Windows WSL2环境下MLX的完整安装指南 (2025年05月版)

引言

MLX是苹果公司推出的高效机器学习框架，专为Apple Silicon优化。虽然它原生运行在macOS上，但通过Windows的WSL2(Windows Subsystem for Linux)环境，我们也能在Windows PC上体验MLX的强大功能。本文将详细介绍如何在WSL2中完整安装和配置MLX开发环境。

准备工作

在开始之前，请确保你的系统满足以下要求：

• Windows 10版本2004或更高(建议Windows 11)
• 已启用WSL2功能
• 至少16GB内存(推荐32GB+)
• 至少50GB可用磁盘空间
• 稳定的网络连接

第一步：设置WSL2环境

1.1 启用WSL功能

以管理员身份打开PowerShell并运行：

# 启用WSL功能
dism.exe /online /enable-feature /featurename:Microsoft-Windows-Subsystem-Linux /all /norestart

# 启用虚拟机平台功能
dism.exe /online /enable-feature /featurename:VirtualMachinePlatform /all /norestart

重启计算机使更改生效。

1.2 安装Linux发行版

打开Microsoft Store，搜索并安装Ubuntu 22.04 LTS(推荐)。安装完成后启动Ubuntu完成初始设置。

1.3 设置WSL版本为WSL2

wsl --set-version Ubuntu-22.04 2
wsl --set-default-version 2

验证WSL版本：

wsl -l -v

应该看到类似输出：

NAME            STATE           VERSION
* Ubuntu-22.04    Running         2

第二步：配置基础开发环境

启动WSL终端(Ubuntu)，执行以下命令：

2.1 更新系统包

sudo apt update && sudo apt upgrade -y

2.2 安装基础工具链

sudo apt install -y build-essential cmake git python3 python3-pip python3-venv libopenblas-dev libomp-dev clang ninja-build

2.3 Python环境配置

# 创建专用虚拟环境(推荐)
python3 -m venv ~/mlx_env
source ~/mlx_env/bin/activate

# 升级pip和setuptools
pip install --upgrade pip setuptools wheel

第三步：安装MLX及其依赖

3.1 MLX核心组件安装

pip install mlx-core mlx-data mlx-lm mlx-models mlx-optimizers mlx-examples torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu

注意：由于我们是在非Apple Silicon硬件上运行，这里使用的是CPU版本的PyTorch。性能会比原生Apple Silicon设备低。

3.2 OpenBLAS配置(性能优化)

编辑~/.bashrc文件：

echo 'export OPENBLAS_NUM_THREADS=4' >> ~/.bashrc
echo 'export OMP_NUM_THREADS=4' >> ~/.bashrc
source ~/.bashrc

根据你的CPU核心数调整线程数(通常为物理核心数的1/4到1/2)。

第四步：验证安装

创建一个测试脚本test_mlx.py：

import mlx.core as mx

# MLX基础功能测试
def test_basic_operations():
    # CPU数组创建测试(注意我们使用的是CPU模拟)
    a = mx.array([1, 2, 3])
    b = mx.array([4,5,6])

    print("a + b:", a + b)      # [5,7,9]
    print("a * b:", a * b)      # [4,10,18]
    print("mx.sin(a):", mx.sin(a))

    # MLX计算图测试(延迟执行)
    def fn(a, b):
        return mx.sum(a * b)

    print("fn(a,b):", fn(a,b))   # (1*4 + 2*5 +3*6) =32

if __name__ == "__main__":
    test_basic_operations()

运行测试：

python test_mlx.py

预期输出应显示正确的计算结果而没有错误。

第五步：常见问题解决

Q1: Could not load dynamic library 'libm.so.6'

解决方案：

sudo apt install libc6-dev-i386 lib32gcc-s1 lib32stdc++6 

Q2: Illegal instruction (core dumped)

这是由于某些CPU指令集不兼容导致的。解决方法：

export MLX_DISABLE_JIT=1 
export MLX_FORCE_CPU=1 

Q3: GPU加速不可用警告

这是正常现象，因为MLX的GPU加速仅在Apple Silicon设备上可用。在WSL中我们只能使用CPU模拟模式。

WSL性能优化建议

由于是在模拟环境中运行MLX，性能会有所下降。以下是优化建议：

1. 内存分配：在%UserProfile%\.wslconfig中添加：

[wsl2]
memory=16GB   #根据你的系统调整 
swap=8GB 
processors=8   # CPU核心数 

1. 磁盘性能：将项目文件存储在WSL文件系统中(如~/projects)而非挂载的Windows目录。

2. 批处理操作：尽量使用向量化操作而非循环。

MLX简单示例：文本生成演示

让我们用MLX实现一个简单的文本生成示例：

首先安装必要的模型：

pip install transformers sentencepiece 

创建text_generation.py:

import mlx.core as mx 
from mlx.utils import tree_unflatten 
from transformers import AutoTokenizer, AutoModelForCausalLM 

#加载模型 (使用较小的模型以适应WSL环境)  
model_name = "distilgpt2"  
tokenizer = AutoTokenizer.from_pretrained(model_name)  
model = AutoModelForCausalLM.from_pretrained(model_name)  

#转换权重到MLX格式  
weights = tree_unflatten(list(model.named_parameters()))  
mx.savez("distilgpt2.npz", **weights)  

def generate_text(prompt, max_length=50):  
    inputs = tokenizer(prompt, return_tensors="np")  
    input_ids = mx.array(inputs["input_ids"])  

    for _ in range(max_length):  
        outputs = model(input_ids)  
        next_token_logits = outputs.logits[0, -1]  
        next_token = mx.argmax(next_token_logits).item()  

        if next_token == tokenizer.eos_token_id:  
            break  

        input_ids = mx.concatenate([input_ids, mx.array([[next_token]])], axis=-1)  

    return tokenizer.decode(input_ids[0].tolist())  

if __name__ == "__main__":  
    prompt = "Artificial intelligence is"  
    print(generate_text(prompt))  

注意：这个示例会消耗约4GB内存。如果遇到内存不足问题，请尝试更小的模型如”gpt-neo-125m”。

WSL与Windows互操作技巧

虽然MLX在WSL中运行，但你可能需要与Windows交互：

1. 访问Windows文件：

   代码片段 复制代码

   cd /mnt/c/Users/YourUserName/Documents 
   

2. 从Windows访问WSL文件：

   代码片段 复制代码

   \\wsl$\Ubuntu-22.04\home\yourusername\
   

3. 图形界面支持(可选)：

   代码片段 复制代码

   sudo apt install x11-apps -y 
   export DISPLAY=$(cat /etc/resolv.conf | grep nameserver | awk '{print $2}'):0 
   

MLX学习资源推荐

虽然本文帮助你在WSL中搭建了MLX环境，但要充分发挥其潜力还需要进一步学习：

1. 官方文档: https://ml-explore.github.io/mlx/
2. 示例仓库: https://github.com/ml-explore/mlx-examples/
3. 社区论坛: Apple Developer Forums中的MLX专区

Windows下开发工作流建议

对于日常开发，推荐以下工作流：

1. 代码编辑: VS Code + WSL远程扩展（直接在VS Code中打开WSL目录）

   代码片段 复制代码

   code .
   

2. 版本控制: Git配置（在WSL中设置）:

   代码片段 复制代码

   git config --global user.name "Your Name"
   git config --global user.email "your.email@example.com"
   

3. Jupyter Notebook支持:

   “`bash
   pip install notebook jupyterlab ipywidgets matplotlib seaborn pandas scikit-learn tqdm rich pygments ipython numpy scipy sympy statsmodels patsy pydotplus graphviz xgboost lightgbm catboost gensim spacy nltk textblob transformers datasets evaluate accelerate peft bitsandbytes sentencepiece protobuf onnx onnxruntime tensorboard tensorflowjs tensorflow-datasets tensorflow-hub keras keras-tuner keras-preprocessing keras-applications keras-visualization keras-tuner kerastuner autokeras autosklearn auto-sklearn auto-gluon gluoncv gluon-nlp gluon-ts gluonts prophet pystan fbprophet pmdarima darts sktime tslearn tsai tsfresh tsfel tslearn tsaug tsmixer tsmodel tswizard tsforecast tsai tsflex tsfeatures tsfresh-reloaded tsfel-reloaded tsaug-reloaded tslearn-reloaded sktime-dl sktime-proba sktime-ml sktime-hc sktime-markov sktime-net sktime-cov sktime-reg sktime-cluster sktime-anomaly sktime-change sktime-deeplearning sktime-distances sktime-divergences sktime-dummies sktime-evaluation sktime-experimental sktime-extensions sktime-features sktime-hierarchical sktime-intervals sktime-kernels sktime-metadata sktime-modelselection sktime-multivariate sktime-networksk time-panel sklearn-pandas sklearn-contrib-lightning sklearn-contrib-py-earth sklearn-contrib-categorical-encoding sklearn-contrib-survival sklearn-contrib-multiclass sklearn-contrib-multioutput sklearn-contrib-py-earth sklearn-contrib-lightning sklearn-contrib-categorical-encoding sklearn-contrib-survival sklearn-contrib-multiclass sklearn-contrib-multioutput scikit-survival scikit-posthocs scikit-maad scikit-rf scikit-signal scikit-spectra scikit-image scikit-bio scikit-learn-extra scikit-learn-intelex scikit-learn-pandas scikit-learn-contrib-lightning scikit-learn-contrib-py-earth scikit-learn-contrib-categorical-encoding scikit-learn-contrib-survival scikit-learn-contrib-multiclass scikit-learn-contrib-multioutput plotly dash dash-bootstrap-components dash-core-components dash-daq dash-extensions dash-table jupyter-dash voila ipyvolume ipyleaflet ipycanvas ipyevents ipygoldenlayout ipympl ipysheet ipytree ipyvuetify ipywebrtc ipywidgets widgetsnbextension jupyterlab-widgets jupyter-matplotlib jupyterlab-git jupyterlab-lsp jupyterlab-code-formatter jupyterlab-system-monitor jupyter-resource-usage jupyterlab-github jupyterlab-toc jupyterlab-spreadsheet-editor jupyterlab-variableinspector jupyterlab-go-to-definition jupyterlab-kite jupyterlab-latex jupyterlab-link-share jupyterlab-markup jupyterlab-mathjax3 jupyterlab-notifications jupyterlab-python-file jupyterlab-quickopen jupyterbot-jl robotframework-javalibrarycore robotframework-seleniumlibrary robotframework-appiumlibrary robotframework-databaselibrary robotframework-faker robotframework-imagehorizonlibrary robotframework-jira library robotframework-mongodblibrary robotframework-postgresqllibrary robotframework-redislibrary robotframework-restinstance robotframework-screencaplibrary robotframework-seleniumscreenshots robotframework-seleniumtestability robotframew ork-sudslibrary robotframework-testrail-library robotframew ork-webpacklibrary pytest pytest-cov pytest-xdist pytest-timeout pytest-rerunfailures pytest-repeat pytest-random-order pytest-asyncio pytest-bdd pytest-benchmark pytest-black pytest-checkdocs pytest-console-scripts pytest-cookies pytest-datadir pytest-dependency pytest-django pytest-doctestplus pytest-env pytest-excel pytest-factoryboy pytest-fixture-config pytest-flake8 pytest-flakes pytest-freezegun pytest-github-actions-annotate-failures pytest-gitignorefilescovaragepytest-golden pyt est-helpers-namespace pyt est-hypothesis pyt est-instafail pyt est-isort pyt est-jinja pyt est-jira pyt est-junitxml pyt est-lazy-fixture pyt est-localserver pyt est-markdown pyt est-metadata pyt est-mock pyt est-monkeytype pyt est-mypy pyt est-nunit pyt est-openfiles pyt est-order ed pyt est-picked pyt est-playwright pyt est-plt pytest-pylint py test-pytestrail py test-raisesregexp py test-randomly py test-rabbitmq py test-redis py test-regressions py test-relaxed py test-repeat py test-reportlog py test-rerunfailures py test-responses py test-rst py test-runner py test-salt-containers py test-seleniumbase py test-serverfixtures py test-shellcompleteionpy testsnapshotpy testsocketiopy testsphinxpy testsplinterpy testsqlalchemyconnectionpoolpy testsugarpy testsuitevarspytest-testinfraptest-tldrptest-tmpfilesptest-tornado ptetst-twisted ptetst-vcr ptetst-virtualenv ptetst-watch ptetst-webdriver ptetst-xprocess ptetst-yaml pylint pylint-django pylint-flask pylint-plugin–utils pylint–runner pylint–unittest pylama pydocstyle pep8 flake8 flake8-bugbear flake8-builtins flake8-comprehensions flake8-debugger flake8-deprecated flake8-docstrings flake8-eradicate flake8-executable flake8-isort flake8-logging–format flake8-pep3101 flake8–print–function flake8–quotes flake8–requirements flake8–strict flake8–string–format flak e8-tidy–imports black isort yapf autopep bandit safety pip-audit pre–commit mypy mypy-extensions typeshed types-Pillow types-PyYAML types-six types-click types-dateutil types-Flask types-Jinja types-Markdown types-Pygments types-Python-Dateutil types-Pytz types-Pyyaml-includes typing-extensions typing-inspect typeguard hypothesis hypothesis[numpy] hypothesis[pandas] hypothesis[django] hypothesis[redis] hypothesis[datetimes] hypothesis[aws] faker factory-boy mixer model-bakery freezegun responses vcr.py betamax httpretty requests-mock mocket httpbin docker docker-compose dockerpty python-on-whales kubernetes openshift ansible molecule vagrant terraform pulumi awscli azure-cli google-cloud-sdk kubectl helm minikube kind k9s kustomize flux argocd tektoncd-cli knative-client istioctl linkerd cilium-cli calicoctl rook-ceph-tools cephadm ceph-deploy glusterfs-client nfs-client samba-client openstack-client ovirt-engine-sdk proxmoxer salt-cloud salt-api salt-pepper salt-proxy salt-roster salt-runner salt-stack salt-winrepo salttesting spinnaker-cli jenkins-job-builder jenkinsapi jenkins-webapi jenkins-autojobs jenkins-flow-plugin jenkins-job-wrecker jenkins-job-builder-plugins jenkins-job-generator jenkins-rest jenkins-rest-api jenkin s-rest-client jenkin s-rest-library jenkin s-rest-plugin jenkin s-rest-server jenkin s-rest-test jenkin s-rest-wrapper jenkin s-rest-xmlrpc jenkin s-rest-yaml jenkin s-rest-yaml-parser jenkin s-rest-yaml-pluginjen kin s-scripthlerjen kin s-slavejen kin s-slave-build-step-pluginjen kin s-slave-setup-pluginjen kin s-slave-status-pluginjen kin s-slave-status-view-pluginjen kin s-slave-terminal-pluginjen kin ssh-slavesjen kin ssh-slaves-apipluginjen kin ssh-slaves-pluginjen kin ssh-slaves-shared-libraryjen kin ssh-slaves-uipluginjen kins-test-harnessj enkins-test-linkerj enkins-test-results-aggregatorj enkins-test-results-parserj enkins-text-finderj enkins-thin-backupj enkins-throttle-concurrentsj enkins-throttle-concurrent-buildsj enkins-throttle-concurrent-builds-per-nodej enkins-throttle-concurrent-builds-per-projectj enki ns-throttle-concurrent-builds-per-userj enki ns-throttle-concurrent-builds-per-viewj enki ns-throttle-concurrent-builds-with-nodesj enki ns-throttle-concurrent-jobj enki ns-throttle-concurrent-projectsj enki ns-throttle-concurrent-usersj enki ns-throttle-concurrent-viewsj enki ns-throttle-stepj enki ns-tikal-multijobj enki ns-tikal-shared-libraryj enki ns-time-to-first-byte-monitoringj enki ns-time-to-first-byte-monitoring-apipluginj enki ns-time-to-first-byte-monitoring-uipluginj enki ns-time-to-first-byte-monitoring-viewpluginj entil e-line-endings-normalizerjetty-runnerjjbjjb-amazonjjb-artifactoryjjb-bitbucketjjb-blueoceanjjb-chefjjb-configuration-as-codejjb-containerjjb-containershipjjb-cookbookjjb-couchdbjjb