https://github.com/danindiana/tensorcatgpu
Can CPU tensorflow build CUDNN tensorflow?
https://github.com/danindiana/tensorcatgpu
bash-scripting cudnn-based-implementations nvidia-gpu tensorflow
Last synced: about 2 months ago
JSON representation
Can CPU tensorflow build CUDNN tensorflow?
- Host: GitHub
- URL: https://github.com/danindiana/tensorcatgpu
- Owner: danindiana
- License: mit
- Created: 2019-02-23T05:21:20.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2023-02-20T10:12:32.000Z (over 3 years ago)
- Last Synced: 2025-01-09T06:37:46.427Z (over 1 year ago)
- Topics: bash-scripting, cudnn-based-implementations, nvidia-gpu, tensorflow
- Language: Shell
- Homepage:
- Size: 8.79 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# TensorCatGPU
A repository exploring the challenges and solutions for installing NVIDIA GPU-accelerated TensorFlow on Ubuntu systems.
## Overview
Installing NVIDIA driver + dependencies for GPU-compiled TensorFlow (NDGPUTF) on Ubuntu is really challenging. Is this difficulty itself a kind of intelligence test?
This repository explores the question: **Can TensorFlow CPU be used to bootstrap a fault-tolerant GPU-supported TensorFlow installation?**
## Table of Contents
- [Architecture](#architecture)
- [Installation Workflow](#installation-workflow)
- [Bootstrap Concept](#bootstrap-concept)
- [System Requirements](#system-requirements)
- [Repository Contents](#repository-contents)
- [Getting Started](#getting-started)
- [Neural Network Approach](#neural-network-approach)
## Architecture
### TensorFlow GPU Stack Dependencies
```mermaid
graph TB
subgraph "Application Layer"
TF[TensorFlow GPU]
end
subgraph "Deep Learning Libraries"
CUDNN[cuDNN Library]
end
subgraph "CUDA Toolkit"
CUDA[CUDA Runtime
libcudart.so]
CUDALIB[CUDA Libraries]
end
subgraph "Driver Layer"
NVDRIVER[NVIDIA Driver
nvidia-smi]
end
subgraph "Hardware"
GPU[GPU Hardware
GeForce GTX 1050]
end
TF -->|depends on| CUDNN
TF -->|depends on| CUDA
CUDNN -->|depends on| CUDA
CUDA -->|depends on| NVDRIVER
NVDRIVER -->|controls| GPU
style TF fill:#4CAF50
style CUDNN fill:#2196F3
style CUDA fill:#FF9800
style NVDRIVER fill:#F44336
style GPU fill:#9C27B0
```
### Component Dependencies
```mermaid
graph LR
subgraph "Installation Components"
A[Ubuntu 18.04+]
B[Python 3.x]
C[pip/virtualenv]
end
subgraph "NVIDIA Stack"
D[NVIDIA Driver]
E[CUDA Toolkit 10.0+]
F[cuDNN 7.4+]
end
subgraph "TensorFlow Options"
G[TensorFlow CPU]
H[TensorFlow GPU]
end
A --> B
B --> C
C --> G
A --> D
D --> E
E --> F
F --> H
G -.->|Bootstrap Helper?| H
style G fill:#81C784
style H fill:#4CAF50
```
## Installation Workflow
### Standard GPU TensorFlow Installation Process
```mermaid
flowchart TD
Start([Start Installation]) --> CheckOS{Ubuntu 18.04+
Compatible?}
CheckOS -->|No| UpgradeOS[Upgrade OS]
CheckOS -->|Yes| CheckGPU{NVIDIA GPU
Present?}
UpgradeOS --> CheckGPU
CheckGPU -->|No| CPUOnly[Install TensorFlow CPU Only]
CheckGPU -->|Yes| InstallDriver[Install NVIDIA Driver]
InstallDriver --> VerifyDriver{nvidia-smi
Working?}
VerifyDriver -->|No| TroubleshootDriver[Troubleshoot Driver
- Secure Boot
- Kernel Modules
- Conflicts]
VerifyDriver -->|Yes| InstallCUDA[Install CUDA Toolkit]
TroubleshootDriver --> VerifyDriver
InstallCUDA --> VerifyCUDA{CUDA Libraries
Found?}
VerifyCUDA -->|No| FixCUDAPath[Fix LD_LIBRARY_PATH]
VerifyCUDA -->|Yes| GetCUDNN{cuDNN Available?}
FixCUDAPath --> VerifyCUDA
GetCUDNN -->|No| RegisterNVIDIA[Register NVIDIA Developer
Download cuDNN]
GetCUDNN -->|Yes| InstallCUDNN[Install cuDNN]
RegisterNVIDIA --> InstallCUDNN
InstallCUDNN --> SetupPython[Setup Python virtualenv]
SetupPython --> InstallTFGPU[pip install tensorflow-gpu]
InstallTFGPU --> TestImport{TensorFlow
Imports?}
TestImport -->|No| DebugTF[Debug:
- Version conflicts
- Missing libraries
- Path issues]
TestImport -->|Yes| TestGPU{GPU Detected
by TF?}
DebugTF --> InstallTFGPU
TestGPU -->|No| CheckCompatibility[Check Version Compatibility:
TF - CUDA - cuDNN]
TestGPU -->|Yes| Success([✓ Success!
GPU TensorFlow Ready])
CheckCompatibility --> InstallTFGPU
CPUOnly --> End([TensorFlow CPU Installation])
style Start fill:#4CAF50
style Success fill:#4CAF50
style End fill:#81C784
style TroubleshootDriver fill:#FF9800
style DebugTF fill:#FF9800
style CheckCompatibility fill:#FF9800
```
## Bootstrap Concept
### Using TensorFlow CPU to Bootstrap GPU Installation
```mermaid
flowchart LR
subgraph "Phase 1: CPU Bootstrap"
A[Install TensorFlow CPU] --> B[Create Diagnostic Script]
B --> C[Analyze System State]
C --> D[Generate Installation Plan]
end
subgraph "Phase 2: Intelligence Layer"
D --> E[Neural Network Analysis]
E --> F{Predict Success
Probability}
end
subgraph "Phase 3: GPU Installation"
F -->|High Confidence| G[Automated GPU Setup]
F -->|Low Confidence| H[Guided Manual Setup]
G --> I[Verify Installation]
H --> I
I --> J{Validation
Passed?}
end
subgraph "Phase 4: Fault Tolerance"
J -->|No| K[ML-Based Debugging]
J -->|Yes| L[Production Ready]
K --> M[Suggest Fixes]
M --> I
end
style A fill:#81C784
style E fill:#2196F3
style G fill:#4CAF50
style L fill:#4CAF50
style K fill:#FF9800
```
### Decision Tree: Installation Strategy
```mermaid
graph TD
Start{What's Your Goal?}
Start -->|Just Learning TF| CPU[Use TensorFlow CPU
✓ Quick Setup
✓ No GPU Required]
Start -->|Production/Training| GPU{Have NVIDIA GPU?}
GPU -->|No| CPU
GPU -->|Yes| Experience{Installation
Experience?}
Experience -->|Beginner| Bootstrap[Use Bootstrap Approach
1. Install TF-CPU
2. Run diagnostics
3. Follow guided setup]
Experience -->|Advanced| Manual[Manual Installation
1. NVIDIA Driver
2. CUDA Toolkit
3. cuDNN
4. TensorFlow-GPU]
Bootstrap --> Validate{Installation
Successful?}
Manual --> Validate
Validate -->|Yes| Success[✓ GPU-Accelerated
TensorFlow Ready]
Validate -->|No| Debug[Use tf_env.txt
for Debugging]
Debug --> Retry{Try Again?}
Retry -->|Yes| Bootstrap
Retry -->|No| CPU
style CPU fill:#81C784
style Success fill:#4CAF50
style Bootstrap fill:#2196F3
style Debug fill:#FF9800
```
## System Requirements
### Test System Specifications
The repository was tested on the following system:
```
OS: Ubuntu 18.10 cosmic
Kernel: x86_64 Linux 4.18.0-15-generic
CPU: AMD Ryzen 5 1600X Six-Core @ 12x 4GHz
GPU: GeForce GTX 1050
RAM: 16 GB (8732MiB available)
Resolution: 2560x1080
DE: LXDE
WM: OpenBox
```
### Required NVIDIA Components
| Component | Version | Notes |
|-----------|---------|-------|
| NVIDIA Driver | Latest | Use `nvidia-smi` to verify |
| CUDA Toolkit | 10.0+ | Multiple versions can coexist |
| cuDNN | 7.4.2+ | Requires NVIDIA Developer Account |
| TensorFlow-GPU | 1.12.0+ | Must match CUDA/cuDNN versions |
## Repository Contents
### Files
- **`README.md`** - This file
- **`tf_env.txt`** - Environment diagnostic output showing:
- System configuration
- CUDA library locations
- TensorFlow installation status
- Virtual environment setup
- nvidia-smi output
- **`cp_pbashscript.sh`** - Copy utility with progress bar
- Useful for copying large CUDA/cuDNN files
- Shows real-time progress using `strace` and `awk`
- **`LICENSE`** - Repository license
### Key Findings from Diagnostics
The `tf_env.txt` file demonstrates two states:
1. **Before Installation**:
- TensorFlow not found
- nvidia-smi failing
- No virtualenv
2. **After Installation**:
- TensorFlow 1.12.0 working
- Virtualenv active
- Still some driver issues (common on VMs)
## Getting Started
### Option 1: CPU TensorFlow (Quick Start)
```bash
# Install TensorFlow CPU
pip install tensorflow
# Verify installation
python -c "import tensorflow as tf; print(tf.__version__)"
```
Reference: [TensorFlow Installation Guide](https://www.tensorflow.org/install/pip)
### Option 2: GPU TensorFlow (Advanced)
1. **Install NVIDIA Driver**
```bash
# Check current driver
nvidia-smi
# If needed, install driver
sudo apt install nvidia-driver-XXX
```
2. **Install CUDA Toolkit**
```bash
# Download from NVIDIA website
# Install CUDA 10.0+
```
3. **Install cuDNN**
- Register at NVIDIA Developer Program
- Download `libcudnn7-dev_7.4.2.24-1+cuda10.0_amd64.deb`
- Install the package
4. **Install TensorFlow GPU**
```bash
pip install tensorflow-gpu
```
5. **Verify GPU Detection**
```bash
python -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
```
### Using the Bootstrap Approach
1. Start with CPU TensorFlow to validate Python environment
2. Use diagnostic scripts to check system compatibility
3. Install GPU components incrementally
4. Use TensorFlow CPU to test each step
## Neural Network Approach
### Conceptual Framework
**Question**: Can a neural network guide the TensorFlow GPU installation process?
**Proposed Architecture**: Convolutional Neural Network (CNN)
#### Input Features
- OS version and kernel
- System uptime and packages
- Shell and desktop environment
- CPU specifications
- GPU model and capabilities
- RAM availability
- Current driver status
- CUDA library presence
#### Output
- Installation commands sequence
- Probability of success
- Potential failure points
- Recommended troubleshooting steps
#### Training Data Sources
- Successful installation logs
- Failed installation attempts
- System configurations
- Version compatibility matrices
### GPT-4 Generated Insight
> *"The neural network that would need to be generated for the Ubuntu Linux system would be a Convolutional Neural Network (CNN). The input to the CNN would be information about the OS, kernel, uptime, packages, shell, resolution, DE, WM, CPU, GPU and RAM of the system. The output of the network would be installation commands for setting up a bare metal GPU supported TensorFlow instance."*
## Research Questions
This repository explores several interesting questions:
1. **Is installation difficulty a feature?**
- Does the complexity serve as a barrier to entry?
- Could AI assistance democratize GPU computing?
2. **Can TensorFlow CPU bootstrap TensorFlow GPU?**
- Use CPU version for diagnostics
- Generate installation scripts
- Predict compatibility issues
- Automated troubleshooting
3. **Cross-platform bootstrapping**
- Can same approach work for AMD GPU (ROCm)?
- Generalize to other accelerators (TPU, Intel GPU)?
4. **What are Neural Networks good for?**
- Can they solve their own installation problems?
- Meta-learning: AI helping install AI frameworks
## Contributing
Contributions welcome! Areas of interest:
- [ ] Automated diagnostic scripts
- [ ] Neural network training data collection
- [ ] Installation success predictor model
- [ ] AMD ROCm bootstrap approach
- [ ] Docker containerization strategy
- [ ] Version compatibility database
## License
See `LICENSE` file for details.
## Further Reading
- [TensorFlow Official Installation Guide](https://www.tensorflow.org/install/pip)
- [NVIDIA CUDA Toolkit Documentation](https://docs.nvidia.com/cuda/)
- [cuDNN Installation Guide](https://docs.nvidia.com/deeplearning/cudnn/install-guide/)
- Review `tf_env.txt` for detailed environment diagnostics
---
**Note**: NVIDIA Developer Account required for downloading cuDNN libraries.