Repository navigation
[serge] Fix 4 integration tests for model minicpm3 failing with other (other (4)) - #49041
sergereview[bot] wants to merge 1 commit into
Conversation
…her` (other (4))
|
[For maintainers] Suggested jobs to run (before merge) run-slow: minicpm3 |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
CI recapDashboard: View test results in Grafana |
ArthurZucker
left a comment
There was a problem hiding this comment.
makes sense.
BTW SERGE can you open a PR to their repo to remove the remote code?
|
Thanks for the fix! The |
Original CI failure
4 integration tests for modelminicpm3failing withother(other (4))tests/models/minicpm3/test_modeling_minicpm3.py::MiniCPM3IntegrationTest::test_minicpm3_4b_generation[multi-gpu] (other, seen 7/7)CI traceback —
tests/models/minicpm3/test_modeling_minicpm3.py::MiniCPM3IntegrationTest::test_minicpm3_4b_generationWhere to watch it:
The MiniCPM3 integration tests crash during
from_pretrainedbecauseMiniCPM3Attentioninherits ayarn_apply_mscalehelper that unconditionally readsrope_parameters["factor"].The MiniCPM3-4B checkpoint uses
rope_type="longrope"and does not setfactor, so the access raisesKeyError.The helper only needs
factorwhenmscale_all_dimis set, so move the read inside that branch.This override in the modular file fixes the crash without touching other models that share the copied helper.
Relates to #49034
✅ Verified on GPU
serge ran the targeted
@slowtest(s) on a GPU runner and opened this PR only after they passed with this patch. Each targeted test was run 5× on both the pre-patch and patched trees to rule out flakiness — the result held on every run.Possibly related
Existing issues/PRs mentioning
test_minicpm3_4b_generation(keyword match — not verified to share a root cause):This change was produced automatically by serge from a CI failure report. The patch was generated by an LLM and applied by serge; review before merging.
serge
v0.1.0· model:moonshotai/Kimi-K2.7-Code· 20 LLM turns · 19 tool calls · 66.0s · 563470 in / 11294 out tokens