custom config filter

[shunting314]

custom config filter

Motivations

1. user may want to be aggressively filter small block sizes for their kernel but don't want to make such call for all kernels
2. would be handy to do investigation like avoid TMA here not use TMA for symm mem tensor #1846

[jansel]

Could we use autotune_config_overrides for this?

[yf225]

related discussion: #1415

[shunting314]

Could we use autotune_config_overrides for this?

I think they target different things.

autotune_config_overrides would override fields of a helion.Config. It does not drop a hl.Config entirely during autotuning.

While the custom config filter can ignore configs either because block-size is too small or we want to avoid using TMA for some reasons.

The filter can cover all cases mentioned in #1415 btw

[jansel]

Maybe we could extend the existing API? Another other ideas considered?

[shunting314]

Maybe we could extend the existing API? Another other ideas considered?

I think a filter function is more flexible then an overriding dict. We can do more things with the filter function. E.g. with overriding function, we can not skip config with too small block size.

One idea to consolidate the two APIs is, we let the filter function return Config | None rather than bool. If the filter return None, autotuner ignore the config; if the filter function returns a Config, it would be a potentially changed Config with overriding. Basically the filter function provides a super set functionality of the existing config overriding API. wdyt?

[jansel]

Let's go with the Config | None path. Re-request review once this is updated.

[shunting314]

Updated the API to allowing both filtering & overriding a config.

Should we remove Settings.autotune_config_overrides? @jansel . I can do that in a followup PR if we want to consolidate to a single API.