1from transformers import AutoTokenizer, AutoModelForCausalLM
2import torch
3
4tokenizer = AutoTokenizer.from_pretrained("lightblue/karasu-7B-chat-plus")
5model = AutoModelForCausalLM.from_pretrained("lightblue/karasu-7B-chat-plus", torch_dtype=torch.bfloat16, device_map="auto")
6
7pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
8
9messages = [{"role": "system", "content": "あなたはAIアシスタントです。"}]
10messages.append({"role": "user", "content": "イギリスの首相は誰ですか？"})
11
12prompt = tokenizer.apply_chat_template(conversation=messages, add_generation_prompt=True, tokenize=False)
13
14pipe(prompt, max_new_tokens=100, do_sample=False, temperature=0.0, return_full_text=False)
1from vllm import LLM, SamplingParams
2
3sampling_params = SamplingParams(temperature=0.0, max_tokens=100)
4llm = LLM(model="lightblue/karasu-7B-chat-plus")
5
6messages = [{"role": "system", "content": "あなたはAIアシスタントです。"}]
7messages.append({"role": "user", "content": "イギリスの首相は誰ですか？"})
8prompt = llm.llm_engine.tokenizer.apply_chat_template(conversation=messages, add_generation_prompt=True, tokenize=False)
9prompts = [prompt]
10
11outputs = llm.generate(prompts, sampling_params)
12for output in outputs:
13 prompt = output.prompt
14 generated_text = output.outputs[0].text
15 print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
