nemo-automodel-launcher-config

To configure NeMo AutoModel launches, you create YAML launcher blocks for Slurm or SkyPilot to define distributed training parameters across on-prem HPC clusters and cloud runtimes. This ensures consistent environment setup and multi-node coordination.

Calculating multi-node world size for distributed training is handled automatically through the Skill's launcher configuration, which properly manages WORLD_SIZE, MASTER_ADDR, and MASTER_PORT environment variables for both Slurm and SkyPilot runs.

Yes, you can use SkyPilot to launch NeMo AutoModel jobs on cloud runtimes by creating a minimal YAML snippet under the skypilot: block. This provides a unified approach to starting cloud-based distributed training with proper environment setup.

Yes, the configuration supports interactive launches using torchrun with multi-node coordination. You can start interactive NeMo AutoModel sessions that ensure consistent environment setup and distributed launch parameters across your compute resources.

To kick off a two-node Slurm job for NeMo AutoModel training, you configure a slurm: YAML block with the appropriate launcher parameters. The configuration handles multi-node coordination, including WORLD_SIZE calculation and MASTER_ADDR setup for distributed training.