Fix bug in MMVQ kernel #446
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Thank you for the fix!🍻
…On Fri, 23 May 2025, 12:17 Kawrakow, ***@***.***> wrote:
After a very long bug hunt, this PR should hopefully fix #389
<#389>, #398
<#398>, #425
<#425>.
Thanks to everybody who tested my previous bug fix attempts!
Huge kudos to @ciprianveg <https://github.com/ciprianveg> who was
instrumental in finding the bug!
The bug was in the CUDA matrix-vector multiplication kernel (a.k.a.,
MMVQ). It only shows up when the kernel processes 2 or 3 tokens. Hence, it
was not observed during TG, and only showed up during PP when an expert in
a MoE model ended up with having to process just 2 or 3 tokens from the
batch (which is rare).
I believe all other changes I made in #442
<#442> are not necessary,
but please test this PR to confirm.
Closes #389 <#389>
Closes #398 <#398>
Closes #425 <#425>
------------------------------
You can view, comment on, or merge this pull request online at:
#446
Commit Summary
- 193a15b
<193a15b>
Fix bug in MMVQ kernel
File Changes
(1 file <https://github.com/ikawrakow/ik_llama.cpp/pull/446/files>)
- *M* ggml/src/ggml-cuda/mmvq.cu
<https://github.com/ikawrakow/ik_llama.cpp/pull/446/files#diff-215515d65e174fb02240522a4bb36f5c8f974d129f7a8d1aa6026a4dbd8dff12>
(5)
Patch Links:
- https://github.com/ikawrakow/ik_llama.cpp/pull/446.patch
- https://github.com/ikawrakow/ik_llama.cpp/pull/446.diff
—
Reply to this email directly, view it on GitHub
<#446>, or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AJTBYK7WCU4ARPNJHW3ML4D273RTLAVCNFSM6AAAAAB5YG7EMGVHI2DSMVQWIX3LMV43ASLTON2WKOZTGA4DKNZVGEYDGNA>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
|
|
I can happily confirm that this PR seems to have fixed the issues on my end! Thank you! |
|
Amazing, thanks for all your work! |
|
Thank you! |
|
It's working fine now, thank you for your patience |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
After a very long bug hunt, this PR should hopefully fix #389, #398, #425.
Thanks to everybody who tested my previous bug fix attempts!
Huge kudos to @ciprianveg who was instrumental in finding the bug!
The bug was in the CUDA matrix-vector multiplication kernel (a.k.a., MMVQ). It only shows up when the kernel processes 2 or 3 tokens. Hence, it was not observed during TG, and only showed up during PP when an expert in a MoE model ended up with having to process just 2 or 3 tokens from the batch (which is rare).
I believe all other changes I made in #442 are not necessary, but please test this PR to confirm.
Closes #389
Closes #398
Closes #425