feat(graph)!: native Dynamo trace replay -- Graph IR agentic workload lane with unified segment store and session routing

.cursor/rules/python.mdc

@@ -134,7 +134,7 @@ Feature branches use `<username>/feature-name` format, forked from `main`. One P
 - Decorators: `@on_init`, `@on_start`, `@on_stop`, `@on_message`, `@on_command`, `@background_task`, `@on_pull_message`, `@on_request`.
 - Communication: `publish()` for broadcast, `@on_message` to subscribe, `send_command_and_wait_for_response()` for sync.
 - `AIPerfLifecycleMixin` for standalone components: `CREATED` -> `INITIALIZING` -> `INITIALIZED` -> `STARTING` -> `RUNNING` -> `STOPPING` -> `STOPPED`; `FAILED` terminal.
-- `dag_jsonl` input type: conversation DAG benchmarks (fork + spawn modes). See `docs/benchmark-modes/dag.md` for abstractions and authoring.
+- `dag_jsonl` input type: legacy conversation DAG benchmarks (fork + spawn modes). See `docs/benchmark-modes/dag.md`. For new agentic workloads use Graph IR — see `docs/benchmark-modes/agentic.md`.
 - Validator gate convention: unsupported constructs raise `NotImplementedError` with a leading `"<loc>: <reason>"` prefix where `<loc>` identifies the conversation/turn (e.g. `"conversation 'foo' turn 3: ..."`). New validators must follow this shape.
 - Per-turn payload contract: `extra_body` / `max_tokens` / `model` are dispatch-turn only; `raw_tools` is the lone field that walks history (system-prompt-like). Dataset rows author `extra`, not `extra_body`. See `docs/dev/patterns.md` "Per-turn dataset `extra`".
 

.github/copilot-instructions.md

@@ -129,7 +129,7 @@ Feature branches use `<username>/feature-name` format, forked from `main`. One P
 - Decorators: `@on_init`, `@on_start`, `@on_stop`, `@on_message`, `@on_command`, `@background_task`, `@on_pull_message`, `@on_request`.
 - Communication: `publish()` for broadcast, `@on_message` to subscribe, `send_command_and_wait_for_response()` for sync.
 - `AIPerfLifecycleMixin` for standalone components: `CREATED` -> `INITIALIZING` -> `INITIALIZED` -> `STARTING` -> `RUNNING` -> `STOPPING` -> `STOPPED`; `FAILED` terminal.
-- `dag_jsonl` input type: conversation DAG benchmarks (fork + spawn modes). See `docs/benchmark-modes/dag.md` for abstractions and authoring.
+- `dag_jsonl` input type: legacy conversation DAG benchmarks (fork + spawn modes). See `docs/benchmark-modes/dag.md`. For new agentic workloads use Graph IR — see `docs/benchmark-modes/agentic.md`.
 - Validator gate convention: unsupported constructs raise `NotImplementedError` with a leading `"<loc>: <reason>"` prefix where `<loc>` identifies the conversation/turn (e.g. `"conversation 'foo' turn 3: ..."`). New validators must follow this shape.
 - Per-turn payload contract: `extra_body` / `max_tokens` / `model` are dispatch-turn only; `raw_tools` is the lone field that walks history (system-prompt-like). Dataset rows author `extra`, not `extra_body`. See `docs/dev/patterns.md` "Per-turn dataset `extra`".
 

.github/workflows/run-unit-tests.yml

@@ -115,7 +115,7 @@ jobs:
     - name: Run unit tests (Windows)
       if: runner.os == 'Windows'
       run: |
-        uv run pytest tests/unit -n auto --tb=short -m 'not performance and not stress and not slow'
+        uv run pytest tests/unit -n auto --timeout 600 --timeout-method thread --tb=short -m 'not performance and not stress and not slow'
     - name: Run zmq real-transport tests (Windows)
       # Real-socket libzmq tests (no looptime). ``!cancelled()`` lets them run
       # even if the unit step failed, matching the POSIX ``make test-ci``
@@ -131,7 +131,7 @@ jobs:
       # finishing after a manual stop.
       if: ${{ !cancelled() && runner.os == 'Windows' }}
       run: |
-        uv run pytest tests/component_integration -n auto --tb=short -v -m 'not performance and not stress and not slow'
+        uv run pytest tests/component_integration -n auto --timeout 600 --timeout-method thread --tb=short -v -m 'not performance and not stress and not slow'
     - name: Upload results to Codecov
       if: matrix.os == 'linux' && matrix.python-version == '3.12'
       uses: codecov/codecov-action@v5

AGENTS.md

@@ -128,7 +128,7 @@ Feature branches use `<username>/feature-name` format, forked from `main`. One P
 - Decorators: `@on_init`, `@on_start`, `@on_stop`, `@on_message`, `@on_command`, `@background_task`, `@on_pull_message`, `@on_request`.
 - Communication: `publish()` for broadcast, `@on_message` to subscribe, `send_command_and_wait_for_response()` for sync.
 - `AIPerfLifecycleMixin` for standalone components: `CREATED` -> `INITIALIZING` -> `INITIALIZED` -> `STARTING` -> `RUNNING` -> `STOPPING` -> `STOPPED`; `FAILED` terminal.
-- `dag_jsonl` input type: conversation DAG benchmarks (fork + spawn modes). See `docs/benchmark-modes/dag.md` for abstractions and authoring.
+- `dag_jsonl` input type: legacy conversation DAG benchmarks (fork + spawn modes). See `docs/benchmark-modes/dag.md`. For new agentic workloads use Graph IR — see `docs/benchmark-modes/agentic.md`.
 - Validator gate convention: unsupported constructs raise `NotImplementedError` with a leading `"<loc>: <reason>"` prefix where `<loc>` identifies the conversation/turn (e.g. `"conversation 'foo' turn 3: ..."`). New validators must follow this shape.
 - Per-turn payload contract: `extra_body` / `max_tokens` / `model` are dispatch-turn only; `raw_tools` is the lone field that walks history (system-prompt-like). Dataset rows author `extra`, not `extra_body`. See `docs/dev/patterns.md` "Per-turn dataset `extra`".
 

CLAUDE.md

@@ -128,7 +128,7 @@ Feature branches use `<username>/feature-name` format, forked from `main`. One P
 - Decorators: `@on_init`, `@on_start`, `@on_stop`, `@on_message`, `@on_command`, `@background_task`, `@on_pull_message`, `@on_request`.
 - Communication: `publish()` for broadcast, `@on_message` to subscribe, `send_command_and_wait_for_response()` for sync.
 - `AIPerfLifecycleMixin` for standalone components: `CREATED` -> `INITIALIZING` -> `INITIALIZED` -> `STARTING` -> `RUNNING` -> `STOPPING` -> `STOPPED`; `FAILED` terminal.
-- `dag_jsonl` input type: conversation DAG benchmarks (fork + spawn modes). See `docs/benchmark-modes/dag.md` for abstractions and authoring.
+- `dag_jsonl` input type: legacy conversation DAG benchmarks (fork + spawn modes). See `docs/benchmark-modes/dag.md`. For new agentic workloads use Graph IR — see `docs/benchmark-modes/agentic.md`.
 - Validator gate convention: unsupported constructs raise `NotImplementedError` with a leading `"<loc>: <reason>"` prefix where `<loc>` identifies the conversation/turn (e.g. `"conversation 'foo' turn 3: ..."`). New validators must follow this shape.
 - Per-turn payload contract: `extra_body` / `max_tokens` / `model` are dispatch-turn only; `raw_tools` is the lone field that walks history (system-prompt-like). Dataset rows author `extra`, not `extra_body`. See `docs/dev/patterns.md` "Per-turn dataset `extra`".
 

Makefile

@@ -247,13 +247,13 @@ test-ci: #? run the tests using pytest-xdist for CI.
 	@printf "$(bold)$(blue)Running unit and component integration tests (CI mode)...$(reset)\n"
 	@# Run unit tests first with coverage
 	@printf "$(bold)$(blue)Running unit tests...$(reset)\n"
-	@$(activate_venv) && pytest tests/unit -n auto --cov=src/aiperf --cov-branch --cov-report= -m 'not performance and not stress and not slow' --tb=short $(args) || exit_code=$$?; \
+	@$(activate_venv) && pytest tests/unit -n auto --timeout 600 --timeout-method thread --cov=src/aiperf --cov-branch --cov-report= -m 'not performance and not stress and not slow' --tb=short $(args) || exit_code=$$?; \
 	# Run real-socket zmq transport tests (real time + real sockets, no looptime) regardless of unit result \
 	printf "$(bold)$(blue)Running zmq real-transport tests...$(reset)\n"; \
 	$(activate_venv) && pytest tests/zmq --cov=src/aiperf --cov-branch --cov-append --cov-report= -m 'not performance and not stress and not slow' --no-looptime --tb=short $(args) || exit_code=$$((exit_code + $$?)); \
 	# Run component integration tests with coverage append regardless of unit test result \
 	printf "$(bold)$(blue)Running component integration tests...$(reset)\n"; \
-	$(activate_venv) && MALLOC_ARENA_MAX=2 pytest tests/component_integration -n auto --cov=src/aiperf --cov-branch --cov-append --cov-report=html --cov-report=xml --cov-report=term -m 'not performance and not stress and not slow' -v --tb=short $(args) || exit_code=$$((exit_code + $$?)); \
+	$(activate_venv) && MALLOC_ARENA_MAX=2 pytest tests/component_integration -n auto --timeout 600 --timeout-method thread --cov=src/aiperf --cov-branch --cov-append --cov-report=html --cov-report=xml --cov-report=term -m 'not performance and not stress and not slow' -v --tb=short $(args) || exit_code=$$((exit_code + $$?)); \
 	if [[ $$exit_code -eq 0 ]]; then \
 		printf "$(bold)$(green)AIPerf unit and component integration tests (CI mode) passed!$(reset)\n"; \
 	else \

README.md

@@ -119,7 +119,7 @@ Log File: /home/user/Code/aiperf/artifacts/granite4:350m-openai-chat-concurrency
 
 - Scalable multiprocess architecture with 10 services communicating via ZMQ
 - 3 UI modes: `dashboard` (real-time TUI), `simple` (progress bars), `none` (headless)
-- Multiple benchmarking modes: concurrency, request-rate, [request-rate with max concurrency](docs/tutorials/request-rate-concurrency.md), [trace replay](docs/benchmark-modes/trace-replay.md)
+- Multiple benchmarking modes: concurrency, request-rate, [request-rate with max concurrency](docs/tutorials/request-rate-concurrency.md), [trace replay](docs/benchmark-modes/trace-replay.md), [agentic workloads](docs/benchmark-modes/agentic.md)
 - Extensible plugin system for endpoints, datasets, transports, and metrics
 - [Public dataset support](docs/benchmark-datasets.md) including ShareGPT and custom formats
 
@@ -153,6 +153,7 @@ Log File: /home/user/Code/aiperf/artifacts/granite4:350m-openai-chat-concurrency
 
 ### Workloads and Data
 - [Trace Benchmarking](docs/benchmark-modes/trace-replay.md) - Deterministic workload replay
+- [Agentic Workloads](docs/benchmark-modes/agentic.md) - Benchmark multi-step agent workflows: replay recorded agent traces (Dynamo, Weka) or author them in AIPerf's Graph IR
 - [Bailian Traces](docs/tutorials/bailian-trace.md) - Bailian production trace replay
 - [Baseten Traces](docs/tutorials/baseten-trace.md) - Baseten Parquet production trace replay
 - [BurstGPT Traces](docs/tutorials/burst-gpt-trace.md) - BurstGPT real-world bursty traffic trace replay

docs/architecture.md

@@ -54,14 +54,15 @@ The Dataset Manager handles all aspects of input data management during benchmar
 - Parsing and validating input data to ensure it matches the expected format
 - Writing dataset to memory-mapped files, enabling workers to access data directly without message passing
 - Supporting custom dataset types, such as MoonCake traces, for advanced benchmarking scenarios
+- For graph IR workloads, broadcasting a graph-typed `DatasetConfiguredNotification` — `DatasetMetadata.graph` (the trace universe plus per-node prefix-cache map) and `GraphSegmentClientMetadata` (the unified segment store and structural sidecar locations) — instead of conversation entries and a conversation memory-mapped store; the store build itself is owned by `GraphStoreBuilder` (`src/aiperf/dataset/graph/store_build.py`), and the schedule plane and workers read those exact broadcast paths (nothing re-parses the workload)
 - Managing the lifecycle of datasets, including initialization, iteration, and cleanup
 
 ### Timing Manager
 
 The Timing Manager controls and coordinates the timing of requests during benchmarking runs through a credit-based system.
 
 **Key Responsibilities:**
-- Scheduling when each request should be sent based on the selected timing mode (fixed schedule, request-rate, or user-centric rate)
+- Scheduling when each request should be sent based on the selected timing mode (fixed schedule, request-rate, user-centric rate, or graph IR for agentic workloads)
 - Managing precise timing to accurately reproduce real-world or synthetic load patterns
 - Supporting advanced timing scenarios, such as replaying traces with specific inter-arrival times or simulating bursty traffic
 - Ensuring that requests are dispatched to workers at the correct intervals for reliable measurement
@@ -183,6 +184,7 @@ The Timing Manager uses a **credit-based flow control system** to control when r
   - **Fixed schedule mode**: Replays conversation traces at precise timestamps from dataset metadata
   - **Request-rate mode**: Issues credits at a specific rate with configurable arrival patterns (constant, Poisson, gamma, concurrency burst)
   - **User-centric rate mode**: Each session acts as a separate user with calculated gaps between turns
+  - **Graph IR mode**: Runs a dataflow executor per agentic-workload trace, issuing a node's credit once its dependencies complete (selected automatically for graph workloads)
 
 **Flow Control Benefits:**
 - Prevents overwhelming the inference server

docs/benchmark-datasets.md

@@ -72,7 +72,9 @@ This document describes datasets that AIPerf can use to generate stimulus. Addit
 
 ## Exgentic Agent Trace Replay
 
-The Exgentic loaders stream recorded agent sessions directly from Hugging Face. `exgentic` is pinned to v1 revision `70036b93a04e61b0ea2706a68b962f4f26774587`; `exgentic_v2` is independently pinned to v2 revision `4b8ad4ab198438e5a170f9171c19c6a2cf7c1814`. Each replays successful, positive-token chat call snapshots. Recorded messages, system instructions, tool definitions, output-token limits, request controls, and call start times are preserved. Tools are not executed, and live responses are not added to later requests. Every request carries the source session as `x-dynamo-session-id` for Dynamo agentic tracing while AIPerf retains its own request correlation ID.
+The Exgentic loaders stream recorded agent sessions directly from Hugging Face. `exgentic` is pinned to v1 revision `70036b93a04e61b0ea2706a68b962f4f26774587`; `exgentic_v2` is independently pinned to v2 revision `4b8ad4ab198438e5a170f9171c19c6a2cf7c1814`. Each replays successful, positive-token chat call snapshots. Recorded messages, system instructions, tool definitions, output-token limits, request controls, and call start times are preserved. Tools are not executed, and live responses are not added to later requests. The loaders no longer stamp a session header themselves; AIPerf keeps a stable per-conversation correlation ID, and pairing the run with `--session-routing dynamo_headers` sends `X-Dynamo-Session-ID` for Dynamo session affinity and agentic tracing.
+
+This changes fixed-schedule affinity grouping. `--session-routing dynamo_headers` keys affinity on the live per-conversation correlation ID, and in `--fixed-schedule` replay the loader explodes one recorded agent session into many single-turn span-conversations — so each span now gets its own key and routes independently, and Dynamo prefix-affinity no longer groups a recorded session's spans onto one worker. The removed stamp instead put the shared recorded session ID on every span, grouping them (but that recorded ID also collided across dataset-recycled replays, binding unrelated live conversations to a single router entry). The two behaviors serve different fidelity goals; restoring source-session grouping without the recycling collision would need a dataset-provided affinity key (possible future work).
 
 Provide a finite materialization bound through `--num-conversations`, `--num-dataset-entries`, or `--request-count`. `--benchmark-duration` limits request issuance, not dataset setup.