Documentation
How Kernova works
A high-level guide to the profile → optimize → verify workflow. Full setup guides are provided to early-access teams.
Workflow
01 · Profile
Capture a representative workload
Run a short training or inference loop under the Kernova profiler. It records operator-level timings, tensor shapes, and memory traffic.
02 · Optimize
Generate candidate kernels
The engine targets the hottest operators, proposes fused or retuned kernels in Triton or CUDA, and autotunes them for your GPU.
03 · Verify
Validate before you ship
Every candidate is checked against the original outputs within configurable numerical tolerances, then benchmarked against the baseline.
04 · Deploy
Drop in the result
Accepted kernels are delivered as source with a validation report, so your team can review, version, and deploy them like any other code.
Supported environments
- Frameworks
- PyTorch 2.x
- Kernel languages
- Triton, CUDA C++
- Hardware
- NVIDIA Ampere and Hopper (A100, H100)
- Precision
- FP32, BF16, FP16
FAQ
Do I need to change my model code?
No. Profiling attaches to existing workloads, and optimized kernels are integrated where the original operators run.
What happens if a kernel changes my outputs?
It is rejected. Only candidates that pass numerical validation are reported.
Where is the full API reference?
Detailed setup guides and API references are shared with early-access teams during onboarding.
Want the full guides? Request early access.
Supported environments reflect the planned scope. Update them to match what the product supports today.