---
base_model: google/gemma-2-9b
tags:
- text-generation-inference
- transformers
- torch
- mlx_lm

license: apache-2.0
language:
- en
---

# 推論用コード
Hugging Faceにアップロードしたモデルを用いてELYZA-tasks-100-TVの出力を得るためのコードです。 
このコードで生成されたjsonlファイルは課題の成果として提出可能なフォーマットになっております。

```
!pip install -U bitsandbytes
!pip install -U transformers
!pip install -U accelerate
!pip install -U datasets
!pip install -U peft
```

```
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import json
from pathlib import Path
from typing import Dict, Any, Optional
from tqdm import tqdm
import time
from datetime import datetime

class GPUPredictions:
    def __init__(self,
                 model_id="testmoto/gemma-2-lora-dpo-moe-1",
                 adapter_path=None,
                 max_tokens=1024,
                 temp=0.0,
                 top_p=0.9,
                 seed=3407):
        self.model_id = model_id
        self.adapter_path = adapter_path
        self.max_tokens = max_tokens
        self.temp = temp
        self.top_p = top_p
        self.seed = seed
        
        print(f"Loading model: {model_id}")
        torch.cuda.empty_cache()
        
        # GPU設定
        n_gpus = torch.cuda.device_count()
        max_memory = {i: "20GiB" for i in range(n_gpus)}
        max_memory["cpu"] = "100GiB"
        
        # トークナイザーの初期化
        self.tokenizer = AutoTokenizer.from_pretrained(
            model_id,
            trust_remote_code=True
        )
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token
        
        try:
            self.model = AutoModelForCausalLM.from_pretrained(
                model_id,
                torch_dtype=torch.float16,
                device_map="auto",
                max_memory=max_memory,
                low_cpu_mem_usage=True,
                trust_remote_code=True
            )
        except Exception as e:
            print(f"First loading attempt failed: {str(e)}")
            print("Trying alternative loading method...")
            self.model = AutoModelForCausalLM.from_pretrained(
                model_id,
                torch_dtype=torch.float16,
                device_map="balanced",
                low_cpu_mem_usage=True,
                trust_remote_code=True
            )
        
        if adapter_path:
            print(f"Loading adapter from {adapter_path}")
            self.model.load_adapter(adapter_path)

        # Generate設定
        self.gen_config = {
            "max_new_tokens": max_tokens,
            "temperature": temp,
            "top_p": top_p,
            "do_sample": temp > 0,
            "pad_token_id": self.tokenizer.pad_token_id,
            "eos_token_id": self.tokenizer.eos_token_id
        }

        print("Model loaded successfully")
        self.device = next(self.model.parameters()).device
        print(f"Model is on device: {self.device}")

    @torch.inference_mode()
    def generate_response(self, prompt: str) -> str:
        """効率的な応答生成"""
        try:
            inputs = self.tokenizer(prompt, return_tensors="pt", padding=True)
            inputs = {k: v.to(self.device) for k, v in inputs.items()}
            
            with torch.cuda.amp.autocast():
                outputs = self.model.generate(
                    **inputs,
                    **self.gen_config
                )
            
            response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
            
            if prompt in response:
                response = response[len(prompt):].strip()
            
            return response
            
        except Exception as e:
            print(f"Error during generation: {str(e)}")
            raise

    def load_tasks(self, file_path: str) -> list:
        """ELYZAタスクの読み込み"""
        datasets = []
        with open(file_path, "r") as f:
            for line in f:
                if line.strip():
                    datasets.append(json.loads(line.strip()))
        return datasets

    def run_inference(self, input_file="elyza-tasks-100-TV_0.jsonl", output_file="gpu_results.jsonl"):
        """バッチ処理による効率的な推論実行"""
        tasks = self.load_tasks(input_file)
        results = []
        
        start_time = time.time()
        execution_date = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
        print(f"Execution started at: {execution_date}")
        print(f"Total tasks: {len(tasks)}")
        print("-" * 50)
        
        for task in tqdm(tasks, desc="Processing tasks"):
            task_start_time = time.time()
            
            prompt = f"""### Instruction: 
{task['input']}
<eos>
### Response: """
            
            try:
                response = self.generate_response(prompt)
                try:
                    answer = response.split('### Response: ')[-1]
                except:
                    answer = response
                
                task_end_time = time.time()
                task_duration = task_end_time - task_start_time
                
                result = {
                    "task_id": task["task_id"],
                    "input": task["input"],
                    "output": answer
                }
                results.append(result)
                
                print(f"\nTask {task['task_id']} completed in {task_duration:.2f} seconds")
                print(f"Input: {task['input'][:100]}...")
                print(f"Output: {answer[:100]}...")
                print("-" * 50)
                
                with open(output_file, 'a', encoding='utf-8') as f:
                    json.dump(result, f, ensure_ascii=False)
                    f.write('\n')
                
                if task["task_id"] % 5 == 0:
                    torch.cuda.empty_cache()
                
            except Exception as e:
                print(f"Error processing task {task['task_id']}: {str(e)}")
                continue
        
        total_time = time.time() - start_time
        avg_time = total_time / len(tasks)
        
        summary = {
            "execution_date": execution_date,
            "total_tasks": len(tasks),
            "total_time": round(total_time, 2),
            "average_time_per_task": round(avg_time, 2),
            "model_id": self.model_id,
            "adapter_used": self.adapter_path is not None
        }
        
        print("\nExecution Summary:")
        print(f"Total execution time: {total_time:.2f} seconds")
        print(f"Average time per task: {avg_time:.2f} seconds")
        print(f"Results saved to: {output_file}")
        
        summary_file = output_file.replace('.jsonl', '_summary.json')
        with open(summary_file, 'w', encoding='utf-8') as f:
            json.dump(summary, f, ensure_ascii=False, indent=2)
        
        return results
```
    
```    
from GPUPredictions import GPUPredictions

predictor = GPUPredictions(
    model_id="testmoto/gemma-2-lora-dpo-moe-1"
)
results = predictor.run_inference(
    input_file="elyza-tasks-100-TV_0.jsonl",
    output_file="llm_2024_elyza_tv_0.jsonl"
)
```