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_pytorch_mcore_fsdp_integration CI test that clones a pinned Megatron-LM commit, creates a mock BPE vocab, sets relevant NVTE_* env vars, and runs a short Megatron-FSDP GPT training job with FP8, CPU offloading, and DCP checkpointing flags.

The two blocking issues flagged in earlier review rounds have been addressed: unset CUDA_DEVICE_MAX_CONNECTIONS is now a standalone shell statement (no longer swallowing python3), and the Python invocation uses direct backslash line-continuations rather than a bash -c "${COMMAND}" string that would split at newlines.

Confidence Score: 5/5

Safe to merge — all previously identified blocking issues have been resolved and only a minor P2 CI-speed suggestion remains.

The two P0/P1 issues from prior review rounds (unset swallowing python3, and the bash -c newline-splitting bug) are both fixed. The only remaining observation is a P2 suggestion to shallow-clone Megatron-LM to speed up CI; this does not affect correctness.

No files require special attention.

Vulnerabilities

No security concerns identified. The script clones from the official NVIDIA Megatron-LM repository and checks out a pinned commit SHA, which limits supply-chain risk.

Important Files Changed

Filename	Overview
qa/L1_pytorch_mcore_fsdp_integration/test.sh	New CI test script; previous blocking issues (unset swallowing python3, bash -c newline splitting) are fixed; one P2 suggestion to shallow-clone Megatron-LM for CI speed
qa/L1_pytorch_mcore_fsdp_integration/.gitignore	Correctly ignores the cloned Megatron-LM directory and generated vocab.json
qa/L1_pytorch_mcore_fsdp_integration/merges.txt	Stub BPE merges file (version header only) consistent with the sibling test's approach for mock data

Sequence Diagram

sequenceDiagram
    participant CI as CI Runner
    participant GH as GitHub (Megatron-LM)
    participant FS as Filesystem
    participant GPU as GPU (GB200)

    CI->>FS: Check if Megatron-LM dir exists
    alt Not present
        CI->>GH: git clone Megatron-LM
        CI->>GH: git checkout 8cbc45b (pinned)
    end
    CI->>FS: Write mock vocab.json (4096 tokens)
    CI->>FS: unset CUDA_DEVICE_MAX_CONNECTIONS
    CI->>FS: export NVTE_* env vars
    CI->>GPU: python3 -m torch.distributed.launch pretrain_gpt.py (FSDP, FP8, BF16)
    GPU-->>CI: Train 10 iters + final eval
    CI-->>CI: Exit 0 (pass) or non-zero (fail)

Loading

_{Reviews (8): Last reviewed commit: "Remove CPU initialization, add FW args." | 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.