Split out of #723.
What goes wrong
When a GPU cannot run any offered CUDA torch build, the desktop app repeats the whole CUDA install on every launch: about 45 seconds of download and install, a failed verification, and a restore. The reporter's setup log shows six cycles.
Why
A CPU result that came from a failure is treated as unsettled, so setup tries again next time. That is right for a network drop. It is wrong for a GPU that has no kernels in any available build, which cannot succeed until a new app version offers different wheels.
Constraints
- A transient failure (network, disk) should still be retried.
- "No kernel image is available" has to be matched as that phrase, not a loose substring, or other failures get treated as permanent.
Split out of #723.
What goes wrong
When a GPU cannot run any offered CUDA torch build, the desktop app repeats the whole CUDA install on every launch: about 45 seconds of download and install, a failed verification, and a restore. The reporter's setup log shows six cycles.
Why
A CPU result that came from a failure is treated as unsettled, so setup tries again next time. That is right for a network drop. It is wrong for a GPU that has no kernels in any available build, which cannot succeed until a new app version offers different wheels.
Constraints