Refactor private attributes in Pydantic models to use PrivateAttr

Description

In Pydantic v2, private attributes (those starting with underscore) should be declared using PrivateAttr() for stability and to follow the official API. While bare annotations like _attr: Type | None = None may work incidentally, they are not part of the guaranteed Pydantic API and can break with minor version updates.

Background

This issue was identified during review of PR #799:

• PR: https://github.com/NVIDIA/cloudai/pull/799
• Comment: https://github.com/NVIDIA/cloudai/pull/799#discussion_r2787752019

Affected Files

The following files contain Pydantic models with private attributes that should be refactored to use PrivateAttr:

Workloads (19 files)

1. src/cloudai/workloads/vllm/vllm.py - _docker_image, _hf_model
2. src/cloudai/workloads/ucc_test/ucc.py - _docker_image
3. src/cloudai/workloads/slurm_container/slurm_container.py - _docker_image
4. src/cloudai/workloads/nixl_perftest/nixl_perftest.py - _docker_image
5. src/cloudai/workloads/osu_bench/osu_bench.py - _osu_image
6. src/cloudai/workloads/triton_inference/triton_inference.py - _server_image, _client_image
7. src/cloudai/workloads/nemo_run/nemo_run.py - _docker_image
8. src/cloudai/workloads/nemo_launcher/nemo_launcher.py - _docker_image, _python_executable
9. src/cloudai/workloads/nixl_kvbench/nixl_kvbench.py - _docker_image
10. src/cloudai/workloads/nixl_bench/nixl_bench.py - _nixl_image
11. src/cloudai/workloads/nccl_test/nccl.py - _docker_image
12. src/cloudai/workloads/jax_toolbox/nemotron.py - _docker_image
13. src/cloudai/workloads/jax_toolbox/grok.py - _docker_image
14. src/cloudai/workloads/jax_toolbox/gpt.py - _docker_image
15. src/cloudai/workloads/megatron_run/megatron_run.py - _docker_image
16. src/cloudai/workloads/deepep/deepep.py - _docker_image
17. src/cloudai/workloads/chakra_replay/chakra_replay.py - _docker_image
18. src/cloudai/workloads/ddlb/ddlb.py - _docker_image
19. src/cloudai/workloads/megatron_bridge/megatron_bridge.py - _docker_image, _python_executable, _megatron_bridge_repo
20. src/cloudai/workloads/ai_dynamo/ai_dynamo.py - _docker_image, _hf_model

Systems (1 file)

21. src/cloudai/systems/kubernetes/kubernetes_system.py - _core_v1, _batch_v1, _custom_objects_api, _genai_perf_completed

Recommended Changes

For each affected file:

1. Add PrivateAttr to the pydantic imports:

   from pydantic import Field, PrivateAttr
   

2. Replace bare annotations with PrivateAttr:

   # Before
   _docker_image: DockerImage | None = None
   
   # After
   _docker_image: DockerImage | None = PrivateAttr(default=None)
   

Good Example

src/cloudai/systems/runai/runai_system.py already follows this pattern correctly:

from pydantic import Field, PrivateAttr

class RunAISystem(KubernetesSystem):
    _api_client: Optional[RunAIRestClient] = PrivateAttr(default=None)

Note

src/cloudai/_core/installables.py uses dataclass field(), not Pydantic, so it doesn't need to be changed.

References

• Pydantic v2 Private Attributes Documentation: https://docs.pydantic.dev/latest/concepts/models/#private-model-attributes

Read src/cloudai/systems/runai/runai_system.py for the existing PrivateAttr pattern, then inspect the listed workload and Kubernetes model files. Replace each affected bare private-attribute annotation with PrivateAttr while preserving its default and type. Done means all listed Pydantic attributes use the official pattern and src/cloudai/_core/installables.py remains unchanged.