perf-moe-hardware-configs
Pass
Audited by Gen Agent Trust Hub on May 19, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill serves as a static documentation and reference guide for hardware-specific MoE training configurations, covering metrics like throughput (TFLOPS) and Model Flops Utilization (MFU).
- [SAFE]: No executable scripts, binary files, or automated command execution patterns were identified in the provided files.
- [SAFE]: The environment variables mentioned (e.g.,
CUDA_DEVICE_MAX_CONNECTIONS,PYTORCH_CUDA_ALLOC_CONF) are standard configuration parameters for optimizing PyTorch and NCCL performance on NVIDIA GPUs. - [SAFE]: All content is presented as technical guidelines for model parallelism (TP, EP, PP) and recompute strategies, posing no risk of prompt injection or data exfiltration.
Audit Metadata