Search code examples
pythontensorflowapple-m1

TensorFlow is not using my M1 MacBook GPU during training

I have installed tensorflow-macos and while training this is my CPU usage and GPU usage .

Can I make Tensorflow run on GPU anyway?

Solution

  • I've been setting up my new M1 machine today and was looking for a test such as that provided by Aman Anand already here. It successfully runs on the GPU after following the standard instructions provided in #153 using a miniforge package manager installed using Homebrew and an environment cloned from the YAML file in the #153 guide.

    Process running on GPU

    I also ran the smaller simpler snippet as followed which only runs on CPU, '% GPU' == 0%:

    import numpy as np
import tensorflow as tf

### Aman's code to enable the GPU
#from tensorflow.python.compiler.mlcompute import mlcompute
#tf.compat.v1.disable_eager_execution()
#mlcompute.set_mlc_device(device_name='gpu')
#print("is_apple_mlc_enabled %s" % mlcompute.is_apple_mlc_enabled())
#print("is_tf_compiled_with_apple_mlc %s" % #mlcompute.is_tf_compiled_with_apple_mlc())
#print(f"eagerly? {tf.executing_eagerly()}")
#print(tf.config.list_logical_devices())

x = np.random.random((10000, 5))
y = np.random.random((10000, 2))

x2 = np.random.random((2000, 5))
y2 = np.random.random((2000, 2))

inp = tf.keras.layers.Input(shape = (5,))
l1 = tf.keras.layers.Dense(256, activation = 'sigmoid')(inp)
l1 = tf.keras.layers.Dense(256, activation = 'sigmoid')(l1)
l1 = tf.keras.layers.Dense(256, activation = 'sigmoid')(l1)
l1 = tf.keras.layers.Dense(256, activation = 'sigmoid')(l1)
l1 = tf.keras.layers.Dense(256, activation = 'sigmoid')(l1)
o = tf.keras.layers.Dense(2, activation = 'sigmoid')(l1)

model = tf.keras.models.Model(inputs = [inp], outputs = [o])
model.compile(optimizer = "Adam", loss = "mse")

model.fit(x, y, validation_data = (x2, y2), batch_size = 500, epochs = 500)

    Training not using GPU

    Uncommenting the line's added from Aman's code and rerunning makes the GPU work again:

    Training uses GPU again

    If these scripts still don't use the GPU per the activity monitor (set the update rate to 1s in view/update_frequency), go back to the #153 page to start agin from a clean slate and follow the instructions carefully, and be sure to ignore the instructions meant for Intel/X86.

    My steps:

    1. install xcode (from the app store)
    2. install Homebrew (don't forget to set the PATH as recommended just after installation finishes, terminal then needs restarting or your shell profile reloading)
    3. install miniforge ("brew install miniforge")
    4. copy environment.yaml file and clone as new conda environment with the command given in #153.
    5. profit.

    UPDATE 2022-01-26:

    The workflow to install tensorflow on apple silicon has become much easier in the last 6 months, it still relies on miniforge but the packages are distributed through conda and pip from a standard conda env rather than having to create one from the yaml file. These instructions are very easy to follow and should have you going in under 2 minutes. The only exception being that I had to run one additional command to install openblas afterwards through conda to make it work.

    My test above breaks in tensorflow 2.7 because they changed something to do with the mlcompute location for m1 but go on to say that the mlcompute no longer needs to instruct the use of the GPU with the Metal plugin and the test works again by simply removing the references to mlcompute in lines 5-10, and runs on the GPU as can be seen in activity monitor.