├── LICENSE └── README.md /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2018 Roman Ring 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Optimized TensorFlow Wheels 2 | 3 | If you see similar messages when you start TensorFlow then these wheels are for you! 4 | 5 | ``` 6 | The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations. 7 | The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations. 8 | The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations. 9 | The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations. 10 | ``` 11 | 12 | ## Introduction 13 | 14 | The builds enable various performance flags targeting modern CPUs, including SIMD support (AVX2, SSE4, FMA). 15 | If you have a CPU released after ~2013 then you'll likely benefit from these on e.g. data pre-processing. 16 | 17 | Build also enables [XLA](https://www.tensorflow.org/xla/) - an Accelerated Linear Algebra domain-specific just-in-time compiler. 18 | 19 | Additional compute capabilities (5.0, 6.1, 7.0, 7.5) are enabled, meaning the wheels should work well on a wide range of GPUS: from `GTX 7xx` to `RTX 20xx` families. 20 | 21 | ## Available Wheels 22 | 23 | |TensorFlow|Python|CUDA|CuDNN|TensorRT|NCCL|Compute Capability|OS|Link| 24 | |---:|---:|---:|---:|---:|---:|---:|:---:|:---:| 25 | |2.1.0|3.8|10.2|7.6|7.0|2.5|5.0,6.1,7.0,7.5|Linux|[tensorflow-2.1.0-cp38-cp38-linux_x86_64.whl](https://github.com/inoryy/tensorflow-optimized-wheels/releases/download/v2.1.0/tensorflow-2.1.0-cp38-cp38-linux_x86_64.whl)| 26 | |2.1.0|3.7|10.2|7.6|7.0|2.5|5.0,6.1,7.0,7.5|Linux|[tensorflow-2.1.0-cp37-cp37m-linux_x86_64.whl](https://github.com/inoryy/tensorflow-optimized-wheels/releases/download/v2.1.0/tensorflow-2.1.0-cp37-cp37m-linux_x86_64.whl)| 27 | |2.0.0|3.8|10.2|7.6|N/A|2.5|5.0,6.1,7.0|Linux|[tensorflow-2.0.0-cp38-cp38-linux_x86_64.whl](https://github.com/inoryy/tensorflow-optimized-wheels/releases/download/v2.0.0-py3.8/tensorflow-2.0.0-cp38-cp38-linux_x86_64.whl)| 28 | |2.0.0|3.7|10.1|7.5|N/A|2.4|5.0,6.1,7.0|Linux|[tensorflow-2.0.0-cp37-cp37m-linux_x86_64.whl](https://github.com/inoryy/tensorflow-optimized-wheels/releases/download/v2.0.0/tensorflow-2.0.0-cp37-cp37m-linux_x86_64.whl)| 29 | 30 | ## Installation 31 | 32 | Assuming you have all the requirements, you can install the wheel directly via pip: 33 | 34 | ``` 35 | pip install https://github.com/inoryy/tensorflow-optimized-wheels/releases/download/v2.1.0/tensorflow-2.1.0-cp37-cp37m-linux_x86_64.whl 36 | ``` 37 | And verify the installation (notice no warning messages): 38 | 39 | ``` 40 | $ python 41 | Python 3.8.0 | packaged by conda-forge | (default, Nov 22 2019, 19:11:38) 42 | [GCC 7.3.0] :: Anaconda, Inc. on linux 43 | >>> import tensorflow as tf 44 | I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.2 45 | >>> tf.__version__ 46 | '2.1.0' 47 | >>> tf.executing_eagerly() 48 | True 49 | >>> tf.constant([123]) + tf.constant([321]) 50 | I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcuda.so.1 51 | ... 52 | I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7 53 | I tensorflow/core/common_runtime/gpu/gpu_device.cc:1241] Created TensorFlow device (...) -> physical GPU (...) 54 | 55 | ``` 56 | 57 | ## Benchmark 58 | 59 | The wheels are benchmarked by training an MNIST model from [TF Models](https://github.com/tensorflow/models) on a CPU. Results for TF 2.1 are as follows: 60 | 61 | | Build / Time Per Epoch |Mean|Min|Max| 62 | |---:|---:|---:|---:| 63 | | Official | 16.7s | 16s | 19s | 64 | | Optimized | 14.3s | 12s | 17s | 65 | 66 | ## Requests 67 | 68 | If you need a different TensorFlow / CUDA / CuDNN / Python combination feel free to open a GitHub ticket. 69 | --------------------------------------------------------------------------------