Repository navigation
Conversation
This was referenced Oct 8, 2026
Contributor
|
Thanks for the contribution!
中文感谢你的贡献！
|
This was referenced Oct 8, 2026
Draft
adibarra
added this pull request to stack #3808
October 8, 2026 14:35
adibarra
force-pushed
the
refactor/recipe-bundles
branch
from
October 8, 2026 16:03
b8a0abc to
9f1346d
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
from
October 8, 2026 16:25
9f1346d to
fcd8458
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
from
October 8, 2026 17:21
fcd8458 to
3bb06f4
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
from
October 8, 2026 21:28
3bb06f4 to
c68a0ad
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
2 times, most recently
from
October 9, 2026 11:39
030da06 to
f0d9622
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
2 times, most recently
from
October 9, 2026 18:20
ff09ac0 to
34c24b0
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
from
October 9, 2026 18:48
34c24b0 to
9c61ead
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
from
October 9, 2026 23:43
9c61ead to
6a5e0b4
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
2 times, most recently
from
October 10, 2026 13:38
7bb4259 to
40643b0
Compare
adibarra
force-pushed
the
refactor/recipe-bundles
branch
from
October 12, 2026 08:02
40643b0 to
c261feb
Compare
prepare_recipe edited the staged fragment before the binder ran. Its text edits only find a flat recipe's top-level name and six-space watchdog-timeout, so bundle variants kept their recipe names as srtctl's run name and never got the cluster's dist-timeout. The edits now apply to the recipe the binder writes, which is flat for every variant. Slurm job names were unaffected: srtctl names the job from RUNNER_NAME.
… entry Master-config entries whose rows each named their own standalone recipe now select overrides of one <agg|disagg>[-axis]-variants.yaml: base holds what the points share and each override_<topology>_c<conc> holds only its point's deltas, including the job name. Ten AgentX entries (50 recipes) and seventeen fixed-sequence entries (184 recipes) move into 27 bundles. Every moved row's selected variant equals its old recipe after composition, with and without the power telemetry block, and the full-sweep matrices differ only in those rows' srt-recipe strings. Entries whose points mix recipe designs, such as a TP low-latency recipe beside wide-EP ones, keep their standalone files. Section-header comments, commented-out settings and per-point labels that restate the topology are not carried into the bundles. RECIPES.md describes the convention.
… into bundles #3659 split the GB300 GLM-5.2 SGLang AgentX recipes into one file per point. Like the other AgentX families, the three aggregate points select overrides of agg-variants.yaml and the three low-latency 1PxD points overrides of disagg-variants.yaml; each override selects exactly its old recipe. The DEP8/DEP16 2P1D point is a different design and keeps its own file.
The health-check floor and model path of a bound bundle variant are already covered by test_srt_config and the multi-node lane test, and the bundle-shaped health budget test exercises inputs that only reach raise_health_attempts as the binder's flat output.
Its other caller went away when the lane edits moved to the bound recipe.
Every launch renames the bound recipe and fingerprints exclude name, so bundles need not set one per override. Drop a misindented commented-out flag and move the NATS comment above request_plane.
…s from the pareto bundles of the pareto entries by naming those points' standalone recipes, which the bundles replace. Their rows select the matching pareto overrides, so each still runs exactly the pareto point's recipe rather than a copy that could drift; the DEP16 points keep disagg-dep16-mtp6-variants.yaml. RECIPES.md notes that an entry rerunning another entry's points selects its overrides.
adibarra
force-pushed
the
refactor/recipe-bundles
branch
from
October 12, 2026 10:22
c261feb to
efe2f87
Compare
This branch has not been deployed
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
1 participant
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.
Summary
6f82f4e62): the launcher's job edits (job name, health floor, dist-timeout) now run on the bound concrete recipe. Before, they edited the fragment as text, so bundle variants silently kept the recipe's ownname, notinferencex-<runner>, and cleanup could not find them. Existing bundle rows in 39 config keys now submit asinferencex-<runner>. Health floor and dist-timeout are unchanged for them.76e678bed): a master config key with 2+ rows now uses one*-variants.yamlbundle.baseholds what the points share, and each row gets oneoverride_<topology>_c<conc>with only its deltas and its old jobname. The convention is written inRECIPES.md.9 fixed-sequence keys stay standalone because they mix recipe designs, e.g. a TP low-latency recipe next to wide-EP ones: 15-30% of keys differ per point, versus 9% or less in every converted key.
Verification
full-sweep(1767 rows) and--all-evals(1797) differ from the base only insrt-recipestrings. Concrete recipes are equal for all 513 moved matrix rows. Their fingerprints change becausesrt-recipeis part of the row.infx generate: same output for the 17 fixed-sequence keys.name, from the fix.The pre-existing
dsr1-fp8-gb200-dynamo-sglangrow that selects a whole 3-variant zip bundle is left as is.