背景
LlamaFactory 的 LoRA 微调功能非常便捷，微调后的模型，没有直接支持 vllm 推理，故导致推理速度不够快。
LlamaFactory 目前支持通过 VLLM API 进行部署，调用 API 时的响应速度，仍然没有vllm批量推理的速度快。
如果模型是通过 LlamaFactory 微调的，为了确保数据集的一致性，建议在推理时也使用 LlamaFactory 提供的封装数据集。
简介
在上述的背景下，我们使用 LlamaFactory 原生数据集，支持 lora的 vllm 批量推理。
完整代码如下：
import json
import os
from typing import Listfrom vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequestfrom llamafactory.data import get_dataset, get_template_and_fix_tokenizer
from llamafactory.extras.constants import IGNORE_INDEX
from llamafactory.hparams import get_train_args
from llamafactory.model import load_tokenizerdef vllm_infer():model_args, data_args, training_args, finetuning_args, generating_args = (get_train_args())tokenizer = load_tokenizer(model_args)["tokenizer"]template = get_template_and_fix_tokenizer(tokenizer, data_args)eval_dataset = get_dataset(template, model_args, data_args, training_args, finetuning_args.stage, tokenizer)["eval_dataset"]prompts = [item["input_ids"] for item in eval_dataset]prompts = tokenizer.batch_decode(prompts, skip_special_tokens=False)labels = [list(filter(lambda x: x != IGNORE_INDEX, item["labels"]))for item in eval_dataset]labels = tokenizer.batch_decode(labels, skip_special_tokens=True)sampling_params = SamplingParams(temperature=generating_args.temperature,top_k=generating_args.top_k,top_p=generating_args.top_p,max_tokens=2048,)if model_args.adapter_name_or_path:if isinstance(model_args.adapter_name_or_path, list):lora_requests = []for i, _lora_path in enumerate(model_args.adapter_name_or_path):lora_requests.append(LoRARequest(f"lora_adapter_{i}", i, lora_path=_lora_path))else:lora_requests = LoRARequest("lora_adapter_0", 0, lora_path=model_args.adapter_name_or_path)enable_lora = Trueelse:lora_requests = Noneenable_lora = Falsellm = LLM(model=model_args.model_name_or_path,trust_remote_code=True,tokenizer=model_args.model_name_or_path,enable_lora=enable_lora,)outputs = llm.generate(prompts, sampling_params, lora_request=lora_requests)if not os.path.exists(training_args.output_dir):os.makedirs(training_args.output_dir, exist_ok=True)output_prediction_file = os.path.join(training_args.output_dir, "generated_predictions.jsonl")with open(output_prediction_file, "w", encoding="utf-8") as writer:res: List[str] = []for text, pred, label in zip(prompts, outputs, labels):res.append(json.dumps({"prompt": text, "predict": pred.outputs[0].text, "label": label},ensure_ascii=False,))writer.write("\n".join(res))
vllm.yaml 示例:
## model
model_name_or_path: qwen/Qwen2.5-7B-Instruct
# adapter_name_or_path: lora模型### method
stage: sft
do_predict: true
finetuning_type: lora### dataset
dataset_dir: 数据集路径
eval_dataset: 数据集
template: qwen
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16### output
output_dir: output/
overwrite_output_dir: true### eval
predict_with_generate: true
程序调用:
python vllm_infer.py vllm.yaml
程序运行速度：
Processed prompts: 100%|█| 1000/1000 [01:56<00:00, 8.60it/s, est. speed input: 5169.35 toks/s, output: 811.57
总结
本方案在原生 LlamaFactory 数据集的基础上，支持 LoRA 的 vllm 批量推理，能提升了推理效率。
进一步阅读
如果微调模型后，发现使用vllm模型批量效果不太好，可以参考下述文章：
- 基于 LLamafactory 的异步API高效调用实现与速度对比.https://blog.csdn.net/sjxgghg/article/details/144176645
亲测，LLamafactory 部署 模型，然后使用 Async API 调用后评估效果会好一些。
