Fix Qwen3 decode all-reduce buffering and add benchmarks#270
Open
chhwang wants to merge 11 commits into
Open
Conversation
Foundation off main for benchmarking ARK fused-packet all-reduce. Excludes the double-buffer/monotonic-flag/ARKSTUCK instrumentation (those fixed a non-existent hang; ARK's loop kernel owning the GPU is by design). That work is preserved on branch allreduce-packet-repeat-hang. Prerequisites (from #268): - codegen.cpp external-buffer OFFSET fix (+ codegen_test.cpp) - ops_all_reduce.cpp copy input into registered buffer (+ ops_all_reduce_test.cpp) - examples/qwen3 harness (__init__.py, _env.py) bench_allreduce.py rewrite: times rt.run(iter=N) with host wall-clock + rt.barrier() alignment. NO torch.cuda.synchronize / CUDA events / torch GPU ops while the ARK runtime is live (they deadlock behind the persistent loop kernel). Reports ARK us vs mscclpp ceiling (decode ~11.7us, prefill ~188us).
…educe-bench: codegen OFFSET fix + ops copy-into-registered-buffer + 3 C++ regression tests + torch-free bench_allreduce.py + d2h-safe test_allreduce.py equivalence tests.
…educe-bench: codegen OFFSET fix + ops copy-into-registered-buffer + 3 C++ regression tests + torch-free bench_allreduce.py + d2h-safe test_allreduce.py equivalence tests.
…alize test_allreduce.py d2h-safe equivalence tests, add bench caveats, and open the PR off main.
… latency, validate TP=2 and TP=8, then open the clean PR.
…sentinel decode latency, and open the clean PR.
…rf logs that include .
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #270 +/- ##
==========================================
+ Coverage 85.78% 85.85% +0.06%
==========================================
Files 129 129
Lines 6495 6498 +3
==========================================
+ Hits 5572 5579 +7
+ Misses 923 919 -4 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
chhwang
changed the title
chhwang
added a commit
that referenced
this pull request
Jun 18, 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
1 participant
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.
Fix Qwen3 decode all-reduce buffering
Preserve external-buffer offsets in generated kernels and copy all-reduce inputs into registered buffers. This lets Qwen3 decode all-reduce run from external tensors safely.
Add C++ regression tests, D2H-safe Qwen3 equivalence tests, and an ARK-safe microbenchmark. Record TP=2 and TP=8 max-rank decode perf evidence tied to
15c53bc.