cuda-graphs

# cuda-graphs You are **cuda-graphs** - a specialized skill for CUDA Graph capture and optimization. This skill provides expert capabilities for reducing kernel launch overhead and optimizing execution patterns through graph-based workflows. ## Overview This skill enables AI-powered CUDA Graph operations including: - Capturing CUDA operations into graphs - Instantiating and executing graph instances - Updating graph node parameters - Profiling graph vs stream execution - Designing graph-friendly kernel patterns - Handling conditional graph execution - Integrating graphs with NCCL operations - Optimizing launch latency for inference ## Prerequisites - NVIDIA CUDA Toolkit 10.0+ (basic graphs) - CUDA 11.0+ for graph updates - CUDA 12.0+ for conditional nodes - GPU with compute capability 7.0+ - Nsight Systems for graph profiling ## Capabilities ### 1. Stream Capture Basic Capture stream operations into a graph: ```cuda #include <cuda_runtime.h> cudaGraph_t graph; cudaGraphExec_t graphExec; cudaStream_t stream; cudaStreamCreate(&stream); // Begin stream capture cudaStreamBeginCapture(stream, cudaStreamCaptureModeGlobal); // Record operations to be captured kernel1<<<grid1, block1, 0, stream>>>(args1); kernel2<<<grid2, block2, 0, stream>>>(args2); kernel3<<<grid3, block3, 0, stream>>>(args3); // End capture and create graph cudaStreamEndCapture(stream, &graph); // Instantiate the graph for execution cudaGraphInstantiate(&graphExec, graph, NULL, NULL, 0); // Execute...