test: Add HSTU (Generative Recommenders) Torch AOTI CI test - #8907
Merged
Merged
Conversation
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
Greptile SummaryAdds an end-to-end HSTU Torch AOTI serving test that:
Confidence Score: 5/5The PR appears safe to merge. No blocking failure remains; the current startup path waits for Triton's readiness endpoint and handles startup failure before launching the client. Important Files Changed
Reviews (12): Last reviewed commit: "Merge branch 'main' into yinggeh/tri-161..." | Re-trigger Greptile |
|
Want your agent to iterate on Greptile's feedback? Try greploops. |
L0_torch_aoti_hstu serves the AOTI package that the CI export stages into a docker volume, using the HSTU runtime baked into the test image. It starts the FlexKV KV-cache server, loads the exported package as version 1 of hstu_gr_ranking_kvcache, and replays the dumped inputs through tritonserver, following Devtech-Compute/distributed-recommender: ci/tritonserver_test.sh.
yinggeh
force-pushed
the
yinggeh/tri-1615-add-hstu-ci-test-to-triton-ci
branch
from
August 12, 2026 01:15
1ec12d3 to
f03c291
Compare
whoisj
previously approved these changes
Aug 12, 2026
whoisj
previously approved these changes
Aug 18, 2026
wait_for_server_ready returns as soon as the process is gone and the failure branch prints the server log, which is where a link error from the image's LD_PRELOAD lands.
whoisj
approved these changes
Aug 19, 2026
8 of 11 tasks
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What does the PR do?
Adds
qa/L0_torch_aoti_hstu/test.sh, an end-to-end test for serving an HSTU(Generative Recommenders) ranking model exported as a Torch AOTI package. It
assembles a
platform: "torch_aoti"model repository from the exported package,starts the FlexKV KV-cache server the model talks to, serves the repository, and
runs the HSTU client against it.
Exporting the package needs the recsys-examples training stack, so it happens in
a separate container in the CI job; the package and the input dump the client
replays arrive here in a volume at
/exported_hstu_model. The companion CI joblands in the internal
tritonserverrepository.Checklist
<commit_type>: <Title>Commit Type:
Related PRs:
Internal
tritonserverMR adding theL0_torch_aoti_hstu--PyTorchjob.Where should the reviewer start?
qa/L0_torch_aoti_hstu/test.shis the only file; read it top to bottom. Twothings that look unusual are deliberate: errexit stays off, because the harness
runs tests under
bash -exand the helpers inqa/common/util.shturn it backon, which would make the client's exit-code check dead and skip teardown; and
teardown ignores the server's exit status, for the reason under Caveats.
Test plan:
Ran the CI job end to end. Green, with the client reporting
max_abs_diff<=0.0625 passed 146/146 batches.Caveats:
tritonserversegfaults during shutdown inside anLD_PRELOADed init hookshipped by the model's runtime, after the client finishes and after SIGINT.
Reported to the runtime's owners. Teardown ignores the server's exit status the
same way
kill_serverinqa/common/util.shdoes.from are all supplied by the CI job.
Background
HSTU serving exercises
platform: "torch_aoti"together with a custom backendruntime and an external KV cache, which no existing L0 test covers.
Related Issues: N/A