Question 60
An Azure ML GPU training job fails because the image lacks the necessary CUDA runtime. Which part of the workload needs attention?
Explanation
Correct Answer
GPU hardware is insufficient when its software stack cannot load the required accelerator libraries for the model.
Incorrect Answers
Register and validate a patched new environment version while retaining the original historical reference. That operation would be relevant if: A security scan finds critical vulnerabilities in an old registered training environment. Its corresponding reason or verification is: Versioned environment upgrades address vulnerabilities while keeping the original dependencies identifiable for audit and past runs. However the problem in this question is: An Azure ML GPU training job fails because the image lacks the necessary CUDA runtime. The requirement instead calls for Select a compatible GPU-capable environment image with the required CUDA and driver compatibility; the question-specific check is: GPU hardware is insufficient when its software stack cannot load the required accelerator libraries for the model.
Align the scoring environment's framework and package versions with model compatibility requirements. A separate source case supports that approach when: A model inference container cannot import the exact framework version used during training, although the model file is present. The supporting rationale or observable result is: Model artifacts depend on their runtime libraries; deploying different incompatible packages can fail even with a valid model path. This question describes a different issue: An Azure ML GPU training job fails because the image lacks the necessary CUDA runtime. Address it through Select a compatible GPU-capable environment image with the required CUDA and driver compatibility, with the required justification or evidence: GPU hardware is insufficient when its software stack cannot load the required accelerator libraries for the model.
Pin dependency versions and register new environment versions when dependencies change. This is useful for the distinct issue: A team updates numpy in a shared environment after a successful model run; recreating the old job now produces different results. The relevant technical rationale or evidence for that issue is: Immutable environment versions keep historical builds reproducible instead of silently changing the packages behind earlier experiments. Here the reported issue is: An Azure ML GPU training job fails because the image lacks the necessary CUDA runtime. The relevant response is: Select a compatible GPU-capable environment image with the required CUDA and driver compatibility. The first result cannot verify this different correction: GPU hardware is insufficient when its software stack cannot load the required accelerator libraries for the model.