skainet-lang-core/sk.ainet.lang.trace

Package-level declarations

Types

CompositeSink

class CompositeSink(sinks: List<OpSink>) : OpSink

Composite sink that fans out traces to multiple sinks in order. Intentionally simple for hot path performance.

NoOpSink

object NoOpSink : OpSink

No-op sink: ignores all traces.

OpAttributeFactory

object OpAttributeFactory

OpAttributeFactory provides methods to generate operation attributes for tracing. It captures metadata about tensor operations for graph construction and optimization.

OpSink

interface OpSink

OpSink consumes operation execution traces emitted by TracingTensorOps. Kept in core so it can be used without depending on graph/tape implementations.

OpTrace

data class OpTrace(val opType: String, val inputs: List<TensorRef>, val outputs: List<TensorRef>, val attributes: Map<String, Any?> = emptyMap())

OpTrace represents a single tensor operation execution with its inputs, outputs, and attributes. This is the core data structure for capturing computation graphs during execution.

TensorRef

data class TensorRef(val id: String, val shape: Shape, val dtype: DType)

TensorRef represents a reference to a tensor in the computation graph. It captures the essential metadata needed for graph construction and optimization.

TraceSession

open class TraceSession

TraceSession manages tensor references during computation graph tracing. It provides methods to convert tensors to TensorRef objects for graph capture.