Skip to main contentModule protocols
Source - DirectRequest
- ForwardPassSnapshot
- Per-iteration forward pass snapshot, mirroring the Python
ForwardPassMetrics
schema in components/src/dynamo/common/forward_pass_metrics.py. - FpmPublisher
- Optional FPM sink used by schedulers.
Wraps
Option<Arc<dyn FpmSink>> for ergonomic passing and no-op default behavior. - G1
- Metadata marker type for kvbm-logical blocks in the mocker’s G1 pool.
- KvEventPublishers
- Shared KV event publisher bundle used by schedulers and KV managers.
- MockEngineArgs
- Configuration arguments for MockEngine
- MockEngineArgsBuilder
- Builder for
MockEngineArgs. - OutputSignal
- Signal for output token generation with completion status
- PrefillCost
- Represents the cost of prefilling content in the cache
- RawKvEvent
- Raw KV event payload used by transport-specific publishers such as the
vLLM-native ZMQ event stream.
- ReasoningConfig
- Configuration for reasoning/thinking token output in the mocker.
- SglangArgs
- SGLang-specific configuration parameters.
- TrtllmArgs
- TensorRT-LLM-specific configuration parameters.
- EngineType
- Engine type for selecting scheduling and KV cache simulation behavior
- G1Backend
- G1 implementation used by the shared vLLM/TRT-LLM mock scheduler.
- KvTransferTimingMode
- Physical KV footprint used to model a coordinated disaggregated transfer.
- MockEngineArgsBuilderError
- Error type for MockEngineArgsBuilder
- MockerEvictionBackend
- Eviction strategy for the kvbm-logical inactive pool.
- MoveBlock
- Represents different block movement operations in the cache
For Use and Promote variants, block hashes are included for KV event publishing
- MoveBlockResponse
- PreemptionMode
- Preemption policy for evicting decode requests under memory pressure
- SchedulingPolicy
- Scheduling policy applied by the shared vLLM scheduler core.
- WorkerType
- Worker type for disaggregated serving configurations
- FpmSink
- Trait for publishing forward pass metrics snapshots.
This abstracts the FPM publishing pipeline so mocker schedulers remain generic.
- KvCacheEventSink
- Trait for publishing KV cache events.
This abstracts the runtime dependency so mocker components can remain generic.
- RawKvEventSink
- Trait for publishing transport-specific raw KV event payloads.
- NumBlocks