%%capture
!pip install -q unsloth
!pip uninstall unsloth -y && pip -q install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install -qU torch
!pip install -qU peft

python

from unsloth import FastLanguageModel
from peft import PeftModel
import torch
import json
from tqdm import tqdm
import re

2. Specify some configs

python

model_id = "llm-jp/llm-jp-3-13b"
adapter_id = "siruku6/llm-jp-3-13b-it_lora2"

# Hugging Face Token を指定。
# 下記の URL から Hugging Face Token を取得できますので下記の HF_TOKEN に入れてください。
# https://huggingface.co/settings/tokens
HF_TOKEN = "" #@param {type:"string"}

# unslothのFastLanguageModelで元のモデルをロード。
dtype = None # Noneにしておけば自動で設定
load_in_4bit = True # 今回は13Bモデルを扱うためTrue

3. Load model and adapter

python

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)

4. Set up inputs for inference

Before running the following cords, please upload the file elyza-tasks-100-TV_0.jsonl in /content/ directory.
If there is /content/elyza-tasks-100-TV_0.jsonl, it's OK! Then, run the following cords.

python

datasets = []
with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
    item = ""
    for line in f:
      line = line.strip()
      item += line
      if item.endswith("}"):
        datasets.append(json.loads(item))
        item = ""

5. Run inference and save results in a jsonl

python

# Inference
FastLanguageModel.for_inference(model)

results = []
for dt in tqdm(datasets):
  input = dt["input"]

  prompt = f"""### 指示\n{input}\n### 回答\n"""

  inputs = tokenizer([prompt], return_tensors = "pt").to(model.device)

  outputs = model.generate(**inputs, max_new_tokens = 512, use_cache = True, do_sample=False, repetition_penalty=1.2)
  prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]

  results.append({"task_id": dt["task_id"], "input": input, "output": prediction})


# Save
json_file_id = re.sub(".*/", "", adapter_id)
with open(f"/content/{json_file_id}_output.jsonl", 'w', encoding='utf-8') as f:
    for result in results:
        json.dump(result, f, ensure_ascii=False)
        f.write('\n')

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.