dgx-harness-engineering

DGX Harness Engineering

An anonymized engineering reference repository for LLM training, evaluation, serving, recovery, and multi-node orchestration workflows.

What this repository contains

• Training and fine-tuning orchestration scripts
• Evaluation and smoke-test harnesses
• Runtime and recovery utilities
• Multi-node / distributed execution helpers
• Dataset-processing source code
• Deployment and service-management examples
• Operational guardrails and verification scripts

What is intentionally excluded

This public package does not contain trained model weights, LoRA adapter weights, checkpoints, training datasets, runtime logs, experiment outputs, private credentials, private keys, or machine-specific secrets.

Privacy and anonymization

Before publication, the source tree is copied to a separate export directory and sanitized. Machine-specific usernames, hostnames, home paths, internal network addresses, SSH targets, and detected credential patterns are replaced or removed in the publication copy.

The original local working tree is not modified by the publication process.

Environment-specific placeholders

Some scripts intentionally retain configuration variables because this repository is an engineering reference rather than a one-click deployment image. Users should provide their own hostnames, addresses, paths, model locations, credentials, and environment settings.

Do not treat documentation/example addresses as production network values.

Security status

The publication pipeline uses a fail-closed residual scan. The package should only be published after:

• residual finding records = 0
• residual occurrences = 0
• residual affected files = 0
• remote SHA-256 verification passes
• independent remote secret / local-identifier scan passes

Repository scope

This repository focuses on harness engineering and operational automation. It does not claim that every script is portable to every GPU, operating system, driver, CUDA version, model architecture, or cluster topology.

Review every command before running it in a different environment.

Safety

Several scripts manage processes, services, GPUs, distributed runtimes, and filesystem artifacts. Inspect paths, process filters, SSH targets, and service names before execution.

License

See LICENSE.

Repository metadata license identifier: other.