IQ-Learn 在 RTX 3090 服务器上的环境配置与踩坑记录

最近在远程服务器上配置 IQ-Learn 的强化学习环境时按照项目的 requirements.txt 直接安装依赖过程中连续遇到了多个老项目兼容性问题。这里把排查和修复过程整理下来方便以后快速复现。1.项目依赖项目的 requirements.txt 如下gym[box2d]0.17.1hydra-core1.0.6stable_baselines31.0tensorboard2.4.0tensorboardX2.1torch1.7.1torchvision0.8.2tornado5.1.1tqdm4.42.1wandbopencv-python4.5.1.48atari-py0.2.6gym_minigrid1.0.2mujoco_py2.0.2项目 README 中说明pytorch 1.4hydra-core1.0hydra-core 1.1 当前不兼容因此真正不能随便升级的是 hydra-core而 torch 虽然写得较老但实际上可以根据 GPU 做适配。2. 原始环境的问题一开始直接执行pip install -r requirements.txt报错主要集中在以下几个方面2.1 mujoco-py 找不到动态库路径报错提示缺少LD_LIBRARY_PATH说明 MuJoCo 路径没有正确导出。2.2 box2d-py 编译失败报错error: command swig failed: No such file or directory说明服务器缺少 swig。2.3 mujoco-py 与 Cython 3.x 不兼容报错涉及 noexcept / Cython.Compiler.Errors.CompileError本质是mujoco-py2.0.2.0不兼容 Cython 3.x2.4 stable_baselines3 / tensorboard / protobuf 冲突报错中出现Descriptors cannot be created directly这是旧 tensorboard 与新 protobuf 不兼容。2.5 tensorboard / stable_baselines3 与 numpy 冲突报错AttributeError: module numpy has no attribute object这是因为旧版 tensorboard 还在使用 np.object。2.6 wandb 版本过新运行 train_iq.py 时wandb 对 Hydra/OmegaConf 配置对象的处理不兼容报错TypeError: first argument must be callable or None2.7 torch 1.7.1 无法驱动 RTX 3090运行训练时出现NVIDIA GeForce RTX 3090 with CUDA capability sm_86 is not compatible with the current PyTorch installation说明 torch1.7.1 太老不支持 sm_86 架构。3. 最终思路单独创建适配 3090 的新环境为了兼容 RTX 3090我没有继续使用原来的 torch1.7.1而是新建了一个环境 irl核心策略是保留项目真正敏感的老依赖hydra-core1.0.6stable_baselines31.0tensorboard2.4.0将 torch / torchvision 升级到支持 3090 的版本torch1.10.2cu113torchvision0.11.3cu113对其他依赖做兼容性回退Cython0.29.37protobuf3.20.3numpy1.23.5wandb0.10.33setuptools59.5.04. 最终安装命令4.1 创建环境conda create -n irl python3.8 -y conda activate irl4.2 安装兼容旧 Hydra/OmegaConf 的 pip 和 setuptoolspip install --no-cache-dir pip24.1 pip install --no-cache-dir setuptools59.5.0说明pip24.1 会拒绝 omegaconf 2.0.x 的旧元数据格式较新的 setuptools 会让旧版 tensorboard/torch.utils.tensorboard 出现 distutils 兼容问题4.3 安装编译工具conda install -n irl -c conda-forge swig cmake4 make -y说明swig 用于 box2d-pycmake4 用于兼容 atari-py0.2.64.4 安装支持 RTX 3090 的 PyTorchpip install --no-cache-dir \ torch1.10.2cu113 \ torchvision0.11.3cu113 \ --extra-index-url https://download.pytorch.org/whl/cu1134.5 安装主依赖pip install --no-cache-dir \ gym[box2d]0.17.1 \ hydra-core1.0.6 \ stable_baselines31.0 \ tensorboard2.4.0 \ tensorboardX2.1 \ tornado5.1.1 \ tqdm4.42.1 \ opencv-python4.5.1.48 \ gym_minigrid1.0.2 \ termcolor4.6 安装兼容性固定版本pip install --no-cache-dir \ Cython0.29.37 \ numpy1.23.5 \ protobuf3.20.3 \ wandb0.10.334.7 配置 MuJoCo 路径export MUJOCO_PY_MUJOCO_PATH/home/duweicheng/.mujoco/mujoco200 export LD_LIBRARY_PATH$LD_LIBRARY_PATH:/home/duweicheng/.mujoco/mujoco200/bin为了避免每次登录都重新设置可以写入 ~/.bashrcecho export MUJOCO_PY_MUJOCO_PATH/home/duweicheng/.mujoco/mujoco200 ~/.bashrc echo export LD_LIBRARY_PATH$LD_LIBRARY_PATH:/home/duweicheng/.mujoco/mujoco200/bin ~/.bashrc source ~/.bashrc4.8 安装 Box2D、MuJoCo、Atari 相关依赖先安装 box2d-pypip install --no-cache-dir box2d-py2.3.8再安装 mujoco-py 的构建依赖pip install --no-cache-dir \ cffi1.17.1 \ glfw2.10.0 \ imageio2.35.1 \ lockfile0.12.2安装 mujoco-py 时关闭构建隔离pip install --no-cache-dir --no-build-isolation mujoco-py2.0.2.0最后安装 Ataripip install --no-cache-dir atari-py0.2.65. 如何验证安装成功5.1 验证 PyTorch 和 3090 是否可用python -c import torch; print(torch:, torch.__version__); print(cuda available:, torch.cuda.is_available()); print(device 0:, torch.cuda.get_device_name(0))以及做一次真实的 CUDA 计算python -c import torch; xtorch.randn(1024,1024, devicecuda); ytorch.randn(1024,1024, devicecuda); zxy; print(cuda matmul ok:, z.shape)如果这两步通过说明PyTorch 版本支持 3090CUDA 驱动和当前 wheel 匹配GPU 计算是可用的5.2 验证主要依赖导入成功python -c from stable_baselines3 import PPO; print(sb3 ok) python -c from torch.utils.tensorboard import SummaryWriter; print(tensorboard ok) python -c import mujoco_py; print(mujoco_py ok) python -c import atari_py; print(atari_py ok)5.3 运行训练命令CUDA_VISIBLE_DEVICES1 python train_iq.py envhopper agentsac expert.demos1 method.lossv0 method.regularizeTrue agent.actor_lr3e-5 seed0如果能进入训练流程就说明整个环境已经可用。6. 这次实际解决了什么问题本次环境配置最终解决了以下问题mujoco-py 缺少 LD_LIBRARY_PATHbox2d-py 需要 swigmujoco-py 与 Cython 3.x 不兼容tensorboard 2.4.0 与 protobuf 5.x 不兼容tensorboard 2.4.0 与 numpy 1.24 不兼容过新 wandb 与旧版 Hydra/OmegaConf 配置对象不兼容torch 1.7.1 不支持 RTX 3090 (sm_86)新版 pip 不兼容旧 omegaconf 2.0.x新版 setuptools 导致旧版 torch.utils.tensorboard 的 distutils 兼容问题atari-py 与 cmake 4.x 不兼容项目还缺少 termcolor 这个未写入 requirements.txt 的依赖7. 最终可用环境版本下面是最终验证通过的核心版本组合python3.8 pip24.0 setuptools59.5.0 torch1.10.2cu113 torchvision0.11.3cu113 gym0.17.1 hydra-core1.0.6 stable_baselines31.0 tensorboard2.4.0 tensorboardX2.1 tornado5.1.1 tqdm4.42.1 opencv-python4.5.1.48 gym_minigrid1.0.2 atari-py0.2.6 mujoco-py2.0.2.0 box2d-py2.3.8 Cython0.29.37 numpy1.23.5 protobuf3.20.3 wandb0.10.33 setuptools59.5.0 termcolor*8. 最终环境打印结果安装并验证通过后环境关键输出如下torch: 1.10.2cu113 cuda available: True device 0: NVIDIA GeForce RTX 3090 cuda matmul ok: torch.Size([1024, 1024]) sb3 ok tensorboard ok mujoco_py ok atari_py ok这说明GPU 可用3090 可被当前 PyTorch 正常驱动stable_baselines3 可导入tensorboard 可导入mujoco_py 可导入atari_py 可导入环境已经满足 IQ-Learn 在 MuJoCo、Atari 等任务上的运行需求。9. 整理好的完整bash指令#!/usr/bin/env bashset -euo pipefailENV_NAMEirlPYTHON_VERSION3.8MUJOCO_PATH/home/duweicheng/.mujoco/mujoco200echo Creating conda environment: ${ENV_NAME}conda create -n ${ENV_NAME} python${PYTHON_VERSION} -yecho Activating environment: ${ENV_NAME}source $(conda info --base)/etc/profile.d/conda.shconda activate ${ENV_NAME}echo Downgrading pip for old hydra/omegaconf compatibilitypip install --no-cache-dir pip24.1pip install --no-cache-dir setuptools59.5.0echo Installing build toolsconda install -n ${ENV_NAME} -c conda-forge swig cmake4 make -yecho Installing PyTorch for RTX 3090pip install --no-cache-dir \torch1.10.2cu113 \torchvision0.11.3cu113 \--extra-index-url https://download.pytorch.org/whl/cu113echo Installing base project dependenciespip install --no-cache-dir \gym[box2d]0.17.1 \hydra-core1.0.6 \stable_baselines31.0 \tensorboard2.4.0 \tensorboardX2.1 \tornado5.1.1 \tqdm4.42.1 \opencv-python4.5.1.48 \gym_minigrid1.0.2 \termcolorecho Installing compatibility-pinned packagespip install --no-cache-dir \Cython0.29.37 \numpy1.23.5 \protobuf3.20.3 \wandb0.10.33echo Exporting MuJoCo environment variables for current sessionexport MUJOCO_PY_MUJOCO_PATH${MUJOCO_PATH}export LD_LIBRARY_PATH${LD_LIBRARY_PATH:-}:${MUJOCO_PATH}/binecho Checking MuJoCo pathtest -d ${MUJOCO_PATH}/binecho Installing Box2Dpip install --no-cache-dir box2d-py2.3.8echo Installing explicit MuJoCo build depspip install --no-cache-dir \cffi1.17.1 \glfw2.10.0 \imageio2.35.1 \lockfile0.12.2echo Installing mujoco-py without build isolationpip install --no-cache-dir --no-build-isolation mujoco-py2.0.2.0echo Installing Atari dependencycmake --versionpip install --no-cache-dir atari-py0.2.6echo Persisting MuJoCo environment variables to ~/.bashrc if missinggrep -qxF export MUJOCO_PY_MUJOCO_PATH${MUJOCO_PATH} ~/.bashrc || \echo export MUJOCO_PY_MUJOCO_PATH${MUJOCO_PATH} ~/.bashrcgrep -qxF export LD_LIBRARY_PATH\$LD_LIBRARY_PATH:${MUJOCO_PATH}/bin ~/.bashrc || \echo export LD_LIBRARY_PATH\$LD_LIBRARY_PATH:${MUJOCO_PATH}/bin ~/.bashrcecho Verifying installationpython -c import torch; print(torch:, torch.__version__); print(cuda available:, torch.cuda.is_available()); print(device 0:, torch.cuda.get_device_name(0))python -c import torch; xtorch.randn(1024,1024, devicecuda); ytorch.randn(1024,1024, devicecuda); zxy; print(cuda matmul ok:, z.shape)python -c from stable_baselines3 import PPO; print(sb3 ok)python -c from torch.utils.tensorboard import SummaryWriter; print(tensorboard ok)python -c import mujoco_py; print(mujoco_py ok)python -c import atari_py; print(atari_py ok)echo Doneecho Activate with: conda activate ${ENV_NAME}echo Run Hopper training with:echo CUDA_VISIBLE_DEVICES1 python train_iq.py envhopper agentsac expert.demos1 method.lossv0 method.regularizeTrue agent.actor_lr3e-5 seed010.现有的环境包Package Version---------------------- ------------absl-py 2.3.1antlr4-python3-runtime 4.8atari-py 0.2.6box2d-py 2.3.8cachetools 4.2.4certifi 2026.4.22cffi 1.17.1charset-normalizer 3.4.7click 8.1.8cloudpickle 1.3.0configparser 7.1.0contourpy 1.1.1cycler 0.12.1Cython 0.29.37docker-pycreds 0.4.0fonttools 4.57.0future 1.0.0gitdb 4.0.12GitPython 3.1.49glfw 2.10.0google-auth 1.35.0google-auth-oauthlib 0.4.6grpcio 1.70.0gym 0.17.1gym-minigrid 1.0.2hydra-core 1.0.6idna 3.13imageio 2.35.1importlib_metadata 8.5.0importlib_resources 6.4.5kiwisolver 1.4.7lockfile 0.12.2Markdown 3.7MarkupSafe 2.1.5matplotlib 3.7.5mujoco-py 2.0.2.0numpy 1.23.5nvidia-ml-py 13.590.48nvitop 1.6.2oauthlib 3.3.1omegaconf 2.0.6opencv-python 4.5.1.48packaging 26.2pandas 2.0.3pathtools 0.1.2pillow 10.4.0pip 24.0promise 2.3protobuf 3.20.3psutil 7.2.2pyasn1 0.6.3pyasn1_modules 0.4.2pycparser 2.23pyglet 1.5.0pyparsing 3.1.4python-dateutil 2.9.0.post0pytz 2026.2PyYAML 6.0.3requests 2.32.4requests-oauthlib 2.0.0rsa 4.9.1scipy 1.10.1sentry-sdk 2.58.0setuptools 59.5.0shortuuid 1.0.13six 1.17.0smmap 5.0.3stable-baselines3 1.0subprocess32 3.5.4tensorboard 2.4.0tensorboard-plugin-wit 1.8.1tensorboardX 2.1termcolor 2.4.0torch 1.10.2cu113torchvision 0.11.3cu113tornado 5.1.1tqdm 4.42.1typing_extensions 4.13.2tzdata 2026.2urllib3 2.2.3wandb 0.10.33Werkzeug 3.0.6wheel 0.44.0zipp 3.20.2hrefblob:https://mp.csdn.net/9b283dfd-623b-4845-afc2-e02622ff2c66 relstylesheet /