File size: 6,987 Bytes
35b3f62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
from vllm import LLM, SamplingParams
import json
import torch
import time
from datetime import datetime, timedelta
import argparse
from tqdm import tqdm
from typing import List, Dict, Any
import concurrent.futures

class VLLMGenerator:
    def __init__(self, model_name: str, n: int = 8, max_tokens: int = 512, 
                 temperature: float = 0.7, top_p: float = 1.0, 
                 frequency_penalty: float = 0.0, presence_penalty: float = 0.0, 
                 stop: List[str] = ['\n\n\n'], batch_size: int = 32):
        self.device_count = torch.cuda.device_count()
        print(f"Initializing with {self.device_count} GPUs")
        self.llm = LLM(
            model=model_name,
            tensor_parallel_size=self.device_count,
            max_model_len=4096,
            gpu_memory_utilization=0.95,
            enforce_eager=True,
            trust_remote_code=True,
            # quantization="bitsandbytes",
            # dtype="half",
            # load_format="bitsandbytes",
            max_num_batched_tokens=4096,
            max_num_seqs=batch_size
        )
        self.sampling_params = SamplingParams(
            n=n,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            frequency_penalty=frequency_penalty,
            presence_penalty=presence_penalty,
            stop=stop,
            logprobs=1
        )
        self.batch_size = batch_size
        self.tokenizer = self.llm.get_tokenizer()
        print(f"Initialization complete. Batch size: {batch_size}")
    
    def parse_response(self, responses):
        all_outputs = []
        for response in responses:
            to_return = []
            for output in response.outputs:
                text = output.text.strip()
                try:
                    logprob = sum(logprob_obj.logprob for item in output.logprobs for logprob_obj in item.values())
                except:
                    logprob = 0  # Fallback if logprobs aren't available
                to_return.append((text, logprob))
            texts = [r[0] for r in sorted(to_return, key=lambda tup: tup[1], reverse=True)]
            all_outputs.append(texts)
        return all_outputs

    def prepare_prompt(self, claim: str, model_name: str) -> str:
        base_prompt = f"Please write a fact-checking article passage to support, refute, indicate not enough evidence, or present conflicting evidence regarding the claim.\nClaim: {claim}"
        
        if "OLMo" in model_name:
            return base_prompt
        else:
            messages = [{"role": "user", "content": base_prompt}]
            return self.tokenizer.apply_chat_template(messages, tokenize=False) + "<|start_header_id|>assistant<|end_header_id|>\n\nPassage: "

    def process_batch(self, batch: List[Dict[str, Any]], model_name: str) -> tuple[List[Dict[str, Any]], float]:
        start_time = time.time()
        prompts = [self.prepare_prompt(example["claim"], model_name) for example in batch]
        
        try:
            results = self.llm.generate(prompts, sampling_params=self.sampling_params)
            outputs = self.parse_response(results)
            
            for example, output in zip(batch, outputs):
                example['hypo_fc_docs'] = output
            
            batch_time = time.time() - start_time
            return batch, batch_time
        except Exception as e:
            print(f"Error processing batch: {str(e)}")
            return batch, time.time() - start_time

# def format_time(seconds: float) -> str:
#     return str(timedelta(seconds=int(seconds)))

# def estimate_completion_time(start_time: float, processed_examples: int, total_examples: int) -> str:
#     elapsed_time = time.time() - start_time
#     examples_per_second = processed_examples / elapsed_time
#     remaining_examples = total_examples - processed_examples
#     estimated_remaining_seconds = remaining_examples / examples_per_second
#     completion_time = datetime.now() + timedelta(seconds=int(estimated_remaining_seconds))
#     return completion_time.strftime("%Y-%m-%d %H:%M:%S")

def main(args):
    total_start_time = time.time()
    print(f"Script started at: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
    
    # Load data
    print("Loading data...")
    with open(args.target_data, 'r', encoding='utf-8') as json_file:
        examples = json.load(json_file)
    print(f"Loaded {len(examples)} examples")
    
    # Initialize generator
    print("Initializing generator...")
    generator = VLLMGenerator(
        model_name=args.model,
        batch_size=32
    )
    
    # Process data in batches
    processed_data = []
    # batch_times = []
    batches = [examples[i:i + generator.batch_size] for i in range(0, len(examples), generator.batch_size)]
    
    print(f"\nProcessing {len(batches)} batches...")
    with tqdm(total=len(examples), desc="Processing examples") as pbar:
        for batch_idx, batch in enumerate(batches, 1):
            processed_batch, batch_time = generator.process_batch(batch, args.model)
            processed_data.extend(processed_batch)
            # batch_times.append(batch_time)
            
            # Update progress and timing information
            # examples_processed = len(processed_data)
            # avg_batch_time = sum(batch_times) / len(batch_times)
            # estimated_completion = estimate_completion_time(total_start_time, examples_processed, len(examples))
            
            # pbar.set_postfix({
            #     'Batch': f"{batch_idx}/{len(batches)}",
            #     'Avg Batch Time': f"{avg_batch_time:.2f}s",
            #     'ETA': estimated_completion
            # })
            # pbar.update(len(batch))
    
    # Calculate and display timing statistics
    # total_time = time.time() - total_start_time
    # avg_batch_time = sum(batch_times) / len(batch_times)
    # avg_example_time = total_time / len(examples)
    
    # print("\nTiming Statistics:")
    # print(f"Total Runtime: {format_time(total_time)}")
    # print(f"Average Batch Time: {avg_batch_time:.2f} seconds")
    # print(f"Average Time per Example: {avg_example_time:.2f} seconds")
    # print(f"Throughput: {len(examples)/total_time:.2f} examples/second")
    
    # Save results
    # print("\nSaving results...")
    with open(args.json_output, "w", encoding="utf-8") as output_json:
        json.dump(processed_data, output_json, ensure_ascii=False, indent=4)
    
    # print(f"Script completed at: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
    # print(f"Total runtime: {format_time(total_time)}")

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument('-i', '--target_data', default='data_store/averitec/dev.json')
    parser.add_argument('-o', '--json_output', default='data_store/hyde_fc.json')
    parser.add_argument('-m', '--model', default="meta-llama/Llama-3.1-8B-Instruct")
    args = parser.parse_args()
    main(args)