Skip to content

Latest commit

 

History

History
34 lines (29 loc) · 2.48 KB

paper_pipeline.md

File metadata and controls

34 lines (29 loc) · 2.48 KB

Paper pipeline and implementation

The main package follows the offline fitting / online serving boundary. Recipes select objectives and data; architecture adapters describe model structure; the backend owns engine-specific execution.

Pipeline stage Implementation Entry point
Native paired traces offline/traces, offline/trajectory, benchmark collector cachebridge serve ... -- python benchmarks/tau_bench/run.py sample ...
Source support offline/support.py load_support; complete trajectory recipe uses supplied initializer support
Centered K/V initialization offline/initialization.py, offline/ridge.py fit_layer for paired synthetic or prepared content-space tensors
Trajectory K initialization offline/keys.py, offline/weighted_keys.py cachebridge fit prepare, then cachebridge fit run
Sequential conditional V calibration offline/calibration/sequential.py Value stage of cachebridge fit run
Fold and export offline/export.py, artifacts.py cachebridge fit run, then cachebridge verify
Source block mapping mapper.py, online/mapping.py, backend source_mapper_adapter.py Prepared SGLang source pipeline
Buffers, delivery, installation online/pipeline.py, backend peer/lifetime modules Backend handoff execution
All-rank publication online/handoff.py, backend handoff.py and ready_publication.py run_handoff / backend transfer
Agent sampling and routing agents/sampling.py, trajectory.py sample_trajectory
Workload evaluation benchmarks/tau_bench Benchmark evaluate, compare.py

cachebridge recipes describes the registered complete recipe. The generic initializer and trajectory-weighted K solver retain distinct objectives and precision. The exported sequential calibration currently solves regularized readout matching. Automatic R² support ranking, energy-constrained calibration, and a dynamic shared multi-request buffer-pool coordinator are not exposed as complete public recipes here. An API or directory name does not imply those method stages are implemented.

The online lifecycle reference modules define ownership contracts. Actual CUDA streams, peer operations and scheduler publication are in the SGLang backend; changing a reference state machine alone does not change GPU execution.

Weights, traces and generated outputs are external inputs/artifacts. Recipes contain configuration and commands, not fitted weights or evaluation outputs.