“container sees only the assigned accelerator while other devices stay hidden” addresses docker device allocation is broader than authorized workload resources, but the current hardware or infrastructure issue is container GPU access requires explicit runtime device assignment. The option “nvidia-smi inside the container lists the granted devices” is more directly tied to the required outcome.

“sample output is correct and GPU activity is attributable to the container” addresses container exit code alone does not prove GPU functional usage, but the current hardware or infrastructure issue is container GPU access requires explicit runtime device assignment. The option “nvidia-smi inside the container lists the granted devices” is more directly tied to the required outcome.

“host and container diagnostics show the intended split and a CUDA test passes” addresses gPU host driver responsibilities were confused with container user-space dependencies, but the current hardware or infrastructure issue is container GPU access requires explicit runtime device assignment. The option “nvidia-smi inside the container lists the granted devices” is more directly tied to the required outcome.

“set the intended `--mount` or bind mapping and file permissions” addresses container bind-mount path and application path do not agree, but the current hardware or infrastructure issue is container CUDA user-space and host driver compatibility were not checked. The option “choose a workload image compatible with the installed qualified driver” is more directly tied to the required outcome.

“run the container with supported `--gpus` allocation and NVIDIA runtime integration” addresses container GPU access requires explicit runtime device assignment, but the current hardware or infrastructure issue is container CUDA user-space and host driver compatibility were not checked. The option “choose a workload image compatible with the installed qualified driver” is more directly tied to the required outcome.

“restrict the container to the correct GPU UUID or device list” addresses docker device allocation is broader than authorized workload resources, but the current hardware or infrastructure issue is container CUDA user-space and host driver compatibility were not checked. The option “choose a workload image compatible with the installed qualified driver” is more directly tied to the required outcome.

“sample output is correct and GPU activity is attributable to the container” addresses container exit code alone does not prove GPU functional usage, but the current hardware or infrastructure issue is container CUDA user-space and host driver compatibility were not checked. The option “a sample CUDA kernel runs from the container without driver mismatch” is more directly tied to the required outcome.

“container process reads the dataset and writes a test artifact to the approved volume” addresses container bind-mount path and application path do not agree, but the current hardware or infrastructure issue is container CUDA user-space and host driver compatibility were not checked. The option “a sample CUDA kernel runs from the container without driver mismatch” is more directly tied to the required outcome.

“nvidia-smi inside the container lists the granted devices” addresses container GPU access requires explicit runtime device assignment, but the current hardware or infrastructure issue is container CUDA user-space and host driver compatibility were not checked. The option “a sample CUDA kernel runs from the container without driver mismatch” is more directly tied to the required outcome.

“run the container with supported `--gpus` allocation and NVIDIA runtime integration” addresses container GPU access requires explicit runtime device assignment, but the current hardware or infrastructure issue is docker device allocation is broader than authorized workload resources. The option “restrict the container to the correct GPU UUID or device list” is more directly tied to the required outcome.

“choose a workload image compatible with the installed qualified driver” addresses container CUDA user-space and host driver compatibility were not checked, but the current hardware or infrastructure issue is docker device allocation is broader than authorized workload resources. The option “restrict the container to the correct GPU UUID or device list” is more directly tied to the required outcome.

“test and align runtime compatibility across the target node classes” addresses heterogeneous host compatibility was not part of image acceptance, but the current hardware or infrastructure issue is docker device allocation is broader than authorized workload resources. The option “restrict the container to the correct GPU UUID or device list” is more directly tied to the required outcome.

“nvidia-smi inside the container lists the granted devices” addresses container GPU access requires explicit runtime device assignment, but the current hardware or infrastructure issue is docker device allocation is broader than authorized workload resources. The option “container sees only the assigned accelerator while other devices stay hidden” is more directly tied to the required outcome.

“sample output is correct and GPU activity is attributable to the container” addresses container exit code alone does not prove GPU functional usage, but the current hardware or infrastructure issue is docker device allocation is broader than authorized workload resources. The option “container sees only the assigned accelerator while other devices stay hidden” is more directly tied to the required outcome.

“a sample CUDA kernel runs from the container without driver mismatch” addresses container CUDA user-space and host driver compatibility were not checked, but the current hardware or infrastructure issue is docker device allocation is broader than authorized workload resources. The option “container sees only the assigned accelerator while other devices stay hidden” is more directly tied to the required outcome.

Useful for container CUDA user-space and host driver compatibility were not checked, but it does not remove the specific condition: container bind-mount path and application path do not agree.

“test and align runtime compatibility across the target node classes” addresses heterogeneous host compatibility was not part of image acceptance, but the current hardware or infrastructure issue is container bind-mount path and application path do not agree. The option “set the intended `--mount` or bind mapping and file permissions” is more directly tied to the required outcome.

“maintain qualified host driver and use compatible CUDA user-space packages in the image” addresses gPU host driver responsibilities were confused with container user-space dependencies, but the current hardware or infrastructure issue is container bind-mount path and application path do not agree. The option “set the intended `--mount` or bind mapping and file permissions” is more directly tied to the required outcome.

“sample output is correct and GPU activity is attributable to the container” addresses container exit code alone does not prove GPU functional usage, but the current hardware or infrastructure issue is container bind-mount path and application path do not agree. The option “container process reads the dataset and writes a test artifact to the approved volume” is more directly tied to the required outcome.

“a sample CUDA kernel runs from the container without driver mismatch” addresses container CUDA user-space and host driver compatibility were not checked, but the current hardware or infrastructure issue is container bind-mount path and application path do not agree. The option “container process reads the dataset and writes a test artifact to the approved volume” is more directly tied to the required outcome.

“host and container diagnostics show the intended split and a CUDA test passes” addresses gPU host driver responsibilities were confused with container user-space dependencies, but the current hardware or infrastructure issue is container bind-mount path and application path do not agree. The option “container process reads the dataset and writes a test artifact to the approved volume” is more directly tied to the required outcome.

“set the intended `--mount` or bind mapping and file permissions” addresses container bind-mount path and application path do not agree, but the current hardware or infrastructure issue is gPU host driver responsibilities were confused with container user-space dependencies. The option “maintain qualified host driver and use compatible CUDA user-space packages in the image” is more directly tied to the required outcome.

“run the container with supported `--gpus` allocation and NVIDIA runtime integration” addresses container GPU access requires explicit runtime device assignment, but the current hardware or infrastructure issue is gPU host driver responsibilities were confused with container user-space dependencies. The option “maintain qualified host driver and use compatible CUDA user-space packages in the image” is more directly tied to the required outcome.

“test and align runtime compatibility across the target node classes” addresses heterogeneous host compatibility was not part of image acceptance, but the current hardware or infrastructure issue is gPU host driver responsibilities were confused with container user-space dependencies. The option “maintain qualified host driver and use compatible CUDA user-space packages in the image” is more directly tied to the required outcome.

“sample output is correct and GPU activity is attributable to the container” addresses container exit code alone does not prove GPU functional usage, but the current hardware or infrastructure issue is gPU host driver responsibilities were confused with container user-space dependencies. The option “host and container diagnostics show the intended split and a CUDA test passes” is more directly tied to the required outcome.

“container process reads the dataset and writes a test artifact to the approved volume” addresses container bind-mount path and application path do not agree, but the current hardware or infrastructure issue is gPU host driver responsibilities were confused with container user-space dependencies. The option “host and container diagnostics show the intended split and a CUDA test passes” is more directly tied to the required outcome.

“nvidia-smi inside the container lists the granted devices” addresses container GPU access requires explicit runtime device assignment, but the current hardware or infrastructure issue is gPU host driver responsibilities were confused with container user-space dependencies. The option “host and container diagnostics show the intended split and a CUDA test passes” is more directly tied to the required outcome.

“test and align runtime compatibility across the target node classes” addresses heterogeneous host compatibility was not part of image acceptance, but the current hardware or infrastructure issue is container isolation was disabled beyond GPU workload needs. The option “grant only required NVIDIA devices and selected data mounts under a constrained security profile” is more directly tied to the required outcome.

“run the container with supported `--gpus` allocation and NVIDIA runtime integration” addresses container GPU access requires explicit runtime device assignment, but the current hardware or infrastructure issue is container isolation was disabled beyond GPU workload needs. The option “grant only required NVIDIA devices and selected data mounts under a constrained security profile” is more directly tied to the required outcome.

“set the intended `--mount` or bind mapping and file permissions” addresses container bind-mount path and application path do not agree, but the current hardware or infrastructure issue is container isolation was disabled beyond GPU workload needs. The option “grant only required NVIDIA devices and selected data mounts under a constrained security profile” is more directly tied to the required outcome.

“container process reads the dataset and writes a test artifact to the approved volume” addresses container bind-mount path and application path do not agree, but the current hardware or infrastructure issue is container isolation was disabled beyond GPU workload needs. The option “negative permission tests fail while approved GPU compute succeeds” is more directly tied to the required outcome.

“container sees only the assigned accelerator while other devices stay hidden” addresses docker device allocation is broader than authorized workload resources, but the current hardware or infrastructure issue is container isolation was disabled beyond GPU workload needs. The option “negative permission tests fail while approved GPU compute succeeds” is more directly tied to the required outcome.

“nvidia-smi inside the container lists the granted devices” addresses container GPU access requires explicit runtime device assignment, but the current hardware or infrastructure issue is container isolation was disabled beyond GPU workload needs. The option “negative permission tests fail while approved GPU compute succeeds” is more directly tied to the required outcome.