[ExecuTorch][WebGPU] Add Gemma 4 MTP export path

[JCNTH]

Stack from ghstack (oldest at bottom):

• [ExecuTorch][WebGPU] Add Gemma 4 MTP and speculative-decode source-closure tests #21678
• [ExecuTorch][WebGPU] Add Gemma 4 speculative decode runtime #21677
• -> [ExecuTorch][WebGPU] Add Gemma 4 MTP export path #21676
• [ExecuTorch][WebGPU] Add Gemma 4 MTP operator and route support #21675
• [ExecuTorch][WebGPU] Add plain Gemma 4 source-closure tests #21674
• [ExecuTorch][WebGPU] Add Gemma 4 plain export and artifact contract #21673
• [ExecuTorch][WebGPU] Add Gemma 4 plain runtime and guarded routes #21672
• [ExecuTorch][Vulkan] Support scoped Gemma symbolic partitioning #21671
• [ExecuTorch][WebGPU] Add shared model runtime prerequisites #21670

Export the official Gemma 4 target plus trained QAT assistant as a source-verifiable k2_round program. The model path preserves BSHD cache ownership, converts donor views to BHKD on the sequence dimension, advances assistant positions correctly, applies Gemma embedding scale, and rewrites only provenance-certified TopK/scatter sites. Publication is transactional and requires a hashed source receipt before combined staging.

Key changes:

• eagle_webgpu_round.py implements donor views, assistant steps, position progression, embedding scale, and residual provenance checks.
• MTP partitioning scopes target/assistant attention and certifies exact TopK/scatter geometry.
• Exporters publish the source receipt last, reject collisions, and roll back every file on failure.
• Manifest pins acquisition, architecture, roles, P512/ctx8960 capacities, and pending versus verified provenance.

Plain Gemma remains unchanged and independently landable.

Co-authored-with: Claude Code.

Differential Revision: D115234081

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21676

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 New Failures

As of commit c4dd7e2 with merge base ceca90f ():

NEW FAILURES - The following jobs have failed:

• Test WebGPU Backend / test-webgpu / test-backend-linux (webgpu, models) / linux-job (gh)
  /pytorch/executorch/backends/webgpu/runtime/ops/to_copy/ToCopy.cpp:217:6: error: redefinition of 'add_bool_to_float_op'
• Test WebGPU Backend / test-webgpu / test-backend-linux (webgpu, operators) / linux-job (gh)
  /pytorch/executorch/backends/webgpu/runtime/ops/to_copy/ToCopy.cpp:217:6: error: redefinition of 'add_bool_to_float_op'

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[meta-cla]

Hi @JCNTH!

Thank you for your pull request.

We require contributors to sign our Contributor License Agreement, and yours needs attention.

You currently have a record in our system, but the CLA is no longer valid, and will need to be resubmitted.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!