reuse grad buffer for layer-wise param allgather

[FDecaYed]

What does this PR do ?

Reuse bucket.grad_data as the receive buffer for layer-wise async param all-gather instead of allocating temporary tensors via torch.empty. During param sync (forward pass), grad_data is idle and always has sufficient bytes since grad_dtype.element_size >= param_dtype.element_size. This eliminates per-bucket temporary allocations and removes the _layerwise_src_buffer bookkeeping.

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"

1. When your PR is ready, click Ready for Review.
2. 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]

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.

[FDecaYed]

/ok to test bec1750

[FDecaYed]

/claude review

[claude]

LGTM

[FDecaYed]

/claude review

[FDecaYed]

/ok to test f6730fc

[deepakn94]

Hi Deyu, I am working on a refactor of the way parameters are assigned in the parameter buffer here: #3812.

Do you still want to merge this?

[FDecaYed]

Hi Deyu, I am working on a refactor of the way parameters are assigned in the parameter buffer here: #3812.

Do you still want to merge this?

@deepakn94 My understanding is once optimizer control layout, inplace RS/AG become possible and this optimization might no longer be needed?

I'm not sure how long it'd take for in-place AG to happen. But since this change is quite small and not conflicting in terms of code/design, I suggest we merge this now. Even without this, the temp buffer code need to be cleaned up later and this PR doesn't make it harder(if not easier)

[deepakn94]

@deepakn94 My understanding is once optimizer control layout, inplace RS/AG become possible and this optimization might no longer be needed?

Yes, that's right.

I'm not sure how long it'd take for in-place AG to happen. But since this change is quite small and not conflicting in terms of code/design, I suggest we merge this now. Even without this, the temp buffer code need to be cleaned up later and this PR doesn't make it harder(if not easier)

I am hopeful I can solicit reviews before the free days, but let's see.

[FDecaYed]

/ok to test dd22546

[FDecaYed]

/claude review

[claude]

LGTM

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/24381405635