Release prior vocabulary-head chunk references before the next projection - #929
Merged
Merged
Conversation
Drop unused prior-chunk tensors before the next head-statistics call. Exercise the real two-chunk CPU caller in five output modes and both grad modes with existing head helpers. This change does not add estimator accounting or claim CUDA reclamation.
Apply repository Ruff formatting without changing test AST or production.
bradhilton
had a problem deploying
to
trainer-rank-gpu-validation
September 21, 2026 10:26 — with 
GitHub Actions
Failure
GitHub Actions
Failure
bradhilton
deployed
to
trainer-rank-gpu-validation
September 21, 2026 11:09 — with
GitHub Actions
Active
GitHub Actions
Active
This was referenced Sep 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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Vocabulary-head projection keeps the previous chunk's logits in local variables while computing the next checkpoint call. Gathered logits and eager top-k selections can also remain referenced after their last use, extending their lifetimes into the next chunk.
Release those references after their final use. Projection calculations, checkpointing and returned outputs are unchanged; all runtime changes are in
art.trainer_rank.The regression covers target, logits, top-k and mixed outputs with gradients enabled and disabled. It checks reference lifetimes at the next chunk and finite backward gradients, and preserves any pre-existing CUDA initialization state.
Validation: all ten targeted lifetime/output/gradient cases passed. Exact-head CPU CI and the hosted two-H200 validation passed, including Ruff, formatting, types, unit tests and distributed GPU checks. The lifetime cases themselves use CPU tensors, so these results do not establish a CUDA peak-memory reduction. Merged-main CPU CI also passed; native memory qualification remains separate.
Related to #848 and #870.