perf: significantly improve rendering performance by reducing GC pressure #687
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
The `renderIterator` function previously caused an extremely large amount of small strings and other objects to be created and abandoned during rendering, which caused performance issues after some time as the GC had to occasionally collect all of these objects. This was exacerbated by using streaming in models, which leads to extremely frequent updates. This commit refactors renderIterator to avoid constructing temporary strings. Instead, the function now performs two passes: 1. A first pass in which the "fragments" to render are aggregated, but the `rendered` string is not yet copied and appended/prepended to. 2. A second pass, which uses a `strings.Builder` to efficiently construct the final output string. This has *significantly* improved crush's performance for me. Whereas before `perf` would show it spending up to 70% (!) of its time in GC-related Go runtime functions, it now spends a trivial amount there. pprof's heap profiling previously showed renderIterator as a massive hotspot, whereas it now doesn't even show up in alloc `top` anymore. The updated function is slightly harder to read. I did spend some time trying different options for making it more readable, and also asking various LLMs about it (using crush!), but ultimately didn't find anything better than the two-pass solution.
038013e to
ddbe51c
Compare
|
By the way, in case anybody else wants to look at stuff like this: There's a second pathological case somewhere when rendering individual lines in ultraviolet. When this occurs there's weird input lag until rendering "catches up". I suspect that something might be off with the logic around changed lines in |
|
Appreciate this, @tazjin! We'll prioritize this one. @aymanbagabas: heads up about UV. |
2 tasks
|
Thanks! I found a low-hanging fruit in ultraviolet and sent a PR for that, but I think there's some more tricky issue hiding somewhere in this. |
|
I've created a tracking ticket for performance issues to avoid spamming this PR with stuff that isn't directly related to this change: #691 |
raphamorim
approved these changes
Aug 11, 2025
aymanbagabas
approved these changes
Aug 11, 2025
aymanbagabas
added a commit
that referenced
this pull request
Aug 11, 2025
scalabl3
pushed a commit
to metaphori-ai/crush
that referenced
this pull request
Aug 19, 2025
The `renderIterator` function previously caused an extremely large amount of small strings and other objects to be created and abandoned during rendering, which caused performance issues after some time as the GC had to occasionally collect all of these objects. This was exacerbated by using streaming in models, which leads to extremely frequent updates. This commit refactors renderIterator to avoid constructing temporary strings. Instead, the function now performs two passes: 1. A first pass in which the "fragments" to render are aggregated, but the `rendered` string is not yet copied and appended/prepended to. 2. A second pass, which uses a `strings.Builder` to efficiently construct the final output string. This has *significantly* improved crush's performance for me. Whereas before `perf` would show it spending up to 70% (!) of its time in GC-related Go runtime functions, it now spends a trivial amount there. pprof's heap profiling previously showed renderIterator as a massive hotspot, whereas it now doesn't even show up in alloc `top` anymore. The updated function is slightly harder to read. I did spend some time trying different options for making it more readable, and also asking various LLMs about it (using crush!), but ultimately didn't find anything better than the two-pass solution.
scalabl3
pushed a commit
to metaphori-ai/crush
that referenced
this pull request
Aug 19, 2025
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.
This is one of a series of changes related to #691
The
renderIteratorfunction previously caused an extremely large amount of small strings and other objects to be created and abandoned during rendering, which caused performance issues after some time as the GC had to occasionally collect all of these objects.This was exacerbated by using streaming in models, which leads to extremely frequent updates.
This commit refactors renderIterator to avoid constructing temporary strings. Instead, the function now performs two passes:
A first pass in which the "fragments" to render are aggregated, but the
renderedstring is not yet copied and appended/prepended to.A second pass, which uses a
strings.Builderto efficiently construct the final output string.This has significantly improved crush's performance for me. Whereas before
perfwould show it spending up to 70% (!) of its time in GC-related Go runtime functions, it now spends a trivial amount there.pprof's heap profiling previously showed renderIterator as a massive hotspot, whereas it now doesn't even show up in alloc
topanymore.The updated function is slightly harder to read. I did spend some time trying different options for making it more readable, and also asking various LLMs about it (using crush!), but ultimately didn't find anything better than the two-pass solution.
Describe your changes
Related issue/discussion:
Checklist before requesting a review
CONTRIBUTING.md