You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We were surprised to discover that our abstraction offers a simple and intuitive API, MatmulScheduler.apply_config(**config_dict), which directly generates a scheduled IR module. Moreover, both the scheduler and the configuration dictionary can be serialized into JSON format, enabling seamless saving and reloading from a database. This functionality significantly improves the reusability and portability of scheduling configurations.
To further enhance this functionality, one proposed improvement is to extend the BaseScheduler class. Specifically, we can update the postproc method to store the config_class and config_dict as attributes of the BaseScheduler object. By doing so, these attributes can also be included in the serialization process, ensuring that all necessary scheduling metadata is preserved. This enhancement would make it easier to save and reload complete scheduling states, improving workflow efficiency and robustness.
And which is very helpful for us to precisely analysis some attributes like weight shape when op is loaded from database.
We were surprised to discover that our abstraction offers a simple and intuitive API,
MatmulScheduler.apply_config(**config_dict), which directly generates a scheduled IR module. Moreover, both the scheduler and the configuration dictionary can be serialized into JSON format, enabling seamless saving and reloading from a database. This functionality significantly improves the reusability and portability of scheduling configurations.To further enhance this functionality, one proposed improvement is to extend the
BaseSchedulerclass. Specifically, we can update thepostprocmethod to store theconfig_classandconfig_dictas attributes of theBaseSchedulerobject. By doing so, these attributes can also be included in the serialization process, ensuring that all necessary scheduling metadata is preserved. This enhancement would make it easier to save and reload complete scheduling states, improving workflow efficiency and robustness.And which is very helpful for us to precisely analysis some attributes like weight shape when op is loaded from database.