API Reference¶

Module Contents¶

class rmm.DeviceBuffer¶

Bases: object

Attributes

nbytes: Gets the size of the buffer in bytes.
ptr: Gets a pointer to the underlying data.
size: Gets the size of the buffer in bytes.

Methods

`capacity`(self)
`copy_from_device`(self, cuda_ary, ...)	Copy from a buffer on host to `self`
`copy_from_host`(self, ary, ...)	Copy from a buffer on host to `self`
`copy_to_host`(self[, ary])	Copy from a `DeviceBuffer` to a buffer on host.
`reserve`(self, size_t new_capacity, ...)
`resize`(self, size_t new_size, ...)
`to_device`(const unsigned char[, ...)	Calls `to_device` function on arguments provided.
`tobytes`(self, Stream stream=DEFAULT_STREAM)

capacity(self) → size_t¶

copy_from_device(self, cuda_ary, Stream stream=DEFAULT_STREAM)¶

Copy from a buffer on host to self

Parameters

cuda_aryobject to copy from that has __cuda_array_interface__
streamCUDA stream to use for copying, default the default stream

Examples

>>> import rmm
>>> db = rmm.DeviceBuffer(size=5)
>>> db2 = rmm.DeviceBuffer.to_device(b"abc")
>>> db.copy_from_device(db2)
>>> hb = db.copy_to_host()
>>> print(hb)
array([97, 98, 99,  0,  0], dtype=uint8)

copy_from_host(self, ary, Stream stream=DEFAULT_STREAM)¶

Copy from a buffer on host to self

Parameters

arybytes-like buffer to copy from
streamCUDA stream to use for copying, default the default stream

Examples

>>> import rmm
>>> db = rmm.DeviceBuffer(size=10)
>>> hb = b"abcdef"
>>> db.copy_from_host(hb)
>>> hb = db.copy_to_host()
>>> print(hb)
array([97, 98, 99,  0,  0,  0,  0,  0,  0,  0], dtype=uint8)

copy_to_host(self, ary=None, Stream stream=DEFAULT_STREAM)¶

Copy from a DeviceBuffer to a buffer on host.

Parameters

arybytes-like buffer to write into
streamCUDA stream to use for copying, default the default stream

Examples

>>> import rmm
>>> db = rmm.DeviceBuffer.to_device(b"abc")
>>> hb = bytearray(db.nbytes)
>>> db.copy_to_host(hb)
>>> print(hb)
bytearray(b'abc')
>>> hb = db.copy_to_host()
>>> print(hb)
bytearray(b'abc')

nbytes¶: Gets the size of the buffer in bytes.

ptr¶: Gets a pointer to the underlying data.

reserve(self, size_t new_capacity, Stream stream=DEFAULT_STREAM) → void¶

resize(self, size_t new_size, Stream stream=DEFAULT_STREAM) → void¶

size¶: Gets the size of the buffer in bytes.

static to_device(const unsigned char[::1] b, Stream stream=DEFAULT_STREAM)¶: Calls to_device function on arguments provided.

tobytes(self, Stream stream=DEFAULT_STREAM) → bytes¶

exception rmm.RMMError(errcode, msg)¶: Bases: Exception

class rmm.RMMNumbaManager(*args, **kwargs)¶

Bases: HostOnlyCUDAMemoryManager

External Memory Management Plugin implementation for Numba. Provides on-device allocation only.

See https://numba.readthedocs.io/en/stable/cuda/external-memory.html for details of the interface being implemented here.

Attributes

interface_version: Returns an integer specifying the version of the EMM Plugin interface supported by the plugin implementation.

Methods

`defer_cleanup`()	Returns a context manager that disables cleanup of mapped or pinned host memory in the current context whilst it is active.
`get_ipc_handle`(memory)	Get an IPC handle for the MemoryPointer memory with offset modified by the RMM memory pool.
`get_memory_info`()	Returns `(free, total)` memory in bytes in the context.
`initialize`()	Perform any initialization required for the EMM plugin instance to be ready to use.
`memalloc`(size)	Allocate an on-device array from the RMM pool.
`memhostalloc`(size[, mapped, portable, wc])	Implements the allocation of pinned host memory.
`mempin`(owner, pointer, size[, mapped])	Implements the pinning of host memory.
`reset`()	Clears up all host memory (mapped and/or pinned) in the current context.

memallocmanaged

get_ipc_handle(memory)¶: Get an IPC handle for the MemoryPointer memory with offset modified by the RMM memory pool.

get_memory_info()¶

Returns (free, total) memory in bytes in the context. May raise NotImplementedError, if returning such information is not practical (e.g. for a pool allocator).

Returns: Memory info
Return type: MemoryInfo

initialize()¶

Perform any initialization required for the EMM plugin instance to be ready to use.

Returns: None

property interface_version¶: Returns an integer specifying the version of the EMM Plugin interface supported by the plugin implementation. Should always return 1 for implementations of this version of the specification.

memalloc(size)¶: Allocate an on-device array from the RMM pool.

rmm.disable_logging()¶: Disable logging if it was enabled previously using rmm.initialize() or rmm.enable_logging().

rmm.enable_logging(log_file_name=None)¶

Enable logging of run-time events for all devices.

Parameters

log_file_name: str, optional: Name of the log file. If not specified, the environment variable RMM_LOG_FILE is used. A ValueError is thrown if neither is available. A separate log file is produced for each device, and the suffix “.dev{id}” is automatically added to the log file name.

Notes

Note that if you use the environment variable CUDA_VISIBLE_DEVICES with logging enabled, the suffix may not be what you expect. For example, if you set CUDA_VISIBLE_DEVICES=1, the log file produced will still have suffix 0. Similarly, if you set CUDA_VISIBLE_DEVICES=1,0 and use devices 0 and 1, the log file with suffix 0 will correspond to the GPU with device ID 1. Use rmm.get_log_filenames() to get the log file names corresponding to each device.

rmm.get_log_filenames()¶

Returns the log filename (or None if not writing logs) for each device in use.

Examples

>>> import rmm
>>> rmm.reinitialize(devices=[0, 1], logging=True, log_file_name="rmm.log")
>>> rmm.get_log_filenames()
{0: '/home/user/workspace/rapids/rmm/python/rmm.dev0.log',
 1: '/home/user/workspace/rapids/rmm/python/rmm.dev1.log'}

rmm.get_versions()¶

rmm.is_initialized()¶: Returns True if RMM has been initialized, False otherwise.

rmm.register_reinitialize_hook(func, *args, **kwargs)¶

Add a function to the list of functions (“hooks”) that will be called before reinitialize().

A user or library may register hooks to perform any necessary cleanup before RMM is reinitialized. For example, a library with an internal cache of objects that use device memory allocated by RMM can register a hook to release those references before RMM is reinitialized, thus ensuring that the relevant device memory resource can be deallocated.

Hooks are called in the reverse order they are registered. This is useful, for example, when a library registers multiple hooks and needs them to run in a specific order for cleanup to be safe. Hooks cannot rely on being registered in a particular order relative to hooks registered by other packages, since that is determined by package import ordering.

Parameters

funccallable: Function to be called before reinitialize()
args, kwargs: Positional and keyword arguments to be passed to func

rmm.reinitialize(pool_allocator=False, managed_memory=False, initial_pool_size=None, maximum_pool_size=None, devices=0, logging=False, log_file_name=None)¶

Finalizes and then initializes RMM using the options passed. Using memory from a previous initialization of RMM is undefined behavior and should be avoided.

Parameters

pool_allocatorbool, default False: If True, use a pool allocation strategy which can greatly improve performance.
managed_memorybool, default False: If True, use managed memory for device memory allocation
initial_pool_sizeint, default None: When pool_allocator is True, this indicates the initial pool size in bytes. By default, 1/2 of the total GPU memory is used. When pool_allocator is False, this argument is ignored if provided.
maximum_pool_sizeint, default None: When pool_allocator is True, this indicates the maximum pool size in bytes. By default, the total available memory on the GPU is used. When pool_allocator is False, this argument is ignored if provided.
devicesint or List[int], default 0: GPU device IDs to register. By default registers only GPU 0.
loggingbool, default False: If True, enable run-time logging of all memory events (alloc, free, realloc). This has a significant performance impact.
log_file_namestr: Name of the log file. If not specified, the environment variable RMM_LOG_FILE is used. A ValueError is thrown if neither is available. A separate log file is produced for each device, and the suffix “.dev{id}” is automatically added to the log file name.

Notes

Note that if you use the environment variable CUDA_VISIBLE_DEVICES with logging enabled, the suffix may not be what you expect. For example, if you set CUDA_VISIBLE_DEVICES=1, the log file produced will still have suffix 0. Similarly, if you set CUDA_VISIBLE_DEVICES=1,0 and use devices 0 and 1, the log file with suffix 0 will correspond to the GPU with device ID 1. Use rmm.get_log_filenames() to get the log file names corresponding to each device.

rmm.rmm_cupy_allocator(nbytes)¶

A CuPy allocator that makes use of RMM.

Examples

>>> import rmm
>>> import cupy
>>> cupy.cuda.set_allocator(rmm.rmm_cupy_allocator)

rmm.unregister_reinitialize_hook(func)¶

Remove func from list of hooks that will be called before reinitialize().

If func was registered more than once, every instance of it will be removed from the list of hooks.

Memory Resources¶

class rmm.mr.BinningMemoryResource(DeviceMemoryResource upstream_mr, int8_t min_size_exponent=-1, int8_t max_size_exponent=-1)¶

Bases: UpstreamResourceAdaptor

Allocates memory from a set of specified “bin” sizes based on a specified allocation size.

If min_size_exponent and max_size_exponent are specified, initializes with one or more FixedSizeMemoryResource bins in the range [2**min_size_exponent, 2**max_size_exponent].

Call add_bin() to add additional bin allocators.

Parameters

upstream_mrDeviceMemoryResource: The memory resource to use for allocations larger than any of the bins.
min_size_exponentsize_t: The base-2 exponent of the minimum size FixedSizeMemoryResource bin to create.
max_size_exponentsize_t: The base-2 exponent of the maximum size FixedSizeMemoryResource bin to create.

Attributes

bin_mrs: Get the list of binned memory resources.
upstream_mr

Methods

`add_bin`(self, size_t allocation_size, ...)	Adds a bin of the specified maximum allocation size to this memory resource.
`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)
`get_upstream`(self)

add_bin(self, size_t allocation_size, DeviceMemoryResource bin_resource=None)¶

Adds a bin of the specified maximum allocation size to this memory resource. If specified, uses bin_resource for allocation for this bin. If not specified, creates and uses a FixedSizeMemoryResource for allocation for this bin.

Allocations smaller than allocation_size and larger than the next smaller bin size will use this fixed-size memory resource.

Parameters

allocation_sizesize_t: The maximum allocation size in bytes for the created bin
bin_resourceDeviceMemoryResource: The resource to use for this bin (optional)

bin_mrs¶: Get the list of binned memory resources.

class rmm.mr.CallbackMemoryResource(allocate_func, deallocate_func)¶

Bases: DeviceMemoryResource

A memory resource that uses the user-provided callables to do memory allocation and deallocation.

CallbackMemoryResource should really only be used for debugging memory issues, as there is a significant performance penalty associated with using a Python function for each memory allocation and deallocation.

Parameters

allocate_func: callable: The allocation function must accept a single integer argument, representing the number of bytes to allocate, and return an integer representing the pointer to the allocated memory.
deallocate_func: callable: The deallocation function must accept two arguments, an integer representing the pointer to the memory to free, and a second integer representing the number of bytes to free.

Examples

>>> import rmm
>>> base_mr = rmm.mr.CudaMemoryResource()
>>> def allocate_func(size):
...     print(f"Allocating {size} bytes")
...     return base_mr.allocate(size)
...
>>> def deallocate_func(ptr, size):
...     print(f"Deallocating {size} bytes")
...     return base_mr.deallocate(ptr, size)
...
>>> rmm.mr.set_current_device_resource(
    rmm.mr.CallbackMemoryResource(allocate_func, deallocate_func)
)
>>> dbuf = rmm.DeviceBuffer(size=256)
Allocating 256 bytes
>>> del dbuf
Deallocating 256 bytes

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)

class rmm.mr.CudaAsyncMemoryResource¶

Bases: DeviceMemoryResource

Memory resource that uses cudaMallocAsync/cudaFreeAsync for allocation/deallocation.

Parameters

initial_pool_sizeint, optional: Initial pool size in bytes. By default, half the available memory on the device is used.
release_threshold: int, optional: Release threshold in bytes. If the pool size grows beyond this value, unused memory held by the pool will be released at the next synchronization point.
enable_ipc: bool, optional: If True, enables export of POSIX file descriptor handles for the memory allocated by this resource so that it can be used with CUDA IPC.

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)

class rmm.mr.CudaMemoryResource¶

Bases: DeviceMemoryResource

Memory resource that uses cudaMalloc/cudaFree for allocation/deallocation.

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)

class rmm.mr.DeviceMemoryResource¶

Bases: object

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)

allocate(self, size_t nbytes)¶

deallocate(self, uintptr_t ptr, size_t nbytes)¶

class rmm.mr.FailureCallbackResourceAdaptor(DeviceMemoryResource upstream_mr, callback)¶

Bases: UpstreamResourceAdaptor

Memory resource that call callback when memory allocation fails.

Parameters

upstreamDeviceMemoryResource: The upstream memory resource.
callbackcallable: Function called when memory allocation fails.

Attributes

upstream_mr

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)
`get_upstream`(self)

class rmm.mr.FixedSizeMemoryResource(DeviceMemoryResource upstream_mr, size_t block_size=1048576, size_t blocks_to_preallocate=128)¶

Bases: UpstreamResourceAdaptor

Memory resource which allocates memory blocks of a single fixed size.

Parameters

upstream_mrDeviceMemoryResource: The DeviceMemoryResource from which to allocate blocks for the pool.
block_sizeint, optional: The size of blocks to allocate (default is 1MiB).
blocks_to_preallocateint, optional: The number of blocks to allocate to initialize the pool.

Notes

Supports only allocations of size smaller than the configured block_size.

Attributes

upstream_mr

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)
`get_upstream`(self)

class rmm.mr.LoggingResourceAdaptor(DeviceMemoryResource upstream_mr, log_file_name=None)¶

Bases: UpstreamResourceAdaptor

Memory resource that logs information about allocations/deallocations performed by an upstream memory resource.

Parameters

upstreamDeviceMemoryResource: The upstream memory resource.
log_file_namestr: Path to the file to which logs are written.

Attributes

upstream_mr

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)
`flush`(self)
`get_file_name`(self)
`get_upstream`(self)

flush(self)¶

get_file_name(self)¶

class rmm.mr.ManagedMemoryResource¶

Bases: DeviceMemoryResource

Memory resource that uses cudaMallocManaged/cudaFree for allocation/deallocation.

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)

class rmm.mr.PoolMemoryResource(DeviceMemoryResource upstream_mr, initial_pool_size=None, maximum_pool_size=None)¶

Bases: UpstreamResourceAdaptor

Coalescing best-fit suballocator which uses a pool of memory allocated from an upstream memory resource.

Parameters

upstream_mrDeviceMemoryResource: The DeviceMemoryResource from which to allocate blocks for the pool.
initial_pool_sizeint, optional: Initial pool size in bytes. By default, half the available memory on the device is used.
maximum_pool_sizeint, optional: Maximum size in bytes, that the pool can grow to.

Attributes

upstream_mr

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)
`get_upstream`(self)
`pool_size`(self)

pool_size(self)¶

class rmm.mr.StatisticsResourceAdaptor(DeviceMemoryResource upstream_mr)¶

Bases: UpstreamResourceAdaptor

Memory resource that tracks the current, peak and total allocations/deallocations performed by an upstream memory resource. Includes the ability to query these statistics at any time.

Parameters

upstreamDeviceMemoryResource: The upstream memory resource.

Attributes

allocation_counts: Gets the current, peak, and total allocated bytes and number of allocations.
upstream_mr

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)
`get_upstream`(self)

allocation_counts¶

Gets the current, peak, and total allocated bytes and number of allocations.

The dictionary keys are current_bytes, current_count, peak_bytes, peak_count, total_bytes, and total_count.

Returns:: dict: Dictionary containing allocation counts and bytes.

class rmm.mr.TrackingResourceAdaptor(DeviceMemoryResource upstream_mr, bool capture_stacks=False)¶

Bases: UpstreamResourceAdaptor

Memory resource that logs tracks allocations/deallocations performed by an upstream memory resource. Includes the ability to query all outstanding allocations with the stack trace, if desired.

Parameters

upstreamDeviceMemoryResource: The upstream memory resource.
capture_stacksbool: Whether or not to capture the stack trace with each allocation.

Attributes

upstream_mr

Methods

`allocate`(self, size_t nbytes)
`deallocate`(self, uintptr_t ptr, size_t nbytes)
`get_allocated_bytes`(self)	Query the number of bytes that have been allocated.
`get_outstanding_allocations_str`(self)	Returns a string containing information about the current outstanding allocations.
`get_upstream`(self)
`log_outstanding_allocations`(self)	Logs the output of get_outstanding_allocations_str to the current RMM log file if enabled.

get_allocated_bytes(self) → size_t¶: Query the number of bytes that have been allocated. Note that this can not be used to know how large of an allocation is possible due to both possible fragmentation and also internal page sizes and alignment that is not tracked by this allocator.

get_outstanding_allocations_str(self) → str¶: Returns a string containing information about the current outstanding allocations. For each allocation, the address, size and optional stack trace are shown.

log_outstanding_allocations(self)¶: Logs the output of get_outstanding_allocations_str to the current RMM log file if enabled.

rmm.mr.disable_logging()¶: Disable logging if it was enabled previously using rmm.initialize() or rmm.enable_logging().

rmm.mr.enable_logging(log_file_name=None)¶

Enable logging of run-time events for all devices.

Parameters

log_file_name: str, optional: Name of the log file. If not specified, the environment variable RMM_LOG_FILE is used. A ValueError is thrown if neither is available. A separate log file is produced for each device, and the suffix “.dev{id}” is automatically added to the log file name.

Notes

Note that if you use the environment variable CUDA_VISIBLE_DEVICES with logging enabled, the suffix may not be what you expect. For example, if you set CUDA_VISIBLE_DEVICES=1, the log file produced will still have suffix 0. Similarly, if you set CUDA_VISIBLE_DEVICES=1,0 and use devices 0 and 1, the log file with suffix 0 will correspond to the GPU with device ID 1. Use rmm.get_log_filenames() to get the log file names corresponding to each device.

rmm.mr.get_current_device_resource() → DeviceMemoryResource¶

Get the memory resource used for RMM device allocations on the current device.

If the returned memory resource is used when a different device is the active CUDA device, behavior is undefined.

rmm.mr.get_current_device_resource_type()¶: Get the memory resource type used for RMM device allocations on the current device.

rmm.mr.get_log_filenames()¶

Returns the log filename (or None if not writing logs) for each device in use.

Examples

>>> import rmm
>>> rmm.reinitialize(devices=[0, 1], logging=True, log_file_name="rmm.log")
>>> rmm.get_log_filenames()
{0: '/home/user/workspace/rapids/rmm/python/rmm.dev0.log',
 1: '/home/user/workspace/rapids/rmm/python/rmm.dev1.log'}

rmm.mr.get_per_device_resource(int device)¶

Get the default memory resource for the specified device.

If the returned memory resource is used when a different device is the active CUDA device, behavior is undefined.

Parameters

deviceint: The ID of the device for which to get the memory resource.

rmm.mr.get_per_device_resource_type(int device)¶

Get the memory resource type used for RMM device allocations on the specified device.

Parameters

deviceint: The device ID

rmm.mr.is_initialized()¶: Check whether RMM is initialized

rmm.mr.set_current_device_resource(DeviceMemoryResource mr)¶

Set the default memory resource for the current device.

Parameters

mrDeviceMemoryResource: The memory resource to set. Must have been created while the current device is the active CUDA device.

rmm.mr.set_per_device_resource(int device, DeviceMemoryResource mr)¶

Set the default memory resource for the specified device.

Parameters

deviceint: The ID of the device for which to get the memory resource.
mrDeviceMemoryResource: The memory resource to set. Must have been created while device was the active CUDA device.