Repository navigation
Fix the Inkling per-expert sync test and collect it in the weekly CPU run - #40725
Merged
Merged
Conversation
Collaborator
Author
|
/rerun-test test_inkling_per_expert_sync.py |
Contributor
|
Results for 🚀 |
The per-expert sync test stubbed `get_moe_expert_parallel_world_size`, `get_moe_expert_parallel_rank` and `get_moe_tensor_parallel_rank` on the model module. The loader stopped importing those names, so the assignments created three attributes nobody reads and the reads went to the real context, failing all four cases with `AttributeError` on the saved-attribute lookup. The rank now publishes the topology it is standing in, and the loader reads it the same way it does in a real process.
It sat under `test/srt/`, which no workflow collects, so it went on failing against a module that had stopped exporting the names it stubbed without anything reporting it. Moving it into the registered tree and giving it a weekly CPU registration makes that class of rot visible.
ch-wan
force-pushed
the
cheng/hot-fix/inkling-per-expert-sync-context
branch
from
September 22, 2026 10:01
af3f3ec to
ff51757
Compare
ch-wan
changed the title
ch-wan
changed the title
ch-wan
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
test/srt/models/test_inkling_per_expert_sync.pyfails all four of its cases onmain.It saves and restores
get_moe_expert_parallel_world_size,get_moe_expert_parallel_rankandget_moe_tensor_parallel_rankonsglang.srt.models.inkling, then reassigns them to drive a simulated EP x MoE-TP grid. The loader stopped importing those names, sosetUpraisesAttributeErrorbefore any case does work, and the assignments would in any event have created three attributes nothing reads.Nothing reported this, because the file sits under
test/srt/, which no workflow collects.Modifications
ep_rank * tp_size + tp_rankundertp_size = ep_size * tp_sizereproduces exactly the grid the test asserts over.lora_compatible_layout_enabledis still a module-level binding and stays a stub.test/registered/unit/models/and carriesregister_cpu_ci(est_time=15, stage="weekly", runner_config="cpu"), so it is collected into the weekly CPU run.Accuracy Tests
No production code changes; this is a test fix.
test_inkling_per_expert_sync.pygoes from 4 failed to 4 passed, from its new path, on CPU.Run in one process with
unit/test_runtime_context.pyto check the published topology is torn down rather than leaked into a neighbour: 170 passed.Speed Tests and Profiling
Not applicable.
Checklist
Review and Merge Process
/tag-and-rerun-ci,/tag-run-ci-label,/rerun-failed-ciCI States
Latest PR Test (Base): ❌ Run #35713663872
Latest PR Test (Extra): ❌ Run #35713663791
Latest PR Test (AMD ROCm 10): ❌ Run #35713663829