Inference: Optimize Prefill Engine Steps for Nemotron by sidsingh-nvidia · Pull Request #4764 · NVIDIA/Megatron-LM

sidsingh-nvidia · 2026-05-12T20:49:28Z

What does this PR do ?

Use padded_prefill_request_count instead of max_requests in the mamba prefill kernels. Speeds up cuda-graphed prefill of nanov3 by 30%.
Add vLLM's autotuning parametesr to _causal_conv1d_varlen_kernel. vLLM uses a smaller block_t of 8, which ensures that all SMs are occupied.
Do not restrict number of allgatherv SMs to 16 with shared expert overlap. Empirically saw better overlap with this.

⚠️ 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.

Issue tracking

For PRs from open-source community contributors:

New features: a linked issue is required. Please open a feature request and reference it here before submitting the PR.
Small updates (bug fixes, minor improvements): a linked issue is recommended and will accelerate the PR review process.

Linked issue:

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.

github-actions · 2026-05-12T20:49:39Z

This PR has been automatically converted to draft because all PRs must start as drafts.

When you are ready for review, click Ready for Review to begin the review process. This will:

Add the oncall reviewer (optional reviewer)
Add required review teams based on your changes

See the contribution guide for more details.

copy-pr-bot · 2026-05-12T20:49:41Z

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.

svcnvidia-nemo-ci · 2026-05-14T22:59:21Z

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/25890498014

Upstream main tip: f92a207 Pulled commits: - f92a207 Update copy-pr-bot.yaml [skip ci] - b3b6719 ci: tolerate git-gc race in /home/runner chown after checkout (NVIDIA#4808) - 9b4074b Inference: Optimize Prefill Engine Steps for Nemotron (NVIDIA#4764) - a53107c chore: Update nightly tests golden values (NVIDIA#4805) - e9a0930 ci: Update workflow to use same commit for build+test (NVIDIA#4787) - 266562f Update owners (NVIDIA#4794) - 98031e1 Bump nvidia-modelopt>=0.44.0 (NVIDIA#4803) - d167123 fix tokenizers in respect to newer transformers (NVIDIA#4608) - dbfc96b Use Protocols to type-check linear_proj submodules of Attention (NVIDIA#3434) Conflict resolutions: - .github/CODEOWNERS: --theirs (upstream granular team mapping) - megatron/core/transformer/attention.py: composed -- kept ours StreamBP imports, dropped dead ModuleSpec/build_module import (replaced by Protocol in NVIDIA#3434), kept ours attn_proj_manager pattern because our fine_grained_activation_offload.py keeps group_offload API while upstream split to static group_commit, but adopted apply_module(self.linear_proj) wrapper from NVIDIA#3434 - megatron/core/transformer/multi_latent_attention.py: same apply_module+attn_proj_manager composition; took ours ChunkRange parameter for StreamBP plus upstream BaseInferenceContext type annotation - pyproject.toml: bumped modelopt >=0.44 (upstream); kept transformer-engine[pytorch,core_cu13]>=2.9.0a0,<2.12.0 pin (custom for CUDA 13 / SM100) - uv.lock: --ours; modelopt bump is non-breaking. uv lock regenerate blocked by stale /home/sjpat/wheelhouse torchcomms path (pre-existing). Gates: - git diff --check: clean - conflict markers: none - py_compile (15 changed .py files): OK - attention.py + multi_latent_attention.py import OK - indexcache: 27/28 pass (same single GPU-env failure as pre-merge base) - transformer gdn/mtp/moe suite: 53 failed / 7 passed / 55 skipped / 5 errors -- identical to base (all failures are cudaErrorDevicesUnavailable from sglang occupying all H200s) - 2-rank torchrun smoke: blocked (no free GPUs) Custom preserved: StreamBP (megatron/core/transformer/streambp.py + tests/unit_tests/transformer/test_streambp.py + tools/streambp_prod_verify.py), IndexCache config + NVFP4 indexer, HISA topk1024 backward, emerging_optimizers v0.2.0 pin, mHC/MTP/MoE composition.

* 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>

tighten gather ops to padded_prefill_count only

fb30931

sidsingh-nvidia requested review from a team as code owners May 12, 2026 20:49

svcnvidia-nemo-ci marked this pull request as draft May 12, 2026 20:49

copy-pr-bot Bot temporarily deployed to public May 12, 2026 20:50 Inactive

copy-pr-bot Bot temporarily deployed to test May 12, 2026 20:50 Inactive

copy-pr-bot Bot temporarily deployed to public May 12, 2026 20:53 Inactive

copy-pr-bot Bot temporarily deployed to public May 12, 2026 21:01 Inactive

tune causal conv1d kernel in line with vLLM

ce6a10b

sidsingh-nvidia changed the title ~~Tighten gather ops to padded_prefill_count only~~ Inference: Optimize Prefill Engine Steps for Nemotron May 12, 2026

sidsingh-nvidia requested review from lmcafee-nvidia and santhnm2 May 13, 2026 17:34

santhnm2 approved these changes May 13, 2026

View reviewed changes

sidsingh-nvidia marked this pull request as ready for review May 13, 2026 17:35

sidsingh-nvidia requested review from a team as code owners May 13, 2026 17:35

Merge branch 'main' into siddharth/optimize-prefill

7eea3cf

svcnvidia-nemo-ci requested a review from a team May 13, 2026 17:35

svcnvidia-nemo-ci added the complexity: low label May 13, 2026

copy-pr-bot Bot temporarily deployed to public May 13, 2026 17:36 Inactive

copy-pr-bot Bot temporarily deployed to public May 13, 2026 17:39 Inactive

fanshiqing approved these changes May 14, 2026

View reviewed changes

svcnvidia-nemo-ci added the Final Review PR is in the "final review" stage label May 14, 2026

sidsingh-nvidia self-assigned this May 14, 2026

sidsingh-nvidia added Run tests Run functional tests labels May 14, 2026

Merge branch 'main' into siddharth/optimize-prefill

236e1ef

copy-pr-bot Bot temporarily deployed to public May 14, 2026 16:57 Inactive

copy-pr-bot Bot temporarily deployed to test May 14, 2026 16:57 Inactive

shanmugamr1992 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

sidsingh-nvidia removed Run tests Run functional tests labels May 14, 2026

Merge branch 'main' into siddharth/optimize-prefill

230e1fd

sidsingh-nvidia enabled auto-merge May 14, 2026 21:23

copy-pr-bot Bot temporarily deployed to public May 14, 2026 21:23 Inactive

copy-pr-bot Bot temporarily deployed to test May 14, 2026 21:24 Inactive

sidsingh-nvidia added this pull request to the merge queue May 14, 2026

Merged via the queue into NVIDIA:main with commit 9b4074b May 15, 2026
68 of 70 checks passed

sidsingh-nvidia deleted the siddharth/optimize-prefill branch May 15, 2026 03:09

svcnvidia-nemo-ci mentioned this pull request May 16, 2026

chore: nightly sync main into dev (16_05_2026) #4831

Closed

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

Inference: Optimize Prefill Engine Steps for Nemotron (NVIDIA#4764)

32fc754

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

Inference: Optimize Prefill Engine Steps for Nemotron (NVIDIA#4764)

f4d81bc

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

Inference: Optimize Prefill Engine Steps for Nemotron (NVIDIA#4764)

00e5aa3

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

Inference: Optimize Prefill Engine Steps for Nemotron (NVIDIA#4764)

f11217d

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

Inference: Optimize Prefill Engine Steps for Nemotron (NVIDIA#4764)

f11762a

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

Navigation Menu

Uh oh!

Inference: Optimize Prefill Engine Steps for Nemotron - #4764

Inference: Optimize Prefill Engine Steps for Nemotron#4764
sidsingh-nvidia merged 8 commits into
NVIDIA:mainfrom
sidsingh-nvidia:siddharth/optimize-prefill

sidsingh-nvidia commented May 12, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented May 12, 2026

Uh oh!

copy-pr-bot Bot commented May 12, 2026

Uh oh!

svcnvidia-nemo-ci commented May 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

Uh oh!

Conversation

sidsingh-nvidia commented May 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do ?

Issue tracking

Contribution process

Pre-checks

Code review

Step 1: Mark PR as "Ready for Review"

Step 2: Final Review

Step 3: Approved

Merge

Uh oh!

github-actions Bot commented May 12, 2026

Uh oh!

copy-pr-bot Bot commented May 12, 2026

Uh oh!

svcnvidia-nemo-ci commented May 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

sidsingh-nvidia commented May 12, 2026 •

edited

Loading