Device & infrastructure testing

One run. Every layer of your AI stack verified.

Vasolai's Unified Test Runner (UTR) is a single, templated framework that tests generated tools, models, and hardware together — GPU, kernel, and framework layers — instead of stitching together separate test suites per vendor.

See test templates Request a test plan

Coverage

Built for the modern AI hardware/software stack

NVIDIA GPU

Device enumeration, memory bandwidth, thermal/throttle behavior, and multi-GPU topology checks.

CUDA

Kernel correctness, occupancy, stream/concurrency validation, and driver/toolkit compatibility matrices.

TensorFlow

Graph/eager execution parity, op-level numerical checks, SavedModel and TF-Serving inference validation.

PyTorch

Autograd correctness, TorchScript/AOT export checks, and distributed training (DDP/FSDP) validation.

The one-run approach

Why teams stitch five test suites together — and why we don't

Most orgs test GPU health, CUDA kernels, and framework behavior in separate, disconnected pipelines. A device can pass its own diagnostic and still produce wrong model output once TensorFlow or PyTorch is layered on top. The Unified Test Runner (UTR) runs all four layers in a single pass against one templated spec, so a failure is traced to its exact origin — hardware, driver, kernel, or model code.

DISCOVER

Detect the stack

UTR fingerprints GPU model, driver/CUDA version, and installed framework versions before writing a single test.

GENERATE

Build the test set from a template

A YAML template expands into hardware, kernel, and model-level test cases matched to what was discovered.

RUN

Execute in one pass

All layers run in a single orchestrated job with shared context, so results can be cross-referenced instead of compared by hand.

VERIFY

Confirm intent match

Output is diffed against the intended spec — the same "does it do exactly what it should" standard we apply to every build.

Templates

Starter test-framework templates

Every engagement starts from a template like this one — a single manifest that defines the whole one-run test job. We adapt it to your environment rather than building from scratch.

# utr.template.yaml — Vasolai Unified Test Runner
run: "one-pass"
target:
  gpu: nvidia-auto-detect
  cuda: ">=12.0"
  frameworks: [tensorflow, pytorch]

stages:
  - name: device
    checks: [enumerate, memory-bandwidth, thermal, topology]
  - name: cuda
    checks: [kernel-correctness, occupancy, stream-concurrency]
  - name: tensorflow
    checks: [graph-eager-parity, op-numerics, serving-inference]
  - name: pytorch
    checks: [autograd, torchscript-export, ddp-sync]

verify:
  against: ./intent-spec.json
  on_fail: trace-to-origin

report: [html, json, junit]
      

GPU health template

A focused manifest for pre-deployment GPU/CUDA acceptance testing on new hardware.

DEVICE + CUDA

Model-parity template

Cross-checks a model's TensorFlow and PyTorch implementations for numerical parity before release.

TF + PYTORCH

Full-stack template

The complete one-run manifest above — device through model verification, in a single job.

ALL LAYERS

Get the full template set