Repository navigation
fix(worker): carry the dp supervisor review follow-ups - #785
Merged
Merged
Conversation
The default named v0.3.1, a build that predates the gateway's attention-dp-aware worker registration, so the --dp-aware flag the renderer emits for the SGLang External DP shape reached a gateway that ignores it. v0.4.0 is the same bundle with that patch carried, which is the image the render is written against. Signed-off-by: thxCode <thxcode0824@gmail.com> (cherry picked from commit 543c210762acacc3f0a0d20b69fedcec9056df2c)
The group fact's partial branch returned before the retention was computed, so an eligible group whose membership was complete, unchanged and current could never restore while the Router held half its endpoints - recorded as a survivor held AwaitingObservation for a whole observation budget behind a gateway that registers one rank per group. The retention is proved by the fresh membership rather than by the view, the same principle the lost-collection path already states, so it is now computed before the view is judged and carries such a group through the half view; the serving answer still publishes the partial reason. Pinned by the recorded shape: a labeled two-rank group with one bound rank retains and restores, and a group whose generation moved does not. Signed-off-by: thxCode <thxcode0824@gmail.com> (cherry picked from commit fb481b9e2f567bba6ad911ce720207b41ebaee0f)
…ter matrix The sglang-gateway cannot run its disaggregation mode beside dp-aware - combining the two produced an invalid decode destination in the recorded single-host container run - which is why the split render never carries the flag. The caveat lived only in the render's source comment after the prefill-decode paragraphs were folded away; the matrix's PD row now states it once where a reader picks a router. Signed-off-by: thxCode <thxcode0824@gmail.com> (cherry picked from commit 1c4dc1ff340a2ef2a89f86b9936c71eadce13183)
The audit's ten approved items: the seven duplicated helper-bullet prohibitions collapse into one Layout line (K1) and the Flow head drops the enumeration orchestration.md owns (K2); the three stale helper sourcing lists are resynced to the measured sets (C2) and both references/ lists name the same five entries (C6); the posture rule enumerates the four values in use and requires one on every title line, the free-text and missing postures are normalized onto it, and the eight case headers missing template fields carry them (C3, C8); the seven case-table Mutates rows follow the headers again (C4); the commit-hash ban narrows to repository history, an input artifact's own revision being an input (C5); the Phase 1 cross-check reads the specs the diff touches instead of one shipped spec (C7); and the tool grant covers the scripts the Flow actually runs (C1). Signed-off-by: thxCode <thxcode0824@gmail.com> (cherry picked from commit c4027f71ed0d7b21f15c6596e3480646e603f99b)
…liation Signed-off-by: thxCode <thxcode0824@gmail.com>
thxCode
force-pushed
the
fix/dp-supervisor-followups
branch
from
October 25, 2026 03:27
422ad0b to
5dd0cf5
Compare
The suite is meant to carry every benchmark scenario, not only the ModelDeployment serving ones it opened with, so the title names the suite and the description scopes today's scenarios rather than locking the identity: a scenario outside serving joins this suite, not a new skill. Signed-off-by: thxCode <thxcode0824@gmail.com>
|
✅ OpenCodeReview: Review complete: 0 finding(s) across 4 selected item(s). |
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.
What type of PR is this?
/kind bug
/kind enhancement
/kind documentation
/kind cleanup
/area worker
/area testing
What this PR does / why we need it:
Four follow-ups from the review rounds of the DP group router supervisor:
model-deployment-router-imageadvances togpustack/llm-router:v0.4.0, the first build carrying the gateway's attention-dp-aware worker registration, so the--dp-awareflag the renderer emits for the SGLang External DP shape reaches a gateway that honors it. The router-objects digest pin moves with it exactly as its own header predicts: every default-image case together, the own-image case not at all.AwaitingObservationfor a whole observation budget behind a gateway that registers one rank per group. The serving answer still publishes the partial reason. Pinned by the recorded shape: a labeled two-rank group with one bound rank retains and restores; a group whose generation moved does not.--pd-disaggregationwith--dp-awareproduced an invalid decode destination in the recorded container run, so the operator never renders the two together.Which issue(s) this PR links to:
None - follow-ups to the review rounds of #783.