Prepare your project, understand how it runs and keep useful evidence. Pick a question from the library or start with the first-launch walkthrough.
Prepare the input, rerun the same code, then verify the result before scaling up.
Read the protocol, examine the proof, then adapt the code to your project.
A diagnostic script, a limited report, the expected results of a 2 × 2 computation and its gradient.
python kernodeck-diagnostic-v1.py --device cpuCPU report provided: verified loss 196. A CUDA check on RTX 5070 is also provided; it does not measure catalog offerings. The script requires a GPU unless you specify CPU.
A small training run, 24 synthetic rows, a continuous/resume comparison and a negative control.
python -B verify_resume.py --output runs/preuve-cpuFrom the extracted folder, with PyTorch and NumPy installed, to a fresh folder. The proof provided compares 10 steps to 5 + 5 on CPU; it does not validate GPU, AMP or distributed training.
From exploratory code to an input you can replay.
Separate parameters, credentials, and traces before sharing or moving code.
State your needs and verify the environment actually provided.
A clear command, a tracked process, interpretable traces.
Define the inputs, outputs, and return codes of your own application.
Scope the measurement and read a CPU/GPU trace without jumping to conclusions.
Distinguish the device, the operation, and the numerical result.
A script and a limited report to verify the first computation.
Pick up the work, choose what comes next, and track the rental.
A mini-project that compares ten continuous steps to five plus five.
Set the context, batch, and concurrency before choosing a configuration.
Distinguish the transfer, its verification, and the preparation of the rental.