← all publishers

NovaSky-AI

@novasky-ai source repo

Report this profile Claim this profile

• 1 skill
• 0 followers
• 2 weeks ago last updated

1. ▌
   Parallelism Strategies · novasky-ai

   Operational guide for choosing and combining parallelism strategies (TP/PP/DP/CP/SP/EP) for the SkyRL Megatron backend. Use when sizing parallelism for a new model, debugging OOM/throughput on a given cluster topology, or extending an existing recipe to a new GPU count. Includes model-size sizing rules, hardware topology mapping, sequence-length thresholds, MoE-specific patterns, and pitfalls.
   1 install