Skip to content

fix(vln): preserve service identity in resumed summaries - #17

Merged
ZhouAo-ZA merged 1 commit into
mainfrom
fix/streamvln-resume-metrics
Nov 2, 2026
Merged

ZhouAo-ZA merged 1 commit into
mainfrom
fix/streamvln-resume-metrics

Conversation

@pppppppppper

Copy link
Copy Markdown
Collaborator

Resumed StreamVLN evaluations can span inference-service restarts, each of which may restart its tensor-batch index at zero. The summary discarded each episode's service-instance identity while flattening traces, so distinct invocations collided and otherwise valid reports failed the occupancy check.

Carry model_service_index into each flattened timing record, retaining the legacy default of zero. Add a regression that summarizes two restarted services using the same batch index and checks that both invocations and their elapsed times remain separate.

Validation on Linux with Python 3.12:

  • Before the fix: the new regression failed; the existing 58 targeted tests passed.
  • After the fix: both navigation benchmark test files passed, 59 tests total.
  • Changed-file Ruff lint/format, whitespace and commit policy passed.

No model inference, simulator workload or GPU benchmark was run for this fix.

Signed-off-by: pppppppppper <185877767+pppppppppper@users.noreply.github.com>
@ZhouAo-ZA
ZhouAo-ZA merged commit 0039f84 into main Nov 2, 2026
5 checks passed
@pppppppppper
pppppppppper deleted the fix/streamvln-resume-metrics branch November 2, 2026 15:38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants