Best end time
-
SAPP Unified Standalone
Unified Parallel Strategy Search
A research system for unified ND and PPB strategy search in large-scale LLM training systems.
This project investigates parallel strategy design for large-model training on large-scale clusters.
It supports end-to-end experimental studies, from ND exploration to PPB analysis, while keeping the
leading configurations easy to compare and inspect.
Research system
Large-scale training studies
Start a study to populate the workspace
Begin with a YAML configuration, choose the framework and device, and launch a unified ND + PPB study. The resulting workspace will summarize the leading strategies and expose detailed PPB artifacts for further analysis.
Feasible candidates
-
Best PPB solver peak
-
PPB cache hits
-
Search time
-
ND strategies
-
PPB-refined candidates
-
Overview
Search overview
Leading strategies after ND generation, performance estimation, and PPB evaluation.
The comparison figure will appear after a completed study.
Summary
Leading configuration
The summary figure will appear after a completed study.
Evaluation note
ND vs PPB counts
ND counts the full feasible search space. PPB counts the leading strategies advanced to the pipeline balance stage.