Repository navigation
[trainer throughput] Cap Nemotron-3.5-Lightning 64K micro-batches at 49,152 tokens - #247
Merged
Micah Yong (micahtyong) merged 1 commit intoNov 17, 2026
Conversation
This was referenced Nov 17, 2026
Micah Yong (micahtyong)
force-pushed
the
micah/nemotron-perf/trainer-4-micro-batch-49152
branch
from
November 17, 2026 02:22
d7e6542 to
d335cac
Compare
Micah Yong (micahtyong)
force-pushed
the
micah/nemotron-perf/trainer-3-expert-parallel-1
branch
from
November 17, 2026 02:33
266c2f3 to
edc6695
Compare
Micah Yong (micahtyong)
changed the base branch from
micah/nemotron-perf/trainer-3-expert-parallel-1
to
main
November 17, 2026 02:46
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Micah Yong (micahtyong)
force-pushed
the
micah/nemotron-perf/trainer-4-micro-batch-49152
branch
from
November 17, 2026 02:58
1a29a48 to
c8c8e9f
Compare
Micah Yong (micahtyong)
marked this pull request as ready for review
November 17, 2026 03:33
Micah Yong (micahtyong)
requested a review
from Kailash Ranganathan (kailash109)
as a code owner
November 17, 2026 03:33
Micah Yong (micahtyong)
deleted the
micah/nemotron-perf/trainer-4-micro-batch-49152
branch
November 17, 2026 03:33
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
max_tokens_per_gpusets how many tokens the trainer packs into one micro-batch. At 65,536, a batch of 16K-token rows fills each micro-batch to the limit. With all experts on every GPU (the previous PR), those full micro-batches push reserved GPU memory to 132 GB of 140 GB, and the step slows down as the allocator runs out of room.This PR lowers
max_tokens_per_gputo 49,152 innemotron35_lightning_30b_a3b_lora_64k. A row longer than that still trains as its own micro-batch, so 64K-token rows work as before. The 256k recipe keeps 65,536.E2E results
Setup: 2x H200, one client, stock Tinker SDK, LoRA rank 32. Short rows are 128 rows of about 1,600 tokens per step, long rows are 8 rows of about 16,000 tokens, and 64K rows are 2 rows of 64K tokens. Per-token log probabilities match the previous PR to within run-to-run noise.