Skip to content

docs(troubleshoot): replace stale embedding batch-size and Triton log guidance (NVBug 6628722) - #2578

Draft
kheiss-uwzoo wants to merge 1 commit into
NVIDIA:mainfrom
kheiss-uwzoo:docs/6628722-troubleshoot-embed-batch
Draft

docs(troubleshoot): replace stale embedding batch-size and Triton log guidance (NVBug 6628722)#2578
kheiss-uwzoo wants to merge 1 commit into
NVIDIA:mainfrom
kheiss-uwzoo:docs/6628722-troubleshoot-embed-batch

Conversation

@kheiss-uwzoo

Copy link
Copy Markdown
Collaborator

Summary

  • NVBug 6628722: replace the no-op EMBEDDER_BATCH_SIZE embedding-startup workaround with image-specific NIM env vars for Helm 2.3.0 (NIM_PIPELINE_MAX_BATCH_SIZE) and development Compose 1.12.0 (NIM_TRITON_MAX_BATCH_SIZE).
  • Clarify that CLI --embed-batch-size and .embed(inference_batch_size=...) do not start the NIM.
  • Reframe the Triton INFO-as-ERROR section as NIM-container log noise (
    imutils.py is inside the NIM image, not NRL source) and keep the ignore-as-info action.

Test plan

  • Confirmed EMBEDDER_BATCH_SIZE is not read in NRL source, CLI, SDK, Helm, or Compose.
  • Confirmed Helm default VL embed image is llama-nemotron-embed-vl-1b-v2:2.3.0 and Compose default tag is 1.12.0.
  • Confirmed
    imutils.py is not in the NRL repository.
  • QA: reproduce the embedding-NIM startup error on RTX 6000 (or equivalent) and recover with the documented NIM env var for that image.
  • QA: confirm Triton INFO lines wrapped as ERROR can still be ignored.

pre-draft: leakage, mkdocs --strict, ::a, ::p, ::r on the diff vs main

Base: upstream/main
Files: docs/docs/extraction/troubleshoot.md

Check Result
Leakage (page roles + see [ CTAs) PASS — troubleshoot.md has no see [ CTAs; conceptual-page greps are not applicable to this file
Allowed paths PASS — 1 documentation file
mkdocs --strict FAIL — exit 1; 4 warnings on untracked leftover pages custom-metadata.md and user-defined-stages.md that are not in this diff; troubleshoot.md had no warnings
::a audit PASS — EMBEDDER_BATCH_SIZE unused (repo grep); Helm 2.3.0 uses NIM_PIPELINE_MAX_BATCH_SIZE; Compose 1.12.0 uses NIM_TRITON_MAX_BATCH_SIZE;
imutils.py not in NRL (95%)
::p polish Applied — deployment-mode recovery, complete-sentence code examples, shortened Triton log dump
::r style 95% — no blocking issues in the changed sections

Code drift (not in this docs PR):
emo_retriever/helm/values.yaml still sets NIM_TRITON_LOG_VERBOSE=1 on
imOperator.vlm_embed. Embedding NIM 2.3.0 removes that variable with no replacement.

PR: this draft

@kheiss-uwzoo kheiss-uwzoo self-assigned this Aug 24, 2026
@kheiss-uwzoo kheiss-uwzoo added 26.08 doc Improvements or additions to documentation labels Aug 24, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

26.08 doc Improvements or additions to documentation

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant