run_llm.py

import os
import sys
import json
import time
import openai
import pickle
import argparse
import requests
from tqdm import tqdm
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, LlamaForCausalLM, LlamaTokenizer

from fastchat.model import load_model, get_conversation_template, add_model_args

from nltk.tag.mapping import _UNIVERSAL_TAGS

import gradio as gr

uni_tags = list(_UNIVERSAL_TAGS)
uni_tags[-1] = 'PUNC'

bio_tags = ['B', 'I', 'O']
chunk_tags = ['ADJP', 'ADVP', 'CONJP', 'INTJ', 'LST', 'NP', 'O', 'PP', 'PRT', 'SBAR', 'UCP', 'VP']

syntags = ['NP', 'S', 'VP', 'ADJP', 'ADVP', 'SBAR', 'TOP', 'PP', 'POS', 'NAC', "''", 'SINV', 'PRN', 'QP', 'WHNP', 'RB', 'FRAG',
 'WHADVP', 'NX', 'PRT', 'VBZ', 'VBP', 'MD', 'NN', 'WHPP', 'SQ', 'SBARQ', 'LST', 'INTJ', 'X', 'UCP', 'CONJP', 'NNP', 'CD', 'JJ',
 'VBD', 'WHADJP', 'PRP', 'RRC', 'NNS', 'SYM', 'CC']

openai.api_key = "OPENAI_API_KEY"


# determinant vs. determiner
# https://wikidiff.com/determiner/determinant
ents_prompt = ['Noun','Verb','Adjective','Adverb','Preposition/Subord','Coordinating Conjunction',# 'Cardinal Number',
    'Determiner',
    'Noun Phrase','Verb Phrase','Adjective Phrase','Adverb Phrase','Preposition Phrase','Conjunction Phrase','Coordinate Phrase','Quantitave Phrase','Complex Nominal',
    'Clause','Dependent Clause','Fragment Clause','T-unit','Complex T-unit',# 'Fragment T-unit',
][7:]
ents = ['NN', 'VB', 'JJ', 'RB', 'IN', 'CC', 'DT', 'NP', 'VP', 'ADJP', 'ADVP', 'PP', 'CONJP', 'CP', 'QP', 'CN', 'C', 'DC', 'FC', 'T', 'CT'][7:]


ents_prompt_uni_tags = ['Verb', 'Noun', 'Pronoun', 'Adjective', 'Adverb', 'Preposition and Postposition', 'Coordinating Conjunction', 
                        'Determiner', 'Cardinal Number', 'Particles or other function words', 
                        'Words that cannot be assigned a POS tag', 'Punctuation']

ents = uni_tags + ents
ents_prompt = ents_prompt_uni_tags + ents_prompt

for i, j in zip(ents, ents_prompt):
    print(i, j)
# raise


model_mapping = {
    # 'gpt3': 'gpt-3',
    'gpt3.5': 'gpt-3.5-turbo-0613',
    'vicuna-7b': 'lmsys/vicuna-7b-v1.3',
    'vicuna-13b': 'lmsys/vicuna-13b-v1.3',
    'vicuna-33b': 'lmsys/vicuna-33b-v1.3',
    'fastchat-t5': 'lmsys/fastchat-t5-3b-v1.0',
    # 'llama2-7b': 'meta-llama/Llama-2-7b-hf',
    # 'llama2-13b': 'meta-llama/Llama-2-13b-hf',
    # 'llama2-70b': 'meta-llama/Llama-2-70b-hf',
    'llama-7b': './llama/hf/7B',
    'llama-13b': './llama/hf/13B',
    'llama-30b': './llama/hf/30B',
    # 'llama-65b': './llama/hf/65B',
    'alpaca': './alpaca-7B',
    # 'koala-7b': 'koala-7b',
    # 'koala-13b': 'koala-13b',
}

for m in model_mapping.keys():
    for eid, ent in enumerate(ents):
        os.makedirs(f'result/prompt1_qa/{m}/ptb/per_ent/{ent}', exist_ok=True)
        
        os.makedirs(f'result/prompt2_instruction/pos_tagging/{m}/ptb', exist_ok=True)
        os.makedirs(f'result/prompt2_instruction/chunking/{m}/ptb', exist_ok=True)
        os.makedirs(f'result/prompt2_instruction/parsing/{m}/ptb', exist_ok=True)

        os.makedirs(f'result/prompt3_structured_prompt/pos_tagging/{m}/ptb', exist_ok=True)
        os.makedirs(f'result/prompt3_structured_prompt/chunking/{m}/ptb', exist_ok=True)
        os.makedirs(f'result/prompt3_structured_prompt/parsing/{m}/ptb', exist_ok=True)


#s = int(sys.argv[1])
#e = int(sys.argv[2])

#s = 0
#e = 1000
with open('sample_uniform_1k_2.txt', 'r') as f:
    selected_idx = f.readlines()
selected_idx = [int(i.strip()) for i in selected_idx]#[s:e]


ptb = []
with open('ptb.jsonl', 'r') as f:
    for l in f:
        ptb.append(json.loads(l))


## Prompt 1
template_all = '''Please output the <Noun, Verb, Adjective, Adverb, Preposition/Subord, Coordinating Conjunction, Cardinal Number, Determiner, Noun Phrase, Verb Phrase, Adjective Phrase, Adverb Phrase, Preposition Phrase, Conjunction Phrase, Coordinate Phrase, Quantitave Phrase, Complex Nominal, Clause, Dependent Clause, Fragment Clause, T-unit, Complex T-unit, Fragment T-unit> in the following sentence without any additional text in json format: "{}"'''
template_single = '''Please output any <{}> in the following sentence one per line without any additional text: "{}"'''

## Prompt 2
prompt2_pos = '''Please pos tag the following sentence using Universal POS tag set without generating any additional text: {}'''
prompt2_chunk = '''Please do sentence chunking for the following sentence as in CoNLL 2000 shared task without generating any addtional text: {}'''
prompt2_parse = '''Generate textual representation of the constituency parse tree of the following sentence using Penn TreeBank tag set without outputing any additional text: {}'''

prompt2_chunk = '''Please chunk the following sentence in CoNLL 2000 format with BIO tags without outputing any additional text: {}'''

## Prompt 3
with open('demonstration_3_42_pos.txt', 'r') as f:
    demon_pos = f.read()
with open('demonstration_3_42_chunk.txt', 'r') as f:
    demon_chunk = f.read()
with open('demonstration_3_42_parse.txt', 'r') as f:
    demon_parse = f.read()


def para(m):
    c = 0
    for n, p in m.named_parameters():
        c += p.numel()
    return c

def main(args=None):

    gid_list = selected_idx[args.start:args.end]


    if 'gpt3' in args.model_path:
        pass

    else:
        path = model_mapping[args.model_path]
        model, tokenizer = load_model(
            path,
            args.device,
            args.num_gpus,
            args.max_gpu_memory,
            args.load_8bit,
            args.cpu_offloading,
            revision=args.revision,
            debug=args.debug,
        )

        whitelist_ids_pos = [tokenizer.encode(word)[1] for word in uni_tags]
        bad_words_ids_pos = [[ids] for ids in range(tokenizer.vocab_size) if ids not in whitelist_ids_pos]

        whitelist_ids_bio = [tokenizer.encode(word)[1] for word in bio_tags]
        bad_words_ids_bio = [[ids] for ids in range(tokenizer.vocab_size) if ids not in whitelist_ids_bio]

        whitelist_ids_chunk = [tokenizer.encode(word)[1] for word in chunk_tags]
        bad_words_ids_chunk = [[ids] for ids in range(tokenizer.vocab_size) if ids not in whitelist_ids_chunk]

        whitelist_ids_parse = [tokenizer.encode(word)[1] for word in syntags]
        bad_words_ids_parse = [[ids] for ids in range(tokenizer.vocab_size) if ids not in whitelist_ids_parse]


    if args.prompt == 1:
        for gid in tqdm(gid_list, desc='Query'):
            text = ptb[gid]['text']

            for eid, ent in enumerate(ents):
                os.makedirs(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/{ent}', exist_ok=True)
                
                if ent == 'NOUN' and not os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/NOUN'):
                    os.system(f'ln -sT ./NN result/prompt1_qa/{args.model_path}/ptb/per_ent/NOUN')
                if ent == 'VERB' and not os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/VERB'):
                    os.system(f'ln -sT ./VB result/prompt1_qa/{args.model_path}/ptb/per_ent/VERB')
                if ent == 'ADJ' and not os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/ADJ'):
                    os.system(f'ln -sT ./JJ result/prompt1_qa/{args.model_path}/ptb/per_ent/ADJ')
                if ent == 'ADV' and not os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/ADV'):
                    os.system(f'ln -sT ./RB result/prompt1_qa/{args.model_path}/ptb/per_ent/ADV')
                if ent == 'CONJ' and not os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/CONJ'):
                    os.system(f'ln -sT ./CC result/prompt1_qa/{args.model_path}/ptb/per_ent/CONJ')
                if ent == 'DET' and not os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/DET'):
                    os.system(f'ln -sT ./DT result/prompt1_qa/{args.model_path}/ptb/per_ent/DET')
                if ent == 'ADP' and not os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/ADP'):
                    os.system(f'ln -sT ./DT result/prompt1_qa/{args.model_path}/ptb/per_ent/IN')

                if os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/{ent}/{gid}.txt'):
                    print(gid, ent, 'skip')
                    continue
                    

                ## Get prompt
                msg = template_single.format(ents_prompt[eid], text)

                ## Run
                if 'gpt3' in args.model_path:
                    if os.path.exists(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/{ent}/{gid}.pkl'):
                        print('Found cache')
                        with open(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/{ent}/{gid}.pkl', 'rb') as f:
                            outputs = pickle.load(f)
                        outputs = outputs['choices'][0]['message']['content']
                    else:
                        outputs = gpt3(msg)
                        if outputs is None:
                            continue
                        time.sleep(0.2)
                    
                else:
                    conv = get_conversation_template(args.model_path)
                    conv.append_message(conv.roles[0], msg)
                    conv.append_message(conv.roles[1], None)
                    conv.system = ''
                    prompt = conv.get_prompt().strip()
                    outputs = fastchat(prompt, model, tokenizer)

                with open(f'result/prompt1_qa/{args.model_path}/ptb/per_ent/{ent}/{gid}.txt', 'w') as f:
                    f.write(outputs)


    if args.prompt == 2:
        for gid in tqdm(gid_list, desc='Query'):
            text = ptb[gid]['text']

            ## POS tagging
            if os.path.exists(f'result/prompt2_instruction/pos_tagging/{args.model_path}/ptb/{gid}.txt'):
                print(gid, 'skip')

            else:
                msg = prompt2_pos.format(text)

                if 'gpt3' in args.model_path:
                    outputs = gpt3(msg)
                    if outputs is None:
                        continue
                    time.sleep(0.2)

                else:
                    conv = get_conversation_template(args.model_path)
                    conv.append_message(conv.roles[0], msg)
                    conv.append_message(conv.roles[1], None)
                    conv.system = ''
                    prompt = conv.get_prompt()

                    outputs = fastchat(prompt, model, tokenizer)

                with open(f'result/prompt2_instruction/pos_tagging/{args.model_path}/ptb/{gid}.txt', 'w') as f:
                    f.write(outputs)


            ## Sentence chunking
            if os.path.exists(f'result/prompt2_instruction/chunking/{args.model_path}/ptb/{gid}.txt'):
                print(gid, 'skip')
            if False:
                pass
            else:
                msg = prompt2_chunk.format(text)

                if 'gpt3' in args.model_path:
                    outputs = gpt3(msg)
                    if outputs is None:
                        continue
                    time.sleep(0.2)

                else:
                    conv = get_conversation_template(args.model_path)
                    conv.append_message(conv.roles[0], msg)
                    conv.append_message(conv.roles[1], None)
                    conv.system = ''
                    prompt = conv.get_prompt()

                    outputs = fastchat(prompt, model, tokenizer)

                print(args.model_path, gid, outputs)
                with open(f'result/prompt2_instruction/chunking/{args.model_path}/ptb/{gid}.txt', 'w') as f:
                    f.write(outputs)

            
            ## Parsing
            if os.path.exists(f'result/prompt2_instruction/parsing/{args.model_path}/ptb/{gid}.txt'):
                print(gid, 'skip')
            
            else:
                msg = prompt2_parse.format(text)

                if 'gpt3' in args.model_path:
                    outputs = gpt3(msg)
                    if outputs is None:
                        continue
                    time.sleep(0.2)

                else:
                    conv = get_conversation_template(args.model_path)
                    conv.append_message(conv.roles[0], msg)
                    conv.append_message(conv.roles[1], None)
                    conv.system = ''
                    prompt = conv.get_prompt()

                    outputs = fastchat(prompt, model, tokenizer)

                with open(f'result/prompt2_instruction/parsing/{args.model_path}/ptb/{gid}.txt', 'w') as f:
                    f.write(outputs)



    if args.prompt == 3:
        for gid in tqdm(gid_list, desc='Query'):
            text = ptb[gid]['text']
            tokens = ptb[gid]['tokens']
            poss = ptb[gid]['uni_poss']

            ## POS tagging
            if os.path.exists(f'result/prompt3_structured_prompt/pos_tagging/{args.model_path}/ptb/{gid}.txt'):
                print(gid, 'skip')
                continue

            prompt = demon_pos + '\n' + 'C: ' + text + '\n' + 'T: '

            if 'gpt3' in args.model_path:
                outputs = gpt3(prompt)
                if outputs is None:
                    continue
                time.sleep(0.2)

            else:
                pred_poss = []
                for _tok, _pos in zip(tokens, poss):
                    prompt = prompt + ' ' + _tok + '_'
                    outputs = structured_prompt(prompt, model, tokenizer, bad_words_ids_pos)
                    prompt = prompt + outputs
                    pred_poss.append(outputs)

            outputs = ' '.join(pred_poss)
            with open(f'result/prompt3_structured_prompt/pos_tagging/{args.model_path}/ptb/{gid}.txt', 'w') as f:
                f.write(outputs)


            ## Chunking
            if os.path.exists(f'result/prompt3_structured_prompt/chunking/{args.model_path}/ptb/{gid}.txt'):
                print(gid, 'skip')
                continue

            prompt = demon_chunk + '\n' + 'C: ' + text + '\n' + 'T: '

            if 'gpt3' in args.model_path:
                outputs = gpt3(prompt)
                print(outputs)
                if outputs is None:
                    continue
                time.sleep(0.2)

            else:
                pred_chunk = []
                for _tok, _pos in zip(tokens, poss):
                    prompt = prompt + ' ' + _tok + '_'

                    # Generate BIO
                    outputs_bio = structured_prompt(prompt, model, tokenizer, bad_words_ids_bio)
                    prompt = prompt + outputs_bio + '-'

                    # Generate tag
                    outputs_chunk = structured_prompt(prompt, model, tokenizer, bad_words_ids_chunk)
                    prompt = prompt + outputs_chunk

                    pred_chunk.append((outputs_bio + '-' + outputs_chunk))

                outputs = ' '.join(pred_chunk)

            with open(f'result/prompt3_structured_prompt/chunking/{args.model_path}/ptb/{gid}.txt', 'w') as f:
                f.write(outputs)

            ## Parsing
            if os.path.exists(f'result/prompt3_structured_prompt/parsing/{args.model_path}/ptb/{gid}.txt'):
                print(gid, 'skip')
                continue

            prompt = demon_parse + '\n' + 'C: ' + text + '\n' + 'T: '

            if 'gpt3' in args.model_path:
                outputs = gpt3(prompt)
                if outputs is None:
                    continue
                time.sleep(0.2)

            else:
                pred_syn = []
                for _tok, _pos in zip(tokens, poss):
                    prompt = prompt + _tok + '_'
                    outputs = structured_prompt(prompt, model, tokenizer, bad_words_ids_parse)
                    pred_syn.append(outputs)

            with open(f'result/prompt3_structured_prompt/parsing/{args.model_path}/ptb/{gid}.txt', 'w') as f:
                f.write(' '.join(pred_syn))


def structured_prompt(prompt, model, tokenizer, bad_words_ids):
    input_ids = tokenizer([prompt]).input_ids
    output_ids = model.generate(
        torch.as_tensor(input_ids).cuda(),
        max_new_tokens=1,
        bad_words_ids=bad_words_ids,
    )

    if model.config.is_encoder_decoder:
        output_ids = output_ids[0]
    else:
        output_ids = output_ids[0][len(input_ids[0]) :]
    outputs = tokenizer.decode(
        output_ids, skip_special_tokens=True, spaces_between_special_tokens=False
    )

    return outputs


def fastchat(prompt, model, tokenizer):
    input_ids = tokenizer([prompt]).input_ids
    output_ids = model.generate(
        torch.as_tensor(input_ids).cuda(),
        do_sample=True,
        temperature=args.temperature,
        repetition_penalty=args.repetition_penalty,
        max_new_tokens=args.max_new_tokens,
    )

    if model.config.is_encoder_decoder:
        output_ids = output_ids[0]
    else:
        output_ids = output_ids[0][len(input_ids[0]) :]
    outputs = tokenizer.decode(
        output_ids, skip_special_tokens=True, spaces_between_special_tokens=False
    )

    #print('Empty system message')
    #print(f"{conv.roles[0]}: {msg}")
    #print(f"{conv.roles[1]}: {outputs}")

    return outputs


def gpt3(prompt):
    try:
        response = openai.ChatCompletion.create(
            model=model_mapping[args.model_path], messages=[{"role": "user", "content": prompt}])

        return response['choices'][0]['message']['content']

    except Exception as err:
        print('Error')
        print(err)

        return None


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    add_model_args(parser)
    parser.add_argument("--temperature", type=float, default=0.7)
    parser.add_argument("--repetition_penalty", type=float, default=1.0)
    parser.add_argument("--max-new-tokens", type=int, default=512)
    parser.add_argument("--debug", action="store_true")
    parser.add_argument("--message", type=str, default="Hello! Who are you?")
    parser.add_argument("--start", type=int, default=0)
    parser.add_argument("--end", type=int, default=1000)
    parser.add_argument("--prompt", required=True, type=int, default=None)
    # parser.add_argument("--system_msg", required=True, type=str, default='default_system_msg')
    args = parser.parse_args()

    # Reset default repetition penalty for T5 models.
    if "t5" in args.model_path and args.repetition_penalty == 1.0:
        args.repetition_penalty = 1.2

    main(args)