Repository navigation
[Feat] simulator: Add empirical and explicit acceptance-progress policies - #49243
Draft
Jeff010101 wants to merge 14 commits into
Draft
Jeff010101 wants to merge 14 commits into
Jeff010101 wants to merge 14 commits into
Conversation
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Jeff010101
force-pushed
the
codex/review-layer-p6
branch
from
November 9, 2026 05:37
7af64d4 to
1ffa515
Compare
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Signed-off-by: Jeff010101 <Jifang.Xing@amd.com>
Jeff010101
force-pushed
the
codex/review-layer-p6
branch
from
November 9, 2026 06:11
1ffa515 to
f2a1a3b
Compare
Jeff010101
added a commit
to Jeff010101/sglang
that referenced
this pull request
Nov 13, 2026
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
P6: Empirical and given-acceptance progress policies
Source head:
codex/review-layer-p6atf2a1a3b8059e13a02a0b237183062df061be5db9.Focused feature base:
b7b7db61ba6eee55fae54e57012c890a04281a87; 1 feature commit, 8 files. Review this feature's fixed diff.Target:
sgl-project/sglang:main. GitHub's full diff also includes unmerged prerequisites; use the focused comparison above to review this layer. Dependencies must land and the branch must be rebased before merge.Dependencies: Requires P5. P8 is a parallel sibling; storage/numeric optimizations are not prerequisites for these progress policies.
Motivation
The native core should remain runnable with SourceGolden while optional progress policies have explicit inputs and independently reviewable support limits.
Modifications
Delivered capability after prerequisites
After the native core, this PR adds two complete policy implementations using the existing progress interface:
empirical: load a checksummed source-bound artifact, fit conditional commit-count distributions from training rows only and expose heldout errors. Training and heldout trajectories cannot overlap.given_mean_acceptance: consume explicitly authored acceptance information as an exact counter ratio or a declared published-precision value, then draw the modeled committed-token count with a pinned seed.Both preserve actual source/runtime identity, draw ordering, retry identity and native output-tail clipping. No alternative cost model, registry or placeholder progress policy is introduced. Existing SourceGolden behavior remains available and has no new default acceptance input.
Dependencies and enablement
Requires the native ADP1/SourceGolden PR. This branch and evidence storage are siblings; neither must merge before the other. Numeric CPU execution is also unnecessary.
Set
progress_kindexplicitly. Empirical mode requiresprogress_artifactwith its original path/SHA and natural TP2 source support; it uses the artifact's declared seed. Given mode requiresgiven_acceptancewith the documented evidence fields and a nonnegativemodel_seed. Relative asset paths resolve against the configuration file. Provide matching external native source/capture/runtime material and the separately installed pinned SDK; selecting a policy does not manufacture compatible evidence.Given AL is additional workload information, not independently predicted acceptance. A mean does not identify a true distribution or cross-request correlations. Heldout diagnostics remain visible and do not create an automatic accuracy threshold. Unsupported topology, identity or artifact fields fail explicitly.
Accuracy Tests
The P1 prerequisite has been rebased onto Quark #49231, advancing the inherited core to
2ad35d45. Its runtime source files are unchanged. Execution evidence below predates this core rebase; matching new-core integration has not been rerun.Minimal acceptance runs
test_dspark_progress.py,test_dspark_given_mean_progress.pyand the native factory controls. The 45 progress tests passed in the earlier Linux matrix and again as affected focused controls after Struct replacement. SDK-dependent fixtures require the installed compatible SDK; a dependency skip is not source-integration evidence.Complete stack counts, current diagnostic status and shared unresolved checks are centralized in the native PR. These policy controls establish mechanics and rejection behavior, not a new workload accuracy or full-MAPE result.
Speed Tests and Profiling
No inference/GPU speedup is claimed; GPU speed benchmarks were not run.
Checklist
Review and Merge Process
Keep this upstream PR Draft until dependency rebases, applicable CI and required maintainer reviews are complete. Existing execution results retain their recorded source/runtime identities; this PR publication adds no new execution or accuracy result.
CI States
Latest PR Test (Base): ⊘ Skipped -- the PR is a draft; mark it ready, then
/rerun-failed-ci.Latest PR Test (Extra): ⊘ Skipped -- the PR is a draft; mark it ready, then
/rerun-failed-ci.Latest PR Test (AMD ROCm 10): ⊘ Skipped -- the PR is a draft; mark it ready, then
/rerun-failed-ci.Tip
Base CI only covers the core cases, and its workload may not exercise your change. Use
/rerun-testto run the tests your change touches, or copy this prompt for an agent: