Repository navigation
Reduce chunked cross-entropy memory test footprint - #8824
Merged
delock merged 2 commits intoOct 15, 2026
Merged
Conversation
Signed-off-by: Masahiro Tanaka <tanaka.masahiro@gmail.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Signed-off-by: Masahiro Tanaka <tanaka.masahiro@gmail.com>
delock
approved these changes
Oct 15, 2026
banxingmjj
pushed a commit
to openanolis/DeepSpeed
that referenced
this pull request
Oct 15, 2026
(deepspeedai#8824 should be merged first) In deepspeedai#8770, we pinned Modal sandbox placement to Oracle Cloud Infrastructure (OCI), where a working NVIDIA driver version was still available, to avoid the CUDA/NCCL failures seen on other Modal hosts. The affected driver has since been rolled out to OCI as well, so this workaround no longer provides a reliable path for our multi-GPU CI workload. Modal has identified the issue and is working on a fix. This PR temporarily disables Modal until the driver issue is resolved, and run the required torch-latest workflow entirely on the existing AWS backend. --------- Signed-off-by: Masahiro Tanaka <tanaka.masahiro@gmail.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The original 8,192-token by 151,936-vocabulary test needs 4.64 GiB for each full FP32 logits-sized tensor. In the reproduced CI stall, a long-lived pytest worker held 28,832 MiB (28.2 GiB) of GPU memory.
This PR reduces the token count from 8,192 to 128 while retaining the 151,936-entry vocabulary: each full FP32 tensor becomes 74.2 MiB, a 64-fold reduction. Preserve loss/gradient tolerances and the two-FP32-tensor memory-savings requirement, and release test-owned CUDA references and cache state on exit.
We now compare gradients in eight-row blocks and calculate the match fraction from integer counts instead of an FP32 mean. The 99.9% threshold is unchanged, but boundary rounding can differ. Check both torch (block_rows=8) and available Triton, instead of selecting just one backend.