Repository navigation
[trainer memory] Split Nemotron-3.5-Lightning 64K experts across both GPUs again - #331
Closed
Micah Yong (micahtyong) wants to merge 1 commit into
Closed
Micah Yong (micahtyong) wants to merge 1 commit into
Micah Yong (micahtyong) wants to merge 1 commit into
Conversation
Reverts #246 until a layout that avoids the per-call gradient copy is measured. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Micah Yong (micahtyong)
requested a review
from Kailash Ranganathan (kailash109)
as a code owner
November 17, 2026 03:32
Collaborator
Author
|
Superseded by #332, which folds the B300 layout into |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
#246 put all 128 experts of
nemotron35-lightning-30b-a3b-lora-64kon both GPUs (expert_model_parallel_size1). Since #257, that layout has to set aside a copy of the adapter gradients during every training call, because the gradient reduce on this layout would otherwise rescale gradients kept from earlier calls. On this recipe the copy is 12.85 GB per GPU. Measured before #257, four clients each training a 64K-token row already peaked at 130 GB of a 140 GB H200, and 16K-token rows at 65,536 tokens per micro-batch reserved 132 GB, so with the copy those cases are expected to run out of memory. A run on main to confirm is in progress.This PR splits the experts across the two GPUs again (
expert_model_parallel_size2), the layout the recipe used before #246. On that layout the reduce leaves kept gradients unchanged, so #257 skips the copy. Training is slower on short-row batches (about 12.2 s instead of about 9.1 s offorward_backwardfor 128 rows of 1,600 tokens on 2x H200), until a layout that keeps the speed without the copy is measured.E2E results
Setup: 2x H200, one client, stock Tinker SDK, LoRA rank 32. Numbers are from earlier runs of these two layouts; this branch is a one-line revert.