mscclpp.default_algos
- mscclpp.default_algos.allgather_multi_nodes(spec)
Build a hierarchical AllGather across nodes and local GPUs.
- Return type:
- Parameters:
spec (AlgoSpec)
- mscclpp.default_algos.allreduce_multi_nodes(spec, thread_block_group_size)
Implements a multi-node AllReduce using a hierarchical approach: 1. Intra-node allreduce 2. Inter-node exchange (exchange reduced data between nodes) 3. Intra-node allreduce
- Return type:
- Parameters:
- mscclpp.default_algos.reducescatter_multi_nodes(spec, thread_block_group_size=1)
Build a hierarchical ReduceScatter across nodes and local GPUs.
- Return type:
- Parameters:
Modules
|
Build a hierarchical AllGather across nodes and local GPUs. |
|
Implements a multi-node AllReduce using a hierarchical approach: 1. |
|
Build a hierarchical ReduceScatter across nodes and local GPUs. |