Skip to content

Bounded candidate log audit

The final 245,760-token C1 candidate remained healthy through preflight, capacity, quality, multimodal, and endurance workloads. The bounded managed log query found no traceback, process restart, corruption report, CUDA error, or out-of-memory event.

Expected non-fatal observations were retained:

  • the SM120 sparse-MLA glm53_nope CPB calibration rejected implausible fits on both ranks and used the runtime's C++ heuristic;
  • optional NCCL network, RMA, GIN, tuner, profiler, and environment plugins were absent, while the local shared-memory/direct collective path operated;
  • Triton emitted deprecation/data-race-analysis warnings during compilation;
  • Torch emitted deprecation warnings for its distributed all-gather API.

No kernel-tune artifact was adopted. The calibration warning is an optimization lead, not proof that a tune improves end-to-end performance; an exact image/model/topology default-versus-tuned A/B remains required.