For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo module
device_graph
This module provides functionality for building and replaying device graphs. A device graph captures a sequence of GPU operations (such as kernel launches, memory copies, and memsets) as a reusable graph that can be replayed at a lower overhead than re-enqueueing each operation individually. The main entry point is DeviceGraph.create(), which hands a DeviceGraphBuilder to a scoped callback.
Graph capture is currently implemented for CUDA and HIP devices only. Creating a graph on any other device, such as an Apple GPU or a CPU, raises.
comptime values
DeviceGraph
comptime DeviceGraph = LegacyDeviceGraph
A captured sequence of device operations that can be replayed.
DeviceGraphBuilder
comptime DeviceGraphBuilder[arena_origin: ImmOrigin] = LegacyDeviceGraphBuilder[arena_origin]
Records operations into a device graph.
Parameters
- arena_origin (
ImmOrigin):
DeviceGraphCache
comptime DeviceGraphCache = LegacyDeviceGraphCache
Reuses instantiated graphs across calls.
DeviceGraphInput
comptime DeviceGraphInput = LegacyDeviceGraphInput
A device graph input that contributes to the graph's cache key.
DeviceGraphInputKind
comptime DeviceGraphInputKind = LegacyDeviceGraphInputKind
How a graph input is stored and keyed.
DeviceGraphNode
comptime DeviceGraphNode[arena_origin: ImmOrigin] = LegacyDeviceGraphNode[arena_origin]
A node of a device graph under construction.
Parameters
- arena_origin (
ImmOrigin):
StableAddr
comptime StableAddr = LegacyStableAddr
A graph-lifetime slot that holds one device address.