chore: bump transformer-engine to 2.14.1

[zpqiu]

Summary

• Bump TransformerEngine to the v2.14.1 release for both automodel and mcore extras.
• Keep the CUDA 12 core extra explicit while resolving the git tag through uv.lock.
• Update uv dependency metadata for the new TE source version.

Related Issues

#2320

Additional Info

The current TE version pinned by RL has a checkpoint resume problem with FusedAdam.

That snapshot includes the FusedAdam DTensor/local-state change, but misses the later DCP compatibility fix. As a result, FusedAdam optimizer states for degenerate sharded params such as shared_expert_gate.weight [1, 2048] were saved with rank-local metadata like (0, 2048) on most ranks and (1, 2048) on a few ranks. Model weights metadata were fine; only optimizer state metadata was corrupted.

The fix is to bump Transformer Engine to a version containing a4a073bbb7c625727b006adacc2a19f4a80cdb61, e.g. TE v2.14 commit f031cf87bd054c7558b887df7bed93975456667f.

Related PR: NVIDIA/TransformerEngine#2795

[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.

[zpqiu]

/ok to test 0d05f0a

[zpqiu]

Hi @terrykong, Sharon has run the nightly tests twice. Here is the CI link: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/50603458. Some tests failed, but they don’t seem to be related to this bump PR. The analysis report generated by CC is below. Do these results look acceptable to you?

[terrykong]

Yea, i think those failures seem unrelated, this PR lgtm