Software Development
Mcore Run On SLURM
The skill demonstrates how to launch distributed Megatron‑LM training jobs on a SLURM cluster, providing a minimal sbatch skeleton and environment‑variable setup for torch.distributed.run. It covers CUDA_DEVICE_MAX_CONNECTIONS rules for various hardware and parallelism modes, container conventions, monitoring, and per‑rank failure diagnosis, enabling data scientists and ML engineers to run large‑scale language model training efficiently.
- Publisher / source
- NVIDIA Skills (Verified)
- Upstream record
- NVIDIA/skills
- Version
- Version 2
- Published
- Publication review
- Approved for publication
- Content fingerprint
09d08fecad809510c20ab45b11ce01e8677d78a6f115dbbefc4340927ab8832c