fix(ci): AUT-1100 keep verl and MCore on main

[svcnemo-autobot]

Background

Run 30245218834 failed because verl could not import a removed Megatron-LM API and skipped Megatron backend registration.

What changed

• Keep verl and Megatron-LM on main.
• Make Energon optional for non-Energon Bridge use.
• Pin Transformers to 5.12.1 in verl suites.
• Respect Megatron-LM’s Python metadata.

Details

Energon runtime code is imported lazily, and its builder tests patch that runtime module. Transformers 5.3 lacked Exaone4_5_ForConditionalGeneration; 5.12.1 is the immutable version installed before the successful local suites. The MCore API fix is NVIDIA/Megatron-LM#6065.

Tested

• Full pre-commit gate, shell syntax, and git diff --check
• Energon builder + module-layout tests: 19 passed
• Clean Python 3.12 pip dry run for Megatron-LM main
• Exact 8-GPU Qwen3, DeepSeek, and MoE suites completed successfully with Transformers 5.12.1 installed before their dependency resolves

[svcnemo-autobot]

/ok to test 7af40e1

[svcnemo-autobot]

/ok to test d9f4f7f

[svcnemo-autobot]

/ok to test d5b4d17

[svcnemo-autobot]

/ok to test 61c2665

[ko3n1g]

@svcnemo-autobot why do we need --ignore-requires-python?

[svcnemo-autobot]

💬 --ignore-requires-python was introduced while testing an older pinned MCore commit. It is not needed now: Megatron-LM main requires Python ≥3.12, and the verl image runs Python 3.12. The flag only bypasses valid package metadata, so I’m removing it from all three scripts and validating in ci-events-agent-write-42f06dbb; parking on workload_phase{Running}.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[svcnemo-autobot]

/ok to test 3134459

[svcnemo-autobot]

/ok to test 66f1c27

[svcnemo-autobot]

/ok to test 4650406