AI Architecture & Systems

Tensor Parallelism

5.7.1Column-Parallel and Row-Parallel Linear Layers#

5.7.2Partitioning Attention Heads and MLPs#

5.7.3Vocabulary, Embedding, and Loss Parallelism#

5.7.4All-Reduce/All-Gather/Reduce-Scatter in TP#

5.7.5KV Head Count, Replication, and Partitioning Constraints#

5.7.6Single-Node and Cross-Node Tensor Parallelism#