The answers you need to choose and prepare your rental.

Start with your project's libraries and the backend they support, then measure the memory of a representative trial. The model, precision, input length, and batch matter more than an isolated card name. Then compare compatible GPUs in the catalog and allow a margin for the most demanding steps.

CUDA is the compute stack used with NVIDIA GPUs; ROCm is the one to verify for the AMD GPUs in the catalog. Your application and its extensions must support the chosen stack. In PyTorch, ROCm notably reuses the torch.cuda namespace: the name of that call is not enough to identify the hardware.

No. Per-card memory is a starting point. During execution, you must also accommodate inputs, temporary tensors, and states specific to your method, such as the inference cache or the optimizer. Run a trial with realistic dimensions and observe the peak memory of the most demanding phase.

No. A program must explicitly distribute its work and, if necessary, its states across the cards. Replicating the model can improve the organization of several workloads without enlarging the memory available for a single replica. Check the parallelism supported by your engine before choosing multiple batches.

A lot includes one GPU for catalog offerings, except B200, which includes two GPUs per lot. You can choose from 1 to 10 lots. The summary distinguishes the number of lots, the total number of cards, and the total price: two B200 lots therefore represent four GPUs.

No. The rental lot is a unit for ordering hardware. The batch is a group of examples that your application processes together. You set the latter in your program based on its memory and its throughput or latency target. Changing one does not automatically configure the other.

Choose the duration based on the complete job: setup, validation, computation, checking results, and export. Three days can frame a targeted trial; seven days let you organize several variants; thirty days suit a more spread-out job. These examples of organization do not predict your application's compute time.

Choose the GPU, a duration of 3, 7, or 30 days, and the number of lots. The configurator displays the total amount in USD before the order is recorded. Check the number of cards per lot, particularly for B200. The amount in crypto and its due date are then specified in the instructions for the chosen payment method.

The Python, PyTorch, notebook, or inference preparation describes the desired starting point for your work. It does not replace your dependency files, your weights, your data, or your licenses. Keep a project-specific installation procedure and a small launch check, then verify the result on your compute environment.

Yes. You use the GPU autonomously and choose the software, models, data, and processing needed for your project. Kernodeck does not inspect their contents. You manage your access, your backups, and the permissions required to use the elements you install or operate.

Keep the code revision, dependency versions, parameters, backend information, and data and weight versions. Add a launch command, a small control input, and an expected output. Setting a random seed does not, by itself, guarantee identical results across hardware or versions.

Use the same Python interpreter as your program to check the PyTorch installation and its backend. Then check device access and the compatibility of the system stack. Do not let the trial silently fall back to CPU if you are trying to validate the GPU. The guide on environments details a minimal check.

Note whether it appears during loading, on the first computation, or after several iterations. Reduce a single factor, such as the batch or the input length, and observe the result. Also look for tensors kept unnecessarily. Freeing unused cache does not free objects that are still alive; deciding to switch cards comes after this diagnosis.

It is enough for some inference uses, accompanied by the appropriate architecture and preprocessing, but it does not describe the entire training. A resume may require the optimizer, the scheduler, the progress, the random states, and the position in the data. Always verify the resume in a new process before relying on a checkpoint for a long job.

Export the results, checkpoints, parameters, and environment information before the due date indicated in your tracking. Allow time to verify the copy and reload at least one backup. Keep a copy in a location you control: the existence of a file in the compute environment does not, by itself, constitute your retention plan.

No. The crypto rental and payment flow does not require a KYC procedure or an ID document. Your first name, last name and email are still required as order contact details. This absence of KYC checks does not mean that the order, its payment or the information needed to track them do not exist.

The available methods are BTC on Bitcoin, ETH on Ethereum, USDT and USDC on Ethereum, TRX and USDT on Tron, SOL on Solana and LTC on Litecoin. Select the exact combination at payment. Then use the address, amount and deadline shown for that payment intent.

The same asset can be offered on several networks, notably USDT. Payment instructions refer to one specific combination; a transfer on another network is not equivalent. Check the network in your wallet and in the order before confirming the transfer, as well as the amount sent to the displayed address.

It signals that you have made the transfer for the payment in question. This signal appears in the tracking, but on its own it does not confirm receipt or allocation of the payment. The order remains subject to verification and reconciliation of the payment. No transaction hash is requested in this form.

Do not automatically reuse an old address or an old amount. Check the instructions and the payment status in your history. If you have already made a transfer, avoid sending a second one to randomly correct the situation; contact support with the order reference and the relevant payment information.

Creating instructions requires an available payment method and a valid rate to calculate the amount. If an error prevents this step, the message shown describes the problem and your order remains viewable. Wait until you have complete instructions before transferring funds; an address from another payment must not be used as a replacement.

The top-up amounts offered are 100, 500, 2,000 and 10,000 USD. Each top-up has its own payment and tracking. The balance is credited after confirmation and allocation of the payment, not when you click “I have paid”. A top-up funds the balance; it is not a GPU rental.

Yes, when the available USD balance covers the full amount of the rental. Choose this payment method in the designated flow. The debit settles the order, while preparation and provisioning are tracked separately. A top-up still awaiting confirmation is not part of the available balance.

These contact details identify your account and allow orders to be tracked. Enter an email address you can check and accurate information. Your email and password are used to log in to Kernodeck; no external account or KYC procedure is required. This information remains separate from the payment and provisioning instructions.

Yes. For a first order, create your Kernodeck account with your first name, last name, email and a password of 15 to 128 characters. If you already have an account, log in; if you are already logged in, no password is requested again for each order. The account lets you find your orders and your balance from another browser. Log out after use on a shared device.

Log in to your Kernodeck account with your email and password, then open the folder and its history. You will find the same orders and the same balance from another browser. Each payment keeps its instructions and its status. A new registration does not recover the folders of another account just because the email matches; no automatic password reset by email is available.

No. Confirmed payment and provisioning are two separate steps. After paying, check the readiness status and then the access details actually provided for the order. A payment notification, a top-up or a balance debit does not prove that your program has started.

The duration you choose describes your rental period. Check the provisioning information and the dates shown in the order tracking to plan your work. Do not treat the order record or clicking "I have paid" as proof that execution has begun. If a date you need is missing, ask support to confirm it.

Contact support with the order in question, the duration or configuration you want and your deadline. The possibility and the terms must be confirmed for that request. Keep preparing your exports according to the period already planned until a change has been confirmed in your tracking.

Provide the order, the time of the issue, the launch command, the relevant versions and the full first error. A minimal example, the input shape and the step where the issue appears help to understand the blockage. Remove access secrets and avoid sending your entire dataset when a reduced example is enough.