[RFC] Add MindSpeed/Megatron Unit Test CI Pipeline

[UsernameFull]

Background

ROLL has added more MindSpeed/Megatron-related functionality, including mcore_adapter, roll/third_party/megatron, Megatron offload, model update, checkpoint handling,and Ascend NPU runtime patches.

These components depend on MindSpeed, Megatron-Core, torch-npu, and the Ascend runtime. They are difficult to validate reliably through local tests or the existing CPU-only unit test pipeline.

This RFC proposes adding a dedicated MindSpeed/Megatron unit test CI pipeline to detect regressions earlier in PRs and main branch updates.

Goals

Add a dedicated CI workflow for MindSpeed/Megatron unit tests that can validate:

• MindSpeed and Megatron-Core installation
• Basic NPU runtime availability
• ROLL Megatron adapter behavior
• Megatron offload-related unit tests
• NPU/Megatron runtime patch compatibility

Non-Goals

This RFC does not aim to:

• Add full end-to-end training jobs
• Add performance benchmark jobs
• Run heavy NPU tests for every unrelated PR
• Replace the existing CPU unit tests or NPU integration tests

Proposal

Add a new GitHub Actions workflow:

.github/workflows/ci-npu-mindspeed.yml