Expand description
Compute server module.
Structs§
- Buffer
Binding - A buffer binding represents a Handle that is bound to managed memory.
- Capture
Status - A read-only view of a device’s captures.
- Communication
Id - An ID unique to any unordered combination of devices.
- Copy
Descriptor - A binding with shape and stride info for non-contiguous reading
- CubeDim
- The number of units across all 3 axis totalling to the number of working units in a cube.
- Device
Captures - The graph captures under way on one device.
- Handle
- Server handle containing the memory handle.
- Kernel
Arguments - Arguments to execute a kernel.
- Memory
Layout - An allocation with associated strides. Strides depend on tensor layout.
- Memory
Layout Descriptor - Descriptor for a new tensor allocation
- Metadata
Binding Info - Binding of a set of scalars of the same type to execute a kernel.
- Profiling
Token - Profiling identification so that the server can support recursive and overlapping profilings.
- Reason
- A reason for an error.
- Server
Utilities - Contains many different types that are useful for server implementations and compute clients.
- Stream
Capture State - Where a stream sits in the graph-capture lifecycle, and the only thing
allowed to move it. Capture is a strict
NoCapture → Prepare → Capture → NoCaptureprogression, driven byprepare,beginandend; each rejects an out-of-order call, so a capture can never start unprepared and two captures can never overlap on one stream. - Tensor
MapBinding - A tensor map used with TMA ops
- Tensor
MapMeta TensorMapmetadata for the opaque proxy used in TMA copies- Weak
Buffer Binding - A
BufferBindingthat does not keep its memory reserved: what a record of a buffer holds when holding the buffer would change what its pool may reuse.
Enums§
- Capture
End - What
StreamCaptureState::endfound when it closed the window: the caller’s own capture, or one belonging to a logical stream that never came back to close it. - Collective
- A collective operation between the devices of one runtime.
- Cube
Count - Specifieds the number of cubes to be dispatched for a kernel.
- Cube
Count Selection - Defines how to select cube count based on the number of cubes required.
- IoError
- Error returned from
create/read/writefunctions. Due to async execution not all errors are able to be caught, so some IO errors will still panic. - Kernel
Resource - A resource passed to a kernel function
- Launch
Error - Kernel Launch Errors.
- Memory
Layout Strategy - Type of allocation, either contiguous or optimized (row-aligned when possible)
- Profile
Error - An error during profiling.
- Reduce
Operation - Different reduce operations.
- Resource
Limit Error - Resource limit errors.
- Server
Error - Error that can happen asynchronously while executing registered kernels.
Traits§
- Memory
Layout Policy - Defines how the memory layout is determined.
- Server
- The compute server is responsible for handling resources and computations over resources.
- Server
Communication - Defines functions for optimized data transfer between servers, supporting custom communication mechanisms such as peer-to-peer communication or specialized implementations.
- Server
Storage - The storage a server allocates from, and the native resources it hands
out. Kept off
Serverso that trait is object-safe: a resource’s type is the backend’s own, and only a caller that names the backend can receive one.