nvidia-nlp-operator

Compile cuda_gemm.cpp directly with nvcc -O3 -lcublas, then run test_gemm.py against the gemm_f16.csv and gemm_f32.csv parameter files in baseline mode. The collector script converts results into a Result Contract 2.0 JSON with latency statistics.

Resolve the actual cuDNN library path from the nvidia.cudnn Python package and link with -l:<actual filename> plus an rpath, since the image may only ship versioned libcudnn.so.* files. Never fall back to an old prebuilt binary after a link failure.

It supports five categories: GEMM (FP16/FP32), Conv2d (FP16/FP32), over 100 long-tail operators from LongTail-Bench, Transformer Encoder/Decoder Block inference latency, and NCCL or OSU-based communication operators requiring multi-node environments.

It uses FP32 inference mode with eval() and torch.inference_mode(), measuring average latency over 1000 iterations after 20 warmup iterations with CUDA synchronization. The project's original test.py is not used because it runs in train mode without proper timing boundaries.

This false positive happens when the top-level CMake fails (it requires cuDNN that GEMM images may lack) and stale CSV baselines are reused. The Skill prevents this by requiring direct nvcc compilation and verifying CSV modification timestamps before collection.

Communication operator tests require a multi-node, multi-GPU environment with SSH passwordless login, InfiniBand or high-speed interconnects, and OpenMPI/UCX installed. Single-GPU machines can run GEMM, Conv2d, long-tail, and Transformer Block tests only.