Repository navigation
Conversation
The gathered PREDECODER_DECODE_OUTPUT_DIR arrays added in #100 are only tested on the gloo backend. Add a multi-GPU test that decodes the same Stim sample files on two NCCL ranks and compares the counts and the saved per-shot arrays with a single-process run. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Signed-off-by: Ivan Basov <ibasov@nvidia.com>
ivanbasov
marked this pull request as ready for review
November 17, 2026 21:35
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
#100 gathers the
PREDECODER_DECODE_OUTPUT_DIRarrays across ranks withtorch.distributed.gather_object, but that path is only tested on the gloo backend (CPU).multi-gpu-testsruns training and inference on generated samples, so it does not reach it either.This adds
TestOfflineStimDecodeNCCLtocode/tests/test_multi_gpu.py. It writes Stim sample files for X and Z (257 shots, so the remainder shot is exercised), decodes them on two NCCL ranks in both decode modes, and compares the reported counts and every saved per-shot array with a single-process run on the same files. Helpers are reused fromtest_offline_stim_decoding.py.Test-only change, 132 added lines.
Verification
Run manually on a node with 4x GB200 (torch 2.12, NCCL 2.29.7) at e382271:
On a single-GPU host the new class is skipped like the rest of the module (
Ran 4 tests, OK (skipped=4)).multi-gpu-testsonly runs onmain, so on this PR the new test is skipped in CI; it will first run there after merge.🤖 Generated with Claude Code