Skip to content

ggml-hexagon: fix swiglu failure at test-backend-ops part2 #17449

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
wants to merge 6 commits into from
Closed

ggml-hexagon: fix swiglu failure at test-backend-ops part2 #17449

wants to merge 6 commits into from
+45 −41

Conversation

@chraac
Copy link
Contributor

@chraac chraac commented Nov 23, 2025 •
edited
Loading

In PR #17344, we fixed some swiglu/silu ops but left hvx_fast_sigmoid_f32 unaddressed. This PR completes the fix by resolving the remaining failed swiglu/silu ops.

Changes

  • Fix hvx_fast_sigmoid_f32 implementation to resolve remaining NaN and accuracy issues

Before

[SILU] NaN at index 231 (HTP0=-nan CPU=88.449669)   SILU(type=f32,ne_a=[128,2,2,2],v=0): �[1;31mFAIL�[0m
  SILU(type=f32,ne_a=[5,7,11,13],v=0): �[1;32mOK�[0m
[SWIGLU] NaN at index 122 (HTP0=nan CPU=-11446.431641)   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,swapped=0): �[1;31mFAIL�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,swapped=0): �[1;32mOK�[0m
[SWIGLU] NMSE = 3.835742624 > 0.000000100   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,swapped=1): �[1;31mFAIL�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,swapped=1): �[1;32mOK�[0m
[SWIGLU] NaN at index 216 (HTP0=nan CPU=-8444.154297)   SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,split): �[1;31mFAIL�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,split): �[1;32mOK�[0m

After

  SILU(type=f32,ne_a=[128,2,2,2],v=0): �[1;32mOK�[0m
  SILU(type=f32,ne_a=[5,7,11,13],v=0): �[1;32mOK�[0m
  SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,swapped=0): �[1;32mOK�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,swapped=0): �[1;32mOK�[0m
  SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,swapped=1): �[1;32mOK�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,swapped=1): �[1;32mOK�[0m
  SWIGLU(type=f32,ne_a=[128,2,2,2],v=0,split): �[1;32mOK�[0m
  SWIGLU(type=f32,ne_a=[5,7,11,13],v=0,split): �[1;32mOK�[0m

@chraac chraac marked this pull request as draft November 23, 2025 16:02
@chraac chraac changed the title [WIP]ggml-hexagon: fix swiglu failure at test-backend-ops part2 ggml-hexagon: fix swiglu failure at test-backend-ops part2 Nov 23, 2025
@chraac chraac marked this pull request as ready for review November 23, 2025 16:06
@github-actions github-actions bot added the ggml changes relating to the ggml tensor library for machine learning label Nov 23, 2025
@loci-dev loci-dev mentioned this pull request Nov 23, 2025
Open
@max-krasnyansky
Copy link
Collaborator

max-krasnyansky commented Nov 23, 2025

In PR #17344, we fixed some swiglu/silu ops but left hvx_fast_sigmoid_f32 unaddressed. This PR completes the fix by resolving the remaining failed swiglu/silu ops.

Changes

  • Fix hvx_fast_sigmoid_f32 implementation to resolve remaining NaN and accuracy issues

@chraac
Looks like these changes already showed in #17212
I just tested and merged that and we're now passing all SWIGLU and SILU tests.
I'll close this one as duplicate. Please reopen if I got that wrong.

Excellent work! Thanks for updates!

@chraac
Copy link
Contributor Author

chraac commented Nov 24, 2025

I'll close this one as duplicate. Please reopen if I got that wrong.

Np, Let's close this PR and merge the other one instead. This PR was created specifically to merge the swiglu fix without including the buffer rework changes.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@max-krasnyansky max-krasnyansky Awaiting requested review from max-krasnyansky max-krasnyansky is a code owner

@lhez lhez Awaiting requested review from lhez lhez is a code owner

Assignees

No one assigned

Labels

ggml changes relating to the ggml tensor library for machine learning

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@chraac @max-krasnyansky