Introduction

The error AttributeError: module 'tensorflow' has no attribute 'GPUOptions' occurs when code written for TensorFlow 1.x is run on TensorFlow 2.x. In TF1, tf.GPUOptions was used to configure GPU memory allocation (memory growth, per-process memory fraction, visible devices). TF2 reorganized the GPU configuration API, moving these settings to tf.config.experimental and later to tf.config.

The Error

1import tensorflow as tf
2
3# TF1 code — fails on TF2
4gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)
5# AttributeError: module 'tensorflow' has no attribute 'GPUOptions'

The Fix: TF2 GPU Configuration

Memory Growth (Allocate as Needed)

The most common use case — prevent TensorFlow from grabbing all GPU memory at startup:

1import tensorflow as tf
2
3# TF2 way — enable memory growth
4gpus = tf.config.list_physical_devices('GPU')
5if gpus:
6    for gpu in gpus:
7        tf.config.experimental.set_memory_growth(gpu, True)

Limit GPU Memory

Set a hard memory limit instead of using a fraction:

1import tensorflow as tf
2
3gpus = tf.config.list_physical_devices('GPU')
4if gpus:
5    tf.config.set_logical_device_configuration(
6        gpus[0],
7        [tf.config.LogicalDeviceConfiguration(memory_limit=4096)]  # 4 GB
8    )

Restrict Visible GPUs

1import tensorflow as tf
2
3# Only use GPU 0 (hide GPU 1, 2, etc.)
4gpus = tf.config.list_physical_devices('GPU')
5if gpus:
6    tf.config.set_visible_devices(gpus[0], 'GPU')

Or use the environment variable:

# Before running the script
export CUDA_VISIBLE_DEVICES=0

Migration Table: TF1 → TF2

Full TF1 → TF2 Migration Example

TF1 (Old)

1import tensorflow as tf
2
3gpu_options = tf.GPUOptions(
4    per_process_gpu_memory_fraction=0.5,
5    allow_growth=True,
6    visible_device_list='0'
7)
8config = tf.ConfigProto(gpu_options=gpu_options)
9sess = tf.Session(config=config)

TF2 (New)

1import tensorflow as tf
2
3gpus = tf.config.list_physical_devices('GPU')
4if gpus:
5    # Restrict to first GPU
6    tf.config.set_visible_devices(gpus[0], 'GPU')
7
8    # Enable memory growth
9    tf.config.experimental.set_memory_growth(gpus[0], True)
10
11    # Or set a hard limit (4 GB)
12    # tf.config.set_logical_device_configuration(
13    #     gpus[0],
14    #     [tf.config.LogicalDeviceConfiguration(memory_limit=4096)]
15    # )
16
17# No session needed — TF2 uses eager execution by default

Using tf.compat.v1 (Temporary Fix)

If you cannot migrate immediately, use the compatibility module:

1import tensorflow as tf
2
3# TF1 compatibility mode
4gpu_options = tf.compat.v1.GPUOptions(allow_growth=True)
5config = tf.compat.v1.ConfigProto(gpu_options=gpu_options)
6sess = tf.compat.v1.Session(config=config)

This is a temporary bridge — plan to migrate to native TF2 APIs.

Checking GPU Availability

1import tensorflow as tf
2
3# List available GPUs
4print(tf.config.list_physical_devices('GPU'))
5# [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
6
7# Check if TensorFlow can see GPUs
8print(f"GPUs available: {len(tf.config.list_physical_devices('GPU'))}")
9
10# Get GPU details
11for gpu in tf.config.list_physical_devices('GPU'):
12    details = tf.config.experimental.get_device_details(gpu)
13    print(f"{gpu.name}: {details}")

Common Pitfalls

• Must configure before any TF operation: GPU configuration (set_memory_growth, set_visible_devices) must be called before any TensorFlow operation. Once TF initializes the GPU runtime, configuration changes raise a RuntimeError.
• set_memory_growth vs memory_limit: You cannot use both on the same GPU. Choose one approach: either let TF grow memory as needed (set_memory_growth) or set a hard cap (LogicalDeviceConfiguration).
• Environment variable precedence: CUDA_VISIBLE_DEVICES is applied before TF sees any GPUs. If you set CUDA_VISIBLE_DEVICES=1, then tf.config.list_physical_devices('GPU') only shows one GPU at index 0 (which is actually physical GPU 1).
• Mixed precision: TF2 mixed precision (tf.keras.mixed_precision) can reduce GPU memory usage significantly. Consider this before hard-limiting memory.
• Docker containers: In Docker, you need --gpus flag: docker run --gpus all. Without it, list_physical_devices('GPU') returns an empty list regardless of configuration.

Summary

• tf.GPUOptions was removed in TF2 — use tf.config APIs instead
• Use tf.config.experimental.set_memory_growth(gpu, True) to enable dynamic memory allocation
• Use tf.config.set_logical_device_configuration to set a hard memory limit
• Use tf.config.set_visible_devices to restrict which GPUs are used
• For temporary compatibility, use tf.compat.v1.GPUOptions
• Configure GPU settings before any TensorFlow operation