tf.device scope not working correctly

[JohnTaylor2000]

Please make sure that this is a bug. As per our
GitHub Policy,
we only address code/doc bugs, performance issues, feature requests and
build/installation issues on GitHub. tag:bug_template

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Y es
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): ppc64le-linux
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device:
• TensorFlow installed from (source or binary)
• TensorFlow version (use command below):
• Python version: python3.7
• Bazel version (if compiling from source):
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version:10.1.243
• GPU model and memory: V100 16GB

You can collect some of this information using our environment capture
script
You can also obtain the TensorFlow version with:

1. TF 1.0: python -c "import tensorflow as tf; print(tf.GIT_VERSION, tf.VERSION)"
2. TF 2.0: python -c "import tensorflow as tf; print(tf.version.GIT_VERSION, tf.version.VERSION)"

Describe the current behavior

tf.device command does not correctly assign a GPU device to tf.keras layers on node with 4 GPUs so cannot implement model parallelism. All layers appear on device GPU:0 with the exception of some IO based on output of tf.debugging.set_log_device_placement(True)

Describe the expected behavior
tf.keras layers are correctly assigned to a device.

Standalone code to reproduce the issue

import tensorflow as tf
from tensorflow import keras

tf.debugging.set_log_device_placement(True)

print("On GPU:1")
inputs = keras.Input(shape=(784,))
with tf.device("/device:GPU:1"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc.
x = keras.layers.Dense(256, activation="relu")(inputs)
print(x)
assert x.device.endswith("/GPU:1")

Provide a reproducible test case that is the bare minimum necessary to generate
the problem. If possible, please share a link to Colab/Jupyter/any notebook.

I have a larger test problem that will run on a 4 GPU node. If you turn off the assert statement, then using nvidia-smi you can see that all memory and computational work is happening on GPU:0 and almost none is assigned to other GPUs. Happy to supply this code if needed.

Other info / logs Include any logs or source code that would be helpful to
diagnose the problem. If including tracebacks, please include the full
traceback. Large logs and files should be attached.

Tensor("dense/Identity:0", shape=(None, 256), dtype=float32)
Traceback (most recent call last):
File "py_test.py", line 11, in
assert x.device.endswith("/GPU:1")
AssertionError

[ravikyram]

@JohnTaylor2000

Will it be possible to share complete code snippet .Please, share the output of tf.config.list_physical_devices().Thanks!

[JohnTaylor2000]

NOTE : uses input data file mnist.npz

import tensorflow as tf
from tensorflow import keras

import time
import socket

import os

tf.debugging.set_log_device_placement(True)

print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))
gpus = tf.config.list_physical_devices('GPU')
print(' gpus = ', gpus)

def get_compiled_model():
    # Make a simple 2-layer densely-connected neural network.
    inputs = keras.Input(shape=(784,))
    print("On GPU:0")
    with tf.device("/device:GPU:0"):
       x = keras.layers.Dense(256, activation="relu")(inputs)
       #assert x.device.endswith("/GPU:0")
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
    print("On GPU:1")
    with tf.device("/device:GPU:1"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc.
       x = keras.layers.Dense(256, activation="relu")(inputs)
       #assert x.device.endswith("/GPU:1")
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
    print("On GPU:2")
    with tf.device("/device:GPU:2"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc.
       x = keras.layers.Dense(256, activation="relu")(inputs)
       #assert x.device.endswith("/GPU:2")
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
       x = keras.layers.Dense(256, activation="relu")(inputs)
    print("On GPU:3")
    with tf.device("/device:GPU:3"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc.
      x = keras.layers.Dense(256, activation="relu")(x)
      #assert x.device.endswith("GPU:3")
      x = keras.layers.Dense(256, activation="relu")(x)
      x = keras.layers.Dense(256, activation="relu")(x)
      x = keras.layers.Dense(256, activation="relu")(x)
      x = keras.layers.Dense(256, activation="relu")(x)
      x = keras.layers.Dense(256, activation="relu")(x)
      outputs = keras.layers.Dense(10)(x)
    model = keras.Model(inputs, outputs)
 
    opt = keras.optimizers.Adam()
    model.compile(
        optimizer=opt,
        loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
        metrics=[keras.metrics.SparseCategoricalAccuracy()],
        experimental_run_tf_function=False,
    )
    return model


def get_dataset():
    batch_size = 32
    num_val_samples = 10000

    # Return the MNIST dataset in the form of a `tf.data.Dataset`.
    path = '/g/g92/jtaylor/workspace/TFnew/mnist.npz'
    (x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data(path)

    # Preprocess the data (these are Numpy arrays)
    x_train = x_train.reshape(-1, 784).astype("float32") / 255
    x_test = x_test.reshape(-1, 784).astype("float32") / 255
    y_train = y_train.astype("float32")
    y_test = y_test.astype("float32")

    # Reserve num_val_samples samples for validation
    x_val = x_train[-num_val_samples:]
    y_val = y_train[-num_val_samples:]
    x_train = x_train[:-num_val_samples]
    y_train = y_train[:-num_val_samples]
    return (
        tf.data.Dataset.from_tensor_slices((x_train, y_train)).batch(batch_size),
        tf.data.Dataset.from_tensor_slices((x_val, y_val)).batch(batch_size),
        tf.data.Dataset.from_tensor_slices((x_test, y_test)).batch(batch_size),
    )

model = get_compiled_model()

# Train the model on all available devices.
train_dataset, val_dataset, test_dataset = get_dataset()
model.fit(train_dataset, epochs=2, validation_data=val_dataset)

# Test the model on all available devices.
model.evaluate(test_dataset)

[JohnTaylor2000]

Output of tf.config.list_physical_devices() as requested:-

Num GPUs Available:  4
 gpus =  [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU'), 
PhysicalDevice(name='/physical_device:GPU:1', device_type='GPU'), 
PhysicalDevice(name='/physical_device:GPU:2', device_type='GPU'), 
PhysicalDevice(name='/physical_device:GPU:3', device_type='GPU')]

[ravikyram]

@JohnTaylor2000

Please, let us know which TF version you are using?
Thanks!

[JohnTaylor2000]

PowerAi Tensorflow version 2.1.0
Also tensorflow 2.2.0 on x86_64

[rmothukuru]

@JohnTaylor2000,
If I understand your requirement correctly, you want to utilize all the GPUs while Training your Model. If that is the case, the best API to use is Distributed Training. Even the Dataset can be Distributed using Distributed Strategy.

Thanks!

[JohnTaylor2000]

The actual model that I am working on is too large to fit into GPU memory. I have a data parallel code using horovod that runs on hundreds of GPUs but now need to use a larger model. To do this I need to spread the layers across multiple GPUs.

[JohnTaylor2000]

Hi all, just wondering if you have been able to run the test code that I have provided and/or need any further help?

[JohnTaylor2000]

Wondering if you can confirm that this is an issue based on my supplied bug report and if you have a solution? This is a roadblock to my research so I am keen to have this resolved - any help greatly appreciated!

[JohnTaylor2000]

Any developments on this issue please?

[JohnTaylor2000]

Any help available? I have had no response since 5 November?

[laplacericky]

I think TF currently does not support model parallelism with keras model like what you have written. However, I believe they do support model parallelism with primitive operations. Also, it is almost like they do not have any docs about model parallelism which is an issue.
Maybe you can try something like wrapping all init, build and call methods inside with tf.device(the device you want).
Example code:

import tensorflow as tf
import numpy as np


tf.debugging.set_log_device_placement(True)

class Dense_in_gpu(tf.keras.layers.Layer):
    def __init__(self, gpu_to_use, units,activation=None,**kwargs):
        self.gpu_to_use=gpu_to_use
        with tf.device(f"/GPU:{self.gpu_to_use}"):
            self.dense=tf.keras.layers.Dense(units,activation=activation)
        super().__init__(**kwargs)

    def call(self, inputs):
        with tf.device(f"/GPU:{self.gpu_to_use}"):
            return self.dense(inputs)


inputs = tf.keras.Input(shape=(784,))
#run in the first gpu
x = Dense_in_gpu(0,64, activation="relu")(inputs)
#run in the second gpu
x = Dense_in_gpu(1,64, activation="relu")(x)
#run in the third gpu
outputs = Dense_in_gpu(2,10)(x)

model = tf.keras.Model(inputs=inputs, outputs=outputs)

model.predict_on_batch(np.random.rand(64,784))

tf_dataset=tf.data.Dataset.from_tensor_slices((np.random.rand(64,784),np.random.rand(64,10)))
tf_dataset=tf_dataset.batch(64)
model.compile(optimizer=tf.keras.optimizers.RMSprop(), loss=tf.keras.losses.MeanSquaredError())
model.fit(tf_dataset)

[JohnTaylor2000]

Thank you for the interesting solution to this problem. I am investigating whether this will work on the model that I have developed.

I noticed that the latest versions of tf.keras.layers no longer has the attribute 'device' which means, as you say, that Keras no longer supports model parallelism where we assign layers to a device.

Interestingly, the development of GPUs with much larger memory eg 80GB on the A100 and the proposed new Grace architecture which allow high bandwidth access to CPU memory, will reduce the need for model parallelism. However, this will likely be offset by the desire to build bigger more complex model systems using the Keras functional API.

[Saduf2019]

@JohnTaylor2000
Is this still an issue, could you please try on the latest tf version and let us know.

[JohnTaylor2000]

I have been able to test this using TensorFlow 2.6.0 using a test code. I do not have access to 2.7.0 at the moment.

Still seeing the same problem when using a dense layer tf.keras.layers.Dense the response was
AttributeError: 'KerasTensor' object has no attribute 'device'. I also observed that the code intended to run on 4 GPUs runs primarily on a single GPU ie the 'with tf.device("GPU:3"): ' is ignored.

here is a simple test where the print command will generate the error above:-

with tf.device("GPU:3"):
    x = tf.keras.layers.Dense(num_hidden_units, activation="relu")(x)
    x = tf.keras.layers.Dense(num_hidden_units, activation="relu")(x)
    x = tf.keras.layers.Dense(num_hidden_units, activation="relu")(x)
    x = tf.keras.layers.Dense(num_hidden_units, activation="relu")(x)
    x = tf.keras.layers.Dense(num_hidden_units, activation="relu")(x)
    print(x.device)

[mohantym]

Hi @JohnTaylor2000 ! You can use set_visible_devices to disable and enable specific GPU's during operation. Have you tried the same in 2.8 version yet?

[google-ml-butler]

This issue has been automatically marked as stale because it has no recent activity. It will be closed if no further activity occurs. Thank you.

[google-ml-butler]

Closing as stale. Please reopen if you'd like to work on this further.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No