Dynamo Sweeper Integration

AI Simulate’s Sweeper core does not depend on Dynamo. Dynamo owns its optional Planner and Router sweep configuration providers and DynamoReplayRunnerFactory.

Install

The supported prebuilt environment is the dynamo-planner image. The image stages and installs the published aisimulate==0.1.0.dev1 wheel alongside the Dynamo wheels. Dynamo’s Rust workspace resolves aisimulate-core==0.1.0-dev.1 from crates.io.

For Dynamo source development, install the published AI Simulate wheel and build the matching Dynamo bindings:

$
python -m pip install pip "maturin[patchelf]"
$
cd lib/bindings/python
$
maturin develop --uv --release --features aic-forward-pass
$
cd ../../..
$
python -m pip install --no-deps -e .
$
python -m pip install "aisimulate==0.1.0.dev1"
$
python -m pip install -r container/deps/requirements.planner.txt

The repository does not yet provide an ai-dynamo[simulation] extra. Add that extra only after the full AI Simulate and Replay refactor has a supported distribution path.

Run a Dynamo Sweep

Compose the runtime directly:

1
from aisimulate.sweeper import SmartSearchConfig, Sweeper
2
from dynamo.replay.simulation import DynamoReplayRunnerFactory
3
4
config = SmartSearchConfig.from_yaml("smart_sweep.yaml")
5
candidates = Sweeper(
6
    runner_factory=DynamoReplayRunnerFactory(),
7
).run(config)

There is no --enable-dynamo flag. Selecting a Dynamo adapter in configuration and injecting the Dynamo runner are the explicit composition points.

Dynamo Providers

The ai-dynamo distribution registers these entry points:

Each adapter value is a search space, not one concrete Planner or Router config:

1
adapters:
2
  dynamo.router:
3
    search_space:
4
      mode: [kv_router, round_robin]
5
      overlap_score_credit: [0.0, 0.5, 1.0]
6
  dynamo.planner:
7
    search_space:
8
      scaling_policy:
9
        preset: [disabled, load_180_5]
10
      fpm_sampling:
11
        preset: [default]
12
      load_sensitivity:
13
        preset: [default]

Each Planner preset sub-item owns a complete knob set:

Named presets and custom mappings are validated against the complete sub-item. The provider fills family defaults for conditionally inactive predictor knobs before validation.

The Planner provider derives load-predictor parameters from all configured scaling intervals during generate_search_space. Enabled candidates materialize a concrete PlannerConfig runtime hook. The Router provider materializes either round-robin behavior without a hook or a concrete KV-router hook.

Replay Composition

DynamoReplayRunnerFactory converts each serializable ReplaySpec into an invocation of the shared AI Simulate Replayer. It resolves materialized Planner and Router hooks into Dynamo-owned scaling and placement policies, while the Replayer continues to own traffic execution and report generation.

The runner passes trace, fixed, or KV-load-derived closed-loop concurrency through ReplaySpec.concurrency. It applies a goodput SLA only when goal.sla is configured; this SLA is independent of the Planner’s scaling SLA. A dynamo.router:placement_policy@1 hook selects the Dynamo placement policy. Without that hook, the Replayer uses its built-in round-robin policy.