← All questions
System DesignVery Hard· Parallelism· Infrastructure

How do you plan the aspect ratio and parallelism for a...

Given the high-speed NVLink within nodes and slower inter-node InfiniBand, you should avoid tensor parallelism across nodes. Instead, use 8-way tensor parallelism within each 8-GPU node. Use pipeline parallelism to distribute the 80 layers across 4 nodes (20 layers per node). This results in a total of 32 H100 GPUs, which is a suitable allocation for a 70B model, keeping the aspect ratio within a validated range.

Code playground (Python)