Crate monarch_rdma

Modules§

device_selection: This module provides functionality to automatically pair compute devices with the best available RDMA NICs based on PCI topology distance.

DoorBell
Gid
IbvWc: Wrapper around ibv_wc (ibverbs work completion).
IbverbsConfig: Represents ibverbs specific configurations.
RdmaBuffer
RdmaDevice: Represents an RDMA device in the system.
RdmaDomain: Represents a domain for RDMA operations, encapsulating the necessary resources for establishing and managing RDMA connections.
RdmaManagerActor
RdmaMemoryRegionView: Represents a view of a memory region that can be registered with an RDMA device.
RdmaPort
RdmaQpInfo: Contains information needed to establish an RDMA queue pair with a remote endpoint.
RdmaQueuePair: Represents an RDMA Queue Pair (QP) that enables communication between two endpoints.

PollTarget: Enum to specify which completion queue to poll
RdmaManagerMessage: Represents a reference to a remote RDMA buffer that can be accessed via RDMA operations. This struct encapsulates all the information needed to identify and access a memory region on a remote host using RDMA.
RdmaOperation: Enum representing the common RDMA operations.
RdmaQpType: Queue pair type for RDMA operations.

format_gid: Formats a GID (Global Identifier) into a human-readable string.
get_all_devices: Retrieves information about all available RDMA devices in the system.
get_link_layer_str: Converts the given link layer type to a human-readable string.
get_port_phy_state_str: Converts the given physical state to a human-readable string.
get_port_state_str: Converts the given port state to a human-readable string.
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
ibverbs_supported: Checks if ibverbs devices can be retrieved successfully.
is_cuda_available: Safely checks if CUDA is available on the system.
mlx5dv_supported: Checks if mlx5dv (Mellanox device-specific verbs extension) is supported.
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.
pt_cuda_allocator_compatibility: Check if PyTorch CUDA caching allocator has expandable segments enabled.
rdma_supported: Checks if RDMA is fully supported on this system.
resolve_qp_type: Converts RdmaQpType to the corresponding integer enum value in rdmaxcel_sys.
validate_execution_context: Utility to validate execution context.