Python2.7安装Numba是否影响Ubuntu16.04上正常运行的TensorFlow-GPU？

Great questions! Let's break this down step by step for you:

1. Will installing Numba interfere with TensorFlow or break my CUDA installation? #

In most cases, no—installing Numba won't damage your CUDA setup or interfere with TensorFlow under normal circumstances. Here's why:

• Both Numba and TensorFlow rely on the system-level CUDA Runtime API to access GPU resources, rather than modifying the CUDA installation itself. They're consumers of CUDA, not modifiers.
• The only potential issue is version compatibility: if the CUDA version required by Numba is drastically different from what TensorFlow uses, you might run into runtime errors (like missing libraries). But this won't "break" CUDA—uninstalling the incompatible Numba version or adjusting your environment will fix things.

As long as you pick a Numba version that supports your existing CUDA version, you're safe.

2. Will installing Numba affect my working TensorFlow-GPU on Ubuntu 16.04 + Python 2.7? #

This depends entirely on installing a compatible Numba version—since Python 2.7 is no longer supported by newer Numba releases. Follow these steps to avoid issues:

First: Install the right Numba version #

Numba dropped Python 2.7 support starting from version 0.51.0, so you must install the last compatible release:

pip install numba==0.50.1

Check compatibility with your existing setup #

• First, confirm your TensorFlow-GPU version and its required CUDA/cuDNN:

  import tensorflow as tf
  print(tf.__version__)
  

  For Python 2.7, your TensorFlow is likely a 1.x release (since TensorFlow 2.x doesn't support Python 2.7), which typically works with CUDA 9.0 or 10.0 on Ubuntu 16.04.
• Numba 0.50.1 supports CUDA versions from 7.5 to 10.1, so it will align perfectly with most TensorFlow 1.x setups on Ubuntu 16.04.

Verify coexistence after installation #

Run quick tests to ensure both tools work:

• Test TensorFlow GPU:

  import tensorflow as tf
  with tf.device('/gpu:0'):
      a = tf.constant([1.0, 2.0, 3.0], shape=[3], name='a')
      b = tf.constant([1.0, 2.0, 3.0], shape=[3], name='b')
      c = a + b
  with tf.Session() as sess:
      print(sess.run(c))
  

  If it outputs [2.0 4.0 6.0], TensorFlow is still working.

• Test Numba GPU:

  from numba import cuda
  import numpy as np
  
  @cuda.jit
  def add_kernel(x, y, out):
      idx = cuda.grid(1)
      if idx < x.size:
          out[idx] = x[idx] + y[idx]
  
  x = np.array([1,2,3,4], dtype=np.float32)
  y = np.array([5,6,7,8], dtype=np.float32)
  out = np.empty_like(x)
  
  threads_per_block = 256
  blocks_per_grid = (x.size + threads_per_block - 1) // threads_per_block
  
  add_kernel[blocks_per_grid, threads_per_block](x, y, out)
  print(out)
  

  If it outputs [6. 8. 10. 12.], Numba's GPU acceleration is working.

Using Numba-computed matrices in TensorFlow #

Passing Numba's results to TensorFlow is straightforward:

• If you used Numba with regular numpy arrays (like the example above), you can directly convert them to TensorFlow tensors:

  import tensorflow as tf
  numba_result = np.array([[1,2],[3,4]])
  tf_tensor = tf.convert_to_tensor(numba_result, dtype=tf.float32)
  

• If you used Numba's cuda.device_array (GPU-resident memory), first copy it to host memory with device_array.copy_to_host() before converting to a TensorFlow tensor. While this adds a small data transfer cost, it's the most reliable method in Python 2.7.

Key Tips to Avoid Headaches #

• Never upgrade CUDA randomly: TensorFlow 1.x has strict CUDA/cuDNN version locks. Upgrading CUDA will break your existing TensorFlow setup, and you can't upgrade to TensorFlow 2.x since it doesn't support Python 2.7.
• Install LLVM for Numba: On Ubuntu 16.04, you'll need LLVM 6.0 for Numba 0.50.1 to work properly:

  sudo apt-get install llvm-6.0
  

内容的提问来源于stack exchange，提问作者py study