Crate max_sys 

Raw FFI bindings to the Modular MAX C API.

Generated by bindgen from the headers in max/include/max/c/. These are opaque-struct-based bindings — you interact through function pointers, not by reading struct fields directly.

For a safe, ergonomic API, use pyroxide instead.

Structs

M_AsyncCompiledModel

M_AsyncModel

M_AsyncTensor

M_AsyncTensorMap

M_CompileConfig

M_Device

M_RuntimeConfig

M_RuntimeContext

M_Status

M_TensorNameArray

M_TensorSpec

M_WeightsRegistry

Constants

M_AllocatorType_kCaching

M_AllocatorType_kSystem

M_DeviceType_M_ACCELERATOR

M_DeviceType_M_HOST

M_Dtype_M_BFLOAT16

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_BOOL

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT4_E2M1FN

Bits 0 through 3 indicate the kind of FP value.

M_Dtype_M_FLOAT8_E3M4

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT8_E4M3FN

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT8_E4M3FNUZ

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT8_E5M2

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT8_E5M2FNUZ

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT8_E8M0FNU

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT16

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT32

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_FLOAT64

Some slots are left blank here to enable us to support more lower precision types in the future.

M_Dtype_M_INT1

Bit 0 encodes “isSigned”.

M_Dtype_M_INT2

Bit 0 encodes “isSigned”.

M_Dtype_M_INT4

Bit 0 encodes “isSigned”.

M_Dtype_M_INT8

Bit 0 encodes “isSigned”.

M_Dtype_M_INT16

Bit 0 encodes “isSigned”.

M_Dtype_M_INT32

Bit 0 encodes “isSigned”.

M_Dtype_M_INT64

Bit 0 encodes “isSigned”.

M_Dtype_M_INT128

Bit 0 encodes “isSigned”.

M_Dtype_M_UINT1

Bit 0 encodes “isSigned”.

M_Dtype_M_UINT2

Bit 0 encodes “isSigned”.

M_Dtype_M_UINT4

Bit 0 encodes “isSigned”.

M_Dtype_M_UINT8

Bit 0 encodes “isSigned”.

M_Dtype_M_UINT16

Bit 0 encodes “isSigned”.

M_Dtype_M_UINT32

Bit 0 encodes “isSigned”.

M_Dtype_M_UINT64

Bit 0 encodes “isSigned”.

M_Dtype_M_UINT128

Bit 0 encodes “isSigned”.

M_Dtype_M_UNKNOWN

M_Dtype_kIntWidthShift

Bit 0 encodes “isSigned”.

M_Dtype_mIsComplex

M_Dtype_mIsFloat

M_Dtype_mIsInteger

M_Dtype_mIsSigned

Bit 0 encodes “isSigned”.

M_ResultOutputStyle_M_BINARY

M_ResultOutputStyle_M_BINARY_MAX_CHECKPOINT

M_ResultOutputStyle_M_COMPACT

M_ResultOutputStyle_M_FULL

M_ResultOutputStyle_M_NONE

M_ValueType_M_BOOL_VALUE

M_ValueType_M_DICT_VALUE

M_ValueType_M_DOUBLE_VALUE

M_ValueType_M_LIST_VALUE

M_ValueType_M_LONG_VALUE

M_ValueType_M_MOJO_VALUE

M_ValueType_M_NONE_VALUE

M_ValueType_M_PYTHON_MOJO_VALUE

M_ValueType_M_STRING_VALUE

M_ValueType_M_TENSOR_VALUE

M_ValueType_M_TUPLE_VALUE

M_ValueType_M_UNKNOWN_VALUE

Functions

M_borrowTensorInto^⚠

Adds a tensor to the tensor map.

M_captureModelSync^⚠

Captures model execution into a device graph for later replay.

M_compileModel^⚠

Compiles a model.

M_compileModelSync^⚠

Synchronously compiles a model.

M_copyTensorToDevice^⚠

Copies a tensor to a different device.

M_debugVerifyReplayModelSync^⚠

Executes eagerly and verifies that the kernel launch trace matches a previously captured device graph.

M_executeModelSync^⚠

Executes a model synchronously.

M_freeAsyncTensorMap^⚠

Deallocates the memory for the tensor map. No-op if tensorMap is NULL.

M_freeCompileConfig^⚠

Deallocates the memory for the compile config. No-op if config is NULL.

M_freeCompiledModel^⚠

Deallocates the memory for the compiled model. No-op if model is NULL.

M_freeDevice^⚠

Deallocates the memory for a device. No-op if device is NULL.

M_freeModel^⚠

Deallocates the memory for the model. No-op if model is NULL.

M_freeRuntimeConfig^⚠

Deallocates the memory for a runtime config. No-op if config is NULL.

M_freeRuntimeContext^⚠

Deallocates the memory for a runtime context. No-op if context is NULL.

M_freeStatus^⚠

Deallocates the memory for the status object. No-op if status is NULL.

M_freeTensor^⚠

Deallocates the memory for the tensor. No-op if tensor is NULL.

M_freeTensorNameArray^⚠

Deallocates the memory for the array of tensor names. No-op if names is NULL.

M_freeTensorSpec^⚠

Deallocates the memory for the tensor spec. No-op if spec is NULL.

M_freeWeightsRegistry^⚠

Deallocates the memory for the weights registry. No-op if weightsRegistry is NULL.

M_getAcceleratorCount^⚠

Returns the number of available accelerator (GPU) devices.

M_getDeviceId^⚠

Gets the ID of a device.

M_getDeviceIdFromSpec^⚠

Gets the device ID from a tensor specification.

M_getDeviceLabel^⚠

Gets the label of a device (e.g., “cpu” or “gpu”).

M_getDeviceType^⚠

Gets the type of a device.

M_getDeviceTypeFromSpec^⚠

Gets the device type from a tensor specification.

M_getDimAt^⚠

Gets the element at a particular axis.

M_getDtype^⚠

Gets the datatype from the tensor spec.

M_getError^⚠

Gets an error message from the M_Status parameter.

M_getName^⚠

Gets the name of the tensor from the tensor spec.

M_getRank^⚠

Gets the rank from the tensor spec.

M_getTensorByNameFrom^⚠

Gets a tensor from the tensor map by name.

M_getTensorData^⚠

Gets a pointer to underlying data of the tensor.

M_getTensorDevice^⚠

Gets the device on which a tensor resides.

M_getTensorNumElements^⚠

Gets the number of elements for the tensor.

M_getTensorSpec^⚠

Gets a Tensor Spec for the tensor.

M_getTensorType^⚠

Gets the corresponding M_Dtype for the tensor.

M_initModel^⚠

Sets up a model for execution.

M_isDynamicRanked^⚠

Returns if the given spec has a dynamic rank.

M_isError^⚠

Checks if status holds an error value.

M_isHostDevice^⚠

Checks if the device is a host device.

M_newAsyncTensorMap^⚠

Creates a map of tensor names to async tensors.

M_newCompileConfig^⚠

Creates an object you can use to configure model compilation.

M_newDevice^⚠

Creates a device of the specified type and ID.

M_newRuntimeConfig^⚠

Creates a new runtime config.

M_newRuntimeContext^⚠

Creates a runtime context.

M_newStatus^⚠

Creates a new status object.

M_newTensorSpec^⚠

Creates a tensor specification.

M_newWeightsRegistry^⚠

Creates a weights registry from parallel arrays of weight names and data pointers.

M_replayModelSync^⚠

Replays a previously captured device graph.

M_runtimeConfigAddDevice^⚠

Adds a device to be accessible from the runtime.

M_setDebugPrintOptions^⚠

Set the options for debugging printing of tensors when executing a model.

M_setModelPath^⚠

Sets the path to a model.

M_setMojoDefineBool^⚠

Sets a mojo compile-time define with an boolean value.

M_setMojoDefineInt^⚠

Sets a mojo compile-time define with an integer value.

M_setMojoDefineString^⚠

Sets a mojo compile-time define with an string value.

M_synchronizeDevice^⚠

Synchronizes the device, ensuring all operations complete.

M_version^⚠

Gets the MAX Engine version.

M_waitForCompilation^⚠

Blocks execution until the model is compiled.

M_waitForModel^⚠

Blocks execution until the model is initialized.

Type Aliases

M_AllocatorType

Contains an AllocatorType. You can choose between kCaching and kSystem kCaching trades off higher memory usage for better performance. kSystem uses the default system allocator.

M_DeviceType

M_Dtype

M_ResultOutputStyle

Represents the result output style for debug printing.

M_ValueType

Represents the type of a value.