Is this problem already fixed? It seems still remain (Nsight profile and cudaGraphInstantiateFlagDeviceLaunch))

[sakaia]

The following problem is already fixed or still planned to fix? (It seems 1 year passed already)
In my environment, the problem still keeps.
Is there any method to know the notice of this fixation?

Nsight profile and cudaGraphInstantiateFlagDeviceLaunch

Nsight profile and cudaGraphInstantiateFlagDeviceLaunch

I am trying to profile an application that make use of CUDA Graph with self tail launch (e.g. I have a persistent graph that loops forever. I instantiate the graph with cudaGraphInstantiate(&graph_exec, run_graph, cudaGraphInstantiateFlagDeviceLaunch); And then relaunch it inside a kernel: __global__ void looper(executor_data *data) { auto g = cudaGetCurrentGraphExec(); if (g) { int ret = cudaGraphLaunch(g, cudaStreamGraphTailLaunch); } } When I run my application in the profiler I can…

[hwilper]

@liuyis is this fixed in the current version of CUDA graph support?

[liuyis]

The devce graph trace is now supported at the graph-level granularity. However the original post seems to be about a crashing issue, which I’m not sure CUPTI has been tracking it.

@sakaia Did you try the latest Nsys 2024.7 release from Nsight Systems - Get Started | NVIDIA Developer? If the issue still persists, is there a simple reproducer that we can try to repro on our side?