-
Story
-
Resolution: Unresolved
-
Undefined
-
None
-
None
-
4
-
PyTorch Sprint 25, PyTorch Sprint 26
Currently, the best configuration and kernel execution time are printed after each generation round. If there are still active search paths, the next generation round begins. However, there is no mechanism to gracefully stop the auto tuning process and ensure that the best configuration from the run is saved to the on-disk directory used for cached autotuning artifacts. If CTRL+C is used after a successful generation round but before the remaining rounds complete, the best config is not saved. Since the generation rounds can lead to diminishing returns, a user might want to halt the process early to save time. While the best config can be parsed from the autotune logging, a user would then have to hard-code the config rather than relying on the cached artifact in future runs. This issue proposes adding user-driven termination of autotuning with cached results.
- clones
-
AIPCC-8685 Helion Devcontainer
-
- Closed
-
- is cloned by
-
AIPCC-9986 [Helion] Enable early termination of bad configs during optimization
-
- Review
-
-
AIPCC-9508 [Helion] More human readable logging
-
- Closed
-