Add Megatron-FSDP E2E integration test to TE CI/CD (L1).

[cspades]

Description

• Adds Megatron-FSDP E2E integration tests to the TransformerEngine CI/CD.

Details

• Megatron-FSDP has experienced various issues with TransformerEngine in the past months, from race conditions on gradients with deterministic FlashAttention, to FSDP2 / Megatron-FSDP DCP checkpointing or decoupled_grad bugs related to FusedAdam, and other less obvious CPU offloading & Tensor API bugs that are difficult to catch without running Megatron-FSDP. This functional test aims to reduce the frequency of that.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

This PR adds a new L1 CI test (qa/L1_pytorch_mcore_fsdp_integration/) that runs a short Megatron-FSDP GPT pre-training job to catch TE regressions in areas such as FSDP2 checkpointing, CPU offloading, and NCCL UBR. The previously flagged shell correctness issues (unset swallowing python3, multiline bash -c splitting) are resolved in the current version — environment variables are set with proper export/unset statements and python3 is invoked directly with backslash continuations.

Confidence Score: 5/5

Safe to merge; all previously flagged blocking issues are resolved and remaining findings are minor P2 suggestions.

The three prior P0/P1 issues (unset-before-python3 swallowing the training command, multiline bash -c splitting, FP8 arch guard) are all addressed in the current revision. Only P2 items remain: the missing --save that would exercise the fsdp_dtensor checkpoint path, and a cosmetic missing newline in .gitignore. Neither blocks correctness of the training run.

qa/L1_pytorch_mcore_fsdp_integration/test.sh — consider adding --save to exercise the DCP checkpoint path called out in the PR description.

Vulnerabilities

No security concerns identified. The script clones a pinned, publicly known Megatron-LM commit over HTTPS and writes only to temporary/local paths within the TE QA directory.

Important Files Changed

Filename	Overview
qa/L1_pytorch_mcore_fsdp_integration/test.sh	New E2E CI test for Megatron-FSDP; previous thread issues (unset-before-python3, bash -c multiline split) are resolved; DCP checkpointing flag present but no --save dir means that specific path isn't exercised
qa/L1_pytorch_mcore_fsdp_integration/.gitignore	Correctly ignores cloned Megatron-LM dir and generated vocab.json; missing trailing newline (cosmetic)
qa/L1_pytorch_mcore_fsdp_integration/merges.txt	Stub BPE merges file with only the version header; safe for mock-data training which does not invoke the tokenizer

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[test.sh entry] --> B{Megatron-LM cloned?}
    B -- No --> C[git clone NVIDIA/Megatron-LM\ngit checkout MCORE_REF]
    C --> D
    B -- Yes --> D[Create mock vocab.json\n4096-token BPE stub]
    D --> E[unset CUDA_DEVICE_MAX_CONNECTIONS\nexport NVTE_* env vars]
    E --> F[python3 -m torch.distributed.launch\nnproc = nvidia-smi GPU count]
    F --> G[pretrain_gpt.py\n--use-megatron-fsdp\n--fp8-recipe mxfp8\n--cpu-offloading-num-layers 1\n--ckpt-format fsdp_dtensor\n10 train iters]
    G --> H{exit 0?}
    H -- Yes --> I[CI pass]
    H -- No --> J[CI fail / set -e abort]

Loading

_{Reviews (10): Last reviewed commit: "Expose MCore hash/tag as an argument to ..." | Re-trigger Greptile}

[timmoon10]

LGTM, pending CI

[timmoon10]

Pipeline 47956532

[cspades]

Depends on this: NVIDIA/Megatron-LM#4133

This PR correctly uses decoupled_grad depending on the FP8 recipe matching the distributed optimizer logic in Megatron-Core.