open-infra-skills/infra-skills

This skill provides the technical framework for AI infrastructure and ML performance engineers to profile, benchmark, and debug training workloads on Moore Threads MUSA hardware. It focuses on increasing throughput while preserving numerical behavior during the transition from CUDA to MUSA.

Engineers can address distributed hangs, memory pressure, and kernel launch overhead using tools like the Moore Perf System and Moore Perf Compute. The guidance covers critical optimizations including FlashAttention, torch.compile, and FSDP2 to improve hardware utilization.