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

ccl

Provides Mojo FFI bindings for NCCL (NVIDIA) and RCCL (AMD) collective operations.

Selects and loads the correct vendor library at runtime: librccl.so on AMD systems and libnccl.so on NVIDIA systems. Exposes allreduce, allgather, and broadcast collectives, along with communicator initialization helpers and availability probes.

comptime values

CCL_LIBRARY

comptime CCL_LIBRARY = _Global[StringSpan("CCL_LIBRARY"), _init_ccl_dylib]

NCCL_LIBRARY_PATHS

comptime NCCL_LIBRARY_PATHS = List(Path("libnccl.so"), Path("libnccl.so.2"), Path("/usr/lib/x86_64-linux-gnu/libnccl.so"), Path("/usr/lib/x86_64-linux-gnu/libnccl.so.2"), __list_literal__=NoneType(None))

ncclComm_t

comptime ncclComm_t = Optional[Pointer[NoneType, MutUntrackedOrigin]]

RCCL_LIBRARY_PATHS

comptime RCCL_LIBRARY_PATHS = List(Path("librccl.so"), Path("librccl.so.1"), Path("/opt/rocm/lib/librccl.so"), Path("/opt/rocm/lib/librccl.so.1"), __list_literal__=NoneType(None))

Structs

• ​Communicators: Holds NCCL/RCCL communicator handles for a fixed set of GPUs.
• ​ncclDataType_t: Data-type selector for NCCL/RCCL collective calls.
• ​ncclRedOp_t: Reduction operation selector for NCCL/RCCL collective calls.
• ​ncclResult_t: Status code returned by NCCL/RCCL collective operations.

Functions

• ​allgather: Performs an allgather across all GPUs via the vendor CCL library.
• ​allreduce: Per-GPU allreduce for use in multi-threaded contexts.
• ​broadcast: Per-GPU broadcast for use in multi-threaded contexts.
• ​group: Returns a context manager that groups NCCL/RCCL collective calls.
• ​init_comms: Pre-initialize NCCL/RCCL communicators.
• ​is_allgather_available: Reports whether the vendor CCL allgather symbol is loadable at runtime.
• ​is_allreduce_available: Reports whether the vendor CCL allreduce symbol is loadable at runtime.
• ​is_broadcast_available: Reports whether the vendor CCL broadcast symbol is loadable at runtime.
• ​ncclCommInitAll: Initializes NCCL/RCCL communicators for a set of GPUs.
• ​wait_for_comms: Spin-wait until communicators for ngpus have been initialized.