Repository navigation
[test] Benchmark README: batch functions take their size and random numbers from their arguments - #6568
Closed
JazyJiang wants to merge 1 commit into
Closed
[test] Benchmark README: batch functions take their size and random numbers from their arguments#6568JazyJiang wants to merge 1 commit into
JazyJiang wants to merge 1 commit into
Conversation
…umbers from their arguments
Collaborator
Author
|
Folded into #6566 as its second commit (unchanged), so the README note lands with the benchmark worker change. Closing this one. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Related issue
Follow-up to #6565; documentation only.
Background
A model's
batch_fnis called by the maximum search with changing sizes and on every rank of a multi-device measurement with a differently seeded generator. The README did not say so, so a batch function written with a fixed size breaks the search, and one using the global random state gives every rank the same batch, which makes a MoE router route identically on every rank.Changes
tests/kit/benchmark/README.md, "Adding a model": a paragraph thatbatch_fnbuilds the batch at the size it is given and takes its random numbers fromgenerator, why, and that these functions go into the same pull request as the entry (or intoconfig/data.pywhen several models can use one).Not in this PR: code changes; the existing batch functions (
causal_lm_batch,PackedText,LatentVideo) already follow this.User and compatibility impact
None.
Test environment / Test commands and results
Documentation only; no tests run.
Known issues
None.
Risks and rollback
None; revert the commit.
Documentation and release impact
README only.