Skip to content

Gpu Shuffling - #40

Merged
jp-dark merged 12 commits into
mainfrom
as/gpu_shuffle
Sep 24, 2025
Merged

Gpu Shuffling#40
jp-dark merged 12 commits into
mainfrom
as/gpu_shuffle

Conversation

@AaditSaluja

Copy link
Copy Markdown
Collaborator

No description provided.

@AaditSaluja

Copy link
Copy Markdown
Collaborator Author

Adds 2 new shufflinf methods

  • gpu_iobatch/ gpu: Emulates the existing shuffling algorithm just pushing the process of shuffling within an io_batch_size to the GPU. Slightly faster in some scenarios, specially for heavy compute models. Further, with recent trends in hardware, should scale better than cpu shuffling in the future. Bottlenecked by size of gpu memory.

  • gpu_minibatch: Skips the selection of random shuffle_chunk_size sized chunks that are the level 1 of shufflinf algorithm, only shuffling within each mini batch for a more localized version of shuffling. It benefits from runtimes that are within ~4-5% of just not shuffling at all.

I will be adding some tests for these to the testing infrastructure soon

@AaditSaluja
AaditSaluja requested review from jp-dark and ktsitsi August 22, 2025 14:58
@AaditSaluja

Copy link
Copy Markdown
Collaborator Author

With no GPU on the CI runner, gpu shufflinf tests are skipped. Did run them locally and they pass reliably! The branch should be ready to merge

@jp-dark
jp-dark merged commit 8ff15e5 into main Sep 24, 2025
27 checks passed
@jp-dark
jp-dark deleted the as/gpu_shuffle branch September 24, 2025 15:26
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants