Skip to content

Core Features & Platform Capabilities

This document provides a comprehensive overview of PowerMind ComputeAI capabilities, architectural concepts, and operational contracts.

1. Tenant Workspace Model

Every user account on ComputeAI has an isolated workspace directory on the cluster. The platform structures user workspaces automatically:

  • datasets/: Storage area for raw and processed datasets (CSV, TFRecords, PyTorch datasets, HDF5, image folders, zip/tar archives). Accessible at runtime via $DATASET_DIR.
  • code/: Directory for shared Python utility modules or helper packages. Accessible at runtime via $CODE_DIR.
  • jobs/: Automatically maintained directory storing job-specific logs, outputs, and execution metadata. Accessible via $JOB_OUTPUT_DIR.
  • checkpoints/: Structured storage where epoch checkpoints are saved during training runs. Accessible via $CHECKPOINT_DIR.

2. Dataset & File Management

2.1 File Ingestion & Automatic Extraction

You can upload datasets, code libraries, and pre-trained weights to your workspace:

  • Dataset Files: Raw training data and compressed archives (.zip, .tar.gz, .tgz). Compressed archives can be automatically unpacked into the dataset directory upon upload.
  • Code Libraries: Supporting Python packages, custom modules, or pre-processing scripts.
  • Model Binaries: Pre-trained weights files (.onnx, .pt, .pth, .h5, SavedModel bundles).

2.2 Workspace Path Resolution

When training scripts run on cluster worker nodes, they do not need hardcoded physical workstation paths. The platform dynamically injects paths via environment variables (DATASET_DIR, USER_WORKSPACE, CHECKPOINT_DIR, JOB_OUTPUT_DIR). For comprehensive variable definitions, see the Runtime Variables & Contracts Guide.

3. Training Job Orchestration

ComputeAI manages training execution across single-node and multi-node GPU/CPU worker pools.

  • Direct Script Dispatch: You do not need to manually pre-upload training scripts before submission. The platform accepts the script directly when initiating a run.
  • Framework Support: Native support for PyTorch and TensorFlow runtimes with CUDA acceleration.
  • Resource Flexibility: Run workloads on single-node CPU or GPU acceleration, or scale across multi-node clusters.
  • Distributed Coordination: Automated master address, port, and rank coordination for distributed data-parallel workloads.

4. Checkpoint & Resume Lifecycle

To protect training jobs from VM preemption, network disruptions, or hardware faults, ComputeAI features an automated checkpointing and resume system:

  • Periodic Checkpointing: Training scripts write epoch weights to $CHECKPOINT_DIR.
  • Automatic Resume State: When resuming a job, the platform detects the highest completed epoch and injects resume_epoch (or COMPUTEAI_RESUME_EPOCH) so the training loop seamlessly resumes from the next epoch.
  • Checkpoint Observability: View stored checkpoints, file sizes, and loss/accuracy metrics per epoch via the Web UI or CLI.

5. Model Registry & Serving Deployments

Once training finishes, or when bringing an external pre-trained model (BYOB), you can register and deploy it for real-time serving:

  • Platform-Trained Models: Automatically cataloged upon job completion for one-click deployment.
  • Bring Your Own Model (BYOB): Upload external model weights, specify input/output tensor shapes and data types, and deploy instantly.
  • High-Performance Serving: Backed by Triton Inference Server with dynamic batching and low-latency prediction pipelines.

6. Inference Security & API Keys

Model serving endpoints are secured with deployment-scoped API access keys:

  • Key Scoping: Each API key is bound strictly to a specific deployment instance.
  • One-Time Secret Display: Cryptographic tokens are shown once at creation time.
  • Input Validation & Shape Verification: Endpoints validate tensor dimensions, datatypes, and payload structures before model execution.

Next: End-to-End Workflow | Back: Getting Started