Runtime API
The application-facing C++ reference is generated from the same source commit as this documentation. Each entry includes its directly declared public and protected members, template parameters, overloads, source comments and links to the exact header revision.
The scope covers model loading, execution, operator registration, memory planning, profiling, resource variables and platform initialization. Internal kernel helpers, test utilities and FlatBuffer builders are outside this application reference. Extraction uses TF_LITE_STATIC_MEMORY=1 and the default uncompressed runtime configuration; declarations guarded by USE_TFLM_COMPRESSION are not included. Inherited members are documented on their declaring base class. For supported operations and data types, use the operator explorer.
MicroMutableOpResolver registration methods are split alphabetically so each page stays manageable. Where an upstream declaration has no contract comment, the reference shows its signature without inventing parameter descriptions.
Execution
MicroInterpreter- Allocate tensors, invoke a model and access inputs/outputs.
tensorflow/lite/micro/micro_interpreter.h Model- FlatBuffer model view; GetModel returns a view into caller-owned bytes.
tensorflow/lite/schema/schema_generated.h GetModel- Return a model view over the caller-owned FlatBuffer bytes.
tensorflow/lite/schema/schema_generated.h
Operator registration
MicroOpResolver- Interface for builtin and custom operator lookup.
tensorflow/lite/micro/micro_op_resolver.h MicroMutableOpResolver- Fixed-capacity registration template; capacity counts registrations, not model nodes.
tensorflow/lite/micro/micro_mutable_op_resolver.h
Memory
MicroAllocator- Arena allocation and memory-planner integration.
tensorflow/lite/micro/micro_allocator.h RecordingMicroAllocator- Allocation accounting for memory diagnostics.
tensorflow/lite/micro/recording_micro_allocator.h RecordedAllocation- Byte and allocation-count recording result.
tensorflow/lite/micro/recording_micro_allocator.h RecordingMicroInterpreter- Interpreter variant exposing the recording allocator.
tensorflow/lite/micro/recording_micro_interpreter.h MicroMemoryPlanner- Interface for planning reusable buffers.
tensorflow/lite/micro/memory_planner/micro_memory_planner.h GreedyMemoryPlanner- Plan buffers using lifetimes and sizes.
tensorflow/lite/micro/memory_planner/greedy_memory_planner.h LinearMemoryPlanner- Sequential buffer placement without reuse.
tensorflow/lite/micro/memory_planner/linear_memory_planner.h
Profiling and state
MicroProfilerInterface- Profiling callback interface.
tensorflow/lite/micro/micro_profiler_interface.h MicroProfiler- Record and report runtime profiling events.
tensorflow/lite/micro/micro_profiler.h ScopedMicroProfiler- Scope-based profiling event lifetime.
tensorflow/lite/micro/micro_profiler.h MicroResourceVariables- Allocate, read, assign and reset resource variables.
tensorflow/lite/micro/micro_resource_variable.h
Platform and version
InitializeTarget- Platform initialization hook; use the implementation matching the application.
tensorflow/lite/micro/system_setup.h HELIA_RT_VERSION- Runtime header version string; does not identify linked kernel or compiler settings.
tensorflow/lite/micro/helia_rt_version.h TFLITE_SCHEMA_VERSION- FlatBuffer model schema revision expected by the interpreter.
tensorflow/lite/micro/micro_interpreter.h
Tensor and status types
TfLiteTensor- Tensor data, dimensions, quantization and allocation metadata for static-memory builds.
tensorflow/lite/core/c/common.h TfLiteEvalTensor- Evaluation-time tensor data, dimensions and type.
tensorflow/lite/core/c/common.h TfLiteStatus- Status values returned by runtime operations.
tensorflow/lite/core/c/c_api_types.h TfLiteType- Tensor element type identifiers.
tensorflow/compiler/mlir/lite/core/c/tflite_types.h
Version
Section titled “Version”Include tensorflow/lite/micro/helia_rt_version.h and read HELIA_RT_VERSION for the runtime source release string. This macro does not report the linked heliaCORE version or the compiler configuration.
Invocation contract
Section titled “Invocation contract”The interpreter has constructors accepting either an arena or an existing allocator. Both borrow resources from the caller. AllocateTensors() and Invoke() return TfLiteStatus; check the returned status before proceeding.
Tensor accessors expose memory associated with the interpreter. Typed accessors require the requested C++ type to match the tensor type and can return null. See the header for each accessor’s index requirements.