What does "no kernel image is available for execution on the device" mean?

Your PyTorch (or other CUDA library) build has no code for that GPU's generation. It is common on new cards such as the RTX 50 series with older builds. Install a newer build that supports the card, or rent an older card.

What does "Failed to initialize NVML: Driver/library version mismatch" mean?

The machine's NVIDIA driver was updated while the old one was still running. A restart fixes it. On a rented machine it is not your fault and not your code: ask for another machine.

Why does PyTorch's DataLoader crash with a bus error inside Docker?

Docker gives containers only 64 MB of shared memory by default, and PyTorch's data-loading workers pass data through shared memory. Start the container with a bigger --shm-size, or set num_workers=0 as a stopgap.

Can I get a refund if a rented GPU fails?

It depends on the service. Vast.ai's billing docs say there are no refunds once credit is spent. On KernelMachine, tapping "It's broken" stops the rental and returns your money automatically, and if a host drops, your rental resumes elsewhere from its last save and you don't pay for the work lost in between.

For renters · Updated October 13, 2026

Why rented GPU jobs fail, and how to protect your work

KernelMachine is preparing an invite-only beta; no launch date is set. Where this guide describes KernelMachine, it's how it works at launch. Join the waitlist

Short answer: most rented GPU jobs fail in one of five ways: the machine never starts (often while billing runs), the host goes offline mid-job, your files vanish when the rental stops, the card isn't what was listed, or your software doesn't match the card. Save your work to your own storage as you go, test on a short rental first, and choose services that bill from when your app works and refund failures.

What renters complain about most

We read about 60 dated negative or mixed renter reviews of Vast.ai, RunPod, TensorDock, Paperspace and Clore.ai on Trustpilot (October 2026). Trustpilot leans negative, so treat this as a ranking of what goes wrong, not how often.

Complaint	Reviews (about)
The card I had is gone, and a stopped machine can't restart	14
The machine won't start or connect, often while billing runs	12
The host went offline mid-job and work was lost	12
Charged for a broken service, refund refused	11
Slow or no support (usually after a refused refund)	20

Sources: Vast.ai, RunPod and TensorDock on Trustpilot.

1. The machine never starts

The image is huge, the host is slow, or something on the machine is broken. Meanwhile the meter may be running: one Vast.ai renter was charged $1.70 for about 20 minutes of a service that never worked (Trustpilot).

Protect yourself: start with a short test before a long job. Use small, common images that are more likely to be cached. Prefer services that bill from when your app or shell is reachable.

2. The host disappears mid-job

On marketplaces, a host's power cut, router reboot or Windows update can end your rental with no warning. Salad's own docs say jobs "will be interrupted with no warning" (Salad FAQ).

Protect yourself: save checkpoints (your training progress) every 10 to 30 minutes to storage you control, such as S3, Backblaze, Google Drive or a Hugging Face repository. Make your script resume from the latest checkpoint. Then a lost machine costs minutes, not days.

3. Your files vanish when the rental stops

Some services delete the disk when you stop. Others keep it, but the GPU goes to the next renter, and your files are stuck until it frees up. RunPod wrote a whole guide on it, titled "Where did my files go?" (RunPod blog), and its docs explain why a stopped machine can come back with no GPU (RunPod docs).

Protect yourself: sync outputs to your own storage as they're made (the free tool rclone handles most clouds). Never rely on a rented disk as the only copy.

4. The card isn't what was listed

Renters report hosts who quietly cut performance, machines that lost GPUs mid-rental, and downloads at about 1 MB/s, roughly 1% of the listed speed (Vast.ai on Trustpilot). In 2024, io.net found about 1.8 million fake GPUs on its network (Cointelegraph).

Protect yourself: in the first minutes, run nvidia-smi to see the cards and their memory, run a short benchmark of your own job, and test a download. If anything is off, stop at once and report it.

5. Your software doesn't match the card

Many failures show up as an error message. Here's what the common ones mean.

Error message	What it means	What to do
`no kernel image is available for execution on the device`	Your PyTorch (or other CUDA) build has no code for this GPU's generation. Common on RTX 50-series cards with older builds.	Install a build that supports the card (for RTX 50, PyTorch builds for CUDA 12.8 or newer), or rent an older card.
`CUDA out of memory`	The model, batch or image size doesn't fit in GPU memory.	Smaller batch, lower precision, or a bigger card. See how much GPU memory you need.
`Failed to initialize NVML: Driver/library version mismatch`	The machine's NVIDIA driver was updated while the old one was still loaded (Hyperstack).	A restart fixes it. On a rental, it's the host's problem: switch machines.
`Failed to initialize NVML: Unknown Error` inside a container	The container lost access to its GPU, often after a routine system reload on the host (NVIDIA).	Restart the container or switch machines.
`Bus error` or "DataLoader worker killed"	Docker gives containers only 64 MB of shared memory by default, and PyTorch's data loader needs more.	Start the container with a bigger `--shm-size`, or set `num_workers=0` as a stopgap.
`No devices were found` from `nvidia-smi`	The driver can't see the card. On RTX 50 cards, this happens with the wrong driver variant (NVIDIA forums).	Nothing you can fix on a rental. Switch machines.

6. The refund is refused

Vast.ai's billing docs say there are no refunds once credit is spent (Vast.ai billing). In the reviews above, slow support usually showed up after a refused refund, and TensorDock sits at 1.6 out of 5 on Trustpilot, with machines that never start and support that never answers (Trustpilot).

Protect yourself: read the refund policy before topping up, top up small amounts at first, and keep notes (times, machine ID, error messages) when something breaks.

A protection checklist

Run a short test job before the long one.
Check the card, its memory and download speed in the first five minutes.
Save checkpoints and outputs to your own storage every 10 to 30 minutes.
Make your script resume from the last checkpoint.
Use a template image built for the card you rented.
Set a spend cap or a stop time.
Know the refund policy and file-keeping rules before you start.

How KernelMachine handles each one

Failure	What KernelMachine does
Machine never starts	Billing starts only when your rental is ready. We check the card works first, and downloads and set-up are free.
Machine drops out mid-job	Top hosts are offered rentals first. Home machines can still lose power or internet. If it happens, your rental can recover from its last complete save on another available card of the same model, never the one that failed. You don't pay for the work lost in between, and the host isn't paid for it. Offline time is never charged.
Files lost at stop	Files on a stopped rental are deleted, so save your work elsewhere as you go. A rental left unused gets an email after 30 minutes and stops after an hour.
Card not as listed	Every machine passes automatic checks before renters see it, and most also a 48-hour trial. Machines that fail renters are removed.
Bill bigger than planned	No storage or transfer fees. Set a time limit when you start and the rent page shows the most it can cost; the rental stops by itself, with an email 10 minutes before.
Refund refused	"It's broken" stops the rental and refunds it automatically.