perf-workload-profiling

Measure per-iteration latency and throughput for a GPU training loop by instrumenting the code with manual timing and CUDA event benchmarking, ensuring proper warmup handling and pre-allocation of timing resources to capture accurate data-load metrics.

Profiling standalone CUDA kernels requires CUDA event-based timing to capture per-iteration statistics, utilizing proper warmup handling and avoiding intra-iteration synchronization to ensure precise kernel benchmarking measurements.

Add NVTX annotations to label profiler timelines by instrumenting GPU workloads with NVTX labeling, providing clearer visualization of training loops and standalone kernel operations within the profiler timeline.

GPU workload timing with CUDA events requires avoiding intra-iteration synchronization to ensure accurate measurements, along with proper warmup handling and pre-allocation of timing resources for reliable benchmarking.

Benchmark TensorRT-LLM workloads by applying manual timing for training loops and CUDA event-based timing for standalone kernels, capturing per-iteration latency, throughput, and data loading metrics with NVTX profiler timeline support.