs)
+
+        return unique_answer_scores
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in tqdm(pred_list):
+            pred_answer = self.answer_processor(entry["pred_answer"])
+            unique_answer_scores = self._compute_answer_scores(entry["gt_answers"])
+            score = unique_answer_scores.get(pred_answer, 0.0)
+            pred_scores.append(score)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class STVQAAccuracyEvaluator:
+    def __init__(self):
+        self.answer_processor = EvalAIAnswerProcessor()
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in pred_list:
+            pred_answer = self.answer_processor(entry["pred_answer"])
+            gts = [self.answer_processor(a) for a in entry["gt_answers"]]
+            score = 1.0 if pred_answer in gts else 0.0
+            pred_scores.append(score)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class STVQAANLSEvaluator:
+    def __init__(self):
+        import editdistance  # install with `pip install editdistance`
+
+        self.get_edit_distance = editdistance.eval
+
+    def get_anls(self, s1, s2):
+        s1 = s1.lower().strip()
+        s2 = s2.lower().strip()
+        iou = 1 - self.get_edit_distance(s1, s2) / max(len(s1), len(s2))
+        anls = iou if iou >= 0.5 else 0.0
+        return anls
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in pred_list:
+            anls = max(
+                self.get_anls(entry["pred_answer"], gt) for gt in entry["gt_answers"]
+            )
+            pred_scores.append(anls)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class TextCapsBleu4Evaluator:
+    def __init__(self):
+        # The following script requires Java 1.8.0 and pycocotools installed.
+        # The pycocoevalcap can be installed with pip as
+        # pip install git+https://github.com/ronghanghu/coco-caption.git@python23
+        # Original pycocoevalcap code is at https://github.com/tylin/coco-caption
+        # but has no python3 support yet.
+        try:
+            from pycocoevalcap.bleu.bleu import Bleu
+            from pycocoevalcap.tokenizer.ptbtokenizer import PTBTokenizer
+        except ModuleNotFoundError:
+            print(
+                "Please install pycocoevalcap module using "
+                "pip install git+https://github.com/ronghanghu/coco-caption.git@python23"  # noqa
+            )
+            raise
+
+        self.tokenizer = PTBTokenizer()
+        self.scorer = Bleu(4)
+
+    def eval_pred_list(self, pred_list):
+        # Create reference and hypotheses captions.
+        gts = {}
+        res = {}
+        for idx, entry in enumerate(pred_list):
+            gts[idx] = [{"caption": a} for a in entry["gt_answers"]]
+            res[idx] = [{"caption": entry["pred_answer"]}]
+
+        gts = self.tokenizer.tokenize(gts)
+        res = self.tokenizer.tokenize(res)
+        score, _ = self.scorer.compute_score(gts, res)
+
+        bleu4 = score[3]  # score is (Bleu-1, Bleu-2, Bleu-3, Bleu-4)
+        return bleu4

LLaVA/llava/eval/model_vqa.py

@@ -0,0 +1,101 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava.conversation import conv_templates, SeparatorStyle
+from llava.model.builder import load_pretrained_model
+from llava.utils import disable_torch_init
+from llava.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
+
+from PIL import Image
+import math
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+    for line in tqdm(questions):
+        idx = line["question_id"]
+        image_file = line["image"]
+        qs = line["text"]
+        cur_prompt = qs
+        if model.config.mm_use_im_start_end:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+        image = Image.open(os.path.join(args.image_folder, image_file)).convert('RGB')
+        image_tensor = process_images([image], image_processor, model.config)[0]
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor.unsqueeze(0).half().cuda(),
+                image_sizes=[image.size],
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                top_p=args.top_p,
+                num_beams=args.num_beams,
+                # no_repeat_ngram_size=3,
+                max_new_tokens=1024,
+                use_cache=True)
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    args = parser.parse_args()
+
+    eval_model(args)

LLaVA/llava/eval/model_vqa_loader.py

@@ -0,0 +1,144 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava.conversation import conv_templates, SeparatorStyle
+from llava.model.builder import load_pretrained_model
+from llava.utils import disable_torch_init
+from llava.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
+from torch.utils.data import Dataset, DataLoader
+
+from PIL import Image
+import math
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+# Custom dataset class
+class CustomDataset(Dataset):
+    def __init__(self, questions, image_folder, tokenizer, image_processor, model_config):
+        self.questions = questions
+        self.image_folder = image_folder
+        self.tokenizer = tokenizer
+        self.image_processor = image_processor
+        self.model_config = model_config
+
+    def __getitem__(self, index):
+        line = self.questions[index]
+        image_file = line["image"]
+        qs = line["text"]
+        if self.model_config.mm_use_im_start_end:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        image = Image.open(os.path.join(self.image_folder, image_file)).convert('RGB')
+        image_tensor = process_images([image], self.image_processor, self.model_config)[0]
+
+        input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt')
+
+        return input_ids, image_tensor, image.size
+
+    def __len__(self):
+        return len(self.questions)
+
+
+def collate_fn(batch):
+    input_ids, image_tensors, image_sizes = zip(*batch)
+    input_ids = torch.stack(input_ids, dim=0)
+    image_tensors = torch.stack(image_tensors, dim=0)
+    return input_ids, image_tensors, image_sizes
+
+
+# DataLoader
+def create_data_loader(questions, image_folder, tokenizer, image_processor, model_config, batch_size=1, num_workers=4):
+    assert batch_size == 1, "batch_size must be 1"
+    dataset = CustomDataset(questions, image_folder, tokenizer, image_processor, model_config)
+    data_loader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False, collate_fn=collate_fn)
+    return data_loader
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+
+    if 'plain' in model_name and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode:
+        args.conv_mode = args.conv_mode + '_mmtag'
+        print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.')
+
+    data_loader = create_data_loader(questions, args.image_folder, tokenizer, image_processor, model.config)
+
+    for (input_ids, image_tensor, image_sizes), line in tqdm(zip(data_loader, questions), total=len(questions)):
+        idx = line["question_id"]
+        cur_prompt = line["text"]
+
+        input_ids = input_ids.to(device='cuda', non_blocking=True)
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor.to(dtype=torch.float16, device='cuda', non_blocking=True),
+                image_sizes=image_sizes,
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                top_p=args.top_p,
+                num_beams=args.num_beams,
+                max_new_tokens=args.max_new_tokens,
+                use_cache=True)
+
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        # ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    parser.add_argument("--max_new_tokens", type=int, default=128)
+    args = parser.parse_args()
+
+    eval_model(args)

LLaVA/llava/eval/model_vqa_mmbench.py

@@ -0,0 +1,160 @@
+import argparse
+import torch
+import os
+import json
+import pandas as pd
+from tqdm import tqdm
+import shortuuid
+
+from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava.conversation import conv_templates, SeparatorStyle
+from llava.model.builder import load_pretrained_model
+from llava.utils import disable_torch_init
+from llava.mm_utils import tokenizer_image_token, process_images, load_image_from_base64, get_model_name_from_path
+
+from PIL import Image
+import math
+
+
+all_options = ['A', 'B', 'C', 'D']
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def is_none(value):
+    if value is None:
+        return True
+    if type(value) is float and math.isnan(value):
+        return True
+    if type(value) is str and value.lower() == 'nan':
+        return True
+    if type(value) is str and value.lower() == 'none':
+        return True
+    return False
+
+def get_options(row, options):
+    parsed_options = []
+    for option in options:
+        option_value = row[option]
+        if is_none(option_value):
+            break
+        parsed_options.append(option_value)
+    return parsed_options
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = pd.read_table(os.path.expanduser(args.question_file))
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+
+    if 'plain' in model_name and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode:
+        args.conv_mode = args.conv_mode + '_mmtag'
+        print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.')
+
+    for index, row in tqdm(questions.iterrows(), total=len(questions)):
+        options = get_options(row, all_options)
+        cur_option_char = all_options[:len(options)]
+
+        if args.all_rounds:
+            num_rounds = len(options)
+        else:
+            num_rounds = 1
+
+        for round_idx in range(num_rounds):
+            idx = row['index']
+            question = row['question']
+            hint = row['hint']
+            image = load_image_from_base64(row['image'])
+            if not is_none(hint):
+                question = hint + '\n' + question
+            for option_char, option in zip(all_options[:len(options)], options):
+                question = question + '\n' + option_char + '. ' + option
+            qs = cur_prompt = question
+            if model.config.mm_use_im_start_end:
+                qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+            else:
+                qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+            if args.single_pred_prompt:
+                if args.lang == 'cn':
+                    qs = qs + '\n' + "请直接回答选项字母。"
+                else:
+                    qs = qs + '\n' + "Answer with the option's letter from the given choices directly."
+
+            conv = conv_templates[args.conv_mode].copy()
+            conv.append_message(conv.roles[0], qs)
+            conv.append_message(conv.roles[1], None)
+            prompt = conv.get_prompt()
+
+            input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+            image_tensor = process_images([image], image_processor, model.config)[0]
+
+            with torch.inference_mode():
+                output_ids = model.generate(
+                    input_ids,
+                    images=image_tensor.unsqueeze(0).half().cuda(),
+                    image_sizes=[image.size],
+                    do_sample=True if args.temperature > 0 else False,
+                    temperature=args.temperature,
+                    top_p=args.top_p,
+                    num_beams=args.num_beams,
+                    # no_repeat_ngram_size=3,
+                    max_new_tokens=1024,
+                    use_cache=True)
+
+            outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+
+            ans_id = shortuuid.uuid()
+            ans_file.write(json.dumps({"question_id": idx,
+                                    "round_id": round_idx,
+                                    "prompt": cur_prompt,
+                                    "text": outputs,
+                                    "options": options,
+                                    "option_char": cur_option_char,
+                                    "answer_id": ans_id,
+                                    "model_id": model_name,
+                                    "metadata": {}}) + "\n")
+            ans_file.flush()
+
+            # rotate options
+            options = options[1:] + options[:1]
+            cur_option_char = cur_option_char[1:] + cur_option_char[:1]
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    parser.add_argument("--all-rounds", action="store_true")
+    parser.add_argument("--single-pred-prompt", action="store_true")
+    parser.add_argument("--lang", type=str, default="en")
+    args = parser.parse_args()
+
+    eval_model(args)

LLaVA/llava/eval/summarize_gpt_review.py

@@ -0,0 +1,60 @@
+import json
+import os
+from collections import defaultdict
+
+import numpy as np
+
+import argparse
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-d', '--dir', default=None)
+    parser.add_argument('-v', '--version', default=None)
+    parser.add_argument('-s', '--select', nargs='*', default=None)
+    parser.add_argument('-f', '--files', nargs='*', default=[])
+    parser.add_argument('-i', '--ignore', nargs='*', default=[])
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+    args = parse_args()
+
+    if args.ignore is not None:
+        args.ignore = [int(x) for x in args.ignore]
+
+    if len(args.files) > 0:
+        review_files = args.files
+    else:
+        review_files = [x for x in os.listdir(args.dir) if x.endswith('.jsonl') and (x.startswith('gpt4_text') or x.startswith('reviews_') or x.startswith('review_') or 'review' in args.dir)]
+
+    for review_file in sorted(review_files):
+        config = os.path.basename(review_file).replace('gpt4_text_', '').replace('.jsonl', '')
+        if args.select is not None and any(x not in config for x in args.select):
+            continue
+        if '0613' in config:
+            version = '0613'
+        else:
+            version = '0314'
+        if args.version is not None and args.version != version:
+            continue
+        scores = defaultdict(list)
+        print(config)
+        with open(os.path.join(args.dir, review_file) if args.dir is not None else review_file) as f:
+            for review_str in f:
+                review = json.loads(review_str)
+                if review['question_id'] in args.ignore:
+                    continue
+                if 'category' in review:
+                    scores[review['category']].append(review['tuple'])
+                    scores['all'].append(review['tuple'])
+                else:
+                    if 'tuple' in review:
+                        scores['all'].append(review['tuple'])
+                    else:
+                        scores['all'].append(review['score'])
+        for k, v in sorted(scores.items()):
+            stats = np.asarray(v).mean(0).tolist()
+            stats = [round(x, 3) for x in stats]
+            # print(k, stats, round(stats[1]/stats[0]*100, 1))
+            print(k, round(stats[1]/stats[0]*100, 1), round(stats[0] * 10, 1), round(stats[1] * 10, 1))
+        print('=================================')

LLaVA/llava/eval/webpage/styles.css

@@ -0,0 +1,105 @@
+body {
+    font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
+    background-color: #f8f9fa;
+}
+
+.navbar-dark .navbar-nav .nav-link {
+    color: #f1cf68;
+    font-size: 1.1rem;
+    padding: 0.5rem 0.6rem;
+}
+
+.card-header {
+    font-weight: bold;
+}
+
+.card {
+    box-shadow: 0 4px 8px rgba(0, 0, 0, 0.1);
+    transition: 0.3s;
+}
+
+.card:hover {
+    box-shadow: 0 8px 16px rgba(0, 0, 0, 0.2);
+}
+
+button {
+    transition: background-color 0.3s;
+}
+
+button:hover {
+    background-color: #007bff;
+}
+
+@media (max-width: 767px) {
+    .form-row .form-group {
+        margin-bottom: 10px;
+    }
+}
+
+/* Extra styles */
+
+.expandable-card .card-text-container {
+    max-height: 200px;
+    overflow-y: hidden;
+    position: relative;
+}
+
+.expandable-card.expanded .card-text-container {
+    max-height: none;
+}
+
+.expand-btn {
+    position: relative;
+    display: none;
+    background-color: rgba(255, 255, 255, 0.8);
+    color: #510c75;
+    border-color: transparent;
+}
+
+.expand-btn:hover {
+    background-color: rgba(200, 200, 200, 0.8);
+    text-decoration: none;
+    border-color: transparent;
+    color: #510c75;
+}
+
+.expand-btn:focus {
+    outline: none;
+    text-decoration: none;
+}
+
+.expandable-card:not(.expanded) .card-text-container:after {
+    content: "";
+    position: absolute;
+    bottom: 0;
+    left: 0;
+    width: 100%;
+    height: 90px;
+    background: linear-gradient(rgba(255, 255, 255, 0.2), rgba(255, 255, 255, 1));
+}
+
+.expandable-card:not(.expanded) .expand-btn {
+    margin-top: -40px;
+}
+
+.card-body {
+    padding-bottom: 5px;
+}
+
+.vertical-flex-layout {
+    justify-content: center;
+    align-items: center;
+    height: 100%;
+    display: flex;
+    flex-direction: column;
+    gap: 5px;
+}
+
+.figure-img {
+    max-width: 100%;
+    height: auto;
+}
+
+.adjustable-font-size {
+    font-size: calc(0.5rem + 2vw);
+}

LLaVA/llava/model/__init__.py

@@ -0,0 +1,6 @@
+try:
+    from .language_model.llava_llama import LlavaLlamaForCausalLM, LlavaConfig
+    from .language_model.llava_mpt import LlavaMptForCausalLM, LlavaMptConfig
+    from .language_model.llava_mistral import LlavaMistralForCausalLM, LlavaMistralConfig
+except:
+    pass

LLaVA/llava/model/apply_delta.py

@@ -0,0 +1,48 @@
+"""
+Usage:
+python3 -m fastchat.model.apply_delta --base ~/model_weights/llama-7b --target ~/model_weights/vicuna-7b --delta lmsys/vicuna-7b-delta
+"""
+import argparse
+
+import torch
+from tqdm import tqdm
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from llava import LlavaLlamaForCausalLM
+
+
+def apply_delta(base_model_path, target_model_path, delta_path):
+    print("Loading base model")
+    base = AutoModelForCausalLM.from_pretrained(
+        base_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+
+    print("Loading delta")
+    delta = LlavaLlamaForCausalLM.from_pretrained(delta_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+    delta_tokenizer = AutoTokenizer.from_pretrained(delta_path)
+
+    print("Applying delta")
+    for name, param in tqdm(delta.state_dict().items(), desc="Applying delta"):
+        if name not in base.state_dict():
+            assert name in ['model.mm_projector.weight', 'model.mm_projector.bias'], f'{name} not in base model'
+            continue
+        if param.data.shape == base.state_dict()[name].shape:
+            param.data += base.state_dict()[name]
+        else:
+            assert name in ['model.embed_tokens.weight', 'lm_head.weight'], \
+                f'{name} dimension mismatch: {param.data.shape} vs {base.state_dict()[name].shape}'
+            bparam = base.state_dict()[name]
+            param.data[:bparam.shape[0], :bparam.shape[1]] += bparam
+
+    print("Saving target model")
+    delta.save_pretrained(target_model_path)
+    delta_tokenizer.save_pretrained(target_model_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-model-path", type=str, required=True)
+    parser.add_argument("--target-model-path", type=str, required=True)
+    parser.add_argument("--delta-path", type=str, required=True)
+
+    args = parser.parse_args()
+
+    apply_delta(args.base_model_path, args.target_model_path, args.delta_path)

LLaVA/llava/model/builder.py

@@ -0,0 +1,167 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+import os
+import warnings
+import shutil
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig
+import torch
+from llava.model import *
+from llava.constants import DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+
+def load_pretrained_model(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="auto", device="cuda", use_flash_attn=False, **kwargs):
+    kwargs = {"device_map": device_map, **kwargs}
+
+    if device != "cuda":
+        kwargs['device_map'] = {"": device}
+
+    if load_8bit:
+        kwargs['load_in_8bit'] = True
+    elif load_4bit:
+        kwargs['load_in_4bit'] = True
+        kwargs['quantization_config'] = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4'
+        )
+    else:
+        kwargs['torch_dtype'] = torch.float16
+
+    if use_flash_attn:
+        kwargs['attn_implementation'] = 'flash_attention_2'
+
+    if 'llava' in model_name.lower():
+        # Load LLaVA model
+        if 'lora' in model_name.lower() and model_base is None:
+            warnings.warn('There is `lora` in model name but no `model_base` is provided. If you are loading a LoRA model, please provide the `model_base` argument. Detailed instruction: https://github.com/haotian-liu/LLaVA#launch-a-model-worker-lora-weights-unmerged.')
+        if 'lora' in model_name.lower() and model_base is not None:
+            from llava.model.language_model.llava_llama import LlavaConfig
+            lora_cfg_pretrained = LlavaConfig.from_pretrained(model_path)
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            print('Loading LLaVA from base model...')
+            model = LlavaLlamaForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=lora_cfg_pretrained, **kwargs)
+            token_num, tokem_dim = model.lm_head.out_features, model.lm_head.in_features
+            if model.lm_head.weight.shape[0] != token_num:
+                model.lm_head.weight = torch.nn.Parameter(torch.empty(token_num, tokem_dim, device=model.device, dtype=model.dtype))
+                model.model.embed_tokens.weight = torch.nn.Parameter(torch.empty(token_num, tokem_dim, device=model.device, dtype=model.dtype))
+
+            print('Loading additional LLaVA weights...')
+            if os.path.exists(os.path.join(model_path, 'non_lora_trainables.bin')):
+                non_lora_trainables = torch.load(os.path.join(model_path, 'non_lora_trainables.bin'), map_location='cpu')
+            else:
+                # this is probably from HF Hub
+                from huggingface_hub import hf_hub_download
+                def load_from_hf(repo_id, filename, subfolder=None):
+                    cache_file = hf_hub_download(
+                        repo_id=repo_id,
+                        filename=filename,
+                        subfolder=subfolder)
+                    return torch.load(cache_file, map_location='cpu')
+                non_lora_trainables = load_from_hf(model_path, 'non_lora_trainables.bin')
+            non_lora_trainables = {(k[11:] if k.startswith('base_model.') else k): v for k, v in non_lora_trainables.items()}
+            if any(k.startswith('model.model.') for k in non_lora_trainables):
+                non_lora_trainables = {(k[6:] if k.startswith('model.') else k): v for k, v in non_lora_trainables.items()}
+            model.load_state_dict(non_lora_trainables, strict=False)
+
+            from peft import PeftModel
+            print('Loading LoRA weights...')
+            model = PeftModel.from_pretrained(model, model_path)
+            print('Merging LoRA weights...')
+            model = model.merge_and_unload()
+            print('Model is loaded...')
+        elif model_base is not None:
+            # this may be mm projector only
+            print('Loading LLaVA from base model...')
+            if 'mpt' in model_name.lower():
+                if not os.path.isfile(os.path.join(model_path, 'configuration_mpt.py')):
+                    shutil.copyfile(os.path.join(model_base, 'configuration_mpt.py'), os.path.join(model_path, 'configuration_mpt.py'))
+                tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=True)
+                cfg_pretrained = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+                model = LlavaMptForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+                cfg_pretrained = AutoConfig.from_pretrained(model_path)
+                model = LlavaLlamaForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs)
+
+            mm_projector_weights = torch.load(os.path.join(model_path, 'mm_projector.bin'), map_location='cpu')
+            mm_projector_weights = {k: v.to(torch.float16) for k, v in mm_projector_weights.items()}
+            model.load_state_dict(mm_projector_weights, strict=False)
+        else:
+            if 'mpt' in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
+                model = LlavaMptForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)
+            elif 'mistral' in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_path)
+                model = LlavaMistralForCausalLM.from_pretrained(
+                    model_path,
+                    low_cpu_mem_usage=True,
+                    **kwargs
+                )
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+                model = LlavaLlamaForCausalLM.from_pretrained(
+                    model_path,
+                    low_cpu_mem_usage=True,
+                    **kwargs
+                )
+    else:
+        # Load language model
+        if model_base is not None:
+            # PEFT model
+            from peft import PeftModel
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            model = AutoModelForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, **kwargs)
+            print(f"Loading LoRA weights from {model_path}")
+            model = PeftModel.from_pretrained(model, model_path)
+            print(f"Merging weights")
+            model = model.merge_and_unload()
+            print('Convert to FP16...')
+            model.to(torch.float16)
+        else:
+            use_fast = False
+            if 'mpt' in model_name.lower():
+                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
+                model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, trust_remote_code=True, **kwargs)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+                model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)
+
+    image_processor = None
+
+    if 'llava' in model_name.lower():
+        mm_use_im_start_end = getattr(model.config, "mm_use_im_start_end", False)
+        mm_use_im_patch_token = getattr(model.config, "mm_use_im_patch_token", True)
+        if mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+        if mm_use_im_start_end:
+            tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+        model.resize_token_embeddings(len(tokenizer))
+
+        vision_tower = model.get_vision_tower()
+        if not vision_tower.is_loaded:
+            vision_tower.load_model(device_map=device_map)
+        if device_map != 'auto':
+            vision_tower.to(device=device_map, dtype=torch.float16)
+        image_processor = vision_tower.image_processor
+
+    if hasattr(model.config, "max_sequence_length"):
+        context_len = model.config.max_sequence_length
+    else:
+        context_len = 2048
+
+    return tokenizer, model, image_processor, context_len

LLaVA/llava/model/consolidate.py

@@ -0,0 +1,29 @@
+"""
+Usage:
+python3 -m llava.model.consolidate --src ~/model_weights/llava-7b --dst ~/model_weights/llava-7b_consolidate
+"""
+import argparse
+
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from llava.model import *
+from llava.model.utils import auto_upgrade
+
+
+def consolidate_ckpt(src_path, dst_path):
+    print("Loading model")
+    auto_upgrade(src_path)
+    src_model = AutoModelForCausalLM.from_pretrained(src_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+    src_tokenizer = AutoTokenizer.from_pretrained(src_path, use_fast=False)
+    src_model.save_pretrained(dst_path)
+    src_tokenizer.save_pretrained(dst_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--src", type=str, required=True)
+    parser.add_argument("--dst", type=str, required=True)
+
+    args = parser.parse_args()
+
+    consolidate_ckpt(args.src, args.dst)

LLaVA/llava/model/llava_arch.py

@@ -0,0 +1,368 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from abc import ABC, abstractmethod
+
+import torch
+import torch.nn as nn
+
+from .multimodal_encoder.builder import build_vision_tower
+from .multimodal_projector.builder import build_vision_projector
+
+from llava.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+from llava.mm_utils import get_anyres_image_grid_shape
+
+
+class LlavaMetaModel:
+
+    def __init__(self, config):
+        super(LlavaMetaModel, self).__init__(config)
+
+        if hasattr(config, "mm_vision_tower"):
+            self.vision_tower = build_vision_tower(config, delay_load=True)
+            self.mm_projector = build_vision_projector(config)
+
+            if 'unpad' in getattr(config, 'mm_patch_merge_type', ''):
+                self.image_newline = nn.Parameter(
+                    torch.empty(config.hidden_size, dtype=self.dtype)
+                )
+
+    def get_vision_tower(self):
+        vision_tower = getattr(self, 'vision_tower', None)
+        if type(vision_tower) is list:
+            vision_tower = vision_tower[0]
+        return vision_tower
+
+    def initialize_vision_modules(self, model_args, fsdp=None):
+        vision_tower = model_args.vision_tower
+        mm_vision_select_layer = model_args.mm_vision_select_layer
+        mm_vision_select_feature = model_args.mm_vision_select_feature
+        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
+        mm_patch_merge_type = model_args.mm_patch_merge_type
+
+        self.config.mm_vision_tower = vision_tower
+
+        if self.get_vision_tower() is None:
+            vision_tower = build_vision_tower(model_args)
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.vision_tower = [vision_tower]
+            else:
+                self.vision_tower = vision_tower
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                vision_tower = self.vision_tower[0]
+            else:
+                vision_tower = self.vision_tower
+            vision_tower.load_model()
+
+        self.config.use_mm_proj = True
+        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
+        self.config.mm_hidden_size = vision_tower.hidden_size
+        self.config.mm_vision_select_layer = mm_vision_select_layer
+        self.config.mm_vision_select_feature = mm_vision_select_feature
+        self.config.mm_patch_merge_type = mm_patch_merge_type
+
+        if getattr(self, 'mm_projector', None) is None:
+            self.mm_projector = build_vision_projector(self.config)
+
+            if 'unpad' in mm_patch_merge_type:
+                embed_std = 1 / torch.sqrt(torch.tensor(self.config.hidden_size, dtype=self.dtype))
+                self.image_newline = nn.Parameter(
+                    torch.randn(self.config.hidden_size, dtype=self.dtype) * embed_std
+                )
+        else:
+            # In case it is frozen by LoRA
+            for p in self.mm_projector.parameters():
+                p.requires_grad = True
+
+        if pretrain_mm_mlp_adapter is not None:
+            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
+            def get_w(weights, keyword):
+                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}
+
+            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))
+
+
+def unpad_image(tensor, original_size):
+    """
+    Unpads a PyTorch tensor of a padded and resized image.
+
+    Args:
+    tensor (torch.Tensor): The image tensor, assumed to be in CxHxW format.
+    original_size (tuple): The original size of PIL image (width, height).
+
+    Returns:
+    torch.Tensor: The unpadded image tensor.
+    """
+    original_width, original_height = original_size
+    current_height, current_width = tensor.shape[1:]
+
+    original_aspect_ratio = original_width / original_height
+    current_aspect_ratio = current_width / current_height
+
+    if original_aspect_ratio > current_aspect_ratio:
+        scale_factor = current_width / original_width
+        new_height = int(original_height * scale_factor)
+        padding = (current_height - new_height) // 2
+        unpadded_tensor = tensor[:, padding:current_height - padding, :]
+    else:
+        scale_factor = current_height / original_height
+        new_width = int(original_width * scale_factor)
+        padding = (current_width - new_width) // 2
+        unpadded_tensor = tensor[:, :, padding:current_width - padding]
+
+    return unpadded_tensor
+
+
+class LlavaMetaForCausalLM(ABC):
+
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+    def encode_images(self, images):
+        image_features = self.get_model().get_vision_tower()(images)
+        image_features = self.get_model().mm_projector(image_features)
+        return image_features
+
+    def prepare_inputs_labels_for_multimodal(
+        self, input_ids, position_ids, attention_mask, past_key_values, labels,
+        images, image_sizes=None
+    ):
+        vision_tower = self.get_vision_tower()
+        if vision_tower is None or images is None or input_ids.shape[1] == 1:
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels
+
+        if type(images) is list or images.ndim == 5:
+            if type(images) is list:
+                images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images]
+            concat_images = torch.cat([image for image in images], dim=0)
+            image_features = self.encode_images(concat_images)
+            split_sizes = [image.shape[0] for image in images]
+            image_features = torch.split(image_features, split_sizes, dim=0)
+            mm_patch_merge_type = getattr(self.config, 'mm_patch_merge_type', 'flat')
+            image_aspect_ratio = getattr(self.config, 'image_aspect_ratio', 'square')
+            if mm_patch_merge_type == 'flat':
+                image_features = [x.flatten(0, 1) for x in image_features]
+            elif mm_patch_merge_type.startswith('spatial'):
+                new_image_features = []
+                for image_idx, image_feature in enumerate(image_features):
+                    if image_feature.shape[0] > 1:
+                        base_image_feature = image_feature[0]
+                        image_feature = image_feature[1:]
+                        height = width = self.get_vision_tower().num_patches_per_side
+                        assert height * width == base_image_feature.shape[0]
+                        if image_aspect_ratio == 'anyres':
+                            num_patch_width, num_patch_height = get_anyres_image_grid_shape(image_sizes[image_idx], self.config.image_grid_pinpoints, self.get_vision_tower().config.image_size)
+                            image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
+                        else:
+                            raise NotImplementedError
+                        if 'unpad' in mm_patch_merge_type:
+                            image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+                            image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+                            image_feature = unpad_image(image_feature, image_sizes[image_idx])
+                            image_feature = torch.cat((
+                                image_feature,
+                                self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
+                            ), dim=-1)
+                            image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+                        else:
+                            image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+                            image_feature = image_feature.flatten(0, 3)
+                        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+                    else:
+                        image_feature = image_feature[0]
+                        if 'unpad' in mm_patch_merge_type:
+                            image_feature = torch.cat((
+                                image_feature,
+                                self.model.image_newline[None].to(image_feature.device)
+                            ), dim=0)
+                    new_image_features.append(image_feature)
+                image_features = new_image_features
+            else:
+                raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}")
+        else:
+            image_features = self.encode_images(images)
+
+        # TODO: image start / end is not implemented here to support pretraining.
+        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
+            raise NotImplementedError
+
+        # Let's just add dummy tensors if they do not exist,
+        # it is a headache to deal with None all the time.
+        # But it is not ideal, and if you have a better idea,
+        # please open an issue / submit a PR, thanks.
+        _labels = labels
+        _position_ids = position_ids
+        _attention_mask = attention_mask
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+        else:
+            attention_mask = attention_mask.bool()
+        if position_ids is None:
+            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
+        if labels is None:
+            labels = torch.full_like(input_ids, IGNORE_INDEX)
+
+        # remove the padding using attention_mask -- FIXME
+        _input_ids = input_ids
+        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
+        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
+
+        new_input_embeds = []
+        new_labels = []
+        cur_image_idx = 0
+        for batch_idx, cur_input_ids in enumerate(input_ids):
+            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
+            if num_images == 0:
+                cur_image_features = image_features[cur_image_idx]
+                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
+                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
+                new_input_embeds.append(cur_input_embeds)
+                new_labels.append(labels[batch_idx])
+                cur_image_idx += 1
+                continue
+
+            image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
+            cur_input_ids_noim = []
+            cur_labels = labels[batch_idx]
+            cur_labels_noim = []
+            for i in range(len(image_token_indices) - 1):
+                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
+                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
+            split_sizes = [x.shape[0] for x in cur_labels_noim]
+            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
+            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
+            cur_new_input_embeds = []
+            cur_new_labels = []
+
+            for i in range(num_images + 1):
+                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
+                cur_new_labels.append(cur_labels_noim[i])
+                if i < num_images:
+                    cur_image_features = image_features[cur_image_idx]
+                    cur_image_idx += 1
+                    cur_new_input_embeds.append(cur_image_features)
+                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
+
+            cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]
+
+            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
+            cur_new_labels = torch.cat(cur_new_labels)
+
+            new_input_embeds.append(cur_new_input_embeds)
+            new_labels.append(cur_new_labels)
+
+        # Truncate sequences to max length as image embeddings can make the sequence longer
+        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
+        if tokenizer_model_max_length is not None:
+            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
+            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
+
+        # Combine them
+        max_len = max(x.shape[0] for x in new_input_embeds)
+        batch_size = len(new_input_embeds)
+
+        new_input_embeds_padded = []
+        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
+        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
+        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
+
+        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
+            cur_len = cur_new_embed.shape[0]
+            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
+                new_input_embeds_padded.append(torch.cat((
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
+                    cur_new_embed
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, -cur_len:] = cur_new_labels
+                    attention_mask[i, -cur_len:] = True
+                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+            else:
+                new_input_embeds_padded.append(torch.cat((
+                    cur_new_embed,
+                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
+                ), dim=0))
+                if cur_len > 0:
+                    new_labels_padded[i, :cur_len] = cur_new_labels
+                    attention_mask[i, :cur_len] = True
+                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
+
+        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
+
+        if _labels is None:
+            new_labels = None
+        else:
+            new_labels = new_labels_padded
+
+        if _attention_mask is None:
+            attention_mask = None
+        else:
+            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
+
+        if _position_ids is None:
+            position_ids = None
+
+        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
+
+    def initialize_vision_tokenizer(self, model_args, tokenizer):
+        if model_args.mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+        if model_args.mm_use_im_start_end:
+            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+            if num_new_tokens > 0:
+                input_embeddings = self.get_input_embeddings().weight.data
+                output_embeddings = self.get_output_embeddings().weight.data
+
+                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+
+                input_embeddings[-num_new_tokens:] = input_embeddings_avg
+                output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = True
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+
+            if model_args.pretrain_mm_mlp_adapter:
+                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
+                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
+                assert num_new_tokens == 2
+                if input_embeddings.shape == embed_tokens_weight.shape:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
+                elif embed_tokens_weight.shape[0] == num_new_tokens:
+                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
+                else:
+                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
+        elif model_args.mm_use_im_patch_token:
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = False
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False

LLaVA/llava/model/make_delta.py

@@ -0,0 +1,52 @@
+"""
+Usage:
+python3 -m llava.model.make_delta --base ~/model_weights/llama-7b --target ~/model_weights/llava-7b --delta ~/model_weights/llava-7b-delta --hub-repo-id liuhaotian/llava-7b-delta
+"""
+import argparse
+
+import torch
+from tqdm import tqdm
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from llava.model.utils import auto_upgrade
+
+
+def make_delta(base_model_path, target_model_path, delta_path, hub_repo_id):
+    print("Loading base model")
+    base = AutoModelForCausalLM.from_pretrained(
+        base_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+
+    print("Loading target model")
+    auto_upgrade(target_model_path)
+    target = AutoModelForCausalLM.from_pretrained(target_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+
+    print("Calculating delta")
+    for name, param in tqdm(target.state_dict().items(), desc="Calculating delta"):
+        if name not in base.state_dict():
+            assert name in ['model.mm_projector.weight', 'model.mm_projector.bias'], f'{name} not in base model'
+            continue
+        if param.data.shape == base.state_dict()[name].shape:
+            param.data -= base.state_dict()[name]
+        else:
+            assert name in ['model.embed_tokens.weight', 'lm_head.weight'], f'{name} dimension mismatch: {param.data.shape} vs {base.state_dict()[name].shape}'
+            bparam = base.state_dict()[name]
+            param.data[:bparam.shape[0], :bparam.shape[1]] -= bparam
+
+    print("Saving delta")
+    if hub_repo_id:
+        kwargs = {"push_to_hub": True, "repo_id": hub_repo_id}
+    else:
+        kwargs = {}
+    target.save_pretrained(delta_path, **kwargs)
+    target_tokenizer = AutoTokenizer.from_pretrained(target_model_path)
+    target_tokenizer.save_pretrained(delta_path, **kwargs)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-model-path", type=str, required=True)
+    parser.add_argument("--target-model-path", type=str, required=True)
+    parser.add_argument("--delta-path", type=str, required=True)
+    parser.add_argument("--hub-repo-id", type=str, default=None)
+    args = parser.parse_args()
+
+    make_delta(args.base_model_path, args.target_model_path, args.delta_path, args.hub_repo_id)

LLaVA/llava/model/utils.py

@@ -0,0 +1,20 @@
+from transformers import AutoConfig
+
+
+def auto_upgrade(config):
+    cfg = AutoConfig.from_pretrained(config)
+    if 'llava' in config and 'llava' not in cfg.model_type:
+        assert cfg.model_type == 'llama'
+        print("You are using newer LLaVA code base, while the checkpoint of v0 is from older code base.")
+        print("You must upgrade the checkpoint to the new code base (this can be done automatically).")
+        confirm = input("Please confirm that you want to upgrade the checkpoint. [Y/N]")
+        if confirm.lower() in ["y", "yes"]:
+            print("Upgrading checkpoint...")
+            assert len(cfg.architectures) == 1
+            setattr(cfg.__class__, "model_type", "llava")
+            cfg.architectures[0] = 'LlavaLlamaForCausalLM'
+            cfg.save_pretrained(config)
+            print("Checkpoint upgraded.")
+        else:
+            print("Checkpoint upgrade aborted.")
+            exit(1)

LLaVA/llava/serve/cli.py

@@ -0,0 +1,126 @@
+import argparse
+import torch
+
+from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava.conversation import conv_templates, SeparatorStyle
+from llava.model.builder import load_pretrained_model
+from llava.utils import disable_torch_init
+from llava.mm_utils import process_images, tokenizer_image_token, get_model_name_from_path
+
+from PIL import Image
+
+import requests
+from PIL import Image
+from io import BytesIO
+from transformers import TextStreamer
+
+
+def load_image(image_file):
+    if image_file.startswith('http://') or image_file.startswith('https://'):
+        response = requests.get(image_file)
+        image = Image.open(BytesIO(response.content)).convert('RGB')
+    else:
+        image = Image.open(image_file).convert('RGB')
+    return image
+
+
+def main(args):
+    # Model
+    disable_torch_init()
+
+    model_name = get_model_name_from_path(args.model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(args.model_path, args.model_base, model_name, args.load_8bit, args.load_4bit, device=args.device)
+
+    if "llama-2" in model_name.lower():
+        conv_mode = "llava_llama_2"
+    elif "mistral" in model_name.lower():
+        conv_mode = "mistral_instruct"
+    elif "v1.6-34b" in model_name.lower():
+        conv_mode = "chatml_direct"
+    elif "v1" in model_name.lower():
+        conv_mode = "llava_v1"
+    elif "mpt" in model_name.lower():
+        conv_mode = "mpt"
+    else:
+        conv_mode = "llava_v0"
+
+    if args.conv_mode is not None and conv_mode != args.conv_mode:
+        print('[WARNING] the auto inferred conversation mode is {}, while `--conv-mode` is {}, using {}'.format(conv_mode, args.conv_mode, args.conv_mode))
+    else:
+        args.conv_mode = conv_mode
+
+    conv = conv_templates[args.conv_mode].copy()
+    if "mpt" in model_name.lower():
+        roles = ('user', 'assistant')
+    else:
+        roles = conv.roles
+
+    image = load_image(args.image_file)
+    image_size = image.size
+    # Similar operation in model_worker.py
+    image_tensor = process_images([image], image_processor, model.config)
+    if type(image_tensor) is list:
+        image_tensor = [image.to(model.device, dtype=torch.float16) for image in image_tensor]
+    else:
+        image_tensor = image_tensor.to(model.device, dtype=torch.float16)
+
+    while True:
+        try:
+            inp = input(f"{roles[0]}: ")
+        except EOFError:
+            inp = ""
+        if not inp:
+            print("exit...")
+            break
+
+        print(f"{roles[1]}: ", end="")
+
+        if image is not None:
+            # first message
+            if model.config.mm_use_im_start_end:
+                inp = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + inp
+            else:
+                inp = DEFAULT_IMAGE_TOKEN + '\n' + inp
+            image = None
+        
+        conv.append_message(conv.roles[0], inp)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(model.device)
+        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+        keywords = [stop_str]
+        streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor,
+                image_sizes=[image_size],
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                max_new_tokens=args.max_new_tokens,
+                streamer=streamer,
+                use_cache=True)
+
+        outputs = tokenizer.decode(output_ids[0]).strip()
+        conv.messages[-1][-1] = outputs
+
+        if args.debug:
+            print("\n", {"prompt": prompt, "outputs": outputs}, "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-file", type=str, required=True)
+    parser.add_argument("--device", type=str, default="cuda")
+    parser.add_argument("--conv-mode", type=str, default=None)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--max-new-tokens", type=int, default=512)
+    parser.add_argument("--load-8bit", action="store_true")
+    parser.add_argument("--load-4bit", action="store_true")
+    parser.add_argument("--debug", action="store_true")
+    args = parser.parse_args()
+    main(args)

LLaVA/llava/serve/controller.py

@@ -0,0 +1,298 @@
+"""
+A controller manages distributed workers.
+It sends worker addresses to clients.
+"""
+import argparse
+import asyncio
+import dataclasses
+from enum import Enum, auto
+import json
+import logging
+import time
+from typing import List, Union
+import threading
+
+from fastapi import FastAPI, Request
+from fastapi.responses import StreamingResponse
+import numpy as np
+import requests
+import uvicorn
+
+from llava.constants import CONTROLLER_HEART_BEAT_EXPIRATION
+from llava.utils import build_logger, server_error_msg
+
+
+logger = build_logger("controller", "controller.log")
+
+
+class DispatchMethod(Enum):
+    LOTTERY = auto()
+    SHORTEST_QUEUE = auto()
+
+    @classmethod
+    def from_str(cls, name):
+        if name == "lottery":
+            return cls.LOTTERY
+        elif name == "shortest_queue":
+            return cls.SHORTEST_QUEUE
+        else:
+            raise ValueError(f"Invalid dispatch method")
+
+
+@dataclasses.dataclass
+class WorkerInfo:
+    model_names: List[str]
+    speed: int
+    queue_length: int
+    check_heart_beat: bool
+    last_heart_beat: str
+
+
+def heart_beat_controller(controller):
+    while True:
+        time.sleep(CONTROLLER_HEART_BEAT_EXPIRATION)
+        controller.remove_stable_workers_by_expiration()
+
+
+class Controller:
+    def __init__(self, dispatch_method: str):
+        # Dict[str -> WorkerInfo]
+        self.worker_info = {}
+        self.dispatch_method = DispatchMethod.from_str(dispatch_method)
+
+        self.heart_beat_thread = threading.Thread(
+            target=heart_beat_controller, args=(self,), daemon=True)
+        self.heart_beat_thread.start()
+
+        logger.info("Init controller")
+
+    def register_worker(self, worker_name: str, check_heart_beat: bool,
+                        worker_status: dict):
+        if worker_name not in self.worker_info:
+            logger.info(f"Register a new worker: {worker_name}")
+        else:
+            logger.info(f"Register an existing worker: {worker_name}")
+
+        if not worker_status:
+            worker_status = self.get_worker_status(worker_name)
+        if not worker_status:
+            return False
+
+        self.worker_info[worker_name] = WorkerInfo(
+            worker_status["model_names"], worker_status["speed"], worker_status["queue_length"],
+            check_heart_beat, time.time())
+
+        logger.info(f"Register done: {worker_name}, {worker_status}")
+        return True
+
+    def get_worker_status(self, worker_name: str):
+        try:
+            r = requests.post(worker_name + "/worker_get_status", timeout=5)
+        except requests.exceptions.RequestException as e:
+            logger.error(f"Get status fails: {worker_name}, {e}")
+            return None
+
+        if r.status_code != 200:
+            logger.error(f"Get status fails: {worker_name}, {r}")
+            return None
+
+        return r.json()
+
+    def remove_worker(self, worker_name: str):
+        del self.worker_info[worker_name]
+
+    def refresh_all_workers(self):
+        old_info = dict(self.worker_info)
+        self.worker_info = {}
+
+        for w_name, w_info in old_info.items():
+            if not self.register_worker(w_name, w_info.check_heart_beat, None):
+                logger.info(f"Remove stale worker: {w_name}")
+
+    def list_models(self):
+        model_names = set()
+
+        for w_name, w_info in self.worker_info.items():
+            model_names.update(w_info.model_names)
+
+        return list(model_names)
+
+    def get_worker_address(self, model_name: str):
+        if self.dispatch_method == DispatchMethod.LOTTERY:
+            worker_names = []
+            worker_speeds = []
+            for w_name, w_info in self.worker_info.items():
+                if model_name in w_info.model_names:
+                    worker_names.append(w_name)
+                    worker_speeds.append(w_info.speed)
+            worker_speeds = np.array(worker_speeds, dtype=np.float32)
+            norm = np.sum(worker_speeds)
+            if norm < 1e-4:
+                return ""
+            worker_speeds = worker_speeds / norm
+            if True:  # Directly return address
+                pt = np.random.choice(np.arange(len(worker_names)),
+                    p=worker_speeds)
+                worker_name = worker_names[pt]
+                return worker_name
+
+            # Check status before returning
+            while True:
+                pt = np.random.choice(np.arange(len(worker_names)),
+                    p=worker_speeds)
+                worker_name = worker_names[pt]
+
+                if self.get_worker_status(worker_name):
+                    break
+                else:
+                    self.remove_worker(worker_name)
+                    worker_speeds[pt] = 0
+                    norm = np.sum(worker_speeds)
+                    if norm < 1e-4:
+                        return ""
+                    worker_speeds = worker_speeds / norm
+                    continue
+            return worker_name
+        elif self.dispatch_method == DispatchMethod.SHORTEST_QUEUE:
+            worker_names = []
+            worker_qlen = []
+            for w_name, w_info in self.worker_info.items():
+                if model_name in w_info.model_names:
+                    worker_names.append(w_name)
+                    worker_qlen.append(w_info.queue_length / w_info.speed)
+            if len(worker_names) == 0:
+                return ""
+            min_index = np.argmin(worker_qlen)
+            w_name = worker_names[min_index]
+            self.worker_info[w_name].queue_length += 1
+            logger.info(f"names: {worker_names}, queue_lens: {worker_qlen}, ret: {w_name}")
+            return w_name
+        else:
+            raise ValueError(f"Invalid dispatch method: {self.dispatch_method}")
+
+    def receive_heart_beat(self, worker_name: str, queue_length: int):
+        if worker_name not in self.worker_info:
+            logger.info(f"Receive unknown heart beat. {worker_name}")
+            return False
+
+        self.worker_info[worker_name].queue_length = queue_length
+        self.worker_info[worker_name].last_heart_beat = time.time()
+        logger.info(f"Receive heart beat. {worker_name}")
+        return True
+
+    def remove_stable_workers_by_expiration(self):
+        expire = time.time() - CONTROLLER_HEART_BEAT_EXPIRATION
+        to_delete = []
+        for worker_name, w_info in self.worker_info.items():
+            if w_info.check_heart_beat and w_info.last_heart_beat < expire:
+                to_delete.append(worker_name)
+
+        for worker_name in to_delete:
+            self.remove_worker(worker_name)
+
+    def worker_api_generate_stream(self, params):
+        worker_addr = self.get_worker_address(params["model"])
+        if not worker_addr:
+            logger.info(f"no worker: {params['model']}")
+            ret = {
+                "text": server_error_msg,
+                "error_code": 2,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+
+        try:
+            response = requests.post(worker_addr + "/worker_generate_stream",
+                json=params, stream=True, timeout=5)
+            for chunk in response.iter_lines(decode_unicode=False, delimiter=b"\0"):
+                if chunk:
+                    yield chunk + b"\0"
+        except requests.exceptions.RequestException as e:
+            logger.info(f"worker timeout: {worker_addr}")
+            ret = {
+                "text": server_error_msg,
+                "error_code": 3,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+
+
+    # Let the controller act as a worker to achieve hierarchical
+    # management. This can be used to connect isolated sub networks.
+    def worker_api_get_status(self):
+        model_names = set()
+        speed = 0
+        queue_length = 0
+
+        for w_name in self.worker_info:
+            worker_status = self.get_worker_status(w_name)
+            if worker_status is not None:
+                model_names.update(worker_status["model_names"])
+                speed += worker_status["speed"]
+                queue_length += worker_status["queue_length"]
+
+        return {
+            "model_names": list(model_names),
+            "speed": speed,
+            "queue_length": queue_length,
+        }
+
+
+app = FastAPI()
+
+
+@app.post("/register_worker")
+async def register_worker(request: Request):
+    data = await request.json()
+    controller.register_worker(
+        data["worker_name"], data["check_heart_beat"],
+        data.get("worker_status", None))
+
+
+@app.post("/refresh_all_workers")
+async def refresh_all_workers():
+    models = controller.refresh_all_workers()
+
+
+@app.post("/list_models")
+async def list_models():
+    models = controller.list_models()
+    return {"models": models}
+
+
+@app.post("/get_worker_address")
+async def get_worker_address(request: Request):
+    data = await request.json()
+    addr = controller.get_worker_address(data["model"])
+    return {"address": addr}
+
+
+@app.post("/receive_heart_beat")
+async def receive_heart_beat(request: Request):
+    data = await request.json()
+    exist = controller.receive_heart_beat(
+        data["worker_name"], data["queue_length"])
+    return {"exist": exist}
+
+
+@app.post("/worker_generate_stream")
+async def worker_api_generate_stream(request: Request):
+    params = await request.json()
+    generator = controller.worker_api_generate_stream(params)
+    return StreamingResponse(generator)
+
+
+@app.post("/worker_get_status")
+async def worker_api_get_status(request: Request):
+    return controller.worker_api_get_status()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=21001)
+    parser.add_argument("--dispatch-method", type=str, choices=[
+        "lottery", "shortest_queue"], default="shortest_queue")
+    args = parser.parse_args()
+    logger.info(f"args: {args}")
+
+    controller = Controller(args.dispatch_method)
+    uvicorn.run(app, host=args.host, port=args.port, log_level="info")

LLaVA/llava/serve/gradio_web_server.py

@@ -0,0 +1,479 @@
+import argparse
+import datetime
+import json
+import os
+import time
+
+import gradio as gr
+import requests
+
+from llava.conversation import (default_conversation, conv_templates,
+                                   SeparatorStyle)
+from llava.constants import LOGDIR
+from llava.utils import (build_logger, server_error_msg,
+    violates_moderation, moderation_msg)
+import hashlib
+
+
+logger = build_logger("gradio_web_server", "gradio_web_server.log")
+
+headers = {"User-Agent": "LLaVA Client"}
+
+no_change_btn = gr.Button()
+enable_btn = gr.Button(interactive=True)
+disable_btn = gr.Button(interactive=False)
+
+priority = {
+    "vicuna-13b": "aaaaaaa",
+    "koala-13b": "aaaaaab",
+}
+
+
+def get_conv_log_filename():
+    t = datetime.datetime.now()
+    name = os.path.join(LOGDIR, f"{t.year}-{t.month:02d}-{t.day:02d}-conv.json")
+    return name
+
+
+def get_model_list():
+    ret = requests.post(args.controller_url + "/refresh_all_workers")
+    assert ret.status_code == 200
+    ret = requests.post(args.controller_url + "/list_models")
+    models = ret.json()["models"]
+    models.sort(key=lambda x: priority.get(x, x))
+    logger.info(f"Models: {models}")
+    return models
+
+
+get_window_url_params = """
+function() {
+    const params = new URLSearchParams(window.location.search);
+    url_params = Object.fromEntries(params);
+    console.log(url_params);
+    return url_params;
+    }
+"""
+
+
+def load_demo(url_params, request: gr.Request):
+    logger.info(f"load_demo. ip: {request.client.host}. params: {url_params}")
+
+    dropdown_update = gr.Dropdown(visible=True)
+    if "model" in url_params:
+        model = url_params["model"]
+        if model in models:
+            dropdown_update = gr.Dropdown(value=model, visible=True)
+
+    state = default_conversation.copy()
+    return state, dropdown_update
+
+
+def load_demo_refresh_model_list(request: gr.Request):
+    logger.info(f"load_demo. ip: {request.client.host}")
+    models = get_model_list()
+    state = default_conversation.copy()
+    dropdown_update = gr.Dropdown(
+        choices=models,
+        value=models[0] if len(models) > 0 else ""
+    )
+    return state, dropdown_update
+
+
+def vote_last_response(state, vote_type, model_selector, request: gr.Request):
+    with open(get_conv_log_filename(), "a") as fout:
+        data = {
+            "tstamp": round(time.time(), 4),
+            "type": vote_type,
+            "model": model_selector,
+            "state": state.dict(),
+            "ip": request.client.host,
+        }
+        fout.write(json.dumps(data) + "\n")
+
+
+def upvote_last_response(state, model_selector, request: gr.Request):
+    logger.info(f"upvote. ip: {request.client.host}")
+    vote_last_response(state, "upvote", model_selector, request)
+    return ("",) + (disable_btn,) * 3
+
+
+def downvote_last_response(state, model_selector, request: gr.Request):
+    logger.info(f"downvote. ip: {request.client.host}")
+    vote_last_response(state, "downvote", model_selector, request)
+    return ("",) + (disable_btn,) * 3
+
+
+def flag_last_response(state, model_selector, request: gr.Request):
+    logger.info(f"flag. ip: {request.client.host}")
+    vote_last_response(state, "flag", model_selector, request)
+    return ("",) + (disable_btn,) * 3
+
+
+def regenerate(state, image_process_mode, request: gr.Request):
+    logger.info(f"regenerate. ip: {request.client.host}")
+    state.messages[-1][-1] = None
+    prev_human_msg = state.messages[-2]
+    if type(prev_human_msg[1]) in (tuple, list):
+        prev_human_msg[1] = (*prev_human_msg[1][:2], image_process_mode)
+    state.skip_next = False
+    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5
+
+
+def clear_history(request: gr.Request):
+    logger.info(f"clear_history. ip: {request.client.host}")
+    state = default_conversation.copy()
+    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5
+
+
+def add_text(state, text, image, image_process_mode, request: gr.Request):
+    logger.info(f"add_text. ip: {request.client.host}. len: {len(text)}")
+    if len(text) <= 0 and image is None:
+        state.skip_next = True
+        return (state, state.to_gradio_chatbot(), "", None) + (no_change_btn,) * 5
+    if args.moderate:
+        flagged = violates_moderation(text)
+        if flagged:
+            state.skip_next = True
+            return (state, state.to_gradio_chatbot(), moderation_msg, None) + (
+                no_change_btn,) * 5
+
+    text = text[:1536]  # Hard cut-off
+    if image is not None:
+        text = text[:1200]  # Hard cut-off for images
+        if '<image>' not in text:
+            # text = '<Image><image></Image>' + text
+            text = text + '\n<image>'
+        text = (text, image, image_process_mode)
+        state = default_conversation.copy()
+    state.append_message(state.roles[0], text)
+    state.append_message(state.roles[1], None)
+    state.skip_next = False
+    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5
+
+
+def http_bot(state, model_selector, temperature, top_p, max_new_tokens, request: gr.Request):
+    logger.info(f"http_bot. ip: {request.client.host}")
+    start_tstamp = time.time()
+    model_name = model_selector
+
+    if state.skip_next:
+        # This generate call is skipped due to invalid inputs
+        yield (state, state.to_gradio_chatbot()) + (no_change_btn,) * 5
+        return
+
+    if len(state.messages) == state.offset + 2:
+        # First round of conversation
+        if "llava" in model_name.lower():
+            if 'llama-2' in model_name.lower():
+                template_name = "llava_llama_2"
+            elif "mistral" in model_name.lower() or "mixtral" in model_name.lower():
+                if 'orca' in model_name.lower():
+                    template_name = "mistral_orca"
+                elif 'hermes' in model_name.lower():
+                    template_name = "chatml_direct"
+                else:
+                    template_name = "mistral_instruct"
+            elif 'llava-v1.6-34b' in model_name.lower():
+                template_name = "chatml_direct"
+            elif "v1" in model_name.lower():
+                if 'mmtag' in model_name.lower():
+                    template_name = "v1_mmtag"
+                elif 'plain' in model_name.lower() and 'finetune' not in model_name.lower():
+                    template_name = "v1_mmtag"
+                else:
+                    template_name = "llava_v1"
+            elif "mpt" in model_name.lower():
+                template_name = "mpt"
+            else:
+                if 'mmtag' in model_name.lower():
+                    template_name = "v0_mmtag"
+                elif 'plain' in model_name.lower() and 'finetune' not in model_name.lower():
+                    template_name = "v0_mmtag"
+                else:
+                    template_name = "llava_v0"
+        elif "mpt" in model_name:
+            template_name = "mpt_text"
+        elif "llama-2" in model_name:
+            template_name = "llama_2"
+        else:
+            template_name = "vicuna_v1"
+        new_state = conv_templates[template_name].copy()
+        new_state.append_message(new_state.roles[0], state.messages[-2][1])
+        new_state.append_message(new_state.roles[1], None)
+        state = new_state
+
+    # Query worker address
+    controller_url = args.controller_url
+    ret = requests.post(controller_url + "/get_worker_address",
+            json={"model": model_name})
+    worker_addr = ret.json()["address"]
+    logger.info(f"model_name: {model_name}, worker_addr: {worker_addr}")
+
+    # No available worker
+    if worker_addr == "":
+        state.messages[-1][-1] = server_error_msg
+        yield (state, state.to_gradio_chatbot(), disable_btn, disable_btn, disable_btn, enable_btn, enable_btn)
+        return
+
+    # Construct prompt
+    prompt = state.get_prompt()
+
+    all_images = state.get_images(return_pil=True)
+    all_image_hash = [hashlib.md5(image.tobytes()).hexdigest() for image in all_images]
+    for image, hash in zip(all_images, all_image_hash):
+        t = datetime.datetime.now()
+        filename = os.path.join(LOGDIR, "serve_images", f"{t.year}-{t.month:02d}-{t.day:02d}", f"{hash}.jpg")
+        if not os.path.isfile(filename):
+            os.makedirs(os.path.dirname(filename), exist_ok=True)
+            image.save(filename)
+
+    # Make requests
+    pload = {
+        "model": model_name,
+        "prompt": prompt,
+        "temperature": float(temperature),
+        "top_p": float(top_p),
+        "max_new_tokens": min(int(max_new_tokens), 1536),
+        "stop": state.sep if state.sep_style in [SeparatorStyle.SINGLE, SeparatorStyle.MPT] else state.sep2,
+        "images": f'List of {len(state.get_images())} images: {all_image_hash}',
+    }
+    logger.info(f"==== request ====\n{pload}")
+
+    pload['images'] = state.get_images()
+
+    state.messages[-1][-1] = "▌"
+    yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 5
+
+    try:
+        # Stream output
+        response = requests.post(worker_addr + "/worker_generate_stream",
+            headers=headers, json=pload, stream=True, timeout=10)
+        for chunk in response.iter_lines(decode_unicode=False, delimiter=b"\0"):
+            if chunk:
+                data = json.loads(chunk.decode())
+                if data["error_code"] == 0:
+                    output = data["text"][len(prompt):].strip()
+                    state.messages[-1][-1] = output + "▌"
+                    yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 5
+                else:
+                    output = data["text"] + f" (error_code: {data['error_code']})"
+                    state.messages[-1][-1] = output
+                    yield (state, state.to_gradio_chatbot()) + (disable_btn, disable_btn, disable_btn, enable_btn, enable_btn)
+                    return
+                time.sleep(0.03)
+    except requests.exceptions.RequestException as e:
+        state.messages[-1][-1] = server_error_msg
+        yield (state, state.to_gradio_chatbot()) + (disable_btn, disable_btn, disable_btn, enable_btn, enable_btn)
+        return
+
+    state.messages[-1][-1] = state.messages[-1][-1][:-1]
+    yield (state, state.to_gradio_chatbot()) + (enable_btn,) * 5
+
+    finish_tstamp = time.time()
+    logger.info(f"{output}")
+
+    with open(get_conv_log_filename(), "a") as fout:
+        data = {
+            "tstamp": round(finish_tstamp, 4),
+            "type": "chat",
+            "model": model_name,
+            "start": round(start_tstamp, 4),
+            "finish": round(finish_tstamp, 4),
+            "state": state.dict(),
+            "images": all_image_hash,
+            "ip": request.client.host,
+        }
+        fout.write(json.dumps(data) + "\n")
+
+title_markdown = ("""
+# 🌋 LLaVA: Large Language and Vision Assistant
+[[Project Page](https://llava-vl.github.io)] [[Code](https://github.com/haotian-liu/LLaVA)] [[Model](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)] | 📚 [[LLaVA](https://arxiv.org/abs/2304.08485)] [[LLaVA-v1.5](https://arxiv.org/abs/2310.03744)] [[LLaVA-v1.6](https://llava-vl.github.io/blog/2024-01-30-llava-1-6/)]
+""")
+
+tos_markdown = ("""
+### Terms of use
+By using this service, users are required to agree to the following terms:
+The service is a research preview intended for non-commercial use only. It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes. The service may collect user dialogue data for future research.
+Please click the "Flag" button if you get any inappropriate answer! We will collect those to keep improving our moderator.
+For an optimal experience, please use desktop computers for this demo, as mobile devices may compromise its quality.
+""")
+
+
+learn_more_markdown = ("""
+### License
+The service is a research preview intended for non-commercial use only, subject to the model [License](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) of LLaMA, [Terms of Use](https://openai.com/policies/terms-of-use) of the data generated by OpenAI, and [Privacy Practices](https://chrome.google.com/webstore/detail/sharegpt-share-your-chatg/daiacboceoaocpibfodeljbdfacokfjb) of ShareGPT. Please contact us if you find any potential violation.
+""")
+
+block_css = """
+
+#buttons button {
+    min-width: min(120px,100%);
+}
+
+"""
+
+def build_demo(embed_mode, cur_dir=None, concurrency_count=10):
+    textbox = gr.Textbox(show_label=False, placeholder="Enter text and press ENTER", container=False)
+    with gr.Blocks(title="LLaVA", theme=gr.themes.Default(), css=block_css) as demo:
+        state = gr.State()
+
+        if not embed_mode:
+            gr.Markdown(title_markdown)
+
+        with gr.Row():
+            with gr.Column(scale=3):
+                with gr.Row(elem_id="model_selector_row"):
+                    model_selector = gr.Dropdown(
+                        choices=models,
+                        value=models[0] if len(models) > 0 else "",
+                        interactive=True,
+                        show_label=False,
+                        container=False)
+
+                imagebox = gr.Image(type="pil")
+                image_process_mode = gr.Radio(
+                    ["Crop", "Resize", "Pad", "Default"],
+                    value="Default",
+                    label="Preprocess for non-square image", visible=False)
+
+                if cur_dir is None:
+                    cur_dir = os.path.dirname(os.path.abspath(__file__))
+                gr.Examples(examples=[
+                    [f"{cur_dir}/examples/extreme_ironing.jpg", "What is unusual about this image?"],
+                    [f"{cur_dir}/examples/waterview.jpg", "What are the things I should be cautious about when I visit here?"],
+                ], inputs=[imagebox, textbox])
+
+                with gr.Accordion("Parameters", open=False) as parameter_row:
+                    temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.2, step=0.1, interactive=True, label="Temperature",)
+                    top_p = gr.Slider(minimum=0.0, maximum=1.0, value=0.7, step=0.1, interactive=True, label="Top P",)
+                    max_output_tokens = gr.Slider(minimum=0, maximum=1024, value=512, step=64, interactive=True, label="Max output tokens",)
+
+            with gr.Column(scale=8):
+                chatbot = gr.Chatbot(
+                    elem_id="chatbot",
+                    label="LLaVA Chatbot",
+                    height=650,
+                    layout="panel",
+                )
+                with gr.Row():
+                    with gr.Column(scale=8):
+                        textbox.render()
+                    with gr.Column(scale=1, min_width=50):
+                        submit_btn = gr.Button(value="Send", variant="primary")
+                with gr.Row(elem_id="buttons") as button_row:
+                    upvote_btn = gr.Button(value="👍  Upvote", interactive=False)
+                    downvote_btn = gr.Button(value="👎  Downvote", interactive=False)
+                    flag_btn = gr.Button(value="⚠️  Flag", interactive=False)
+                    #stop_btn = gr.Button(value="⏹️  Stop Generation", interactive=False)
+                    regenerate_btn = gr.Button(value="🔄  Regenerate", interactive=False)
+                    clear_btn = gr.Button(value="🗑️  Clear", interactive=False)
+
+        if not embed_mode:
+            gr.Markdown(tos_markdown)
+            gr.Markdown(learn_more_markdown)
+        url_params = gr.JSON(visible=False)
+
+        # Register listeners
+        btn_list = [upvote_btn, downvote_btn, flag_btn, regenerate_btn, clear_btn]
+        upvote_btn.click(
+            upvote_last_response,
+            [state, model_selector],
+            [textbox, upvote_btn, downvote_btn, flag_btn]
+        )
+        downvote_btn.click(
+            downvote_last_response,
+            [state, model_selector],
+            [textbox, upvote_btn, downvote_btn, flag_btn]
+        )
+        flag_btn.click(
+            flag_last_response,
+            [state, model_selector],
+            [textbox, upvote_btn, downvote_btn, flag_btn]
+        )
+
+        regenerate_btn.click(
+            regenerate,
+            [state, image_process_mode],
+            [state, chatbot, textbox, imagebox] + btn_list
+        ).then(
+            http_bot,
+            [state, model_selector, temperature, top_p, max_output_tokens],
+            [state, chatbot] + btn_list,
+            concurrency_limit=concurrency_count
+        )
+
+        clear_btn.click(
+            clear_history,
+            None,
+            [state, chatbot, textbox, imagebox] + btn_list,
+            queue=False
+        )
+
+        textbox.submit(
+            add_text,
+            [state, textbox, imagebox, image_process_mode],
+            [state, chatbot, textbox, imagebox] + btn_list,
+            queue=False
+        ).then(
+            http_bot,
+            [state, model_selector, temperature, top_p, max_output_tokens],
+            [state, chatbot] + btn_list,
+            concurrency_limit=concurrency_count
+        )
+
+        submit_btn.click(
+            add_text,
+            [state, textbox, imagebox, image_process_mode],
+            [state, chatbot, textbox, imagebox] + btn_list
+        ).then(
+            http_bot,
+            [state, model_selector, temperature, top_p, max_output_tokens],
+            [state, chatbot] + btn_list,
+            concurrency_limit=concurrency_count
+        )
+
+        if args.model_list_mode == "once":
+            demo.load(
+                load_demo,
+                [url_params],
+                [state, model_selector],
+                js=get_window_url_params
+            )
+        elif args.model_list_mode == "reload":
+            demo.load(
+                load_demo_refresh_model_list,
+                None,
+                [state, model_selector],
+                queue=False
+            )
+        else:
+            raise ValueError(f"Unknown model list mode: {args.model_list_mode}")
+
+    return demo
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="0.0.0.0")
+    parser.add_argument("--port", type=int)
+    parser.add_argument("--controller-url", type=str, default="http://localhost:21001")
+    parser.add_argument("--concurrency-count", type=int, default=16)
+    parser.add_argument("--model-list-mode", type=str, default="once",
+        choices=["once", "reload"])
+    parser.add_argument("--share", action="store_true")
+    parser.add_argument("--moderate", action="store_true")
+    parser.add_argument("--embed", action="store_true")
+    args = parser.parse_args()
+    logger.info(f"args: {args}")
+
+    models = get_model_list()
+
+    logger.info(args)
+    demo = build_demo(args.embed, concurrency_count=args.concurrency_count)
+    demo.queue(
+        api_open=False
+    ).launch(
+        server_name=args.host,
+        server_port=args.port,
+        share=args.share
+    )

LLaVA/llava/serve/model_worker.py

@@ -0,0 +1,288 @@
+"""
+A model worker executes the model.
+"""
+import argparse
+import asyncio
+import json
+import time
+import threading
+import uuid
+
+from fastapi import FastAPI, Request, BackgroundTasks
+from fastapi.responses import StreamingResponse
+import requests
+import torch
+import uvicorn
+from functools import partial
+
+from llava.constants import WORKER_HEART_BEAT_INTERVAL
+from llava.utils import (build_logger, server_error_msg,
+    pretty_print_semaphore)
+from llava.model.builder import load_pretrained_model
+from llava.mm_utils import process_images, load_image_from_base64, tokenizer_image_token
+from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from transformers import TextIteratorStreamer
+from threading import Thread
+
+
+GB = 1 << 30
+
+worker_id = str(uuid.uuid4())[:6]
+logger = build_logger("model_worker", f"model_worker_{worker_id}.log")
+global_counter = 0
+
+model_semaphore = None
+
+
+def heart_beat_worker(controller):
+
+    while True:
+        time.sleep(WORKER_HEART_BEAT_INTERVAL)
+        controller.send_heart_beat()
+
+
+class ModelWorker:
+    def __init__(self, controller_addr, worker_addr,
+                 worker_id, no_register,
+                 model_path, model_base, model_name,
+                 load_8bit, load_4bit, device, use_flash_attn=False):
+        self.controller_addr = controller_addr
+        self.worker_addr = worker_addr
+        self.worker_id = worker_id
+        if model_path.endswith("/"):
+            model_path = model_path[:-1]
+        if model_name is None:
+            model_paths = model_path.split("/")
+            if model_paths[-1].startswith('checkpoint-'):
+                self.model_name = model_paths[-2] + "_" + model_paths[-1]
+            else:
+                self.model_name = model_paths[-1]
+        else:
+            self.model_name = model_name
+
+        self.device = device
+        logger.info(f"Loading the model {self.model_name} on worker {worker_id} ...")
+        self.tokenizer, self.model, self.image_processor, self.context_len = load_pretrained_model(
+            model_path, model_base, self.model_name, load_8bit, load_4bit, device=self.device, use_flash_attn=use_flash_attn)
+        self.is_multimodal = 'llava' in self.model_name.lower()
+
+        if not no_register:
+            self.register_to_controller()
+            self.heart_beat_thread = threading.Thread(
+                target=heart_beat_worker, args=(self,), daemon=True)
+            self.heart_beat_thread.start()
+
+    def register_to_controller(self):
+        logger.info("Register to controller")
+
+        url = self.controller_addr + "/register_worker"
+        data = {
+            "worker_name": self.worker_addr,
+            "check_heart_beat": True,
+            "worker_status": self.get_status()
+        }
+        r = requests.post(url, json=data)
+        assert r.status_code == 200
+
+    def send_heart_beat(self):
+        logger.info(f"Send heart beat. Models: {[self.model_name]}. "
+                    f"Semaphore: {pretty_print_semaphore(model_semaphore)}. "
+                    f"global_counter: {global_counter}")
+
+        url = self.controller_addr + "/receive_heart_beat"
+
+        while True:
+            try:
+                ret = requests.post(url, json={
+                    "worker_name": self.worker_addr,
+                    "queue_length": self.get_queue_length()}, timeout=5)
+                exist = ret.json()["exist"]
+                break
+            except requests.exceptions.RequestException as e:
+                logger.error(f"heart beat error: {e}")
+            time.sleep(5)
+
+        if not exist:
+            self.register_to_controller()
+
+    def get_queue_length(self):
+        if model_semaphore is None:
+            return 0
+        else:
+            return args.limit_model_concurrency - model_semaphore._value + (len(
+                model_semaphore._waiters) if model_semaphore._waiters is not None else 0)
+
+    def get_status(self):
+        return {
+            "model_names": [self.model_name],
+            "speed": 1,
+            "queue_length": self.get_queue_length(),
+        }
+
+    @torch.inference_mode()
+    def generate_stream(self, params):
+        tokenizer, model, image_processor = self.tokenizer, self.model, self.image_processor
+
+        prompt = params["prompt"]
+        ori_prompt = prompt
+        images = params.get("images", None)
+        num_image_tokens = 0
+        if images is not None and len(images) > 0 and self.is_multimodal:
+            if len(images) > 0:
+                if len(images) != prompt.count(DEFAULT_IMAGE_TOKEN):
+                    raise ValueError("Number of images does not match number of <image> tokens in prompt")
+
+                images = [load_image_from_base64(image) for image in images]
+                image_sizes = [image.size for image in images]
+                images = process_images(images, image_processor, model.config)
+
+                if type(images) is list:
+                    images = [image.to(self.model.device, dtype=torch.float16) for image in images]
+                else:
+                    images = images.to(self.model.device, dtype=torch.float16)
+
+                replace_token = DEFAULT_IMAGE_TOKEN
+                if getattr(self.model.config, 'mm_use_im_start_end', False):
+                    replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
+                prompt = prompt.replace(DEFAULT_IMAGE_TOKEN, replace_token)
+
+                num_image_tokens = prompt.count(replace_token) * model.get_vision_tower().num_patches
+            else:
+                images = None
+                image_sizes = None
+            image_args = {"images": images, "image_sizes": image_sizes}
+        else:
+            images = None
+            image_args = {}
+
+        temperature = float(params.get("temperature", 1.0))
+        top_p = float(params.get("top_p", 1.0))
+        max_context_length = getattr(model.config, 'max_position_embeddings', 2048)
+        max_new_tokens = min(int(params.get("max_new_tokens", 256)), 1024)
+        stop_str = params.get("stop", None)
+        do_sample = True if temperature > 0.001 else False
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(self.device)
+        keywords = [stop_str]
+        # stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
+        streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=15)
+
+        max_new_tokens = min(max_new_tokens, max_context_length - input_ids.shape[-1] - num_image_tokens)
+
+        if max_new_tokens < 1:
+            yield json.dumps({"text": ori_prompt + "Exceeds max token length. Please start a new conversation, thanks.", "error_code": 0}).encode() + b"\0"
+            return
+
+        thread = Thread(target=model.generate, kwargs=dict(
+            inputs=input_ids,
+            do_sample=do_sample,
+            temperature=temperature,
+            top_p=top_p,
+            max_new_tokens=max_new_tokens,
+            streamer=streamer,
+            use_cache=True,
+            **image_args
+        ))
+        thread.start()
+
+        generated_text = ori_prompt
+        for new_text in streamer:
+            generated_text += new_text
+            if generated_text.endswith(stop_str):
+                generated_text = generated_text[:-len(stop_str)]
+            yield json.dumps({"text": generated_text, "error_code": 0}).encode() + b"\0"
+
+    def generate_stream_gate(self, params):
+        try:
+            for x in self.generate_stream(params):
+                yield x
+        except ValueError as e:
+            print("Caught ValueError:", e)
+            ret = {
+                "text": server_error_msg,
+                "error_code": 1,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+        except torch.cuda.CudaError as e:
+            print("Caught torch.cuda.CudaError:", e)
+            ret = {
+                "text": server_error_msg,
+                "error_code": 1,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+        except Exception as e:
+            print("Caught Unknown Error", e)
+            ret = {
+                "text": server_error_msg,
+                "error_code": 1,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+
+
+app = FastAPI()
+
+
+def release_model_semaphore(fn=None):
+    model_semaphore.release()
+    if fn is not None:
+        fn()
+
+
+@app.post("/worker_generate_stream")
+async def generate_stream(request: Request):
+    global model_semaphore, global_counter
+    global_counter += 1
+    params = await request.json()
+
+    if model_semaphore is None:
+        model_semaphore = asyncio.Semaphore(args.limit_model_concurrency)
+    await model_semaphore.acquire()
+    worker.send_heart_beat()
+    generator = worker.generate_stream_gate(params)
+    background_tasks = BackgroundTasks()
+    background_tasks.add_task(partial(release_model_semaphore, fn=worker.send_heart_beat))
+    return StreamingResponse(generator, background=background_tasks)
+
+
+@app.post("/worker_get_status")
+async def get_status(request: Request):
+    return worker.get_status()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=21002)
+    parser.add_argument("--worker-address", type=str,
+        default="http://localhost:21002")
+    parser.add_argument("--controller-address", type=str,
+        default="http://localhost:21001")
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--model-name", type=str)
+    parser.add_argument("--device", type=str, default="cuda")
+    parser.add_argument("--multi-modal", action="store_true", help="Multimodal mode is automatically detected with model name, please make sure `llava` is included in the model path.")
+    parser.add_argument("--limit-model-concurrency", type=int, default=5)
+    parser.add_argument("--stream-interval", type=int, default=1)
+    parser.add_argument("--no-register", action="store_true")
+    parser.add_argument("--load-8bit", action="store_true")
+    parser.add_argument("--load-4bit", action="store_true")
+    parser.add_argument("--use-flash-attn", action="store_true")
+    args = parser.parse_args()
+    logger.info(f"args: {args}")
+
+    if args.multi_modal:
+        logger.warning("Multimodal mode is automatically detected with model name, please make sure `llava` is included in the model path.")
+
+    worker = ModelWorker(args.controller_address,
+                         args.worker_address,
+                         worker_id,
+                         args.no_register,
+                         args.model_path,
+                         args.model_base,
+                         args.model_name,
+                         args.load_8bit,
+                         args.load_4bit,
+                         args.device,
+                         use_flash_attn=args.use_flash_attn)
+    uvicorn.run(app, host=args.host, port=args.port, log_level="info")

LLaVA/llava/serve/register_worker.py

@@ -0,0 +1,26 @@
+"""
+Manually register workers.
+
+Usage:
+python3 -m fastchat.serve.register_worker --controller http://localhost:21001 --worker-name http://localhost:21002
+"""
+
+import argparse
+
+import requests
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--controller-address", type=str)
+    parser.add_argument("--worker-name", type=str)
+    parser.add_argument("--check-heart-beat", action="store_true")
+    args = parser.parse_args()
+
+    url = args.controller_address + "/register_worker"
+    data = {
+        "worker_name": args.worker_name,
+        "check_heart_beat": args.check_heart_beat,
+        "worker_status": None,
+    }
+    r = requests.post(url, json=data)
+    assert r.status_code == 200

LLaVA/llava/serve/sglang_worker.py

@@ -0,0 +1,244 @@
+"""
+A model worker executes the model.
+"""
+import argparse
+import asyncio
+from concurrent.futures import ThreadPoolExecutor
+import json
+import time
+import threading
+import uuid
+
+from fastapi import FastAPI, Request, BackgroundTasks
+from fastapi.responses import StreamingResponse
+import requests
+import re
+import uvicorn
+from functools import partial
+
+from llava.constants import WORKER_HEART_BEAT_INTERVAL
+from llava.utils import (build_logger, server_error_msg,
+    pretty_print_semaphore)
+from llava.mm_utils import process_images, load_image_from_base64, tokenizer_image_token, expand2square
+from llava.constants import DEFAULT_IMAGE_TOKEN
+
+import sglang as sgl
+from sglang.backend.runtime_endpoint import RuntimeEndpoint
+
+
+GB = 1 << 30
+
+worker_id = str(uuid.uuid4())[:6]
+logger = build_logger("model_worker", f"model_worker_{worker_id}.log")
+global_counter = 0
+
+model_semaphore = None
+
+
+def heart_beat_worker(controller):
+    while True:
+        time.sleep(WORKER_HEART_BEAT_INTERVAL)
+        controller.send_heart_beat()
+
+
+@sgl.function
+def pipeline(s, prompt, max_tokens):
+    for p in prompt:
+        if type(p) is str:
+            s += p
+        else:
+            s += sgl.image(p)
+    s += sgl.gen("response", max_tokens=max_tokens)
+
+
+class ModelWorker:
+    def __init__(self, controller_addr, worker_addr, sgl_endpoint,
+                 worker_id, no_register, model_name):
+        self.controller_addr = controller_addr
+        self.worker_addr = worker_addr
+        self.worker_id = worker_id
+
+        # Select backend
+        backend = RuntimeEndpoint(sgl_endpoint)
+        sgl.set_default_backend(backend)
+        model_path = backend.model_info["model_path"]
+
+        if model_path.endswith("/"):
+            model_path = model_path[:-1]
+        if model_name is None:
+            model_paths = model_path.split("/")
+            if model_paths[-1].startswith('checkpoint-'):
+                self.model_name = model_paths[-2] + "_" + model_paths[-1]
+            else:
+                self.model_name = model_paths[-1]
+        else:
+            self.model_name = model_name
+
+        logger.info(f"Loading the SGLANG model {self.model_name} on worker {worker_id} ...")
+
+        if not no_register:
+            self.register_to_controller()
+            self.heart_beat_thread = threading.Thread(
+                target=heart_beat_worker, args=(self,), daemon=True)
+            self.heart_beat_thread.start()
+
+    def register_to_controller(self):
+        logger.info("Register to controller")
+
+        url = self.controller_addr + "/register_worker"
+        data = {
+            "worker_name": self.worker_addr,
+            "check_heart_beat": True,
+            "worker_status": self.get_status()
+        }
+        r = requests.post(url, json=data)
+        assert r.status_code == 200
+
+    def send_heart_beat(self):
+        logger.info(f"Send heart beat. Models: {[self.model_name]}. "
+                    f"Semaphore: {pretty_print_semaphore(model_semaphore)}. "
+                    f"global_counter: {global_counter}")
+
+        url = self.controller_addr + "/receive_heart_beat"
+
+        while True:
+            try:
+                ret = requests.post(url, json={
+                    "worker_name": self.worker_addr,
+                    "queue_length": self.get_queue_length()}, timeout=5)
+                exist = ret.json()["exist"]
+                break
+            except requests.exceptions.RequestException as e:
+                logger.error(f"heart beat error: {e}")
+            time.sleep(5)
+
+        if not exist:
+            self.register_to_controller()
+
+    def get_queue_length(self):
+        if model_semaphore is None:
+            return 0
+        else:
+            return args.limit_model_concurrency - model_semaphore._value + (len(
+                model_semaphore._waiters) if model_semaphore._waiters is not None else 0)
+
+    def get_status(self):
+        return {
+            "model_names": [self.model_name],
+            "speed": 1,
+            "queue_length": self.get_queue_length(),
+        }
+
+    async def generate_stream(self, params):
+        ori_prompt = prompt = params["prompt"]
+        images = params.get("images", None)
+        if images is not None and len(images) > 0:
+            if len(images) > 0:
+                if len(images) != prompt.count(DEFAULT_IMAGE_TOKEN):
+                    raise ValueError("Number of images does not match number of <image> tokens in prompt")
+
+                images = [load_image_from_base64(image) for image in images]
+
+                # FIXME: for image-start/end token
+                # replace_token = DEFAULT_IMAGE_TOKEN
+                # if getattr(self.model.config, 'mm_use_im_start_end', False):
+                #     replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
+                # prompt = prompt.replace(DEFAULT_IMAGE_TOKEN, replace_token)
+                prompt = prompt.replace(' ' + DEFAULT_IMAGE_TOKEN + '\n', DEFAULT_IMAGE_TOKEN)
+                prompt_split = prompt.split(DEFAULT_IMAGE_TOKEN)
+                prompt = []
+                for i in range(len(prompt_split)):
+                    prompt.append(prompt_split[i])
+                    if i < len(images):
+                        prompt.append(images[i])
+        else:
+            prompt = [prompt]
+
+        temperature = float(params.get("temperature", 1.0))
+        top_p = float(params.get("top_p", 1.0))
+        # max_context_length = getattr(model.config, 'max_position_embeddings', 2048)
+        max_new_tokens = min(int(params.get("max_new_tokens", 256)), 1024)
+        stop_str = params.get("stop", None)
+        stop_str = [stop_str] if stop_str is not None else None
+
+        print({'prompt': prompt, 'max_new_tokens': max_new_tokens, 'temperature': temperature, 'top_p': top_p})
+        state = pipeline.run(prompt, max_new_tokens, temperature=temperature, top_p=top_p, stream=True)
+
+        generated_text = ori_prompt
+        async for text_outputs in state.text_async_iter(var_name="response"):
+            generated_text += text_outputs
+            yield json.dumps({"text": generated_text, "error_code": 0}).encode() + b"\0"
+
+    async def generate_stream_gate(self, params):
+        try:
+            async for x in self.generate_stream(params):
+                yield x
+        except ValueError as e:
+            print("Caught ValueError:", e)
+            ret = {
+                "text": server_error_msg,
+                "error_code": 1,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+        except Exception as e:
+            print("Caught Unknown Error", e)
+            ret = {
+                "text": server_error_msg,
+                "error_code": 1,
+            }
+            yield json.dumps(ret).encode() + b"\0"
+
+
+app = FastAPI()
+
+
+def release_model_semaphore(fn=None):
+    model_semaphore.release()
+    if fn is not None:
+        fn()
+
+
+@app.post("/worker_generate_stream")
+async def generate_stream(request: Request):
+    global model_semaphore, global_counter
+    global_counter += 1
+    params = await request.json()
+
+    if model_semaphore is None:
+        model_semaphore = asyncio.Semaphore(args.limit_model_concurrency)
+    await model_semaphore.acquire()
+    worker.send_heart_beat()
+    generator = worker.generate_stream_gate(params)
+    background_tasks = BackgroundTasks()
+    background_tasks.add_task(partial(release_model_semaphore, fn=worker.send_heart_beat))
+    return StreamingResponse(generator, background=background_tasks)
+
+
+@app.post("/worker_get_status")
+async def get_status(request: Request):
+    return worker.get_status()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=21002)
+    parser.add_argument("--worker-address", type=str,
+        default="http://localhost:21002")
+    parser.add_argument("--controller-address", type=str,
+        default="http://localhost:21001")
+    parser.add_argument("--model-name", type=str)
+    parser.add_argument("--sgl-endpoint", type=str)
+    parser.add_argument("--limit-model-concurrency", type=int, default=5)
+    parser.add_argument("--stream-interval", type=int, default=1)
+    parser.add_argument("--no-register", action="store_true")
+    args = parser.parse_args()
+    logger.info(f"args: {args}")
+
+    worker = ModelWorker(args.controller_address,
+                         args.worker_address,
+                         args.sgl_endpoint,
+                         worker_id,
+                         args.no_register,
+                         args.model_name)
+    uvicorn.run(app, host=args.host, port=args.port, log_level="info")

LLaVA/llava/serve/test_message.py

@@ -0,0 +1,62 @@
+import argparse
+import json
+
+import requests
+
+from llava.conversation import default_conversation
+
+
+def main():
+    if args.worker_address:
+        worker_addr = args.worker_address
+    else:
+        controller_addr = args.controller_address
+        ret = requests.post(controller_addr + "/refresh_all_workers")
+        ret = requests.post(controller_addr + "/list_models")
+        models = ret.json()["models"]
+        models.sort()
+        print(f"Models: {models}")
+
+        ret = requests.post(controller_addr + "/get_worker_address",
+            json={"model": args.model_name})
+        worker_addr = ret.json()["address"]
+        print(f"worker_addr: {worker_addr}")
+
+    if worker_addr == "":
+        return
+
+    conv = default_conversation.copy()
+    conv.append_message(conv.roles[0], args.message)
+    prompt = conv.get_prompt()
+
+    headers = {"User-Agent": "LLaVA Client"}
+    pload = {
+        "model": args.model_name,
+        "prompt": prompt,
+        "max_new_tokens": args.max_new_tokens,
+        "temperature": 0.7,
+        "stop": conv.sep,
+    }
+    response = requests.post(worker_addr + "/worker_generate_stream", headers=headers,
+            json=pload, stream=True)
+
+    print(prompt.replace(conv.sep, "\n"), end="")
+    for chunk in response.iter_lines(chunk_size=8192, decode_unicode=False, delimiter=b"\0"):
+        if chunk:
+            data = json.loads(chunk.decode("utf-8"))
+            output = data["text"].split(conv.sep)[-1]
+            print(output, end="\r")
+    print("")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--controller-address", type=str, default="http://localhost:21001")
+    parser.add_argument("--worker-address", type=str)
+    parser.add_argument("--model-name", type=str, default="facebook/opt-350m")
+    parser.add_argument("--max-new-tokens", type=int, default=32)
+    parser.add_argument("--message", type=str, default=
+        "Tell me a story with more than 1000 words.")
+    args = parser.parse_args()
+
+    main()

LLaVA/llava/train/llama_flash_attn_monkey_patch.py

@@ -0,0 +1,115 @@
+from typing import Optional, Tuple
+import warnings
+
+import torch
+
+import transformers
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
+
+try:
+    from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func
+except ImportError:
+    from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func as flash_attn_unpadded_qkvpacked_func
+from flash_attn.bert_padding import unpad_input, pad_input
+
+
+def forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.Tensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    if output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
+
+    bsz, q_len, _ = hidden_states.size()
+
+    query_states = (
+        self.q_proj(hidden_states)
+        .view(bsz, q_len, self.num_heads, self.head_dim)
+        .transpose(1, 2)
+    )
+    key_states = (
+        self.k_proj(hidden_states)
+        .view(bsz, q_len, self.num_key_value_heads, self.head_dim)
+        .transpose(1, 2)
+    )
+    value_states = (
+        self.v_proj(hidden_states)
+        .view(bsz, q_len, self.num_key_value_heads, self.head_dim)
+        .transpose(1, 2)
+    )  # shape: (b, num_heads, s, head_dim)
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
+
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    query_states, key_states = apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, position_ids
+    )
+
+    if past_key_value is not None:
+        # reuse k, v
+        key_states = torch.cat([past_key_value[0], key_states], dim=2)
+        value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    # repeat k/v heads if n_kv_heads < n_heads
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    # Transform the data into the format required by flash attention
+    qkv = torch.stack([query_states, key_states, value_states], dim=2)
+    qkv = qkv.transpose(1, 3)  # shape: [b, s, 3, num_heads, head_dim]
+    key_padding_mask = attention_mask
+
+    if key_padding_mask is None:
+        qkv = qkv.reshape(-1, 3, self.num_heads, self.head_dim)
+        cu_q_lens = torch.arange(
+            0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=qkv.device
+        )
+        max_s = q_len
+        output = flash_attn_unpadded_qkvpacked_func(
+            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
+        )
+        output = output.view(bsz, q_len, -1)
+    else:
+        qkv = qkv.reshape(bsz, q_len, -1)
+        qkv, indices, cu_q_lens, max_s = unpad_input(qkv, key_padding_mask)
+        qkv = qkv.view(-1, 3, self.num_heads, self.head_dim)
+        output_unpad = flash_attn_unpadded_qkvpacked_func(
+            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
+        )
+        output_unpad = output_unpad.reshape(-1, self.num_heads * self.head_dim)
+        output = pad_input(output_unpad, indices, bsz, q_len)
+
+    return self.o_proj(output), None, past_key_value
+
+
+# Disable the transformation of the attention mask in LlamaModel as the flash attention
+# requires the attention mask to be the same as the key_padding_mask
+def _prepare_decoder_attention_mask(
+    self, attention_mask, input_shape, inputs_embeds, past_key_values_length
+):
+    # [bsz, seq_len]
+    return attention_mask
+
+
+def replace_llama_attn_with_flash_attn():
+    cuda_major, cuda_minor = torch.cuda.get_device_capability()
+    if cuda_major < 8:
+        warnings.warn(
+            "Flash attention is only supported on A100 or H100 GPU during training due to head dim > 64 backward."
+            "ref: https://github.com/HazyResearch/flash-attention/issues/190#issuecomment-1523359593"
+        )
+    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (
+        _prepare_decoder_attention_mask
+    )
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = forward

LLaVA/llava/train/llama_xformers_attn_monkey_patch.py

@@ -0,0 +1,129 @@
+"""
+Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments
+"""
+
+import logging
+import math
+from typing import Optional, Tuple
+
+import torch
+import transformers.models.llama.modeling_llama
+from torch import nn
+
+try:
+    import xformers.ops
+except ImportError:
+    logging.error("xformers not found! Please install it before trying to use it.")
+
+
+def replace_llama_attn_with_xformers_attn():
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward
+
+
+def xformers_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    # pylint: disable=duplicate-code
+    bsz, q_len, _ = hidden_states.size()
+
+    query_states = (
+        self.q_proj(hidden_states)
+        .view(bsz, q_len, self.num_heads, self.head_dim)
+        .transpose(1, 2)
+    )
+    key_states = (
+        self.k_proj(hidden_states)
+        .view(bsz, q_len, self.num_heads, self.head_dim)
+        .transpose(1, 2)
+    )
+    value_states = (
+        self.v_proj(hidden_states)
+        .view(bsz, q_len, self.num_heads, self.head_dim)
+        .transpose(1, 2)
+    )
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    (
+        query_states,
+        key_states,
+    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, position_ids
+    )
+    # [bsz, nh, t, hd]
+
+    if past_key_value is not None:
+        # reuse k, v, self_attention
+        key_states = torch.cat([past_key_value[0], key_states], dim=2)
+        value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    # We only apply xformers optimizations if we don't need to output the whole attention matrix
+    if not output_attentions:
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
+        # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
+        if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
+            # input and output should be of form (bsz, q_len, num_heads, head_dim)
+            attn_output = xformers.ops.memory_efficient_attention(
+                query_states, key_states, value_states, attn_bias=None
+            )
+        else:
+            # input and output should be of form (bsz, q_len, num_heads, head_dim)
+            attn_output = xformers.ops.memory_efficient_attention(
+                query_states,
+                key_states,
+                value_states,
+                attn_bias=xformers.ops.LowerTriangularMask(),
+            )
+        attn_weights = None
+    else:
+        attn_weights = torch.matmul(
+            query_states, key_states.transpose(2, 3)
+        ) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(
+                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
+            )
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(
+            attn_weights, dim=-1, dtype=torch.float32
+        ).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+    attn_output = self.o_proj(attn_output)
+    return attn_output, attn_weights, past_key_value

LLaVA/llava/train/llava_trainer.py

@@ -0,0 +1,255 @@
+import os
+import torch
+import torch.nn as nn
+
+from torch.utils.data import Sampler
+
+from transformers import Trainer
+from transformers.trainer import (
+    is_sagemaker_mp_enabled,
+    get_parameter_names,
+    has_length,
+    ALL_LAYERNORM_LAYERS,
+    logger,
+)
+from typing import List, Optional
+
+
+def maybe_zero_3(param, ignore_status=False, name=None):
+    from deepspeed import zero
+    from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+    if hasattr(param, "ds_id"):
+        if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
+            if not ignore_status:
+                print(name, 'no ignore status')
+        with zero.GatheredParameters([param]):
+            param = param.data.detach().cpu().clone()
+    else:
+        param = param.detach().cpu().clone()
+    return param
+
+
+def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
+    to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
+    to_return = {k: maybe_zero_3(v, ignore_status=True, name=k).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def split_to_even_chunks(indices, lengths, num_chunks):
+    """
+    Split a list of indices into `chunks` chunks of roughly equal lengths.
+    """
+
+    if len(indices) % num_chunks != 0:
+        return [indices[i::num_chunks] for i in range(num_chunks)]
+
+    num_indices_per_chunk = len(indices) // num_chunks
+
+    chunks = [[] for _ in range(num_chunks)]
+    chunks_lengths = [0 for _ in range(num_chunks)]
+    for index in indices:
+        shortest_chunk = chunks_lengths.index(min(chunks_lengths))
+        chunks[shortest_chunk].append(index)
+        chunks_lengths[shortest_chunk] += lengths[index]
+        if len(chunks[shortest_chunk]) == num_indices_per_chunk:
+            chunks_lengths[shortest_chunk] = float("inf")
+
+    return chunks
+
+
+def get_modality_length_grouped_indices(lengths, batch_size, world_size, generator=None):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    assert all(l != 0 for l in lengths), "Should not have zero length."
+    if all(l > 0 for l in lengths) or all(l < 0 for l in lengths):
+        # all samples are in the same modality
+        return get_length_grouped_indices(lengths, batch_size, world_size, generator=generator)
+    mm_indices, mm_lengths = zip(*[(i, l) for i, l in enumerate(lengths) if l > 0])
+    lang_indices, lang_lengths = zip(*[(i, -l) for i, l in enumerate(lengths) if l < 0])
+
+    mm_shuffle = [mm_indices[i] for i in get_length_grouped_indices(mm_lengths, batch_size, world_size, generator=None)]
+    lang_shuffle = [lang_indices[i] for i in get_length_grouped_indices(lang_lengths, batch_size, world_size, generator=None)]
+    megabatch_size = world_size * batch_size
+    mm_megabatches = [mm_shuffle[i : i + megabatch_size] for i in range(0, len(mm_shuffle), megabatch_size)]
+    lang_megabatches = [lang_shuffle[i : i + megabatch_size] for i in range(0, len(lang_shuffle), megabatch_size)]
+
+    last_mm = mm_megabatches[-1]
+    last_lang = lang_megabatches[-1]
+    additional_batch = last_mm + last_lang
+    megabatches = mm_megabatches[:-1] + lang_megabatches[:-1]
+    megabatch_indices = torch.randperm(len(megabatches), generator=generator)
+    megabatches = [megabatches[i] for i in megabatch_indices]
+
+    if len(additional_batch) > 0:
+        megabatches.append(sorted(additional_batch))
+
+    return [i for megabatch in megabatches for i in megabatch]
+
+
+def get_length_grouped_indices(lengths, batch_size, world_size, generator=None, merge=True):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    indices = torch.randperm(len(lengths), generator=generator)
+    megabatch_size = world_size * batch_size
+    megabatches = [indices[i : i + megabatch_size].tolist() for i in range(0, len(lengths), megabatch_size)]
+    megabatches = [sorted(megabatch, key=lambda i: lengths[i], reverse=True) for megabatch in megabatches]
+    megabatches = [split_to_even_chunks(megabatch, lengths, world_size) for megabatch in megabatches]
+
+    return [i for megabatch in megabatches for batch in megabatch for i in batch]
+
+
+class LengthGroupedSampler(Sampler):
+    r"""
+    Sampler that samples indices in a way that groups together features of the dataset of roughly the same length while
+    keeping a bit of randomness.
+    """
+
+    def __init__(
+        self,
+        batch_size: int,
+        world_size: int,
+        lengths: Optional[List[int]] = None,
+        generator=None,
+        group_by_modality: bool = False,
+    ):
+        if lengths is None:
+            raise ValueError("Lengths must be provided.")
+
+        self.batch_size = batch_size
+        self.world_size = world_size
+        self.lengths = lengths
+        self.generator = generator
+        self.group_by_modality = group_by_modality
+
+    def __len__(self):
+        return len(self.lengths)
+
+    def __iter__(self):
+        if self.group_by_modality:
+            indices = get_modality_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        else:
+            indices = get_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        return iter(indices)
+
+
+class LLaVATrainer(Trainer):
+
+    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
+        if self.train_dataset is None or not has_length(self.train_dataset):
+            return None
+
+        if self.args.group_by_modality_length:
+            lengths = self.train_dataset.modality_lengths
+            return LengthGroupedSampler(
+                self.args.train_batch_size,
+                world_size=self.args.world_size * self.args.gradient_accumulation_steps,
+                lengths=lengths,
+                group_by_modality=True,
+            )
+        else:
+            return super()._get_train_sampler()
+
+    def create_optimizer(self):
+        """
+        Setup the optimizer.
+
+        We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the
+        Trainer's init through `optimizers`, or subclass and override this method in a subclass.
+        """
+        if is_sagemaker_mp_enabled():
+            return super().create_optimizer()
+
+        opt_model = self.model
+
+        if self.optimizer is None:
+            decay_parameters = get_parameter_names(opt_model, ALL_LAYERNORM_LAYERS)
+            decay_parameters = [name for name in decay_parameters if "bias" not in name]
+            if self.args.mm_projector_lr is not None:
+                projector_parameters = [name for name, _ in opt_model.named_parameters() if "mm_projector" in name]
+                optimizer_grouped_parameters = [
+                    {
+                        "params": [
+                            p for n, p in opt_model.named_parameters() if (n in decay_parameters and n not in projector_parameters and p.requires_grad)
+                        ],
+                        "weight_decay": self.args.weight_decay,
+                    },
+                    {
+                        "params": [
+                            p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n not in projector_parameters and p.requires_grad)
+                        ],
+                        "weight_decay": 0.0,
+                    },
+                    {
+                        "params": [
+                            p for n, p in opt_model.named_parameters() if (n in decay_parameters and n in projector_parameters and p.requires_grad)
+                        ],
+                        "weight_decay": self.args.weight_decay,
+                        "lr": self.args.mm_projector_lr,
+                    },
+                    {
+                        "params": [
+                            p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n in projector_parameters and p.requires_grad)
+                        ],
+                        "weight_decay": 0.0,
+                        "lr": self.args.mm_projector_lr,
+                    },
+                ]
+            else:
+                optimizer_grouped_parameters = [
+                    {
+                        "params": [
+                            p for n, p in opt_model.named_parameters() if (n in decay_parameters and p.requires_grad)
+                        ],
+                        "weight_decay": self.args.weight_decay,
+                    },
+                    {
+                        "params": [
+                            p for n, p in opt_model.named_parameters() if (n not in decay_parameters and p.requires_grad)
+                        ],
+                        "weight_decay": 0.0,
+                    },
+                ]
+
+            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args)
+
+            self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+            if optimizer_cls.__name__ == "Adam8bit":
+                import bitsandbytes
+
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+
+                skipped = 0
+                for module in opt_model.modules():
+                    if isinstance(module, nn.Embedding):
+                        skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
+                        logger.info(f"skipped {module}: {skipped/2**20}M params")
+                        manager.register_module_override(module, "weight", {"optim_bits": 32})
+                        logger.debug(f"bitsandbytes: will optimize {module} in fp32")
+                logger.info(f"skipped: {skipped/2**20}M params")
+
+        return self.optimizer
+
+    def _save_checkpoint(self, model, trial, metrics=None):
+        if getattr(self.args, 'tune_mm_mlp_adapter', False):
+            from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+            checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
+
+            run_dir = self._get_output_dir(trial=trial)
+            output_dir = os.path.join(run_dir, checkpoint_folder)
+
+            # Only save Adapter
+            keys_to_match = ['mm_projector', 'vision_resampler']
+            if getattr(self.args, "use_im_start_end", False):
+                keys_to_match.extend(['embed_tokens', 'embed_in'])
+
+            weight_to_save = get_mm_adapter_state_maybe_zero_3(self.model.named_parameters(), keys_to_match)
+
+            if self.args.local_rank == 0 or self.args.local_rank == -1:
+                self.model.config.save_pretrained(output_dir)
+                torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin'))
+        else:
+            super(LLaVATrainer, self)._save_checkpoint(model, trial, metrics)
+
+    def _save(self, output_dir: Optional[str] = None, state_dict=None):
+        if getattr(self.args, 'tune_mm_mlp_adapter', False):
+            pass
+        else:
+            super(LLaVATrainer, self)._save(output_dir, state_dict)

LLaVA/llava/train/train.py

@@ -0,0 +1,991 @@
+# Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright:
+# Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright:
+#    Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+import os
+import copy
+from dataclasses import dataclass, field
+import json
+import logging
+import pathlib
+from typing import Dict, Optional, Sequence, List
+
+import torch
+
+import transformers
+import tokenizers
+
+from llava.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from torch.utils.data import Dataset
+from llava.train.llava_trainer import LLaVATrainer
+
+from llava import conversation as conversation_lib
+from llava.model import *
+from llava.mm_utils import tokenizer_image_token
+
+from PIL import Image
+
+
+local_rank = None
+
+
+def rank0_print(*args):
+    if local_rank == 0:
+        print(*args)
+
+
+from packaging import version
+IS_TOKENIZER_GREATER_THAN_0_14 = version.parse(tokenizers.__version__) >= version.parse('0.14')
+
+
+@dataclass
+class ModelArguments:
+    model_name_or_path: Optional[str] = field(default="facebook/opt-125m")
+    version: Optional[str] = field(default="v0")
+    freeze_backbone: bool = field(default=False)
+    tune_mm_mlp_adapter: bool = field(default=False)
+    vision_tower: Optional[str] = field(default=None)
+    mm_vision_select_layer: Optional[int] = field(default=-1)   # default to the last layer
+    pretrain_mm_mlp_adapter: Optional[str] = field(default=None)
+    mm_projector_type: Optional[str] = field(default='linear')
+    mm_use_im_start_end: bool = field(default=False)
+    mm_use_im_patch_token: bool = field(default=True)
+    mm_patch_merge_type: Optional[str] = field(default='flat')
+    mm_vision_select_feature: Optional[str] = field(default="patch")
+
+
+@dataclass
+class DataArguments:
+    data_path: str = field(default=None,
+                           metadata={"help": "Path to the training data."})
+    lazy_preprocess: bool = False
+    is_multimodal: bool = False
+    image_folder: Optional[str] = field(default=None)
+    image_aspect_ratio: str = 'square'
+
+
+@dataclass
+class TrainingArguments(transformers.TrainingArguments):
+    cache_dir: Optional[str] = field(default=None)
+    optim: str = field(default="adamw_torch")
+    remove_unused_columns: bool = field(default=False)
+    freeze_mm_mlp_adapter: bool = field(default=False)
+    mpt_attn_impl: Optional[str] = field(default="triton")
+    model_max_length: int = field(
+        default=512,
+        metadata={
+            "help":
+            "Maximum sequence length. Sequences will be right padded (and possibly truncated)."
+        },
+    )
+    double_quant: bool = field(
+        default=True,
+        metadata={"help": "Compress the quantization statistics through double quantization."}
+    )
+    quant_type: str = field(
+        default="nf4",
+        metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."}
+    )
+    bits: int = field(
+        default=16,
+        metadata={"help": "How many bits to use."}
+    )
+    lora_enable: bool = False
+    lora_r: int = 64
+    lora_alpha: int = 16
+    lora_dropout: float = 0.05
+    lora_weight_path: str = ""
+    lora_bias: str = "none"
+    mm_projector_lr: Optional[float] = None
+    group_by_modality_length: bool = field(default=False)
+
+
+def maybe_zero_3(param, ignore_status=False, name=None):
+    from deepspeed import zero
+    from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+    if hasattr(param, "ds_id"):
+        if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
+            if not ignore_status:
+                logging.warning(f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}")
+        with zero.GatheredParameters([param]):
+            param = param.data.detach().cpu().clone()
+    else:
+        param = param.detach().cpu().clone()
+    return param
+
+
+# Borrowed from peft.utils.get_peft_model_state_dict
+def get_peft_state_maybe_zero_3(named_params, bias):
+    if bias == "none":
+        to_return = {k: t for k, t in named_params if "lora_" in k}
+    elif bias == "all":
+        to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k}
+    elif bias == "lora_only":
+        to_return = {}
+        maybe_lora_bias = {}
+        lora_bias_names = set()
+        for k, t in named_params:
+            if "lora_" in k:
+                to_return[k] = t
+                bias_name = k.split("lora_")[0] + "bias"
+                lora_bias_names.add(bias_name)
+            elif "bias" in k:
+                maybe_lora_bias[k] = t
+        for k, t in maybe_lora_bias:
+            if bias_name in lora_bias_names:
+                to_return[bias_name] = t
+    else:
+        raise NotImplementedError
+    to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()}
+    return to_return
+
+
+def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
+    to_return = {k: t for k, t in named_params if "lora_" not in k}
+    if require_grad_only:
+        to_return = {k: t for k, t in to_return.items() if t.requires_grad}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
+    to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def find_all_linear_names(model):
+    cls = torch.nn.Linear
+    lora_module_names = set()
+    multimodal_keywords = ['mm_projector', 'vision_tower', 'vision_resampler']
+    for name, module in model.named_modules():
+        if any(mm_keyword in name for mm_keyword in multimodal_keywords):
+            continue
+        if isinstance(module, cls):
+            names = name.split('.')
+            lora_module_names.add(names[0] if len(names) == 1 else names[-1])
+
+    if 'lm_head' in lora_module_names: # needed for 16-bit
+        lora_module_names.remove('lm_head')
+    return list(lora_module_names)
+
+
+def safe_save_model_for_hf_trainer(trainer: transformers.Trainer,
+                                   output_dir: str):
+    """Collects the state dict and dump to disk."""
+
+    if getattr(trainer.args, "tune_mm_mlp_adapter", False):
+        # Only save Adapter
+        keys_to_match = ['mm_projector']
+        if getattr(trainer.args, "use_im_start_end", False):
+            keys_to_match.extend(['embed_tokens', 'embed_in'])
+
+        weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match)
+        trainer.model.config.save_pretrained(output_dir)
+
+        current_folder = output_dir.split('/')[-1]
+        parent_folder = os.path.dirname(output_dir)
+        if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
+            if current_folder.startswith('checkpoint-'):
+                mm_projector_folder = os.path.join(parent_folder, "mm_projector")
+                os.makedirs(mm_projector_folder, exist_ok=True)
+                torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin'))
+            else:
+                torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin'))
+        return
+
+    if trainer.deepspeed:
+        torch.cuda.synchronize()
+        trainer.save_model(output_dir)
+        return
+
+    state_dict = trainer.model.state_dict()
+    if trainer.args.should_save:
+        cpu_state_dict = {
+            key: value.cpu()
+            for key, value in state_dict.items()
+        }
+        del state_dict
+        trainer._save(output_dir, state_dict=cpu_state_dict)  # noqa
+
+
+def smart_tokenizer_and_embedding_resize(
+    special_tokens_dict: Dict,
+    tokenizer: transformers.PreTrainedTokenizer,
+    model: transformers.PreTrainedModel,
+):
+    """Resize tokenizer and embedding.
+
+    Note: This is the unoptimized version that may make your embedding size not be divisible by 64.
+    """
+    num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
+    model.resize_token_embeddings(len(tokenizer))
+
+    if num_new_tokens > 0:
+        input_embeddings = model.get_input_embeddings().weight.data
+        output_embeddings = model.get_output_embeddings().weight.data
+
+        input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+            dim=0, keepdim=True)
+        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+            dim=0, keepdim=True)
+
+        input_embeddings[-num_new_tokens:] = input_embeddings_avg
+        output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+
+def _tokenize_fn(strings: Sequence[str],
+                 tokenizer: transformers.PreTrainedTokenizer) -> Dict:
+    """Tokenize a list of strings."""
+    tokenized_list = [
+        tokenizer(
+            text,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ) for text in strings
+    ]
+    input_ids = labels = [
+        tokenized.input_ids[0] for tokenized in tokenized_list
+    ]
+    input_ids_lens = labels_lens = [
+        tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item()
+        for tokenized in tokenized_list
+    ]
+    return dict(
+        input_ids=input_ids,
+        labels=labels,
+        input_ids_lens=input_ids_lens,
+        labels_lens=labels_lens,
+    )
+
+
+def _mask_targets(target, tokenized_lens, speakers):
+    # cur_idx = 0
+    cur_idx = tokenized_lens[0]
+    tokenized_lens = tokenized_lens[1:]
+    target[:cur_idx] = IGNORE_INDEX
+    for tokenized_len, speaker in zip(tokenized_lens, speakers):
+        if speaker == "human":
+            target[cur_idx+2:cur_idx + tokenized_len] = IGNORE_INDEX
+        cur_idx += tokenized_len
+
+
+def _add_speaker_and_signal(header, source, get_conversation=True):
+    """Add speaker and start/end signal on each round."""
+    BEGIN_SIGNAL = "### "
+    END_SIGNAL = "\n"
+    conversation = header
+    for sentence in source:
+        from_str = sentence["from"]
+        if from_str.lower() == "human":
+            from_str = conversation_lib.default_conversation.roles[0]
+        elif from_str.lower() == "gpt":
+            from_str = conversation_lib.default_conversation.roles[1]
+        else:
+            from_str = 'unknown'
+        sentence["value"] = (BEGIN_SIGNAL + from_str + ": " +
+                             sentence["value"] + END_SIGNAL)
+        if get_conversation:
+            conversation += sentence["value"]
+    conversation += BEGIN_SIGNAL
+    return conversation
+
+
+def preprocess_multimodal(
+    sources: Sequence[str],
+    data_args: DataArguments
+) -> Dict:
+    is_multimodal = data_args.is_multimodal
+    if not is_multimodal:
+        return sources
+
+    for source in sources:
+        for sentence in source:
+            if DEFAULT_IMAGE_TOKEN in sentence['value']:
+                sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '').strip()
+                sentence['value'] = DEFAULT_IMAGE_TOKEN + '\n' + sentence['value']
+                sentence['value'] = sentence['value'].strip()
+                if "mmtag" in conversation_lib.default_conversation.version:
+                    sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '<Image>' + DEFAULT_IMAGE_TOKEN + '</Image>')
+            replace_token = DEFAULT_IMAGE_TOKEN
+            if data_args.mm_use_im_start_end:
+                replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
+            sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, replace_token)
+
+    return sources
+
+
+def preprocess_llama_2(
+    sources,
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+
+    if has_image:
+        input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+
+    assert conv.sep_style == conversation_lib.SeparatorStyle.LLAMA_2
+
+    # Mask targets
+    sep = "[/INST] "
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        rounds = conversation.split(conv.sep2)
+        cur_len = 1
+        target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            if has_image:
+                round_len = len(tokenizer_image_token(rou, tokenizer))
+                instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2
+            else:
+                round_len = len(tokenizer(rou).input_ids)
+                instruction_len = len(tokenizer(parts[0]).input_ids) - 2
+
+            target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_v1(
+    sources,
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+
+    if has_image:
+        input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+
+    assert conv.sep_style == conversation_lib.SeparatorStyle.TWO
+
+    # Mask targets
+    sep = conv.sep + conv.roles[1] + ": "
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        rounds = conversation.split(conv.sep2)
+        cur_len = 1
+        target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            if has_image:
+                round_len = len(tokenizer_image_token(rou, tokenizer))
+                instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2
+            else:
+                round_len = len(tokenizer(rou).input_ids)
+                instruction_len = len(tokenizer(parts[0]).input_ids) - 2
+
+            if i != 0 and not tokenizer.legacy and IS_TOKENIZER_GREATER_THAN_0_14:
+                round_len -= 1
+                instruction_len -= 1
+
+            target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_mpt(
+    sources,
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+
+    if has_image:
+        input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+    assert conv.sep_style == conversation_lib.SeparatorStyle.MPT
+
+    # Mask targets
+    sep = conv.sep + conv.roles[1]
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum())
+
+        rounds = conversation.split(conv.sep)
+        re_rounds = [conv.sep.join(rounds[:3])] # system + user + gpt
+        for conv_idx in range(3, len(rounds), 2):
+            re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx+2]))    # user + gpt
+        cur_len = 0
+        target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(re_rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            if has_image:
+                round_len = len(tokenizer_image_token(rou, tokenizer))
+                instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 1
+            else:
+                round_len = len(tokenizer(rou).input_ids)
+                instruction_len = len(tokenizer(parts[0]).input_ids) - 1
+
+            if i != 0 and getattr(tokenizer, 'legacy', False) and IS_TOKENIZER_GREATER_THAN_0_14:
+                round_len += 1
+                instruction_len += 1
+
+            target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_plain(
+    sources: Sequence[str],
+    tokenizer: transformers.PreTrainedTokenizer,
+) -> Dict:
+    # add end signal and concatenate together
+    conversations = []
+    for source in sources:
+        assert len(source) == 2
+        assert DEFAULT_IMAGE_TOKEN in source[0]['value']
+        source[0]['value'] = DEFAULT_IMAGE_TOKEN
+        conversation = source[0]['value'] + source[1]['value'] + conversation_lib.default_conversation.sep
+        conversations.append(conversation)
+    # tokenize conversations
+    input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations]
+    targets = copy.deepcopy(input_ids)
+    for target, source in zip(targets, sources):
+        tokenized_len = len(tokenizer_image_token(source[0]['value'], tokenizer))
+        target[:tokenized_len] = IGNORE_INDEX
+
+    return dict(input_ids=input_ids, labels=targets)
+
+
+def preprocess(
+    sources: Sequence[str],
+    tokenizer: transformers.PreTrainedTokenizer,
+    has_image: bool = False
+) -> Dict:
+    """
+    Given a list of sources, each is a conversation list. This transform:
+    1. Add signal '### ' at the beginning each sentence, with end signal '\n';
+    2. Concatenate conversations together;
+    3. Tokenize the concatenated conversation;
+    4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
+    """
+    if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN:
+        return preprocess_plain(sources, tokenizer)
+    if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.LLAMA_2:
+        return preprocess_llama_2(sources, tokenizer, has_image=has_image)
+    if conversation_lib.default_conversation.version.startswith("v1"):
+        return preprocess_v1(sources, tokenizer, has_image=has_image)
+    if conversation_lib.default_conversation.version == "mpt":
+        return preprocess_mpt(sources, tokenizer, has_image=has_image)
+    # add end signal and concatenate together
+    conversations = []
+    for source in sources:
+        header = f"{conversation_lib.default_conversation.system}\n\n"
+        conversation = _add_speaker_and_signal(header, source)
+        conversations.append(conversation)
+    # tokenize conversations
+    def get_tokenize_len(prompts):
+        return [len(tokenizer_image_token(prompt, tokenizer)) for prompt in prompts]
+
+    if has_image:
+        input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations]
+    else:
+        conversations_tokenized = _tokenize_fn(conversations, tokenizer)
+        input_ids = conversations_tokenized["input_ids"]
+
+    targets = copy.deepcopy(input_ids)
+    for target, source in zip(targets, sources):
+        if has_image:
+            tokenized_lens = get_tokenize_len([header] + [s["value"] for s in source])
+        else:
+            tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source], tokenizer)["input_ids_lens"]
+        speakers = [sentence["from"] for sentence in source]
+        _mask_targets(target, tokenized_lens, speakers)
+
+    return dict(input_ids=input_ids, labels=targets)
+
+
+class LazySupervisedDataset(Dataset):
+    """Dataset for supervised fine-tuning."""
+
+    def __init__(self, data_path: str,
+                 tokenizer: transformers.PreTrainedTokenizer,
+                 data_args: DataArguments):
+        super(LazySupervisedDataset, self).__init__()
+        list_data_dict = json.load(open(data_path, "r"))
+
+        rank0_print("Formatting inputs...Skip in lazy mode")
+        self.tokenizer = tokenizer
+        self.list_data_dict = list_data_dict
+        self.data_args = data_args
+
+    def __len__(self):
+        return len(self.list_data_dict)
+
+    @property
+    def lengths(self):
+        length_list = []
+        for sample in self.list_data_dict:
+            img_tokens = 128 if 'image' in sample else 0
+            length_list.append(sum(len(conv['value'].split()) for conv in sample['conversations']) + img_tokens)
+        return length_list
+
+    @property
+    def modality_lengths(self):
+        length_list = []
+        for sample in self.list_data_dict:
+            cur_len = sum(len(conv['value'].split()) for conv in sample['conversations'])
+            cur_len = cur_len if 'image' in sample else -cur_len
+            length_list.append(cur_len)
+        return length_list
+
+    def __getitem__(self, i) -> Dict[str, torch.Tensor]:
+        sources = self.list_data_dict[i]
+        if isinstance(i, int):
+            sources = [sources]
+        assert len(sources) == 1, "Don't know why it is wrapped to a list"  # FIXME
+        if 'image' in sources[0]:
+            image_file = self.list_data_dict[i]['image']
+            image_folder = self.data_args.image_folder
+            processor = self.data_args.image_processor
+            image = Image.open(os.path.join(image_folder, image_file)).convert('RGB')
+            if self.data_args.image_aspect_ratio == 'pad':
+                def expand2square(pil_img, background_color):
+                    width, height = pil_img.size
+                    if width == height:
+                        return pil_img
+                    elif width > height:
+                        result = Image.new(pil_img.mode, (width, width), background_color)
+                        result.paste(pil_img, (0, (width - height) // 2))
+                        return result
+                    else:
+                        result = Image.new(pil_img.mode, (height, height), background_color)
+                        result.paste(pil_img, ((height - width) // 2, 0))
+                        return result
+                image = expand2square(image, tuple(int(x*255) for x in processor.image_mean))
+                image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            else:
+                image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            sources = preprocess_multimodal(
+                copy.deepcopy([e["conversations"] for e in sources]),
+                self.data_args)
+        else:
+            sources = copy.deepcopy([e["conversations"] for e in sources])
+        data_dict = preprocess(
+            sources,
+            self.tokenizer,
+            has_image=('image' in self.list_data_dict[i]))
+        if isinstance(i, int):
+            data_dict = dict(input_ids=data_dict["input_ids"][0],
+                             labels=data_dict["labels"][0])
+
+        # image exist in the data
+        if 'image' in self.list_data_dict[i]:
+            data_dict['image'] = image
+        elif self.data_args.is_multimodal:
+            # image does not exist in the data, but the model is multimodal
+            crop_size = self.data_args.image_processor.crop_size
+            data_dict['image'] = torch.zeros(3, crop_size['height'], crop_size['width'])
+        return data_dict
+
+
+@dataclass
+class DataCollatorForSupervisedDataset(object):
+    """Collate examples for supervised fine-tuning."""
+
+    tokenizer: transformers.PreTrainedTokenizer
+
+    def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
+        input_ids, labels = tuple([instance[key] for instance in instances]
+                                  for key in ("input_ids", "labels"))
+        input_ids = torch.nn.utils.rnn.pad_sequence(
+            input_ids,
+            batch_first=True,
+            padding_value=self.tokenizer.pad_token_id)
+        labels = torch.nn.utils.rnn.pad_sequence(labels,
+                                                 batch_first=True,
+                                                 padding_value=IGNORE_INDEX)
+        input_ids = input_ids[:, :self.tokenizer.model_max_length]
+        labels = labels[:, :self.tokenizer.model_max_length]
+        batch = dict(
+            input_ids=input_ids,
+            labels=labels,
+            attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
+        )
+
+        if 'image' in instances[0]:
+            images = [instance['image'] for instance in instances]
+            if all(x is not None and x.shape == images[0].shape for x in images):
+                batch['images'] = torch.stack(images)
+            else:
+                batch['images'] = images
+
+        return batch
+
+
+def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer,
+                                data_args) -> Dict:
+    """Make dataset and collator for supervised fine-tuning."""
+    train_dataset = LazySupervisedDataset(tokenizer=tokenizer,
+                                data_path=data_args.data_path,
+                                data_args=data_args)
+    data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
+    return dict(train_dataset=train_dataset,
+                eval_dataset=None,
+                data_collator=data_collator)
+
+
+def train(attn_implementation=None):
+    global local_rank
+
+    parser = transformers.HfArgumentParser(
+        (ModelArguments, DataArguments, TrainingArguments))
+    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+    local_rank = training_args.local_rank
+    compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
+
+    bnb_model_from_pretrained_args = {}
+    if training_args.bits in [4, 8]:
+        from transformers import BitsAndBytesConfig
+        bnb_model_from_pretrained_args.update(dict(
+            device_map={"": training_args.device},
+            load_in_4bit=training_args.bits == 4,
+            load_in_8bit=training_args.bits == 8,
+            quantization_config=BitsAndBytesConfig(
+                load_in_4bit=training_args.bits == 4,
+                load_in_8bit=training_args.bits == 8,
+                llm_int8_skip_modules=["mm_projector"],
+                llm_int8_threshold=6.0,
+                llm_int8_has_fp16_weight=False,
+                bnb_4bit_compute_dtype=compute_dtype,
+                bnb_4bit_use_double_quant=training_args.double_quant,
+                bnb_4bit_quant_type=training_args.quant_type # {'fp4', 'nf4'}
+            )
+        ))
+
+    if model_args.vision_tower is not None:
+        if 'mpt' in model_args.model_name_or_path:
+            config = transformers.AutoConfig.from_pretrained(model_args.model_name_or_path, trust_remote_code=True)
+            config.attn_config['attn_impl'] = training_args.mpt_attn_impl
+            model = LlavaMptForCausalLM.from_pretrained(
+                model_args.model_name_or_path,
+                config=config,
+                cache_dir=training_args.cache_dir,
+                **bnb_model_from_pretrained_args
+            )
+        else:
+            model = LlavaLlamaForCausalLM.from_pretrained(
+                model_args.model_name_or_path,
+                cache_dir=training_args.cache_dir,
+                attn_implementation=attn_implementation,
+                torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
+                **bnb_model_from_pretrained_args
+            )
+    else:
+        model = transformers.LlamaForCausalLM.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            attn_implementation=attn_implementation,
+            torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
+            **bnb_model_from_pretrained_args
+        )
+    model.config.use_cache = False
+
+    if model_args.freeze_backbone:
+        model.model.requires_grad_(False)
+
+    if training_args.bits in [4, 8]:
+        from peft import prepare_model_for_kbit_training
+        model.config.torch_dtype=(torch.float32 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
+        model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=training_args.gradient_checkpointing)
+
+    if training_args.gradient_checkpointing:
+        if hasattr(model, "enable_input_require_grads"):
+            model.enable_input_require_grads()
+        else:
+            def make_inputs_require_grad(module, input, output):
+                output.requires_grad_(True)
+            model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
+
+    if training_args.lora_enable:
+        from peft import LoraConfig, get_peft_model
+        lora_config = LoraConfig(
+            r=training_args.lora_r,
+            lora_alpha=training_args.lora_alpha,
+            target_modules=find_all_linear_names(model),
+            lora_dropout=training_args.lora_dropout,
+            bias=training_args.lora_bias,
+            task_type="CAUSAL_LM",
+        )
+        if training_args.bits == 16:
+            if training_args.bf16:
+                model.to(torch.bfloat16)
+            if training_args.fp16:
+                model.to(torch.float16)
+        rank0_print("Adding LoRA adapters...")
+        model = get_peft_model(model, lora_config)
+
+    if 'mpt' in model_args.model_name_or_path:
+        tokenizer = transformers.AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            model_max_length=training_args.model_max_length,
+            padding_side="right"
+        )
+    else:
+        tokenizer = transformers.AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            model_max_length=training_args.model_max_length,
+            padding_side="right",
+            use_fast=False,
+        )
+
+    if model_args.version == "v0":
+        if tokenizer.pad_token is None:
+            smart_tokenizer_and_embedding_resize(
+                special_tokens_dict=dict(pad_token="[PAD]"),
+                tokenizer=tokenizer,
+                model=model,
+            )
+    elif model_args.version == "v0.5":
+        tokenizer.pad_token = tokenizer.unk_token
+    else:
+        tokenizer.pad_token = tokenizer.unk_token
+        if model_args.version in conversation_lib.conv_templates:
+            conversation_lib.default_conversation = conversation_lib.conv_templates[model_args.version]
+        else:
+            conversation_lib.default_conversation = conversation_lib.conv_templates["vicuna_v1"]
+
+    if model_args.vision_tower is not None:
+        model.get_model().initialize_vision_modules(
+            model_args=model_args,
+            fsdp=training_args.fsdp
+        )
+        
+        vision_tower = model.get_vision_tower()
+        vision_tower.to(dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device)
+
+        data_args.image_processor = vision_tower.image_processor
+        data_args.is_multimodal = True
+
+        model.config.image_aspect_ratio = data_args.image_aspect_ratio
+        model.config.tokenizer_padding_side = tokenizer.padding_side
+        model.config.tokenizer_model_max_length = tokenizer.model_max_length
+
+        model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter
+        if model_args.tune_mm_mlp_adapter:
+            model.requires_grad_(False)
+            for p in model.get_model().mm_projector.parameters():
+                p.requires_grad = True
+
+        model.config.freeze_mm_mlp_adapter = training_args.freeze_mm_mlp_adapter
+        if training_args.freeze_mm_mlp_adapter:
+            for p in model.get_model().mm_projector.parameters():
+                p.requires_grad = False
+
+        if training_args.bits in [4, 8]:
+            model.get_model().mm_projector.to(dtype=compute_dtype, device=training_args.device)
+
+        model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end
+        model.config.mm_projector_lr = training_args.mm_projector_lr
+        training_args.use_im_start_end = model_args.mm_use_im_start_end
+        model.config.mm_use_im_patch_token = model_args.mm_use_im_patch_token
+        model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer)
+
+    if training_args.bits in [4, 8]:
+        from peft.tuners.lora import LoraLayer
+        for name, module in model.named_modules():
+            if isinstance(module, LoraLayer):
+                if training_args.bf16:
+                    module = module.to(torch.bfloat16)
+            if 'norm' in name:
+                module = module.to(torch.float32)
+            if 'lm_head' in name or 'embed_tokens' in name:
+                if hasattr(module, 'weight'):
+                    if training_args.bf16 and module.weight.dtype == torch.float32:
+                        module = module.to(torch.bfloat16)
+
+    data_module = make_supervised_data_module(tokenizer=tokenizer,
+                                              data_args=data_args)
+    trainer = LLaVATrainer(model=model,
+                    tokenizer=tokenizer,
+                    args=training_args,
+                    **data_module)
+
+    if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")):
+        trainer.train(resume_from_checkpoint=True)
+    else:
+        trainer.train()
+    trainer.save_state()
+
+    model.config.use_cache = True
+
+    if training_args.lora_enable:
+        state_dict = get_peft_state_maybe_zero_3(
+            model.named_parameters(), training_args.lora_bias
+        )
+        non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3(
+            model.named_parameters()
+        )
+        if training_args.local_rank == 0 or training_args.local_rank == -1:
+            model.config.save_pretrained(training_args.output_dir)
+            model.save_pretrained(training_args.output_dir, state_dict=state_dict)
+            torch.save(non_lora_state_dict, os.path.join(training_args.output_dir, 'non_lora_trainables.bin'))
+    else:
+        safe_save_model_for_hf_trainer(trainer=trainer,
+                                       output_dir=training_args.output_dir)
+
+
+if __name__ == "__main__":
+    train()

LLaVA/llava/train/train_mem.py

@@ -0,0 +1,4 @@
+from llava.train.train import train
+
+if __name__ == "__main__":
+    train(attn_implementation="flash_attention_2")

LLaVA/llava/train/train_xformers.py

@@ -0,0 +1,13 @@
+# Make it more memory efficient by monkey patching the LLaMA model with xformers attention.
+
+# Need to call this before importing transformers.
+from llava.train.llama_xformers_attn_monkey_patch import (
+    replace_llama_attn_with_xformers_attn,
+)
+
+replace_llama_attn_with_xformers_attn()
+
+from llava.train.train import train
+
+if __name__ == "__main__":
+    train()

LLaVA/playground/data/prompts/complex_reasoning/000_caps.txt

@@ -0,0 +1,18 @@
+A man wearing multiple neck ties making a goofy face.
+A man in a white shirt wearing very many ties.
+a man with ties on poses for a picture
+A man wearing multiple ties on his neck.
+A young man smiles while wearing several ties.
+
+tie: [0.574, 0.298, 0.752, 0.704]
+tie: [0.464, 0.339, 0.639, 0.789]
+tie: [0.349, 0.363, 0.563, 0.732]
+tie: [0.259, 0.255, 0.668, 0.805]
+person: [0.019, 0.065, 0.962, 0.988]
+person: [0.0, 0.24, 0.214, 1.0]
+tie: [0.316, 0.778, 0.443, 0.867]
+tie: [0.386, 0.707, 0.496, 0.801]
+tie: [0.251, 0.354, 0.402, 0.747]
+tie: [0.44, 0.318, 0.63, 0.397]
+tie: [0.566, 0.33, 0.655, 0.543]
+tie: [0.25, 0.409, 0.359, 0.556]

LLaVA/playground/data/prompts/complex_reasoning/000_conv.txt

@@ -0,0 +1,5 @@
+Question:
+What is unusual about this photo?
+===
+Answer:
+In the photo, the man is wearing a total of ten ties around his neck. This is considered unusual because it is not common for someone to wear multiple ties at once. In a typical setting, a person would wear only one tie. The man's goofy facial expression also adds to the unconventional and humorous nature of the image.

LLaVA/playground/data/prompts/complex_reasoning/001_caps.txt

@@ -0,0 +1,18 @@
+A group of people standing outside of a black vehicle with various luggage.
+Luggage surrounds a vehicle in an underground parking area
+People try to fit all of their luggage in an SUV.
+The sport utility vehicle is parked in the public garage, being packed for a trip
+Some people with luggage near a van that is transporting it.
+
+person: [0.681, 0.242, 0.774, 0.694]
+person: [0.63, 0.222, 0.686, 0.516]
+person: [0.444, 0.233, 0.487, 0.34]
+backpack: [0.384, 0.696, 0.485, 0.914]
+backpack: [0.755, 0.413, 0.846, 0.692]
+suitcase: [0.758, 0.413, 0.845, 0.69]
+suitcase: [0.1, 0.497, 0.173, 0.579]
+bicycle: [0.282, 0.363, 0.327, 0.442]
+car: [0.786, 0.25, 0.848, 0.322]
+car: [0.783, 0.27, 0.827, 0.335]
+car: [0.86, 0.254, 0.891, 0.3]
+car: [0.261, 0.101, 0.787, 0.626]

LLaVA/playground/data/prompts/complex_reasoning/001_conv.txt

@@ -0,0 +1,5 @@
+Question:
+What challenges do these people face?
+===
+Answer:
+In the image, a group of people is standing outside a black SUV in a parking area, surrounded by various pieces of luggage, including suitcases and backpacks. They are facing the challenge of fitting all their luggage into the black SUV. There are multiple suitcases and backpacks to be packed, which suggests that the group has a significant amount of belongings to accommodate. They might have to strategize and arrange the luggage efficiently to ensure that everything fits properly into the vehicle. Additionally, they need to consider the comfort of the passengers and visibility while driving, so the placement of the luggage must not obstruct the driver's view or make the passengers uncomfortable during the trip.

LLaVA/playground/data/prompts/complex_reasoning/002_caps.txt

@@ -0,0 +1,7 @@
+There is a movie theater that displays the show times above the doors.
+A red fire hydrant is deep in the snow.
+The fire hydrant is in the snow near a recently plowed sidewalk.
+This city has had a very hard winter with snow.
+A hotel for dogs in the snow in winter.
+
+fire hydrant: [0.326, 0.612, 0.426, 0.72]

LLaVA/playground/data/prompts/complex_reasoning/002_conv.txt

@@ -0,0 +1,5 @@
+Question:
+What challenges might this city face?
+===
+Answer:
+The city faces challenges due to the harsh winter conditions and heavy snowfall. In the image, a red fire hydrant is almost buried deep in the snow, which indicates the significant amount of snow the city has experienced. This can lead to various challenges such as difficulties in transportation, increased risk of accidents, and disruptions to daily life. For example, the recently plowed sidewalk near the fire hydrant shows that the city has to constantly clear snow from roads and sidewalks to maintain access and safety for pedestrians and vehicles. Moreover, emergency services, like firefighters, might face challenges accessing crucial equipment, such as fire hydrants, during emergencies due to the snow accumulation. This highlights the importance of effective snow management strategies and preparedness in such cities to minimize the impact of harsh winter conditions on residents and essential services.

LLaVA/playground/data/prompts/complex_reasoning/system_message.txt

@@ -0,0 +1,10 @@
+You are an AI visual assistant that can analyze a single image. You receive five sentences, each describing the same image you are observing. In addition, specific object locations within the image are given, along with detailed coordinates. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2) with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y, bottom right x, and bottom right y.
+
+The task is to use the provided caption and bounding box information, create a plausible question about the image, and provide the answer in detail.
+
+Create complex questions beyond describing the scene.
+To answer such questions, one should require first understanding the visual content, then based on the background knowledge or reasoning, either explain why the things are happening that way, or provide guides and help to user's request.  Make the question challenging by not including the visual content details in the question so that the user needs to reason about that first.
+
+Instead of directly mentioning the bounding box coordinates, utilize this data to explain the scene using natural language. Include details like object counts, position of the objects, relative position between the objects.  
+
+When using the information from the caption and coordinates, directly explain the scene, and do not mention that the information source is the caption or the bounding box.  Always answer as if you are directly looking at the image.

LLaVA/playground/data/prompts/conversation/001_caps.txt

@@ -0,0 +1,5 @@
+A man is skiing in the open snow covered hills
+A skier is making his way into the snow.
+A skier on the foothills of a huge mountain range.
+A skier looks at mountains while standing near a trail sign.
+a single person skiing by an area with a lot of bushes

LLaVA/playground/data/prompts/conversation/001_conv.txt

@@ -0,0 +1,37 @@
+Question:
+What is the skier doing?
+===
+Answer:
+The skier is skiing in the snow-covered hills, making his way through the snow and enjoying the beautiful mountain scenery. They are also standing near a trail sign, which indicates their location on the mountain.
+===
+Question:
+Are there any other people or skiers visible in the image?
+===
+Answer:
+No, there is only a single person skiing in the image, and they are surrounded by an area with a lot of bushes.
+===
+Question:
+What kind of skiing is the person doing? Cross-country or downhill?
+===
+Answer:
+Based on what I can see in the photo, it is not possible to definitively determine whether the skier is doing cross-country or downhill skiing. However, given the presence of a trail sign and the fact that they are skiing in the foothills, it is more likely that the skier is engaged in cross-country skiing. Downhill skiing typically takes place on more steep slopes and involves ski lifts, whereas cross-country skiing often occurs on relatively flatter terrain with marked trails.
+===
+Question:
+Can you explain cross-country and downhill skiing a bit more?
+===
+Answer:
+Cross-country skiing and downhill skiing are two distinct types of skiing, each with its own techniques, equipment, and terrain preferences. 
+
+1. Cross-Country Skiing:
+Also known as Nordic skiing, cross-country skiing involves traversing relatively flat terrain or gentle slopes using skis and poles. It is a popular recreational and competitive sport that focuses on endurance, balance, and aerobic fitness. Cross-country skiers use their own locomotion to move across snow-covered terrain, rather than relying on ski lifts or other forms of assistance. 
+There are two main techniques in cross-country skiing:
+- Classic: This technique involves skiing in parallel tracks, using a forward-and-backward gliding motion while keeping the skis parallel to each other. The skier uses their poles to propel themselves forward.
+- Skate skiing: This technique is similar to ice skating, where the skier pushes off one ski at an angle and glides on the other. It requires a smooth, groomed surface and is typically faster than the classic technique. 
+Cross-country ski equipment consists of lightweight skis, boots, bindings, and poles. The boots are more flexible than those used in downhill skiing, allowing for better ankle movement and control. 
+
+2. Downhill Skiing:
+Also known as alpine skiing, downhill skiing involves descending slopes at high speeds using skis and poles for balance and control. This sport is more focused on speed, technique, and navigating challenging terrain, including steep slopes, moguls, and even jumps. 
+Downhill skiing can be further categorized into several disciplines, such as slalom, giant slalom, super-G, and downhill racing. Each discipline has its own set of rules, courses, and ski equipment. 
+Downhill ski equipment includes heavier and stiffer skis, boots, bindings, and poles than those used in cross-country skiing. The boots are more rigid to provide better support and control during high-speed descents and sharp turns. 
+
+In summary, cross-country skiing is an endurance-based sport that involves traveling across flat or gently sloping terrain, while downhill skiing is focused on speed and technique as skiers navigate steeper slopes and challenging terrain. Both sports require specialized equipment and techniques, but they offer different experiences and challenges to participants.

LLaVA/playground/data/prompts/detail_description/000_caps.txt

@@ -0,0 +1,18 @@
+A harbor filled with lots of boats next to a building.
+A bicycle parked in front of several boats at a dock.
+A red bicycle in front of a line of docked white yachts
+A bike sits before boats which sit before a long building.
+A bicycle is a convenient means of land transportation when you live on a boat.
+
+bicycle: [0.287, 0.641, 0.507, 0.874]
+bicycle: [0.566, 0.667, 0.63, 0.731]
+boat: [0.318, 0.579, 0.575, 0.724]
+boat: [0.704, 0.607, 0.818, 0.727]
+boat: [0.818, 0.601, 0.942, 0.744]
+boat: [0.002, 0.53, 0.243, 0.71]
+boat: [0.541, 0.611, 0.668, 0.731]
+person: [0.778, 0.527, 0.797, 0.57]
+cup: [0.708, 0.733, 0.724, 0.758]
+boat: [0.236, 0.532, 0.404, 0.64]
+boat: [0.81, 0.632, 0.836, 0.676]
+boat: [0.957, 0.526, 1.0, 0.752]

LLaVA/playground/data/prompts/detail_description/000_conv.txt

@@ -0,0 +1,3 @@
+It is a harbor filled with numerous boats of various sizes docked next to a long building. Among the boats, there are a few white yachts lined up, standing out from the rest. There is a red bicycle prominently parked in front of the line of docked boats, serving as a convenient means of land transportation for those living on the boats. Another bicycle can be seen further back in the scene, near the middle of the harbor.
+
+A person is visible near the right side of the harbor, possibly enjoying the view or attending to their boat. Additionally, there is a cup placed on a surface near the middle of the scene.

LLaVA/playground/data/prompts/detail_description/001_caps.txt

@@ -0,0 +1,18 @@
+A group of people standing outside of a black vehicle with various luggage.
+Luggage surrounds a vehicle in an underground parking area
+People try to fit all of their luggage in an SUV.
+The sport utility vehicle is parked in the public garage, being packed for a trip
+Some people with luggage near a van that is transporting it.
+
+person: [0.681, 0.242, 0.774, 0.694]
+person: [0.63, 0.222, 0.686, 0.516]
+person: [0.444, 0.233, 0.487, 0.34]
+backpack: [0.384, 0.696, 0.485, 0.914]
+backpack: [0.755, 0.413, 0.846, 0.692]
+suitcase: [0.758, 0.413, 0.845, 0.69]
+suitcase: [0.1, 0.497, 0.173, 0.579]
+bicycle: [0.282, 0.363, 0.327, 0.442]
+car: [0.786, 0.25, 0.848, 0.322]
+car: [0.783, 0.27, 0.827, 0.335]
+car: [0.86, 0.254, 0.891, 0.3]
+car: [0.261, 0.101, 0.787, 0.626]

LLaVA/playground/data/prompts/detail_description/001_conv.txt

@@ -0,0 +1,5 @@
+The image is an underground parking area with a black sport utility vehicle (SUV) parked. There are three people in the scene, with one person standing closer to the left side of the vehicle, another person in the middle, and the third person on the right side. They are all working together to pack their luggage into the SUV for a trip. 
+
+In the parking area, various luggage items are scattered around the vehicle. There are two backpacks, one located near the left rear wheel and the other closer to the right side of the vehicle. Additionally, there are two suitcases, one on the right side of the car and another further away near the center of the parking area. A bicycle can also be seen on the left side of the vehicle. 
+
+Other cars are parked around the main SUV, with one car positioned behind it and slightly to the left, another behind and slightly to the right, and the third car further behind on the right side.

LLaVA/playground/data/prompts/detail_description/002_caps.txt

@@ -0,0 +1,15 @@
+A man holds a Wii-mote above his head while another looks on.
+A guy and his friend are playing Nintendo Wii.
+A young man is holding a video game remote over his head.
+two men standing in a room while one plays with a wii mote
+Some guys standing and playing a video game.
+
+couch: [0.697, 0.759, 0.995, 1.0]
+dining table: [0.426, 0.755, 1.0, 0.987]
+person: [0.082, 0.252, 0.342, 1.0]
+person: [0.399, 0.085, 0.742, 0.982]
+remote: [0.477, 0.135, 0.516, 0.187]
+sink: [0.016, 0.501, 0.063, 0.52]
+potted plant: [0.798, 0.384, 0.888, 0.645]
+refrigerator: [0.305, 0.389, 0.414, 0.547]
+chair: [0.72, 0.509, 0.858, 0.725]

LLaVA/playground/data/prompts/detail_description/002_conv.txt

@@ -0,0 +1,3 @@
+The image shows two men standing in a room, engaged in playing a video game on a Nintendo Wii console. One of the men is holding a Wii remote above his head with enthusiasm, while the other man looks on, likely enjoying the friendly competition. 
+
+The room appears to be a living space with a couch located in the background and a dining table nearby. A potted plant can be seen placed close to the couch, and a chair is situated in the middle of the room. The room also features a kitchen area with a sink and a refrigerator visible in the background.

LLaVA/playground/data/prompts/detail_description/system_message.txt

@@ -0,0 +1,7 @@
+You are an AI visual assistant that can analyze a single image. You receive five sentences, each describing the same image you are observing. In addition, specific object locations within the image are given, along with detailed coordinates. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2) with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y, bottom right x, and bottom right y.
+
+Using the provided caption and bounding box information, describe the scene in a detailed manner.
+
+Instead of directly mentioning the bounding box coordinates, utilize this data to explain the scene using natural language. Include details like object counts, position of the objects, relative position between the objects.
+
+When using the information from the caption and coordinates, directly explain the scene, and do not mention that the information source is the caption or the bounding box.  Always answer as if you are directly looking at the image.