Repository navigation
perf: Unroll the threefry2x32 rounds in Nx.Random - #1891
Merged
polvalente merged 3 commits intoOct 22, 2026
Merged
Conversation
The 20 rounds of threefry2x32 ran in two while loops. A backend that evaluates each iteration of a while loop separately, such as EMLX, then does many small device calls for each draw. The rounds have a static count, so the loops now use generators with unroll: true, and the rounds become part of the traced expression. The values do not change.
josevalim
approved these changes
Oct 21, 2026
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Nx.Randomran the 20 rounds of threefry2x32 in twowhileloops. The round count is static, so the loops now use generators withunroll: true, and the rounds become part of the traced expression. The values do not change.EMLX evaluates each iteration of a
whileloop separately, so each draw did many small device calls. On the GPU of an M2 Pro, with EMLX 0.5.0:uniform, shape{2, 49152}fold_inOn EXLA (CPU, same Mac), compile takes longer and a large draw runs faster:
uniform, shape{2, 49152}uniform, shape{16}The sums of the drawn values are equal before and after on both backends.
nx/test/nx/random_test.exsandexla/test/exla/random_test.exspass.I found this in a text-to-speech model whose decoder draws dropout masks at each step. The draws took about 21 ms of a 25 ms step on EMLX.