Environment:Recommenders team Recommenders GPU CUDA Environment

Knowledge Sources	Recommenders gpu_utils.py setup.py GPU extras
Domains	Infrastructure, Deep_Learning, GPU
Last Updated	2026-02-10 00:00 GMT

Overview

NVIDIA GPU environment with CUDA support, TensorFlow 2.8-2.15, PyTorch 2.0+, and nvidia-ml-py for GPU-accelerated recommendation models.

Description

This environment extends the core Python dependencies with GPU-specific packages for deep learning models. It supports two frameworks: TensorFlow (used by NCF, NRMS, DeepRec, Wide&Deep models) and PyTorch (used by SASRec, SSE-PT, EmbeddingDotBias models). TensorFlow models use the `tf.compat.v1` graph execution mode with `GPUOptions(allow_growth=True)` for dynamic GPU memory allocation. GPU detection is provided via `numba.cuda` and `torch.cuda` with fallback logic.

Usage

Use this environment for any deep learning workflow including NCF training/prediction, NRMS news recommendation, DeepRec sequential models, and the GPU-accelerated benchmarking paths. Required when `recommenders[gpu]` extra is installed.

System Requirements

Category	Requirement	Notes
OS	Linux (Ubuntu 24.04 in Docker)	GPU Docker image uses `nvidia/cuda:12.6.1-devel-ubuntu24.04`
Hardware	NVIDIA GPU with CUDA support	Tested on Azure STANDARD_NC6S_V2 (Tesla P100) and GeForce GTX 1660 Ti
VRAM	>= 6 GB	Benchmark reference machine uses 6 GB (GTX 1660 Ti)
RAM	>= 30 GB	Benchmark reference uses 30 GB
CPUs	>= 4	Benchmark reference uses 4 CPUs

Dependencies

GPU Python Packages

`nvidia-ml-py` >= 11.525.84
`tensorflow` >= 2.8.4, != 2.9.0.*, != 2.9.1, != 2.9.2, != 2.10.0.*, < 2.16 (pinned due to security and breaking changes, issue #2073)
`tf-slim` >= 1.1.0
`torch` >= 2.0.1, < 3
`numpy` < 1.25.0 (Python <= 3.8 only, additional GPU constraint)
`spacy` <= 3.7.5 (Python <= 3.8 only)

System Packages (Docker)

CUDA Toolkit (12.6.1 in Docker, or system-installed)
cuDNN (version detected via header files or `torch.backends.cudnn.version()`)

Credentials

No GPU-specific credentials required. Standard NVIDIA driver installation is sufficient.

Quick Install

# Install with GPU extras
pip install "recommenders[gpu]"

# Or install all extras
pip install "recommenders[all]"

Code Evidence

GPU count detection with torch/numba fallback from `recommenders/utils/gpu_utils.py:18-34`:

def get_number_gpus():
    try:
        import torch
        return torch.cuda.device_count()
    except (ImportError, ModuleNotFoundError):
        pass
    try:
        import numba
        return len(numba.cuda.gpus)
    except Exception:
        return 0

CUDA version detection with platform-specific fallback from `recommenders/utils/gpu_utils.py:71-100`:

def get_cuda_version():
    try:
        import torch
        return torch.version.cuda
    except (ImportError, ModuleNotFoundError):
        if sys.platform == "win32":
            candidate = "C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v*\\version.txt"
            path_list = glob.glob(candidate)
            if path_list:
                path = path_list[0]
        elif sys.platform == "linux" or sys.platform == "darwin":
            path = "/usr/local/cuda/version.txt"

TensorFlow GPU memory growth from `recommenders/models/newsrec/models/base_model.py:61-65`:

# set GPU use with on demand growth
gpu_options = tf.compat.v1.GPUOptions(allow_growth=True)
sess = tf.compat.v1.Session(
    config=tf.compat.v1.ConfigProto(gpu_options=gpu_options)
)

TensorFlow version pinning from `setup.py:58`:

"tensorflow>=2.8.4,!=2.9.0.*,!=2.9.1,!=2.9.2,!=2.10.0.*,<2.16",
# Fixed TF due to constant security problems and breaking changes #2073

GPU validation test from `tests/unit/examples/test_notebooks_gpu.py:14-15`:

def test_gpu_vm():
    assert get_number_gpus() >= 1

Common Errors

Error Message	Cause	Solution
`CudaSupportError`	No NVIDIA GPU or drivers not installed	Install NVIDIA drivers and CUDA toolkit
TF models fail with TF > 2.10.1	xDeepFM and SUM models break on newer TF (issue #2018)	Use TF < 2.16 as pinned in setup.py
GPU notebook tests disabled (issue #1883)	Multiple GPU notebook tests known to fail	Check issue tracker for resolution status
`No CUDA available` (logger.info)	GPU clear_memory called without CUDA	Expected on CPU-only machines; no action needed

Compatibility Notes

TensorFlow: All TF-based models use `tf.compat.v1` graph execution mode (eager execution disabled). TF versions 2.9.0-2.9.2 and 2.10.0.x are explicitly excluded due to bugs.
PyTorch: Models auto-detect GPU via `torch.cuda.is_available()` and fall back to CPU.
Docker: GPU Docker image based on `nvidia/cuda:12.6.1-devel-ubuntu24.04`. CPU image uses `buildpack-deps:24.04`.
Azure CI: GPU tests run on Azure STANDARD_NC6S_V2 (6 vCPUs, 112 GB RAM, 1 NVIDIA Tesla P100).
Benchmark reference: 4 CPUs, 30 GB RAM, 1 GeForce GTX 1660 Ti (6 GB VRAM).

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment