Summary
Introduce a versioned typed Agentic Mooncake format and a validated agentic graph that drives Dynamo replay through one dependency-aware workload scheduler. Add local Weka file and directory ingestion, deterministic closed-loop agentic lanes, exact trajectory E2E reporting, aggregated offline and online mock replay, and offline prefill/decode replay.
Motivation
Current agentic replay serializes dependencies as completion-only waits. That cannot faithfully represent overlapping subagent dispatch, post-completion spawn timing, later joins, or complete trajectory accounting. Source-specific lowering and lifecycle state are split across parsers, the workload driver, and report collection, which makes semantic validation and parity difficult. Public Weka traces also need deterministic local hash namespacing and a supported direct ingestion path without an AIPerf runtime dependency.
Proposal
- Replace the headerless Agentic Mooncake format with an explicitly versioned typed schema. This is an intentional compatibility break; the new reader will not implement a legacy execution path.
- Represent every play as validated nodes and typed sequence, spawn, and join edges triggered by logical request dispatch or successful completion.
- Compile the graph once into dense indices and make the workload driver the sole owner of node state, failure propagation, closed-loop lane recycling, and actual dispatch timestamps.
- Add public streaming local Weka ingestion in
dynamo-data-gen, with deterministic directory ordering, source-local ID and hash namespacing, and an optional canonical Agentic Mooncake materializer.
- Lower Claude, Codex, and Dynamo request traces through the same typed edge boundary.
- Support aggregated offline replay, aggregated online mock replay, and offline disaggregated prefill/decode replay. Graph completion in P/D occurs only after successful logical decode completion, never at prefill or handoff completion.
- Report complete and incomplete trajectory counts plus trajectory E2E distributions without requiring detailed per-request capture.
- Validate the change with source-semantic goldens, deterministic direct-versus-materialized equivalence, intentional-nonparity evidence for corrected branch timing, and paired replay performance campaigns for vLLM and SGLang aggregated and P/D modes.
Detailed implementation and campaign progress will be linked from the implementation PR.
Summary
Introduce a versioned typed Agentic Mooncake format and a validated agentic graph that drives Dynamo replay through one dependency-aware workload scheduler. Add local Weka file and directory ingestion, deterministic closed-loop agentic lanes, exact trajectory E2E reporting, aggregated offline and online mock replay, and offline prefill/decode replay.
Motivation
Current agentic replay serializes dependencies as completion-only waits. That cannot faithfully represent overlapping subagent dispatch, post-completion spawn timing, later joins, or complete trajectory accounting. Source-specific lowering and lifecycle state are split across parsers, the workload driver, and report collection, which makes semantic validation and parity difficult. Public Weka traces also need deterministic local hash namespacing and a supported direct ingestion path without an AIPerf runtime dependency.
Proposal
dynamo-data-gen, with deterministic directory ordering, source-local ID and hash namespacing, and an optional canonical Agentic Mooncake materializer.Detailed implementation and campaign progress will be linked from the implementation PR.