chore: AUT-2517 bump MCore main to fde43b6d

[svcnemo-autobot]

Background

Nightly Megatron-Core bump (main lane). Moves .main.commit and the 3rdparty/Megatron-LM gitlink from 8949def8 to fde43b6d (NVIDIA/Megatron-LM main). .dev.commit is unchanged (4a4de865).

What changed

• .main.commit and the gitlink now point to fde43b6d617752433310b4668c6b5c55e503521a.
• Regenerated uv.lock with uv lock --upgrade (uv 0.7.2, matching UV_VERSION). MCore's pyproject.toml is identical at both pins, so the dependency metadata needed no change. Six packages moved: grpcio/grpcio-tools 1.85.0rc1 (prereleases are allowed), httptools 0.9.0, openai 3.28.0, pycparser 3.11 and ruff 0.17.0.
• Test fix: MCore #7469 makes process_mtp_loss read config.attention_cp_layout. test_process_mtp_loss_with_native_packing_metadata now sets it to "zigzag", TransformerConfig's default, in its SimpleNamespace stub. Nothing in Bridge src/ calls process_mtp_loss.

Details

The tests were chosen from the MCore diff (MoE/flex dispatch, MLA, MTP, DSv4, checkpointing). The diffusion unit row was excluded because none of its code uses the changed paths.

Tested

Run locally on H100 (2 GPUs per row) with a validation image built from this head:

• Unit Core shards 0/3, 1/3 and 2/3 passed: 3984, 3965 and 3963 tests. Shard 2/3 passed after the test fix.
• Functional runs that passed: models_deepseek, training_pretrain, recipes_perf_config, recipes_nemotronh, vlm_inference, and the 4 MLM↔Bridge checkpoint runs (llama32_1b and qwen3_4b, both directions).
• Not covered locally:
  • L0_Launch_recipes_deepseek was excluded on maintainer instruction. Its DeepSeek-V4 toy model is missing from the test data, so it ran zero tests.
  • The dist_train case in training_pretrain needs 8 GPUs, so it skipped on 2.
• Main CI is pending.

[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 fcdb71f

[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.