[PyTorch] Deprecate grouped linear and MLP environment gates

[zhongbozhu]

Description

Please include a brief summary of the changes, relevant motivation and context.

Fixes # (issue)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Removes environment variable gates for grouped linear and MLP features.

The PR does not appear safe to merge while unrelated CPU-only forwards can still initialize CUDA through the grouped-MLP fusion callbacks.

Findings

1.  Unrelated forwards initialize CUDA ▶

Summary

The PR removes environment gates for grouped-linear parameter layout and grouped-MLP fusion, updates the Mixtral example accordingly, and adjusts tests and QA launchers.

• Explicit constructor options now select grouped parameters without the former environment variable.
• Eligible grouped-MLP patterns are registered without the former import-time gate.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[GroupedLinear → activation → GroupedLinear] --> B{Recipe and pattern eligible?}
  B -->|No| C[Separate operations]
  B -->|Yes| D{Device and kernels supported?}
  D -->|No| C
  D -->|Yes| E[Fused grouped MLP]

Loading

_{Reviews (4) · Last reviewed commit: "Merge branch 'main' into cleanup/grouped..."}

[zhongbozhu]

/te-ci pytorch L1

[zhongbozhu]

/te-ci pytorch L1

[zhongbozhu]

/te-ci pytorch L1

[greptile-apps]

Unrelated forwards initialize CUDA If a machine has a GPU, the first forward of even an unrelated CPU-only Sequential now queries the current CUDA device before checking whether the operations or recipe could use grouped-MLP fusion. The unary fusion callback also checks support before candidacy. This can initialize CUDA in a process that later forks workers, causing a worker to fail when it tries to use CUDA. Check the recipe and operation pattern before probing the device.

Knowledge Base Used: PyTorch tensor and operator abstractions

[vthumbe1503]

LGTM. Thanks for cleaning these env variables. All @timmoon10's comments seem to be resolved as well. @zhongbozhu I resolved the merge conflicts.

[vthumbe1503]

/te-ci pytorch