The proposed method for offline deployment of an AI stack involves creating a local repository of Python dependencies. By using the 'pip download' command with the '--only-binary=:all:' flag, developers can capture all necessary packages in a local directory, ensuring that the environment remains identical across different machines.
To ensure reproducibility, the process utilizes a 'requirements.lock' file to fix specific package versions. The configuration specifically targets CPU-based versions of libraries like PyTorch by using the '--extra-index-url' flag, which prevents the accidental inclusion of CUDA-enabled dependencies and keeps the repository size manageable.
The deployment workflow requires the target machine to run the same operating system and Python version as the source machine, specifically Debian 12 with Python 3.11. The offline installation is performed using 'pip install --no-index', which forces the system to rely exclusively on the pre-downloaded local wheel files.
The final archive includes the local repository, which is approximately 730 MB in size, alongside project source code. System-level dependencies such as CMake and build-essential must be installed separately on the target machine. Sensitive data, including user sessions and security keys, requires careful handling during the transfer process.
