feat: update grouped_topk to support softmax and sigmoid#3680
Conversation
0f3db83 to
af2ee75
Compare
af2ee75 to
55c2e68
Compare
|
@FrankLeeeee Thanks @zixuanzhang226 could you rebase with the main? I can merge it tommorow. |
55c2e68 to
f84b584
Compare
Hi @zhaochenyang20 , I rebased with main. Thank you! |
|
@zixuanzhang226 Still out of main. And, could you add a description to this PR. |
f84b584 to
9ec7746
Compare
|
@zixuanzhang226 could you rebase this with the main? And, do we need to merge it? @FrankLeeeee |
I rebased it! Thanks! |
|
@zhaochenyang20 I also updated the description! Thank you! |
|
@zhaochenyang20 this pr can be merged |
Hello @zhaochenyang20, would you please merge it? Thank you! |
|
@zixuanzhang226 I don't know why I can't rebase your PR. Could you rebase it and I can run the CI. I don't have access to merge main into your PR. |
|
@zhaochenyang20 squash and merge? |
…)" This reverts commit 0c227ee.
Motivation
This PR is to support both softmax and sigmoid scoring functions in grouped_topk. Also, verified DeepSeek V2/V3/R1 uses biased_grouped_top and updated the corresponding comments.
Ref #2739
Modifications
Checklist