CUDA Graph Compute Function Refactor (precursor for performance improvements) #11042
+221
−185
Conversation
This file contains 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
…meters) to separate function for improved readability.
…aluation and capture to its own function.
… for improved readability.
github-actions
bot
added
Nvidia GPU
|
FYI: setting Status to Draft whilst I investigate the failed tests. |
I think you just need to make the functions |
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Outdated suggestions cannot be applied.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.
Hi All,
I am working on improving llama.cpp's CUDA graph performance on behalf of NVIDIA.
In preliminary testing, we are seeing up to 3% of performance gain by overlapping CPU and GPU work, and by improving CPU -> GPU copy scheduling on a high end system. The changes are likely to be even more impactful on less capable hardware.
To pave the way for these changes (and to provide readable diffs), I first isolated the cosmetic changes in this PR.
This PR does not contain any changes in the logic. It merely slims down the
ggml_backend_cuda_graph_compute()by moving certain loops and other subtasks of the original function into 5 new functions.These changes considerably improve the readability and future maintainability of this part of the CUDA backend.
Should I add prefixes to the new function names, and if so, what do you suggest?
@agray3 @mtavenrath