from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

import torch
import json

from datasets import load_dataset
from huggingface_hub import snapshot_download

id = "llm-jp-3-13b-it-bs4-ac10-step370-lora"
lora_path = snapshot_download(repo_id="jaked97/"+ id)
model_id = "models/models--llm-jp--llm-jp-3-13b/snapshots/cd3823f4c1fcbb0ad2e2af46036ab1b0ca13192a"

tasks = load_dataset("json", data_files="./elyza-tasks-100-TV_0.jsonl", split="train")


prompts = [
    f"""### instruction:
あなたは親切なAIアシスタントです。
### input:
{input}
### output:
""" for input in tasks["input"]]

llm = LLM(
    model=model_id,
    gpu_memory_utilization=0.99,
    quantization="bitsandbytes",
    load_format="bitsandbytes",
    trust_remote_code=True,
    enforce_eager=True,
    enable_lora=True,
    max_lora_rank=64,
)

outputs = llm.generate(
    prompts,
    sampling_params = SamplingParams(
        temperature=0,
        max_tokens=1024,
        min_tokens=1,
        repetition_penalty=1.2,
        skip_special_tokens=True,
        seed=97,
    ),
    lora_request=LoRARequest("sql_adapter", 1, lora_path),
)

with open(f"./{id}_max1024-nf4-vllm.jsonl", 'w', encoding='utf-8') as f:
    for i in range(len(outputs)):
        result = {
            "task_id" : tasks[i]["task_id"],
            "input" : tasks[i]["input"],
            "output" : outputs[i].outputs[0].text
        }
        json.dump(result, f, ensure_ascii=False)
        f.write('\n')

Uploaded model

Developed by: jaked97
License: apache-2.0
Finetuned from model : llm-jp/llm-jp-3-13b

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

jaked97/llm-jp-3-13b-it-bs4-ac10-step370-lora

以下は推論用コードです。

Uploaded model

jaked97/
llm-jp-3-13b-it-bs4-ac10-step370-lora