Mixed Precision Grouped Gemm with zero points and GPT-Q semantics closes #2261

[ankutalev]

Hello!
This MR provides two things:

1. Zero points for default mode
2. GPT-Q semantics

Closes #2261

[thakkarV]

sorry, running a bit behind. we will get to it soon.

[Junkai-Wu]

@ankutalev Thanks for submitting this feature MR. Have you checked the functionality of this feature? Could you post the result of running this feature (example 69) here?

[ankutalev]

@ankutalev Thanks for submitting this feature MR. Have you checked the functionality of this feature? Could you post the result of running this feature (example 69) here?

Yes, I checked - it shows "Disposition Passed" for all scenarios ({shuffled/unshuffled} X {direct convert, no zeros, zeros, gptq}).
Which is not good also - because new gptq semantics dequantizes matrix in different way; the test in examples is weak.

I can provide unit tests if you like.

Also I don't like the way I implemented gptq mode switch, but runtime parameters seems like "not cutlass style"; I will apreciate any advices and suggestions here =)

We are interested in this functionality in main branch, because nobody likes to have patched forks =)

[ankutalev]

@Junkai-Wu Hi! any uppdates here?

[Junkai-Wu]

@ankutalev we are reviewing the changes internally. Will merge this PR once got approved and merged in our internal repo.

[ankutalev]

@ankutalev we are reviewing the changes internally. Will merge this PR once got approved and merged in our internal repo.

Hi! Any news here?

[github-actions]

This PR has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this PR if it is no longer required. Otherwise, please respond with a comment indicating any updates. This PR will be labeled inactive-90d if there is no activity in the next 60 days.