chore(deps): AUT-2529 bump Megatron-LM main to 8c815a4c

[svcnemo-autobot]

Background

Nightly Megatron-Core bump for the main lane. This moves the MCore main pin from e30d863a to 8c815a4c (8c815a4c98725726d27358847d291288b736f6bd), on top of Bridge main at b1bb0a97. .dev.commit is unchanged (4a4de865).

What changed

• .main.commit and the 3rdparty/Megatron-LM gitlink now point to 8c815a4c.
• uv.lock was regenerated with uv 0.7.2 (uv lock --upgrade, matching the docker/Dockerfile.ci pin). The result is 108 lines added and 108 removed, with nine transitive version bumps: ast-serialize, bracex, google-auth, openai, pre-commit, soupsieve, truststore, virtualenv and wrapt. pyproject.toml is unchanged, and uv lock --check passes.

Details

• Between the two pins, MCore made no changes to the attention, TransformerEngine-extension, vision, hybrid or SSM code paths.
• The lock delta contains no GPU or compiled packages.
• Test selection followed the impact of the actual diff: Core unit shards plus the converter and VLM-inference functional launchers.

Tested

Local H100 runs on this exact head, using a docker/Dockerfile.ci validation image built from it (INSTALL_DIFFUSION_DEPS=true was set for validation only):

• Unit tests, Core shard 0/3: 4044 passed.
• Unit tests, Core shard 1/3: 4020 passed, 18 skipped. The first run stalled in a FlashAttention-4 backward pass under the default local environment. The rerun used the same FlashAttention selection CI sets (NVTE_FLASH_ATTN_V2=1, NVTE_FLASH_ATTN_V3=0, NVTE_FLASH_ATTN_V4=0) and passed.
• Unit tests, Core shard 2/3: 4041 passed.
• Functional L0_Launch_converter_generate: 3 passed.
• Functional L1_Launch_vlm_inference: 1 passed.

Main required CI on this PR is still pending and is the remaining check.

[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.

[svcnvidia-nemo-ci]

/ok to test c17e56c

[github-actions]

Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing:

/review

Add model=claude to use a Claude reviewer (the default is model=codex). mode=light|strict selects the review depth; for example, /review model=claude mode=strict. Comment /review help for all options.