使用unsloth实现LoRA微调
前置需安装conda与wsl
1.conda使用一条命令
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda.sh ; bash ~/miniconda.sh -b -p $HOME/miniconda ; eval "$($HOME/miniconda/bin/conda shell.bash hook)" ; echo 'export PATH="$HOME/miniconda/bin:$PATH"' >> ~/.bashrc ; source ~/.bashrc2.wsl
打开功能添加子系统 hyperv平台 虚拟机平台
mircosoft下载ubuntu1.conda可以将每个python项目隔离开来，conda环境隔离的是整个Python解释器和所有包，不只是pip安装的包，可以选择对应的python版本
2.windows10里最好用的linux方式就是wsl，在ai这个方向wsl可以轻松调用nvidia的驱动，WSL2实际上是基于轻量级虚拟化技术的，但开销确实很小，接近原生性能
选择在wsl-Linux上进行本地部署微调环境
下载模型
(base)root@Riemann-Plan:/mnt/c/Users/25671#
(base)root@Riemann-Plan:/mnt/c/Users/25671# conda activate huggingface
(huggingface)root@Riemann-Plan:~# export HF_ENDPOINT=https://hf-mirror.com
(huggingface)root@Riemann-Plan:~#
(huggingface)root@Riemann-Plan:~# huggingface-cli download Qwen/Qwen3-8B使用vllm运行模型并本地调用
vLLM：专为生产环境设计的高性能大模型推理引擎，通过PagedAttention技术实现极致的GPU内存利用率和并发处理能力。
vllm会吃满整个显存，他的计算方式为 kvcache 最大上下文 以及模型本身
Ollama：让本地运行大模型变得像使用Docker一样简单，一键下载、运行各种量化模型的开箱即用工具。
(huggingface)root@Riemann-Plan:~# conda env list
# conda environments:
#
base /root/miniconda
LF /root/miniconda/envs/LF
django /root/miniconda/envs/django
docling /root/miniconda/envs/docling
dots_ocr /root/miniconda/envs/dots_ocr
huggingface * /root/miniconda/envs/huggingface
ocrflux /root/miniconda/envs/ocrflux
unsloth-blackwell /root/miniconda/envs/unsloth-blackwell
vllm_tx /root/miniconda/envs/vllm_tx
(huggingface)root@Riemann-Plan:~# conda activate unsloth-blackwell
(unsloth-blackwell)root@Riemann-Plan:~#
(unsloth-blackwell)root@Riemann-Plan:~# cd /root/.cache/huggingface/hub/models--Qwen--Qwen3-8B/
(unsloth-blackwell)root@Riemann-Plan:~/.cache/huggingface/hub/models--Qwen--Qwen3-8B# ls
blobsrefssnapshots
(unsloth-blackwell)root@Riemann-Plan:~/.cache/huggingface/hub/models--Qwen--Qwen3-8B# cd snapshots/
(unsloth-blackwell)root@Riemann-Plan:~/.cache/huggingface/hub/models--Qwen--Qwen3-8B/snapshots# ls
b968826d9c46dd6066d109eabc6255188de91218
(unsloth-blackwell)root@Riemann-Plan:~/.cache/huggingface/hub/models--Qwen--Qwen3-8B/snapshots#
python -m vllm.entrypoints.openai.api_server \
--model /root/.cache/huggingface/hub/models--Qwen--Qwen3-8B/snapshots/b968826d9c46dd6066d109eabc6255188de91218/ \
--served-model-name Qwen3-8B \
--host 0.0.0.0 \
--port 8000 \
--gpu-memory-utilization 0.93 \
--max-model-len 4096 调用模型
curl http://localhost:8000/v1/chat/completions
-H "Content-Type: application/json"
-d '{
"model": "Qwen3-8B",
"messages": [
{"role": "user", "content": "Solve (x + 2)^2 = 0."}
],
"temperature": 0.7,
"max_tokens": 2048
}'
安装unsloth与jupyter lab
pip install -U vllm --extra-index-url https://download.pytorch.org/whl/cu128
pip install unsloth unsloth_zoo bitsandbytes
pip jupyterlab
jupyter lab --allow-root1.部署了模型推理框架
2.部署unsloth微调框架
3.安装了jupyter
传统-xformers-flash attention
介绍一下
方法比喻特点传统把所有资料都摊在桌上准确但占地方，速度慢xformers只看重点资料速度快但可能遗漏信息Flash Att 2一本一本看，用脑子记住又快又准确，还不占地方batch-size
方案A（2×4）：每次吃2口，吃4次，总共8口才消化
问题：第一口就噎死了
方案B（1×8）：每次吃1口，吃8次，总共8口才消化
结果：能正常吃完，虽然吃的次数多了点
安装llama.cpp
sudo apt install libcurl4-openssl-dev
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build
cmake --build build --config Release -j 8启动vllm模型并dify连接
python -m vllm.entrypoints.openai.api_server --model /root/qwen3math/ --served-model-name qwen3math --host 0.0.0.0 --port 8000 --gpu-memory-utilization 0.93 --max-model-len 4096 阿瓦达阿瓦达a```powershell
from unsloth import FastLanguageModel
import torch
这些参数的含义如下：
local_model_path = "/root/.cache/huggingface/hub/models--Qwen--Qwen3-8B/snapshots/b968826d9c46dd6066d109eabc6255188de91218"
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = local_model_path, # 指定要加载的模型名称或路径
max_seq_length = 2048, # 最大上下文长度（输入序列长度），越大越占用显存
load_in_4bit = True, # 是否以4bit量化加载模型，节省显存
load_in_8bit = False, # 是否以8bit量化加载模型，精度更高但显存占用更多
full_finetuning = False, # 是否进行全参数微调，通常LoRA微调时为False
# token = "hf_...", # 如果加载受限模型，需要填写HuggingFace的访问令牌
)
```powershell
model = FastLanguageModel.get_peft_model(
model,
r = 32, # LoRA秩(rank)，控制可训练参数量。32较大，提升表达能力，适合大模型和复杂任务
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj",],
# 需要注入LoRA的模块，覆盖Transformer的主要线性层，提升微调效果
lora_alpha = 32,# LoRA缩放因子，通常设为rank或2*rank，这里与r一致，保证训练稳定
lora_dropout = 0, # LoRA dropout，防止过拟合。设为0可获得最佳性能（Unsloth已优化）
bias = "none", # 是否训练bias参数。设为"none"节省显存且已优化
use_gradient_checkpointing = "unsloth", # 启用Unsloth优化的梯度检查点，节省显存，支持更长上下文
random_state = 3407, # 随机种子，保证实验可复现
use_rslora = False, # 是否使用rank stabilized LoRA，关闭以减少复杂性
loftq_config = None,# 是否使用LoftQ量化，None表示不启用
)
# 这些设置兼顾了训练效率、显存占用和模型能力，适合大多数LoRA微调场景from datasets import load_dataset
reasoning_dataset = load_dataset("unsloth/OpenMathReasoning-mini", split = "cot")
non_reasoning_dataset = load_dataset("mlabonne/FineTome-100k", split = "train")reasoning_datasetnon_reasoning_datasetdef generate_conversation(examples):
# 获取每个样本的问题和生成的解答
problems= examples["problem"]
solutions = examples["generated_solution"]
conversations = []
# 遍历每个问题和对应的解答，构造成对话格式
for problem, solution in zip(problems, solutions):
conversations.append([
{"role" : "user", "content" : problem}, # 用户提问
{"role" : "assistant", "content" : solution}, # 助手回答
])
# 返回包含对话的字典，键为"conversations"
return { "conversations": conversations, }# 将reasoning_dataset中的每个样本转换为对话格式（user/assistant），并应用聊天模板，得到最终的对话文本列表
reasoning_conversations = tokenizer.apply_chat_template(
reasoning_dataset.map(generate_conversation, batched = True)["conversations"],
tokenize = False,
)reasoning_conversations# 导入 Unsloth 的 standardize_sharegpt 函数，将非推理数据集（ShareGPT风格）标准化为统一的多轮对话格式
from unsloth.chat_templates import standardize_sharegpt
dataset = standardize_sharegpt(non_reasoning_dataset)
# 直接提取标准化后的"conversations"字段，并用tokenizer的apply_chat_template方法转换为最终对话文本
# 由于已经是标准多轮对话格式，无需像奥数竞赛数据集那样手动拼接 user/assistant 结构
non_reasoning_conversations = tokenizer.apply_chat_template(
dataset["conversations"],
tokenize = False,
)non_reasoning_conversationsprint(len(reasoning_conversations))
print(len(non_reasoning_conversations))chat_percentage = 0.25import pandas as pd
non_reasoning_subset = pd.Series(non_reasoning_conversations)
non_reasoning_subset = non_reasoning_subset.sample(
int(len(reasoning_conversations)*(chat_percentage/(1 - chat_percentage))),
random_state = 2407,
)
print(len(reasoning_conversations))
print(len(non_reasoning_subset))
print(len(non_reasoning_subset) / (len(non_reasoning_subset) + len(reasoning_conversations)))data = pd.concat([
pd.Series(reasoning_conversations),
pd.Series(non_reasoning_subset)
])
data.name = "text"
from datasets import Dataset
combined_dataset = Dataset.from_pandas(pd.DataFrame(data))
combined_dataset = combined_dataset.shuffle(seed = 3407)from trl import SFTTrainer, SFTConfig
# 创建 SFTTrainer 实例，用于对大语言模型进行指令微调（Supervised Fine-Tuning）
trainer = SFTTrainer(
model = model, # 传入已加载并加装LoRA的模型
tokenizer = tokenizer, # 传入对应的分词器
train_dataset = combined_dataset, # 训练数据集，已混合推理和聊天数据
eval_dataset = None, # 不设置验证集（可选，可后续添加用于评估）
args = SFTConfig(
dataset_text_field = "text", # 指定数据集中用于训练的文本字段名
per_device_train_batch_size = 1, # 每张GPU上的batch size，2较小，适合大模型和显存有限场景
gradient_accumulation_steps = 8, # 梯度累积步数，等效于总batch size=2*4=8，提升训练稳定性
warmup_steps = 5, # 学习率预热步数，防止初始训练不稳定
# num_train_epochs = 1, # 可选：完整训练1轮（此处注释掉，仅演示max_steps用法）
max_steps = 30, # 最多训练30步，加快演示速度（实际训练可增大）
learning_rate = 2e-4, # 初始学习率，短期演示用较大值，长期训练建议降为2e-5
logging_steps = 1, # 每1步记录一次日志，便于观察训练过程
optim = "adamw_8bit", # 使用8bit AdamW优化器，节省显存，适合大模型
weight_decay = 0.01, # 权重衰减，防止过拟合
lr_scheduler_type = "linear", # 线性学习率调度，常用且简单
seed = 3407, # 随机种子，保证实验可复现
report_to = "none", # 不上报到WandB等平台，纯本地训练
),
)
# 总结：本配置兼顾显存、训练速度和实验复现性，适合Colab等资源有限环境下的快速微调演示trainer_stats = trainer.train()messages = [
{"role" : "user", "content" : "Solve (x + 2)^2 = 0."}
]
text = tokenizer.apply_chat_template(
messages,
tokenize = False,
add_generation_prompt = True, # Must add for generation
enable_thinking = False, # Disable thinking
)
from transformers import TextStreamer
_ = model.generate(
**tokenizer(text, return_tensors = "pt").to("cuda"),
max_new_tokens = 256, # Increase for longer outputs!
temperature = 0.7, top_p = 0.8, top_k = 20, # For non thinking
streamer = TextStreamer(tokenizer, skip_prompt = True),
)messages = [
{"role" : "user", "content" : "Solve (x + 2)^2 = 0."}
]
text = tokenizer.apply_chat_template(
messages,
tokenize = False,
add_generation_prompt = True, # Must add for generation
enable_thinking = True, # Disable thinking
)
from transformers import TextStreamer
_ = model.generate(
**tokenizer(text, return_tensors = "pt").to("cuda"),
max_new_tokens = 1024, # Increase for longer outputs!
temperature = 0.6, top_p = 0.95, top_k = 20, # For thinking
streamer = TextStreamer(tokenizer, skip_prompt = True),
)# Merge to 16bit
if True:
model.save_pretrained_merged("qwen3mathmodel", tokenizer, save_method = "merged_16bit",)
if False: # Pushing to HF Hub
model.push_to_hub_merged("hf/model", tokenizer, save_method = "merged_16bit", token = "")
# Merge to 4bit
if False:
model.save_pretrained_merged("model", tokenizer, save_method = "merged_4bit",)
if False: # Pushing to HF Hub
model.push_to_hub_merged("hf/model", tokenizer, save_method = "merged_4bit", token = "")
# Just LoRA adapters
if False:
model.save_pretrained("model")
tokenizer.save_pretrained("model")
if False: # Pushing to HF Hub
model.push_to_hub("hf/model", token = "")
tokenizer.push_to_hub("hf/model", token = "")
来源：程序园用户自行投稿发布，如果侵权，请联系站长删除
免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！
页:
[1]