cb.fabric-performance-loop

Benchmark Fabric performance by running parity-guarded benchmarks across modes, batch sizes, and inner steps to verify CUDA kernel ownership and scaling behaviors. This process requires current Fabric artifacts, parity gates, and a live progress doc to enforce reproducibility and collect evidence.

Detect performance regressions by capturing and curating performance evidence during profiling tasks. This structured approach measures throughput and resource usage across architectures and backends, ensuring scaling behaves as expected through parity-guarded benchmarking practices.

Yes, a live progress document is required to run Fabric profiling tasks. It establishes a performance plan and logs benchmark results, enforcing checks for reproducibility, evidence collection, and safe benchmarking practices throughout the optimization workflow.

Yes, you can measure throughput, latency, and resource usage across multiple backends and architectures. It applies parity-guarded benchmarking to validate Fabric behavior, ensuring reproducible performance characteristics during scaling and optimization tasks.

Parity gates are required during Fabric benchmarking to validate behavior consistency across modes, batch sizes, and inner steps. They ensure safe benchmarking practices by enforcing reproducibility and accurate evidence collection before applying optimizations or detecting regressions.