Repository navigation
[Docs epic] Close post-v1.2.0 documentation gaps for 26.07 #2118
Description
Activity
- added sub-issues
on Jun 29, 2026 FWIW, here is a full list of PRs merged after code freeze and not included in the 26.04 release. I excluded PRs that were just docs updates, minor dependency bumps and/or GitHub CI updates, benchmarking updates, or very minor PRs acting as follow-ups to major PRs:
- fix: give nemotron-cc stages distinct names #1763
- fix: preserve utf-8 characters in JSONL output #1762
- refactor: extract Ray Serve backend boundary - inference (1/5) #1813
- feat!: Typed Serve Config + Dyamo config stub - inference (2/5) #1820
- Add qwen3 #1827
- [Tutorial] Merge multiple file prefixes generated by
MegatronTokenizerWriter#1427 - feat: Placement Group based subprocess manager - inference (3/5) #1833
- feat: (experimental) Dynamo aggregated backend - inference (4/5) #1834
- Tutorials: Add interleaved getting-started tutorial #1774
- feat(tutorials/readspeech): add interactive Jupyter notebook tutorial #1870
- feat: (experimental) Dynamo Disaggregated serving + validators + router polish - inference (5/5) #1861
- Add generic audio processing pipeline for ASR and TTS data preparation #1679
- Bump python requirements to 3.11 min and 3.13 max #1855
- Add Nemotron Omni Model for Video caption pipeline #1883
- Bump Ray 2.55.1 / CUDA 12.9.1 / torch 2.10 / vllm 0.18; HAProxy ingress #1895
- Lazy import for text module #1888
- Lazy init for openai client #1890
- Add translation #1844
- Add
fit_data_fractionandcache_pathtoKMeansStage#1927 - chore: Update dynamo to 1.1.0 #1957
- fix(filters): escape
-in regex_url character class (#1601) #1880 - Add Skill to Nemo Curator for "Getting Started" #1746
- Nemotron-CLIMB data curation recipe #1727
- Add filtration and PnCWithLLM stages for audio tagging pipeline #1863
- Add long-form audio cutting pipeline for ALM pretraining #1898
- fix(utils): include port 65535 in get_free_port() scan range #2024
- Expose translation control knobs #2038
- Unify task identity: collapse _uuid into deterministic task_id #2036
- fix(audio): download FLEURS via hf_hub_download to avoid HTTP 429 #2057
- Refactor empty/sentinel tasks: EmptyTask class + SentinelTask base #2062
- fix(audio): lazy-load audio stages and fix tutorial notebook paths #1835
- Nemotron OCR SDG Pipeline #1899
- caption quality evaluation #1980
- Update Ray Data Backend to use
computestrategy + Allow min/max/initial workers to be specified #2077 - fix: extract arxiv archives outside download dir #2055
- feat(nemotron-parse): richer inference metrics, engine_kwargs passthrough, fanout stage spec #2054
- feat: support stage config overrides via with_ #2086
- fix(dynamo): multimodal chat processor, hybrid KV cache compat, subprocess env passthrough #2053
- Remove FlashAttn Deps #2107
- [video] Remove model_does_preprocess #2109
- fix: default workflow input extensions by filetype #2045
- Add note about
InferenceServerhang inndd_data_generation_exampletutorial #2114 - [Curator Video] Enable Ray Data stage fusion via cpus=1.0 in CPU video stages #2110
Listed in the order the PRs were merged.
Reacted by Lawrence LaneSupplemental workstream PR mapping is now complete:
- [Docs] Rewrite InferenceServer docs for typed Ray Serve and Dynamo backends #2146 -> draft PR [codex] Rewrite InferenceServer docs for Ray Serve and Dynamo #2147 (typed Ray Serve and Dynamo
InferenceServerguide) - [Docs] Update video caption model matrix and Qwen migration #2145 -> draft PR [codex] Document current video caption models and Qwen migration #2148 (video caption model matrix and Qwen migration)
- [Docs] Close late code-freeze docs and release-note gaps #2144 -> draft PR [codex] Document memory-efficient semantic dedup fitting #2149 (memory-efficient semantic deduplication) plus updated draft PR [codex] publish 26.07 release notes, migration checklist, and cut the 26.07 GA docs train #2143 (late-freeze release notes and final dependency stack)
All four branches pass
fern checkwith 0 errors. #2143 has also been corrected for #2063 now being onmain, so the release documentation no longer says resumability is excluded.- [Docs] Rewrite InferenceServer docs for typed Ray Serve and Dynamo backends #2146 -> draft PR [codex] Rewrite InferenceServer docs for Ray Serve and Dynamo #2147 (typed Ray Serve and Dynamo
Not sure if this is already being tracked somewhere but the last 2 PRs to make it before code freeze were:
- Pipeline resumability via source-level counter checkpointing #2063
- Add support for Slurm arrays #2059
so those should be ready for documentation as well.
Disposition for Sarah latest code-freeze addendum is implemented:
- Pipeline resumability via source-level counter checkpointing #2063 -> existing issue [Docs] Replace legacy resumability guidance and document deterministic task identity #2124 and repaired draft PR [codex] docs: document native pipeline resumability #2134. The PR was rebuilt from current
mainwith documentation commits only, retargeted from the obsolete feature branch, and updated to link array-specific operations. - Add support for Slurm arrays #2059 -> new assigned issue [Docs] Document SLURM array sharding, completion tracking, and retries #2153 and draft PR [codex] Document SLURM job arrays and retries #2154 for SLURM array sharding, completion tracking, scheduler limits, and retries.
- Release-notes PR [codex] publish 26.07 release notes, migration checklist, and cut the 26.07 GA docs train #2143 now includes both Add support for Slurm arrays #2059 and Pipeline resumability via source-level counter checkpointing #2063 and records [codex] docs: document native pipeline resumability #2134/[codex] Document SLURM job arrays and retries #2154 as detailed-page dependencies.
Validation: #2134 and #2154 both pass
fern checkwith 0 errors. #2154 also validates all 10 shell examples withbash -nand documents every backend SLURM-array environment variable.- Pipeline resumability via source-level counter checkpointing #2063 -> existing issue [Docs] Replace legacy resumability guidance and document deterministic task identity #2124 and repaired draft PR [codex] docs: document native pipeline resumability #2134. The PR was rebuilt from current
- added sub-issues
on Jul 2, 2026 Adversarial review follow-ups are now tracked as sub-issues of this epic:
Before the 26.06 docs publication
- [Docs] Integrate 26.07 documentation PRs and stage the version train #2160 — integrate the PR set, reconcile four cross-PR conflict groups, merge [codex] publish 26.07 release notes, migration checklist, and cut the 26.07 GA docs train #2143 last, and stage the v26.06 version train
- [Docs] Align global pipeline execution concepts with current backends #2161 — correct stale global pipeline/resource concepts, removed imports, backend-specific sizing claims, task-lifecycle guidance, and the migration FAQ
- [Docs] Repair audio API examples and add workflow selection guidance #2162 — repair 27 stale audio import occurrences and add audio workflow-selection guidance
Non-blocking post-release IA
- [Docs IA] Modularize large workflow pages and multimodal synthetic navigation #2159 — evaluate modular splits for Inference Server, Translation, captions/preview, and multimodal synthetic-data navigation
PR #2143 was updated directly in commit
86e286845with the two missing release-critical migrations: the removedqwenvideo caption variant and the removedInferenceModelConfigAPI. It now also carries the corresponding Qwen/Nemotron and typed Ray Serve/Dynamo release-note entries.Recommended timing: merge detailed feature PRs first; complete #2161 and #2162 before the documentation publication tag; use #2160 as the final integration gate; address #2159 after release unless a smaller split becomes necessary while resolving another issue. Merging to
maindoes not publish the site, so there is room to reconcile the combined tree before an explicitly authorizeddocs/v*tag.the part that bites in an audit this size is the per-PR triage. scoring each of the 67 in isolation into dedicated-docs vs release-note-bullet vs no-docs surfaces whatever cleared the bar on its own day, but a behavior that drifted a little across a dozen small PRs, none individually docs-worthy, gets bundled into one bullet and its cumulative migration hazard never gets measured. the slow-accumulation change is usually the exact one that reads as breaking to a user, and it's the one the per-item score quietly drops.
Added supplemental workstream for arm64 audio platform constraints:
- [Docs] Document arm64/aarch64 platform constraints for audio curation #2168 — Document that audio curation is unsupported on arm64/aarch64 (intentional dependency gating in
pyproject.toml, not a container regression)
Trigger: nvbugs 6400678 / 6409435 / 6409439 thread (2026-07-07). Audio missing packages on arm64 are expected; Dong Hyuk confirmed docs should reflect this. Sarah requested tutorial/docs updates and nvbug responses (assigned to Aaftab V in Slack).
Disposition: Separate child issue rather than bundling into #2162 (API/example repair) or #2144 (late-freeze release notes). Coordinate with #2162 to avoid conflicting audio-page edits. The datasets/dedup portion of the thread stays with #2164.
Suggested timing: Before 26.06 docs publication, alongside or immediately after #2162.
- [Docs] Document arm64/aarch64 platform constraints for audio curation #2168 — Document that audio curation is unsupported on arm64/aarch64 (intentional dependency gating in
- changed the title
[-][Docs epic] Close post-v1.2.0 documentation gaps for 26.06[/-][+][Docs epic] Close post-v1.2.0 documentation gaps for 26.07[/+]on Jul 22, 2026 Tracker naming update: The docs GA train for this epic is 26.07, not 26.06. 26.07 is not hard-coded in Fern until the GA snapshot (
main/→v26.07/) lands in #2160.Updated in this pass:
- Renamed label
docs-release-26.06→docs-release-26.07 - Retitled the epic ([Docs epic] Close post-v1.2.0 documentation gaps for 26.07 #2118), release-notes ([Docs] Publish 26.07 release notes and migration checklist #2120), integration/staging ([Docs] Integrate 26.07 documentation PRs and stage the version train #2160), and late-freeze ([Docs] Close late code-freeze docs and release-note gaps #2144) issues to use 26.07 train verbiage
Still to do at GA cut (not in tracker yet): create
fern/versions/v26.07/, retargetlatest.yml, and updatefern/docs.ymlredirects/version picker. The unmerged release-notes branch still uses 26.06 filenames/text and should be renamed to 26.07 before merge.- Renamed label
Goal
Close every actionable documentation gap introduced after NeMo Curator v1.2.0, and make the next release's docs/release notes complete enough for users to adopt new features and migrate breaking behavior.
Audit boundary and method
mainbut absent from the v1.2.0 tag, reviewed through PR #2116 (merged 2026-06-29 19:34:32 UTC)git v1.2.0..main, and the currentfern/versions/main/pages/documentation treePlanned documentation workstreams
Child issues are linked above for traceability.
Complete PR disposition
Dedicated docs workstreams required (18)
These PRs add or change user-facing workflows/APIs and need more than a release-note bullet:
#1727, #1835, #1863, #1898, #1899, #1980, #2033, #2036, #2038, #2045, #2054, #2058, #2062, #2070, #2072, #2077, #2086, #2110
Release-notes or migration coverage required (8)
These are smaller fixes, platform constraints, or already-partially-documented removals. They belong in the release/migration task rather than separate feature pages:
#2024, #2049, #2053, #2055, #2057, #2104, #2107, #2109
Documentation already landed; verify in release sweep (9)
No standalone follow-up is currently justified, but the release-notes owner should verify navigation, versioning, and consistency:
#1983, #1961, #1991, #2000, #2029, #2030, #2071, #2096, #2114
No public docs change required (32)
Dependency/security lockfile updates, CI/release plumbing, test-only changes, and internal benchmark-harness maintenance:
#1966, #2001, #2002, #2009, #2021, #2022, #2026, #2032, #2035, #2039, #2041, #2043, #2046, #2047, #2048, #2068, #2079, #2081, #2082, #2083, #2084, #2085, #2087, #2088, #2091, #2099, #2100, #2103, #2105, #2108, #2115, #2116
Supplemental code-freeze disposition
Sara's follow-up list added 22 release-relevant PRs that were outside the original publication-time query.
New dedicated or bundled workstreams
Added to existing workstreams
Bundled release-note coverage (#2144 and #2120)
#1746, #1762, #1763, #1774, #1870, #1880, #1888, #1890
Epic-level acceptance criteria
fern/versions/main/pages/about/release-notes/index.mdxdescribes 26.07 rather than 26.04task_idOut of scope