NVIDIA has detailed the method for reproducing coaching scores from the MLPerf v5.0 benchmarks, particularly specializing in Llama 2 70B LoRA fine-tuning and Llama 3.1 405B pretraining. This initiative follows NVIDIA’s earlier announcement of attaining as much as 2.6x greater efficiency in MLPerf Coaching v5.0, as reported by Sukru Burc Eryilmaz on the NVIDIA weblog. The benchmarks are a part of MLPerf’s complete analysis suite geared toward measuring the efficiency of machine studying fashions.

Stipulations for Benchmarking

To run these benchmarks, particular {hardware} and software program necessities should be met. For Llama 2 70B LoRA, an NVIDIA DGX B200 or GB200 NVL72 system is critical, whereas the Llama 3.1 405B requires at the very least 4 GB200 NVL72 techniques related through InfiniBand. Moreover, substantial disk area is required: 2.5 TB for Llama 3.1 and 300 GB for LoRA fine-tuning.

Cluster and Surroundings Setup

NVIDIA makes use of a cluster setup managed by the NVIDIA Base Command Supervisor (BCM), which requires an setting primarily based on Slurm, Pyxis, and Enroot. Quick native storage configured in RAID0 is beneficial to reduce information bottlenecks. Networking ought to incorporate NVIDIA NVLink and InfiniBand for optimum efficiency.

Executing the Benchmarks

The execution course of includes a number of steps, beginning with constructing a Docker container and downloading obligatory datasets and checkpoints. The benchmarks are run utilizing SLURM, with a configuration file detailing hyperparameters and system settings. The method is designed to be versatile, permitting for changes primarily based on totally different system sizes and necessities.

Analyzing Benchmark Logs

Throughout the benchmarking course of, logs are generated that embody key MLPerf markers. These logs present insights into initialization, coaching progress, and ultimate accuracy. The final word purpose is to attain a goal analysis loss, which indicators the profitable completion of the benchmark.

For extra detailed directions, together with particular scripts and configuration examples, confer with the NVIDIA weblog.

Picture supply: Shutterstock