Views
No views yet
1# ライブラリのインストール
2!pip install transformers==4.46.3 accelerate==1.1.11# モデル・アダプター・トークナイザーの読み込み
2import torch
3from transformers import (
4 AutoModelForCausalLM,
5 AutoTokenizer,
6)
7
8HF_TOKEN = "{your_hf_token}"
9model_id = "google/gemma-2-9b"
10adapter_id = "miya-99999/matsuo_llm_exp07_ckp6000"
11device = "cuda" if torch.cuda.is_available() else "cpu"
12
13tokenizer = AutoTokenizer.from_pretrained(adapter_id, token=HF_TOKEN)
14model = AutoModelForCausalLM.from_pretrained(
15 model_id,
16 device_map=device,
17 token=HF_TOKEN,
18 torch_dtype=torch.bfloat16,
19 use_cache=True
20)
21
22model.load_adapter(adapter_id, token=HF_TOKEN)1# テストデータの読み込み
2import json
3datasets = []
4with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
5 item = ""
6 for line in f:
7 line = line.strip()
8 item += line
9 if item.endswith("}"):
10 datasets.append(json.loads(item))
11 item = ""1# 推論
2from tqdm import tqdm
3results = []
4for data in tqdm(datasets):
5
6 input = data["input"]
7
8 chat = [
9 {"role": "user", "content": input},
10 ]
11
12 tokenized_input = tokenizer.apply_chat_template(chat, add_generation_prompt=True, tokenize=True, return_tensors="pt").to(model.device)
13 with torch.no_grad():
14 outputs = model.generate(
15 tokenized_input,
16 max_new_tokens=512,
17 do_sample=False,
18 repetition_penalty=1.2,
19 pad_token_id=tokenizer.pad_token_id,
20 eos_token_id=[1,107],
21 )[0]
22 output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
23 print("入力：", input)
24 print("出力：", output)
25 print("\n", "="*50, "\n")
26
27 results.append({"task_id": data["task_id"], "input": input, "output": output})1# 推論結果の保存
2import re
3jsonl_id = re.sub(".*/", "", adapter_id)
4with open(f"./{jsonl_id}-outputs.jsonl", 'w', encoding='utf-8') as f:
5 for result in results:
6 json.dump(result, f, ensure_ascii=False) # ensure_ascii=False for handling non-ASCII characters
7 f.write('\n')