iree.turbine.dynamo

class iree.turbine.dynamo.DeviceTensor(size, dtype, raw_data=None, requires_grad=False)

A Tensor accessing memory on a Turbine device.

property buffer_view: HalBufferView

cpu(memory_format=torch.preserve_format) → Tensor

Returns a copy of this object in CPU memory.

If this object is already in CPU memory, then no copy is performed and the original object is returned.

Parameters:

memory_format (torch.memory_format, optional) – the desired memory format of returned Tensor. Default: torch.preserve_format.

property device

Is the torch.device where this Tensor is.

static from_torch(input_tensor: Tensor)

class iree.turbine.dynamo.TurbineMode

Enables PyTorch tensor device= support for Tensor factory functions.

This can be used in a with block to dynamically scope enablement, or it can be enabled globally via the enable() function.

CACHED_IMPLEMENTATIONS: dict = {}

COMPUTE_METHODS = {<built-in method abs of type object>, <built-in method add of type object>, <built-in method mul of type object>, <built-in method sub of type object>}

IMPLEMENTATIONS: dict = {<built-in function _parse_to>: <function _parse_to>, <built-in method arange of type object>: <function device_factory.<locals>._inner_fn.<locals>._filter_impl>, <built-in method empty of type object>: <function device_factory.<locals>._inner_fn.<locals>._filter_impl>, <built-in method ones of type object>: <function device_factory.<locals>._inner_fn.<locals>._filter_impl>, <built-in method rand of type object>: <function device_factory.<locals>._inner_fn.<locals>._filter_impl>, <built-in method zeros of type object>: <function device_factory.<locals>._inner_fn.<locals>._filter_impl>, <method 'to' of 'torch._C.TensorBase' objects>: <function to>}

iree.turbine.dynamo.enable()

Enables PyTorch tensor device= support for Turbine permanently.