ExecuTorch Runtime API Reference

The ExecuTorch C++ API provides an on-device execution framework for exported PyTorch models.

For a tutorial style introduction to the runtime API, check out the runtime tutorial and its simplified version.

For detailed information on how APIs evolve and the deprecation process, please refer to the ExecuTorch API Life Cycle and Deprecation Policy.

Model Loading and Execution

class Program

A deserialized ExecuTorch program binary.

Public Types

enum class Verification : uint8_t

Types of validation that the Program can do before parsing the data.

Values:

enumerator Minimal

Do minimal verification of the data, ensuring that the header appears correct.

Has minimal runtime overhead.

enumerator InternalConsistency

Do full verification of the data, ensuring that internal pointers are self-consistent and that the data has not been truncated or obviously corrupted. May not catch all types of corruption, but should guard against illegal memory operations during parsing.

Will have higher runtime overhead, scaling with the complexity of the proram data.

enum HeaderStatus

Describes the presence of an ExecuTorch program header.

Values:

enumerator CompatibleVersion

An ExecuTorch program header is present, and its version is compatible with this version of the runtime.

enumerator IncompatibleVersion

An ExecuTorch program header is present, but its version is not compatible with this version of the runtime.

enumerator NotPresent

An ExecuTorch program header is not present.

enumerator ShortData

The data provided was too short to find the program header.

Public Functions

Result<const void*> get_constant_buffer_data(size_t buffer_idx, size_t nbytes) const

Get the constant buffer inside Program with index buffer_idx.

Parameters

• buffer_idx – [in] the index of the buffer in the constant_buffer.

• nbytes – [in] the number of bytes to read from the buffer.

Returns

The buffer with corresponding index.

size_t num_methods() const

Returns the number of methods in the program.

Result<const char*> get_method_name(size_t method_index) const

Returns the name of the method at particular index.

Parameters

method_index – [in] The index of the method name to retrieve. Must be less than the value returned by num_methods().

Returns

The name of the requested method. The pointer is owned by the Program, and has the same lifetime as the Program.

Result<Method> load_method(const char *method_name, MemoryManager *memory_manager, EventTracer *event_tracer = nullptr) const

Loads the named method and prepares it for execution.

Parameters

• method_name – [in] The name of the method to load.

• memory_manager – [in] The allocators to use during initialization and execution of the loaded method. If memory_manager.temp_allocator() is null, the runtime will allocate temp memory using et_pal_allocate().

• event_tracer – [in] The event tracer to use for this method run.

Returns

The loaded method on success, or an error on failure.

Result<MethodMeta> method_meta(const char *method_name) const

Gathers metadata for the named method.

Parameters

method_name – [in] The name of the method to get metadata for.

ET_DEPRECATED Result< const char * > get_output_flattening_encoding (const char *method_name="forward") const

DEPRECATED: Get the pytree encoding string for the output. Deprecated as this functionality will eventually move out of the core program into a higher level structure, but that does not exist at this time.

Parameters

method_name – [in] The name of the method to get the encoding for.

Returns

The pytree encoding string for the output

Public Static Functions

static ET_NODISCARD Result< Program > load (DataLoader *loader, Verification verification=Verification::Minimal)

Loads a Program from the provided loader. The Program will hold a pointer to the loader, which must outlive the returned Program instance.

Parameters

• loader – [in] The source to load program data from. The Program will hold a pointer to this loader, which must outlive the returned Program instance.

• verification – [in] The type of verification to do before returning success.

static inline ET_DEPRECATED ET_NODISCARD Result< Program > Load (DataLoader *loader, Verification verification=Verification::Minimal)

DEPRECATED: Use the lowercase load() instead.

static HeaderStatus check_header(const void *data, size_t size)

Looks for an ExecuTorch program header in the provided data.

Parameters

• data – [in] The data from the beginning of a file that might contain an ExecuTorch program.

• size – [in] The size of data in bytes. Must be >= kMinHeadBytes.

Returns

A value describing the presence of a header in the data.

Public Static Attributes

static constexpr size_t kMinHeadBytes = 64

The minimum number of bytes necessary for calls to check_header.

class Method

An executable method of an executorch program. Maps to a python method like forward() on the original nn.Module.

Public Functions

inline Method(Method &&rhs) noexcept

Move ctor. Takes ownership of resources previously owned by rhs, and leaves rhs in an uninitialized state.

ET_NODISCARD Error set_input (const EValue &input_evalue, size_t input_idx)

Sets the internal input value to be equivalent to the to the provided value.

Parameters

• input_evalue – [in] The evalue to copy into the method input. If the evalue is a tensor, the data is copied in most cases, so the tensor passed in here does not always need to outlive this call. But there is a case where the Method will keep a pointer to the tensor’s data. Based on the memory plan of the method, the inputs may not have buffer space pre-allocated for them. In this case the executor will alias the memory of the tensors provided as inputs here rather then deepcopy the input into the memory planned arena.

• input_idx – [in] Zero-based index of the input to set. Must be less than the value returned by inputs_size().

Returns

Error::Ok on success, non-Ok on failure.

ET_NODISCARD Error set_inputs (const executorch::aten::ArrayRef< EValue > &input_evalues)

Sets the values of all method inputs.

See set_input() for a more detailed description of the behavior.

Parameters

input_evalues – [in] The new values for all of the method inputs. The type of each element must match the type of corresponding input. If the value of an element is a tensor, attempts to allow dynamic shape, but the dtype must always agree.

Returns

Error::Ok on success, non-Ok on failure.

ET_NODISCARD Error set_output_data_ptr (void *buffer, size_t size, size_t output_idx)

Sets the data buffer of the specified method output to the provided value.

NOTE: Based on the memory plan of the method, the output tensors may not have buffer space pre-allocated for them, in this case the executor will point those tensors to the buffer provided here, so the user should take care that the life span of this memory outlasts the executor forward.

Parameters

• buffer – [in] The block of memory to point the specified tensor at.

• size – [in] the length of buffer in bytes, must be >= the nbytes of the specified tensor.

• output_idx – [in] The index of the output to set the data_ptr for. Must correspond to a tensor, and that tensor must not have had a buffer allocated by the memory plan.

Returns

Error::Ok on success, non-Ok on failure.

ET_NODISCARD Error get_outputs (EValue *output_evalues, size_t length)

Copies the method’s outputs into the provided array.

WARNING: The output contains shallow copies of internal tensor outputs. Please do not mutate returned Tensor elements.

TODO(T139259264): Add checks to detect output mutation, or deep-copy outputs.

Parameters

• output_evalues – [in] The array to copy the outputs into. The first outputs_size() elements will be set to the corresponding output values. The rest of the array will be set to the EValue value None.

• length – [in] The size of the output_evalues array in elements. Must be greater than or equal to outputs_size().

Returns

Error::Ok on success, non-Ok on failure.

ET_NODISCARD Error get_inputs (EValue *input_evalues, size_t length)

Copies the method’s inputs into the provided array.

WARNING: The input contains shallow copies of internal tensor inputs. Please do not mutate returned Tensor elements.

Parameters

• input_evalues – [in] The array to copy the inputs into. The first inputs_size() elements will be set to the corresponding input values. The rest of the array will be set to the EValue value None.

• length – [in] The size of the input_evalues array in elements. Must be greater than or equal to inputs_size().

Returns

Error::Ok on success, non-Ok on failure.

ET_NODISCARD Error execute ()

Execute the method.

NOTE: Will fail if the method has been partially executed using the step() api.

Returns

Error::Ok on success, non-Ok on failure.

ET_EXPERIMENTAL ET_NODISCARD Error step ()

EXPERIMENTAL: Advances/executes a single instruction in the method.

Return values

• Error::Ok – step succeeded

• non-Ok – step failed

• Error::EndOfMethod – method finished executing successfully

ET_DEPRECATED ET_NODISCARD Error experimental_step ()

DEPRECATED: Use step() instead.

ET_EXPERIMENTAL ET_NODISCARD Error reset_execution ()

EXPERIMENTAL: Resets execution state to the start of the Method. For use with the step() API.

Return values

• Error:Ok – on success

• Error::InvalidState – if called before step-based execution reached the end of the Method. This means it is not possible to recover a Method that failed mid-execution.

ET_DEPRECATED ET_NODISCARD Error experimental_reset_execution ()

DEPRECATED: Use reset_execution() instead.

MethodMeta method_meta() const

Returns the MethodMeta that corresponds to the calling Method.

size_t inputs_size() const

Returns the number of inputs the Method expects.

size_t outputs_size() const

Returns the number of outputs the Method returns.

const EValue &get_output(size_t i) const

Retrieves the output at the specified index.

ET_DEPRECATED const EValue & get_input (size_t i) const

DEPRECATED: Use MethodMeta instead to access metadata, and set_input to update Method inputs.

ET_DEPRECATED EValue & mutable_input (size_t i)

DEPRECATED: Use MethodMeta instead to access metadata, and set_input to update Method inputs.

ET_DEPRECATED EValue & mutable_output (size_t i)

DEPRECATED: Use MethodMeta instead to access metadata, and get_output to retrieve Method outputs.

class MethodMeta

Describes a a method in an ExecuTorch program.

The program used to create a MethodMeta object must outlive the MethodMeta. It is separate from Method so that this information can be accessed without paying the initialization cost of loading the full Method.

Public Functions

const char *name() const

Get the name of this method.

Returns

The method name.

size_t num_inputs() const

Get the number of inputs to this method.

Returns

The number of inputs.

Result<Tag> input_tag(size_t index) const

Get the tag of the specified input.

Parameters

index – [in] The index of the input to look up.

Returns

The tag of input, can only be [Tensor, Int, Bool, Double, String].

Result<TensorInfo> input_tensor_meta(size_t index) const

Get metadata about the specified input.

Parameters

index – [in] The index of the input to look up.

Returns

The metadata on success, or an error on failure. Only valid for tag::Tensor

size_t num_outputs() const

Get the number of outputs to this method.

Returns

The number of outputs.

Result<Tag> output_tag(size_t index) const

Get the tag of the specified output.

Parameters

index – [in] The index of the output to look up.

Returns

The tag of output, can only be [Tensor, Int, Bool, Double, String].

Result<TensorInfo> output_tensor_meta(size_t index) const

Get metadata about the specified output.

Parameters

index – [in] The index of the output to look up.

Returns

The metadata on success, or an error on failure. Only valid for tag::Tensor

size_t num_memory_planned_buffers() const

Get the number of memory-planned buffers this method requires.

Returns

The number of memory-planned buffers.

Result<int64_t> memory_planned_buffer_size(size_t index) const

Get the size in bytes of the specified memory-planned buffer.

Parameters

index – [in] The index of the buffer to look up.

Returns

The size in bytes on success, or an error on failure.

inline ET_DEPRECATED size_t num_non_const_buffers () const

DEPRECATED: Use num_memory_planned_buffers() instead.

inline Result<int64_t> non_const_buffer_size(size_t index) const

DEPRECATED: Use memory_planned_buffer_size() instead.

class DataLoader

Loads from a data source.

See //executorch/extension/data_loader for common implementations.

Public Functions

virtual ET_NODISCARD Result< FreeableBuffer > load (size_t offset, size_t size, const SegmentInfo &segment_info) const =0

Loads data from the underlying data source.

NOTE: This must be thread-safe. If this call modifies common state, the implementation must do its own locking.

Parameters

• offset – The byte offset in the data source to start loading from.

• size – The number of bytes to load.

• segment_info – Information about the segment being loaded.

Returns

a FreeableBuffer that owns the loaded data.

inline virtual ET_NODISCARD Error load_into (size_t offset, size_t size, const SegmentInfo &segment_info, void *buffer) const

Loads data from the underlying data source into the provided buffer.

NOTE: This must be thread-safe. If this call modifies common state, the implementation must do its own locking.

Parameters

• offset – The byte offset in the data source to start loading from.

• size – The number of bytes to load.

• segment_info – Information about the segment being loaded.

• buffer – The buffer to load data into. Must point to at least size bytes of memory.

Returns

an Error indicating if the load was successful.

virtual ET_NODISCARD Result< size_t > size () const =0

Returns the length of the underlying data source, typically the file size.

struct SegmentInfo

Describes the content of the segment.

Public Types

enum class Type

Represents the purpose of the segment.

Values:

enumerator Program

Data for the actual program.

enumerator Constant

Holds constant tensor data.

enumerator Backend

Data used for initializing a backend.

enumerator Mutable

Data used for initializing mutable tensors.

Public Members

Type segment_type

Type of the segment.

size_t segment_index

Index of the segment within the segment list. Undefined for program segments.

const char *descriptor

An optional, null-terminated string describing the segment. For Backend segments, this is the backend ID. Null for other segment types.

class MemoryAllocator

A class that does simple allocation based on a size and returns the pointer to the memory address. It bookmarks a buffer with certain size. The allocation is simply checking space and growing the cur_ pointer with each allocation request.

Simple example:

// User allocates a 100 byte long memory in the heap. uint8_t* memory_pool = malloc(100 * sizeof(uint8_t)); MemoryAllocator allocator(100, memory_pool) // Pass allocator object in the Executor

Underneath the hood, ExecuTorch will call allocator.allocate() to keep iterating cur_ pointer

Subclassed by executorch::runtime::internal::PlatformMemoryAllocator

Public Functions

inline MemoryAllocator(uint32_t size, uint8_t *base_address)

Constructs a new memory allocator of a given size, starting at the provided base_address.

Parameters

• size – [in] The size in bytes of the buffer at base_address.

• base_address – [in] The buffer to allocate from. Does not take ownership of this buffer, so it must be valid for the lifetime of of the MemoryAllocator.

inline virtual void *allocate(size_t size, size_t alignment = kDefaultAlignment)

Allocates size bytes of memory.

Parameters

• size – [in] Number of bytes to allocate.

• alignment – [in] Minimum alignment for the returned pointer. Must be a power of 2.

Return values

nullptr – Not enough memory, or alignment was not a power of 2.

Returns

Aligned pointer to the allocated memory on success.

template<typename T>
inline T *allocateInstance(size_t alignment = alignof(T))

Allocates a buffer large enough for an instance of type T. Note that the memory will not be initialized.

Example:

auto p = memory_allocator->allocateInstance<MyType>();

Parameters

alignment – [in] Minimum alignment for the returned pointer. Must be a power of 2. Defaults to the natural alignment of T.

Return values

nullptr – Not enough memory, or alignment was not a power of 2.

Returns

Aligned pointer to the allocated memory on success.

template<typename T>
inline T *allocateList(size_t size, size_t alignment = alignof(T))

Allocates size number of chunks of type T, where each chunk is of size equal to sizeof(T) bytes.

Parameters

• size – [in] Number of memory chunks to allocate.

• alignment – [in] Minimum alignment for the returned pointer. Must be a power of 2. Defaults to the natural alignment of T.

Return values

nullptr – Not enough memory, or alignment was not a power of 2.

Returns

Aligned pointer to the allocated memory on success.

Public Static Attributes

static constexpr size_t kDefaultAlignment = alignof(void*)

Default alignment of memory returned by this class. Ensures that pointer fields of structs will be aligned. Larger types like long double may not be, however, depending on the toolchain and architecture.

class HierarchicalAllocator

A group of buffers that can be used to represent a device’s memory hierarchy.

Public Functions

inline explicit HierarchicalAllocator(Span<Span<uint8_t>> buffers)

Constructs a new hierarchical allocator with the given array of buffers.

• Memory IDs are based on the index into buffers: buffers[N] will have a memory ID of N.

• buffers.size() must be >= MethodMeta::num_non_const_buffers().

• buffers[N].size() must be >= MethodMeta::non_const_buffer_size(N).

inline ET_DEPRECATED HierarchicalAllocator(uint32_t n_allocators, MemoryAllocator *allocators)

DEPRECATED: Use spans instead.

inline ET_NODISCARD Result< void * > get_offset_address (uint32_t memory_id, size_t offset_bytes, size_t size_bytes)

Returns the address at the byte offset offset_bytes from the given buffer’s base address, which points to at least size_bytes of memory.

Parameters

• memory_id – [in] The ID of the buffer in the hierarchy.

• offset_bytes – [in] The offset in bytes into the specified buffer.

• size_bytes – [in] The amount of memory that should be available at the offset.

Returns

On success, the address of the requested byte offset into the specified buffer. On failure, a non-Ok Error.

class MemoryManager

A container class for allocators used during Method load and execution.

This class consolidates all dynamic memory needs for Method load and execution. This can allow for heap-based as well as heap-less execution (relevant to some embedded scenarios), and overall provides more control over memory use.

This class, however, cannot ensure all allocation is accounted for since kernel and backend implementations are free to use a separate way to allocate memory (e.g., for things like scratch space). But we do suggest that backends and kernels use these provided allocators whenever possible.

Public Functions

inline explicit MemoryManager(MemoryAllocator *method_allocator, HierarchicalAllocator *planned_memory = nullptr, MemoryAllocator *temp_allocator = nullptr)

Constructs a new MemoryManager.

Parameters

• method_allocator – [in] The allocator to use when loading a Method and allocating its internal structures. Must outlive the Method that uses it.

• planned_memory – [in] The memory-planned buffers to use for mutable tensor data when executing a Method. Must outlive the Method that uses it. May be nullptr if the Method does not use any memory-planned tensor data. The sizes of the buffers in this HierarchicalAllocator must agree with the corresponding MethodMeta::num_memory_planned_buffers() and MethodMeta::memory_planned_buffer_size(N) values, which are embedded in the Program.

• temp_allocator – [in] The allocator to use when allocating temporary data during kernel or delegate execution. Must outlive the Method that uses it. May be nullptr if the Method does not use kernels or delegates that allocate temporary data. This allocator will be reset after every kernel or delegate call during execution.

inline ET_DEPRECATED MemoryManager(__attribute__((unused)) MemoryAllocator *constant_allocator, HierarchicalAllocator *non_constant_allocator, MemoryAllocator *runtime_allocator, MemoryAllocator *temporary_allocator)

DEPRECATED: Use the constructor without constant_allocator instead.

TODO(T162089316): Remove this once all users migrate to the new ctor.

inline MemoryAllocator *method_allocator() const

Returns the allocator that the runtime will use to allocate internal structures while loading a Method. Must not be used after its associated Method has been loaded.

inline HierarchicalAllocator *planned_memory() const

Returns the memory-planned buffers to use for mutable tensor data.

inline MemoryAllocator *temp_allocator() const

Returns the allocator to use for allocating temporary data during kernel or delegate execution.

This allocator will be reset after every kernel or delegate call during execution.

Values

struct EValue

Public Functions

inline EValue(executorch::aten::Scalar s)

Construct an EValue using the implicit value of a Scalar.

template<typename T>
inline executorch::aten::optional<T> toOptional() const

Converts the EValue to an optional object that can represent both T and an uninitialized state.

union Payload

union TriviallyCopyablePayload

class Tensor

A minimal Tensor type whose API is a source compatible subset of at::Tensor.

NOTE: Instances of this class do not own the TensorImpl given to it, which means that the caller must guarantee that the TensorImpl lives longer than any Tensor instances that point to it.

See the documention on TensorImpl for details about the return/parameter types used here and how they relate to at::Tensor.

Public Types

using SizesType = TensorImpl::SizesType

The type used for elements of sizes().

using DimOrderType = TensorImpl::DimOrderType

The type used for elements of dim_order().

using StridesType = TensorImpl::StridesType

The type used for elements of strides().

Public Functions

inline TensorImpl *unsafeGetTensorImpl() const

Returns a pointer to the underlying TensorImpl.

NOTE: Clients should be wary of operating on the TensorImpl directly instead of the Tensor. It is easy to break things.

inline size_t nbytes() const

Returns the size of the tensor in bytes.

NOTE: Only the alive space is returned not the total capacity of the underlying data blob.

inline ssize_t size(ssize_t dim) const

Returns the size of the tensor at the given dimension.

NOTE: that size() intentionally does not return SizeType even though it returns an element of an array of SizeType. This is to help make calls of this method more compatible with at::Tensor, and more consistent with the rest of the methods on this class and in ETensor.

inline ssize_t dim() const

Returns the tensor’s number of dimensions.

inline ssize_t numel() const

Returns the number of elements in the tensor.

inline ScalarType scalar_type() const

Returns the type of the elements in the tensor (int32, float, bool, etc).

inline ssize_t element_size() const

Returns the size in bytes of one element of the tensor.

inline const ArrayRef<SizesType> sizes() const

Returns the sizes of the tensor at each dimension.

inline const ArrayRef<DimOrderType> dim_order() const

Returns the order the dimensions are laid out in memory.

inline const ArrayRef<StridesType> strides() const

Returns the strides of the tensor at each dimension.

inline TensorShapeDynamism shape_dynamism() const

Returns the mutability of the shape of the tensor.

template<typename T>
inline const T *const_data_ptr() const

Returns a pointer of type T to the constant underlying data blob.

inline const void *const_data_ptr() const

Returns a pointer to the constant underlying data blob.

template<typename T>
inline T *mutable_data_ptr() const

Returns a pointer of type T to the mutable underlying data blob.

inline void *mutable_data_ptr() const

Returns a pointer to the mutable underlying data blob.

template<typename T> inline ET_DEPRECATED T * data_ptr () const

DEPRECATED: Use const_data_ptr or mutable_data_ptr instead.

inline ET_DEPRECATED void * data_ptr () const

DEPRECATED: Use const_data_ptr or mutable_data_ptr instead.

inline ET_DEPRECATED void set_data (void *ptr) const

DEPRECATED: Changes the data_ptr the tensor aliases. Does not free the previously pointed to data, does not assume ownership semantics of the new ptr. This api does not exist in at::Tensor so kernel developers should avoid it.