VAnta

It explains: 1. GPU load and temperature 2. VRAM usage 3. CPU usage 4. System RAM usage 5. Live GPU and VRAM resource signals

Inspiration

GPU workloads are often launched as separate processes with no shared coordination. On machines with limited VRAM, jobs can compete for memory, fail unexpectedly, and leave users wondering why one job ran before another. We created VAnta for the InfinityX Global Hackathon 2K26 to make GPU workload management visible, predictable, and explainable. Our guiding loop is: Observe → Rule → Queue / Action → Verify

What it does

VAnta is a user-level GPU workload controller for Windows. It:

Monitors GPU utilization, VRAM, temperature, power, CPU, and RAM
Assigns jobs critical, high, normal, or low priority
Queues GPU workloads and launches them when resources are available
Uses an explainable VRAM admission rule with a 512 MB safety reserve
Controls and cancels processes launched by VAnta
Verifies VRAM usage before and after each action
Displays decisions through a live dashboard and decision feed
Stores job history in SQLite
Recovers safely from interrupted processes
Includes a Policy Lab and one-click demo scenario ## How we built it We built VAnta as a native Windows desktop application using Python and PySide6. The system is divided into several layers:
Telemetry: NVIDIA Management Library through pynvml, plus psutil for CPU and RAM
Scheduling: Python domain modules for rules, priorities, queueing, and job lifecycle
Persistence: SQLite for jobs, events, and decision history
Demo workload: CUDA Driver API through ctypes, without requiring PyTorch or the CUDA Toolkit
Packaging: PyInstaller for a standalone Windows application
Testing: pytest for scheduling, admission rules, persistence, recovery, and demo behavior A job is admitted only when enough VRAM remains after reserving a safety buffer: free VRAM - requested VRAM ≥ safety reserve VAnta operates at normal user level. It does not modify the NVIDIA driver, Windows Scheduler, registry, or system services ## Challenges we ran into The biggest challenge was managing GPU workloads safely without controlling the operating system at kernel or driver level. We designed VAnta to control only the child processes it launches. We also had to handle:
Preventing unsafe interruption of running CUDA workloads
Measuring actual VRAM changes after a process starts
Recovering jobs after an unexpected shutdown
Keeping scheduling decisions understandable to users
Demonstrating realistic GPU behavior without requiring a large machine
Testing telemetry and CUDA features when compatible hardware is unavailable ## Accomplishments that we're proud of We built a working end-to-end controller rather than only a monitoring dashboard. VAnta can observe resources, apply rules, queue jobs, launch workloads, verify results, and explain every decision. We are especially proud of:
A functional Windows dashboard
Predictable non-preemptive scheduling
Explainable VRAM admission decisions
A real CUDA pulse workload
Persistent job and event history
Safe process recovery
A Policy Lab for comparing scheduling policies
Automated tests covering the core behavior
A one-click demo that shows queueing, verification, and VRAM rejection ## What we learned We learned that resource scheduling is not only about maximizing GPU utilization. Users also need predictable behavior, safe recovery, and clear explanations for every scheduling decision. We gained practical experience with:
GPU and system telemetry
Process lifecycle management
Queue and priority design
VRAM admission control
Native Windows desktop development
SQLite persistence and crash recovery
Testing scheduling policies with deterministic workloads
Designing within the limits of user-level applications ## What's next for VAnta Our next steps are:
Support multiple GPUs
Add Linux telemetry support
Support remote GPU workers
Add per-user and per-team quotas
Add scheduled workloads
Improve workload estimation using historical data
Explore checkpoint-aware preemption
Add analytics for throughput, wait time, and resource efficiency