Fix recompute checkpointing + training CGs by tdene · Pull Request #3919 · NVIDIA/Megatron-LM

tdene · 2026-03-18T08:40:22Z

What does this PR do ?

⚠️ For major changes (either in lines of code or in its impact), please make sure to first share a design doc with the team. If you're unsure what's the best way to do so, contact the @mcore-oncall.

Contribution process

Pre-checks

I have added relevant unit tests
I have added relevant functional tests
I have added proper typing to my code Typing guidelines
I have added relevant documentation
I have run the autoformatter.sh on my PR

Code review

Feel free to message or comment the @mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!

All PRs start as draft. If you open a non-draft PR, it will be automatically converted to draft.

Step 1: Mark PR as "Ready for Review"

When your PR is ready, click Ready for Review.
An oncall reviewer is auto-assigned and expert reviewers are notified based on your changes.
- Some PRs may jump straight to step 2. This is determined by .github/CODEOWNERS.

⚠️ Only mark as ready once merge-conflicts are resolved and the CI is passing.
Final Review might get declined if these requirements are not fulfilled.

Step 2: Final Review

For PRs that change megatron/core, once all expert reviewers have approved, the Final Review label is applied automatically and final reviewers are assigned.

For PRs outside megatron/core, this step is skipped.

Step 3: Approved

Once all required reviewers have approved, the Approved label is applied automatically.

Merge

Any member of mcore-engineers will be able to merge your PR.

For MRs into `dev` branch

The proposed review process for `dev` branch is under active discussion.

MRs are mergable after one approval by either eharper@nvidia.com or zijiey@nvidia.com.

copy-pr-bot · 2026-03-18T08:40:27Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

tdene · 2026-03-18T09:23:05Z

/claude review

jiemingz · 2026-03-18T13:35:56Z

+    # Skip checkpointing during CUDA graph warmup and capture, matching the behavior of
+    # CheckpointWithoutOutput. The graph captures all ops directly; recomputation cannot
+    # run inside a captured graph.
+    if is_graph_warmup() or is_graph_capturing():


Doesnt this just disable recomputation during graph capture? is_graph_capturing() is true when capturing. Why cant recomputation work with graphs?

Right, this only disables recomputation during graph capture. That's where I'm seeing errors in main right now: the first time we run through a training pass, we error out because we are capturing the graph while also attempting recompute.

Subsequent training passes seem to run fine.

jiemingz

There might be a way to get recompute working with per layer graphs but I am okay with explicitly disabling it like this unless needed.

svcnvidia-nemo-ci · 2026-05-14T03:26:48Z

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/25839862145

svcnvidia-nemo-ci · 2026-05-14T11:56:29Z

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/25858675752

* origin/main: (138 commits) Refactor CUDA graph API: decompose cuda_graph_scope into full_iteration impl, inference scope, and per-layer capture modules (NVIDIA#4292) Add high-priority A2A stream and HybridEP preprocessing SMs (NVIDIA#4694) add is_torch_min_version in fsdp src (NVIDIA#4812) [Main][feat] Support A2A Overlap for Megatron-FSDP (NVIDIA#3797) Reorder mtp_post_process after attention backward in 1F1B schedule plan (NVIDIA#4695) [fix] Use MSC for checking checkpoint existence (NVIDIA#4251) Combine GEMM + SwiGLU fused MLP PRs (3890, 4071, 4095, 4219, 4311, 4324) → main (NVIDIA#4636) Strengthen test_checkpoint to verify distributed checkpoint behavior (NVIDIA#4711) Disable MSC by default; opt in via --enable-msc (NVIDIA#4629) additional tests for nvrx (NVIDIA#4522) Update copy-pr-bot.yaml [skip ci] ci: tolerate git-gc race in /home/runner chown after checkout (NVIDIA#4808) Inference: Optimize Prefill Engine Steps for Nemotron (NVIDIA#4764) chore: Update nightly tests golden values (NVIDIA#4805) ci: Update workflow to use same commit for building docker image and running tests (NVIDIA#4787) Update owners (NVIDIA#4794) Bump nvidia-modelopt>=0.44.0 (NVIDIA#4803) fix tokenizers in respect to newer transformers (NVIDIA#4608) Use Protocols to type-check linear_proj submodules of Attention (NVIDIA#3434) Fix recompute checkpointing + training CGs (NVIDIA#3919) ... # Conflicts: # megatron/core/transformer/moe/moe_utils.py

Signed-off-by: yhgalaxy <yhgalaxy@outlook.com>

Signed-off-by: Jon Barker <jbarker@aws-cmh-slurm-1-vscode-02.cm.cluster>

Signed-off-by: Dmytro Pykhtar <dpykhtar@nvidia.com>

Skip CheckpointFunction during CUDA graph warmup and capture

9c5e82d

claude Bot reviewed Mar 18, 2026

View reviewed changes

Comment thread megatron/core/tensor_parallel/random.py

tdene marked this pull request as ready for review March 18, 2026 09:40

tdene requested review from a team as code owners March 18, 2026 09:40

svcnvidia-nemo-ci requested a review from a team March 18, 2026 09:40

svcnvidia-nemo-ci added this to the Core 0.16 milestone Mar 18, 2026

svcnvidia-nemo-ci added Final Review PR is in the "final review" stage complexity: low labels Mar 18, 2026

copy-pr-bot Bot temporarily deployed to test March 18, 2026 09:41 Inactive

jiemingz reviewed Mar 18, 2026

View reviewed changes

Merge branch 'main' into tde/fix_cg_recompute

993ff52

copy-pr-bot Bot temporarily deployed to test March 19, 2026 00:05 Inactive

jiemingz approved these changes Apr 1, 2026

View reviewed changes

ericharper approved these changes Apr 13, 2026

View reviewed changes

ericharper enabled auto-merge April 13, 2026 13:20

jaredcasper approved these changes May 14, 2026

View reviewed changes

svcnvidia-nemo-ci added Approved All necessary approvals have been made and removed Final Review PR is in the "final review" stage labels May 14, 2026

Merge branch 'main' into tde/fix_cg_recompute

45f497a

copy-pr-bot Bot temporarily deployed to public May 14, 2026 00:43 Inactive

copy-pr-bot Bot temporarily deployed to test May 14, 2026 00:43 Inactive

copy-pr-bot Bot temporarily deployed to public May 14, 2026 00:46 Inactive

copy-pr-bot Bot temporarily deployed to public May 14, 2026 00:47 Inactive

copy-pr-bot Bot temporarily deployed to public May 14, 2026 00:54 Inactive

ericharper added this pull request to the merge queue May 14, 2026

github-merge-queue Bot removed this pull request from the merge queue due to failed status checks May 14, 2026

tdene added this pull request to the merge queue May 14, 2026

Merged via the queue into NVIDIA:main with commit ada8dfe May 14, 2026
75 checks passed

tdene deleted the tde/fix_cg_recompute branch May 14, 2026 13:57

cspades pushed a commit to cspades/Megatron-LM that referenced this pull request May 14, 2026

Fix recompute checkpointing + training CGs (NVIDIA#3919)

aa4d1eb

janEbert pushed a commit to janEbert/Megatron-LM that referenced this pull request Jun 2, 2026

Fix recompute checkpointing + training CGs (NVIDIA#3919)

3baea1b

yhgalaxy pushed a commit to yhgalaxy/Megatron-LM that referenced this pull request Jun 17, 2026

Fix recompute checkpointing + training CGs (NVIDIA#3919)

86db439

Signed-off-by: yhgalaxy <yhgalaxy@outlook.com>

jon-barker pushed a commit to jon-barker/Megatron-LM that referenced this pull request Jul 10, 2026

Fix recompute checkpointing + training CGs (NVIDIA#3919)

ade1b70

Signed-off-by: Jon Barker <jbarker@aws-cmh-slurm-1-vscode-02.cm.cluster>

terminator123 pushed a commit to 021ai/Megatron-LM that referenced this pull request Aug 3, 2026

Fix recompute checkpointing + training CGs (NVIDIA#3919)

9ae8e08

svcnvidia-nemo-ci pushed a commit to dimapihtar/Megatron-LM that referenced this pull request Aug 4, 2026

Fix recompute checkpointing + training CGs (NVIDIA#3919)

0914d1e

Signed-off-by: Dmytro Pykhtar <dpykhtar@nvidia.com>

Navigation Menu

Uh oh!

Fix recompute checkpointing + training CGs - #3919

Fix recompute checkpointing + training CGs#3919
tdene merged 3 commits into
NVIDIA:mainfrom
tdene:tde/fix_cg_recompute

tdene commented Mar 18, 2026

Uh oh!

copy-pr-bot Bot commented Mar 18, 2026

Uh oh!

tdene commented Mar 18, 2026

Uh oh!

Uh oh!

jiemingz Mar 18, 2026 •

edited

Loading

Uh oh!

tdene Mar 18, 2026

Uh oh!

jiemingz left a comment

Uh oh!

svcnvidia-nemo-ci commented May 14, 2026

Uh oh!

Uh oh!

svcnvidia-nemo-ci commented May 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Uh oh!

Conversation

tdene commented Mar 18, 2026

What does this PR do ?

Contribution process

Pre-checks

Code review

Step 1: Mark PR as "Ready for Review"

Step 2: Final Review

Step 3: Approved

Merge

Uh oh!

copy-pr-bot Bot commented Mar 18, 2026

Uh oh!

tdene commented Mar 18, 2026

Uh oh!

Uh oh!

jiemingz Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tdene Mar 18, 2026

Choose a reason for hiding this comment

Uh oh!

jiemingz left a comment

Choose a reason for hiding this comment

Uh oh!

svcnvidia-nemo-ci commented May 14, 2026

Uh oh!

Uh oh!

svcnvidia-nemo-ci commented May 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

jiemingz Mar 18, 2026 •

edited

Loading