Re-exports§
pub use backend::ibverbs::mlx_domain::CudaSegmentScanner;pub use backend::ibverbs::mlx_domain::ScannedSegment;pub use backend::ibverbs::mlx_domain::register_cuda_segment_scanner;pub use rdmaxcel_sys;pub use backend::ibverbs::primitives::*;
Modules§
- backend
- RDMA backend implementations.
- config
- RDMA configuration attributes.
- device_
selection - PCI topology parsing and device discovery utilities for RDMA device selection.
- efa
- EFA (Elastic Fabric Adapter) specific RDMA operations.
- local_
memory - Local memory abstractions for RDMA operations.
Macros§
- cu_
check - register_
ibv_ device_ impl - Submits an
inventoryentry for a concreteIbvDeviceImpl.
Structs§
- Ensure
Rdma Manager - Caller -> owner request to ensure a proc mesh’s per-proc managers are ready.
- EntryId
- Owner-local handle for a
ProcMeshRef. - GetTcp
Actor Ref - Serializable query for resolving the
TcpManagerActorref from a remoteRdmaManagerActor. - Rdma
Action - A batch of RDMA operations submitted as a single unit.
- Rdma
Manager Actor - Rdma
Manager Owner Actor - Rdma
Manager Ready - Owner -> manager-mesh readiness cast, delivered to every per-proc
RdmaManagerActor. Carries the owner ref so each rank can postReadyAckback after its owninit()(RMR-1). - RdmaOp
- A single RDMA operation to be submitted to a backend.
- Rdma
Remote Buffer - Lightweight handle representing a registered RDMA buffer.
- Ready
Ack - Per-proc manager -> owner readiness acknowledgement, one per rank.
- Release
Buffer - Serializable release message for wire transport.
Enums§
- Rdma
Init Error - A typed initialization failure returned by the RDMA manager owner.
- Rdma
Manager Message - Local-only messages for the
RdmaManagerActor. - Rdma
OpType - Type of RDMA operation.
- Rdma
Transport Level - Transport level, ordered slowest to fastest.
Constants§
- RDMA_
MANAGER_ OWNER_ ACTOR_ NAME - Static client-root service name for the RDMA manager owner.
Traits§
- Ensure
Rdma Manager Client - The custom client trait for this message type.
- Ensure
Rdma Manager Handler - The custom handler trait for this message type.
- GetTcp
Actor RefClient - The custom client trait for this message type.
- GetTcp
Actor RefHandler - The custom handler trait for this message type.
- Rdma
Manager Message Client - The custom client trait for this message type.
- Rdma
Manager Message Handler - The custom handler trait for this message type.
- Rdma
Manager Ready Client - The custom client trait for this message type.
- Rdma
Manager Ready Handler - The custom handler trait for this message type.
- Ready
AckClient - The custom client trait for this message type.
- Ready
AckHandler - The custom handler trait for this message type.
- Release
Buffer Client - The custom client trait for this message type.
- Release
Buffer Handler - The custom handler trait for this message type.
Functions§
- get_
rdmaxcel_ error_ message - Helper function to get detailed error messages from RDMAXCEL error codes
- get_
registered_ cuda_ segments - Get all segments that have been registered with MRs for the given PD.
- is_
cuda_ available - Safely checks if CUDA is available on the system.
- print_
device_ info - Print comprehensive RDMA device information for debugging (always prints).
- print_
device_ info_ if_ debug_ enabled - Print comprehensive RDMA device information for debugging. Controlled by MONARCH_DEBUG_RDMA environment variable.
- rdma_
supported - Whether any RDMA backend is available on this system.
- register_
segment_ scanner - Register a segment scanner callback.
- validate_
execution_ context - Utility to validate execution context.
Type Aliases§
- Segment
Scanner Fn - Segment scanner callback type alias for convenience.