版本CUDA(测试版)cuDNNPythonKeras关键变化2.2112.5129.33.10–3.13Keras 3起不再支持 Python 3.9numpy 2.x2.2012.5129.33.9–3.13Keras 3numpy 2.x2.1912.5129.33.9–3.12Keras 3numpy 2.x2.1812.5129.33.9–3.12Keras 3默认编译 NumPy 2.0仍兼容 1.26Hermetic CUDATensorRT 停用2.1712.3128.93.9–3.12Keras 3numpy 1.26.x2.1612.3128.93.9–3.12Keras 3Keras 3 成为默认、tf.estimator 移除、numpy 1.26.x2.1512.2128.93.9–3.11Keras 22.1411.8118.73.9–3.11Keras 2pip install tensorflow[and-cuda]2.1311.8118.63.8–3.11Keras 2F1Score Mesh Lion CosineDecay2.1211.8118.63.8–3.11Keras 2tf.keras.utils.FeatureSpace2.1111.2118.13.7–3.10Keras 2支持 WSL2原生 Windows 起不再支持 GPU 编译2.1011.2118.13.7–3.10Keras 2最后一个支持原生 Windows GPU 的版本## tensorflow # Mac gpu with Apple silicon or AMD GPU: # 最新只支持到2.18.0支持力度参考此网页的更新 # https://pypi.org/project/tensorflow-metal/#description !pip install --upgrade tensorflow2.18.0 2.16使用tensorflow-macos !pip install --upgrade tensorflow-meta1.2.0 # 最新2.18 macOS15.0 看这里(最新支持到tf2.21.0) # https://github.com/IPNP-BIPN/tensorflow-metal-plugin !pip install --upgrade tensorflow tensorflow-metal-plugin import tensorflow as tf # 1 tf.config.list_physical_devices(GPU) # output: [PhysicalDevice(name/physical_device:GPU:0, device_typeGPU)] # 2 x tf.constant([1,2,3]) print(x.device) # output: /job:localhost/replica:0/task:0/device:GPU:0 # Mac gpu with Intel: # https://intel.github.io/intel-extension-for-tensorflow/latest/get_started.html## pytorch # 1.12 开始原生支持苹果m芯片mps import torch # 1 torch.mps.is_available() # 2 x torch.as_tensor([1,2,3]).to(mps) print(x.device) # mps:0# 已知兼容版本最好的组合历史 numpy 1.26.4 pandas 1.5.3 pyspark 3.3.1# linux conda #已验证可用搭配(主要是conda上的cuda库更新太慢了) conda install cudatoolkit11.8.0 cudnn8.9.7.29 tensorflow2.14.0
tensorflow版本差异总结，Mac GPU使用
版本CUDA(测试版)cuDNNPythonKeras关键变化2.2112.5129.33.10–3.13Keras 3起不再支持 Python 3.9numpy 2.x2.2012.5129.33.9–3.13Keras 3numpy 2.x2.1912.5129.33.9–3.12Keras 3numpy 2.x2.1812.5129.33.9–3.12Keras 3默认编译 NumPy 2.0仍兼容 1.26Hermetic CUDATensorRT 停用2.1712.3128.93.9–3.12Keras 3numpy 1.26.x2.1612.3128.93.9–3.12Keras 3Keras 3 成为默认、tf.estimator 移除、numpy 1.26.x2.1512.2128.93.9–3.11Keras 22.1411.8118.73.9–3.11Keras 2pip install tensorflow[and-cuda]2.1311.8118.63.8–3.11Keras 2F1Score Mesh Lion CosineDecay2.1211.8118.63.8–3.11Keras 2tf.keras.utils.FeatureSpace2.1111.2118.13.7–3.10Keras 2支持 WSL2原生 Windows 起不再支持 GPU 编译2.1011.2118.13.7–3.10Keras 2最后一个支持原生 Windows GPU 的版本## tensorflow # Mac gpu with Apple silicon or AMD GPU: # 最新只支持到2.18.0支持力度参考此网页的更新 # https://pypi.org/project/tensorflow-metal/#description !pip install --upgrade tensorflow2.18.0 2.16使用tensorflow-macos !pip install --upgrade tensorflow-meta1.2.0 # 最新2.18 macOS15.0 看这里(最新支持到tf2.21.0) # https://github.com/IPNP-BIPN/tensorflow-metal-plugin !pip install --upgrade tensorflow tensorflow-metal-plugin import tensorflow as tf # 1 tf.config.list_physical_devices(GPU) # output: [PhysicalDevice(name/physical_device:GPU:0, device_typeGPU)] # 2 x tf.constant([1,2,3]) print(x.device) # output: /job:localhost/replica:0/task:0/device:GPU:0 # Mac gpu with Intel: # https://intel.github.io/intel-extension-for-tensorflow/latest/get_started.html## pytorch # 1.12 开始原生支持苹果m芯片mps import torch # 1 torch.mps.is_available() # 2 x torch.as_tensor([1,2,3]).to(mps) print(x.device) # mps:0# 已知兼容版本最好的组合历史 numpy 1.26.4 pandas 1.5.3 pyspark 3.3.1# linux conda #已验证可用搭配(主要是conda上的cuda库更新太慢了) conda install cudatoolkit11.8.0 cudnn8.9.7.29 tensorflow2.14.0
