[Bugfix][Kernel] Bump DeepGEMM pin for SM107 JIT compatibility

[lisawym]

Overview

Bump both DeepGEMM pins to 550036ad, which includes the SM107 runtime JIT workaround merged in vllm-project/DeepGEMM#23. This addresses Rubin startup failures tracked in deepseek-ai/DeepGEMM#461.

Claims

• Use the merged workaround through both CMake vendoring and the standalone installer, with matching dependency pins.
• On physical SM107, DeepGEMM selects the compatible 100f JIT target while preserving device capability reporting and other architectures' compiler targets.

Validation

Check	Result
Pre-commit on both changed files, shell syntax, and diff checks	Passed
Fresh CMake FetchContent check with CUDA compilation disabled	Fetched exactly 550036ad and initialized the expected CUTLASS/DeepJIT submodules
C++17 host harness compiling the fetched JIT header with mocked CUDA calls	Only SM107 changed to 100f; SM90, SM100, SM103, SM120, SM121 and a synthetic SM108 were preserved

Commands run for the changed files:

git diff upstream/main --check
.venv/bin/python -m pre_commit run --files \
  cmake/external_projects/deepgemm.cmake tools/install_deepgemm.sh
bash -n tools/install_deepgemm.sh
bash tools/install_deepgemm.sh --help

Prior Rubin startup and serving checks used a custom build with the same JIT override at an earlier dependency revision, as described in DeepGEMM#23. The combined revision has not received a full CUDA build, GPU serving smoke test, or model accuracy evaluation in this update. Fresh GPU validation and CI are still needed; no performance improvement is claimed.

Details

The pinned CUTLASS 4.2.1 headers lack native SM107 feature guards. DeepGEMM#23 temporarily selects the compatible SM100 family JIT target for physical SM107 devices. Its removal requires validated native SM107 support in DeepGEMM.

The bump from e1f418c also includes DeepGEMM #10/#14 (SM120) and #17/#19/#22 (NVFP4 Mega MoE). Their GPU paths need regression coverage. Related vLLM#59385 proposes an earlier pin for SM120 page-32 support; this revision also includes the SM107 fix.

The original vLLM patch and its application logic have been reverted. Validate using a fresh dependency checkout: CMake can reuse an existing _deps/deepgemm-src, and an externally installed deep_gemm can override the vendored copy.

AI assistance: Codex assisted with implementation, host checks, review, and PR preparation.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment /ci run for upstream CI or /amd-ci run for AMD CI only whenever CI signals are needed.

Once the PR is approved or has the ready label, the PR author can also use the corresponding /ci run, /ci retry, and /ci cancel commands, or their /amd-ci variants. New commits do not start upstream CI automatically.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mgoin]

/ci run

[github-actions]

✅ Triggered Buildkite CI #92476 for commit 2d8e74444a90.

[lisawym]

• Reverted the old patch and rebased onto current upstream main.
• Updated both DeepGEMM pins to 550036ad.

[tlrmchlsmth]

not an issue with this PR but we should make this a single-source of truth so these don't need to stay in sync

[tlrmchlsmth]

/ci run

[github-actions]

✅ @lisawym, CI is now available for this PR.

• /ci run starts upstream CI; /amd-ci run starts AMD CI only.
• Your branch must contain every commit currently on its upstream target branch. Merge or rebase onto the latest target branch, then rerun the command. Append --allow-stale to a run command to test an outdated branch at your own risk.
• /ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
• /amd-ci retry retries failed jobs in AMD CI for the current PR head. Use /amd-ci run when the current head has no AMD CI build.
• /ci cancel cancels scheduled or running CI builds for this PR branch; /amd-ci cancel does the same for AMD CI only.

[github-actions]

✅ Triggered Buildkite CI #92627 for commit be2405a79716.