Nexa Compute & Nexa Forge

AI Infrastructure Platform with Managed API Service

A complete AI foundry platform for orchestrating data generation, model distillation, training, and evaluation on ephemeral GPU compute.

---

Quick Start

1. Setup Environment

We use uv for dependency management. The repo is configured to work with the Homebrew install at /opt/homebrew/bin/uv.

# optional: create a project virtualenv
/opt/homebrew/bin/uv venv .venv
source .venv/bin/activate

# install all runtime + dev dependencies deterministically
/opt/homebrew/bin/uv pip sync requirements/requirements-dev.lock

# install git hooks
pre-commit install

2. Training Workflow

Prerequisites: You must bring your own compute from your preferred vendor (AWS, GCP, Lambda, Prime Intellect, etc.).

1. Spin up a GPU instance (Ubuntu 22.04 recommended).
2. Copy .env.example to .env and configure your services (WandB, HuggingFace, S3).

Deploy a Training Node: SSH into your node and run the turn-key deployment script. This will sync your code, install dependencies, and set up a persistent workspace.

./nexa_infra/scripts/provision/deploy.sh ubuntu@gpu-node-ip

Start Training (Remote or Local): Once attached to the remote session (tmux), you can start training immediately:

# Run V1 Stability Plan
python nexa_train/train.py --config-mode v1 --run-name my_stability_run

# Run V2 Performance Plan (Distributed)
torchrun --nproc_per_node=8 nexa_train/train.py --config-mode v2 --dry-run true

3. Run Infrastructure

Start Backend & Dashboard:

# Using the orchestrator script
./nexa_infra/scripts/orchestration/start_forge.sh

---

Project Structure

Nexa_compute/
├── nexa_data/           # Data Engineering (MS/MS, Tool Use, Distillation)
├── nexa_train/          # Training Engine (Axolotl, HF Trainer)
├── nexa_distill/        # Knowledge Distillation Pipeline
├── nexa_eval/           # Evaluation & LLM-as-a-Judge
├── nexa_inference/      # vLLM Serving & Tool Controller
├── nexa_infra/          # IaC (Terraform), Monitoring, Provisioning
├── nexa_ui/             # Dashboards (Streamlit/Next.js)
├── src/
│   └── nexa_compute/
│       ├── api/         # FastAPI backend
│       ├── cli/         # CLI Entrypoint
│       ├── core/        # Core Primitives (DAG, Registry, Artifacts)
│       ├── data/        # DataOps (Versioning, Lineage)
│       ├── models/      # ModelOps (Registry, Versioning)
│       ├── monitoring/  # Observability (Alerts, Metrics, Drift)
│       └── orchestration/ # Workflow Engine (Scheduler, Templates)
├── docs/
│   ├── compute_plans/   # Training Configuration Templates (V1/V2/V3)
│   ├── pipelines/       # Detailed Architecture Docs
│   ├── platform/        # Platform Guide & Best Practices
│   ├── api/             # API Reference
│   └── projects/        # Active Research Projects
├── sdk/                 # Python Client SDK
└── pyproject.toml       # Dependencies & Config

---

Core Features

Compute Engine

• Unified Training CLI: nexa_train/train.py supports flexible overrides and configuration modes (V1 Stability, V2 Performance, V3 Full).
• Infrastructure as Code: Terraform modules for AWS GPU clusters.
• Observability: Distributed tracing (OpenTelemetry), Prometheus metrics, and real-time cost tracking.
• Automated Provisioning: One-command deployment to bare metal or cloud instances with Spot instance support.

Managed API (Nexa Forge)

• Workflows: Declarative pipeline orchestration (DAGs) with resume capability.
• 6 Job Types: Generate, Audit, Distill, Train, Evaluate, Deploy.
• Worker Orchestration: Pull-based job queue for ephemeral workers.
• Security: SHA256 API keys and metered billing.

MLOps & DataOps

• Model Registry: Full lineage tracking from dataset to deployed model.
• Data Versioning: Content-addressable storage for datasets.
• Monitoring: Automated drift detection and A/B testing framework.

---

Documentation

For detailed instructions on how the platform works and what each component does, please refer to the documentation:

• Documentation Map: Central index for all documentation.
• Platform Guide: Overview of platform capabilities.
• API Reference: API endpoints and usage.
• Infrastructure Guide: Docker, Provisioning, and Hardware.
• Training Pipeline: Configuration and Execution.
• Data Refinery: MS/MS and Synthetic Data.
• Compute Plans: Run Configurations.

---

Contributing

We welcome contributions! Please review our guidelines before submitting pull requests.

See docs/conventions/ for:

• Coding Standards
• Data Organization
• Naming Conventions

Development Commands

1. Linting: ruff check .
2. Testing: pytest tests/
3. Infrastructure: Validate Terraform with terraform validate.

---

Tags: machine-learning, distributed-training, infrastructure-as-code, mlops, knowledge-distillation, fastapi, pytorch, spectral-analysis