Install cuDNN frontend in GitHub builds

[sudhakarsingh27]

Summary

• Install the declared nvidia-cudnn-frontend>=1.25.0 build dependency in every GitHub Actions build environment.
• Keep the workflow requirement aligned with pyproject.toml.

Root cause

TransformerEngine now resolves cuDNN frontend headers from the installed Python distribution. The GitHub Actions builds use --no-build-isolation, so build-system requirements are not installed automatically and metadata generation fails with PackageNotFoundError.

Validation

Scenario	Result
Workflow YAML parsing	Passed with PyYAML
Diff hygiene	git diff --check passed
Package availability	Python 3.10 and 3.12 x86_64 wheels are published for the currently resolved 1.26.0 release
GitHub Actions	Build run 28997378989 pending

[ptrendx]

I think there is an issue here that we shouldn't just WAR - the cudnn frontend should be installed as a dependency of TE, the fact that it does not do that is a problem.

[timmoon10]

It seems we are still experiencing build failures because NCCL EP added a NCCL dependency: #3127

[vcherepanov-nv]

LGTM, thanks!

[greptile-apps]

Greptile Summary

This PR adds nvidia-cudnn-frontend>=1.25.0 to the pip install step in all four GitHub Actions build jobs (core, pytorch, jax, all) so that the package is present when building with --no-build-isolation. The constraint matches what is already declared in the root pyproject.toml build-system requirements.

• All four build jobs (core, pytorch, jax, all) now install the dependency before the pip install --no-build-isolation step, closing the gap that caused PackageNotFoundError during metadata generation.
• The quoted form "nvidia-cudnn-frontend>=1.25.0" is used consistently, correctly escaping the > operator from shell interpretation.

Confidence Score: 5/5

Safe to merge — adds a single missing build dependency across four CI jobs with no logic changes.

The change is a one-line addition per job, the version constraint matches pyproject.toml exactly, and the quoting is correct for shell safety. No test, runtime, or security concerns are introduced.

No files require special attention.

Important Files Changed

Filename	Overview
.github/workflows/build.yml	Adds nvidia-cudnn-frontend>=1.25.0 to pip install steps in all four build jobs; constraint is consistent with pyproject.toml and quoting is correct for shell safety.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[GitHub Actions Build Job] --> B[Dependencies Step]
    B --> C[apt-get / docker setup]
    C --> D["pip install build tools\ncmake, pybind11, ninja, etc."]
    D --> E["pip install nvidia-cudnn-frontend>=1.25.0\n✅ Added by this PR"]
    E --> F["pip install --no-build-isolation .\nreads pyproject.toml build-system.requires"]
    F --> G{cudnn-frontend\navailable?}
    G -->|Before PR: NO| H["❌ PackageNotFoundError\nMetadata generation fails"]
    G -->|After PR: YES| I["✅ Build succeeds\ntransformer_engine installed"]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[GitHub Actions Build Job] --> B[Dependencies Step]
    B --> C[apt-get / docker setup]
    C --> D["pip install build tools\ncmake, pybind11, ninja, etc."]
    D --> E["pip install nvidia-cudnn-frontend>=1.25.0\n✅ Added by this PR"]
    E --> F["pip install --no-build-isolation .\nreads pyproject.toml build-system.requires"]
    F --> G{cudnn-frontend\navailable?}
    G -->|Before PR: NO| H["❌ PackageNotFoundError\nMetadata generation fails"]
    G -->|After PR: YES| I["✅ Build succeeds\ntransformer_engine installed"]

Loading

_{Reviews (1): Last reviewed commit: "Install cuDNN frontend in GitHub builds" | Re-trigger Greptile}