Regression build failing with modern CUDA

[tgrogers]

Now that regressions are setup, we need to get them to pass with workloads of importance.
I have prioritized the following workloads to start:

gpu-app-collection/src/Makefile

Line 7 in fec0850

all: GPU_Microbenchmark microbench rodinia_2.0-ft mnist_cudnn cutlass mlperf_inference mlperf_training rodinia-3.1 pannotia proxy-apps heterosync ispass-2009 lonestargpu-2.0 polybench custom_apps

Right now, I think all of the following are failing:

mnist_cudnn cutlass mlperf_inference mlperf_training

Docker makes it much easier to reproduce the errors - what the regressions system does is listed here:

https://github.com/accel-sim/gpu-app-collection/blob/dev/.github/workflows/test-build.yml

It starts with the docker image

nvidia/cuda:12.6.3-cudnn-devel-ubuntu24.04

gpu-app-collection/.github/workflows/test-build.yml

Line 27 in fec0850

apt-get update

You can test this on any machine that has docker installed. You don't even need a GPU.