Repository navigation
llama : keep input embeddings on layer 0 preventing unneeded scheduling thread wakeups - #30152
Closed
deadprogram wants to merge 1 commit into
Closed
deadprogram wants to merge 1 commit into
deadprogram wants to merge 1 commit into
Conversation
This PR is address performance degredation due to unexpected consequences to PR ggml-org#29622. It results in moving moving small operations for thread execution unecessarily, which this PR addresses by looking a little more carefully before doing so, so that the mixed models that ggml-org#29622 is intended to support still work as well. Should address ggml-org#30018 and ggml-org#30033 Signed-off-by: deadprogram <ron@hybridgroup.com>
deadprogram
marked this pull request as draft
October 8, 2026 13:06
deadprogram
marked this pull request as ready for review
October 8, 2026 13:08
deadprogram
changed the title
Contributor
Author
|
I had to change the title for this PR to correctly match what I was trying to do here. Thanks for looking! |
deadprogram
marked this pull request as draft
October 8, 2026 14:09
Member
|
Thanks for the report, but it's not the correct fix. Please test #30160 |
Contributor
Author
|
Testing #30160 now thanks for quick response @ggerganov now closing this PR. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Overview
This PR is to address performance degredation due to unexpected consequences to PR #29622. It resulted in small operations being moved for thread execution unecessarily. This PR addresses it by looking a little more carefully before doing so, and keeping input embeding together whicle still letting mixed models do whatever they need so that #29622 still works as well.
Should address #30018 and #30033
Additional information
Replaces PR #30151 from my personal account.
Should address #30018 and #30033 not sure if any others.
The following shows the automated benchmark results from testing this PR:
RTX 4070 Laptop GPU, CUDA, Q4_K_M models,
llama-bench -ngl 99 -p 512 -n 128 -r 5, with-tas the thread count below. Values are tokens per second.master is upstream 03aa006. b11399 is the release before PR 29622. pp512 varies about 3 percent from run to run.
Requirements
Generative tools used to diagnose the problem and for producing the benchmarks to determine if this PR actually did fix the problem.
UPDATE: reedited again to try to clarify.