Refactor GroupedLinear quantization dispatch

[negvet]

Description

Move grouped quantization helpers into a dedicated module and replace separate split APIs with one. Keeps quantizer validation at construction only for CustomRecipe while skipping validation for the built-in recipes.
Followup for #2817

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]

Greptile Summary

The PR extracts GroupedLinear split-quantization and validation into a dedicated helper module while consolidating forward and backward split APIs.

• Dispatches native, hybrid, identity, debug, and Python quantization through one helper.
• Computes optional per-expert bias gradients through the unified split operation.
• Limits grouped quantizer compatibility validation to CustomRecipe generations.
• Expands output, gradient, dispatch, and recipe-transition test coverage.

Confidence Score: 5/5

The PR appears safe to merge, with no concrete changed-code-triggered failures identified.

The unified dispatch preserves the previous native, debug, hybrid, identity, fallback, bias-gradient, and CPU-offload branch behavior, while custom-generation validation remains retried after failed transitions.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/module/_grouped_quantization.py	Introduces the unified grouped split-quantization, optional bias-gradient, hybrid dispatch, and compatibility-validation helpers without an identified correctness regression.
transformer_engine/pytorch/module/grouped_linear.py	Replaces local helper implementations with the dedicated module and restricts generation validation to custom recipes while preserving reachable dispatch behavior.
tests/pytorch/test_hybrid_quantization.py	Expands coverage for native and fallback dispatch, hybrid source policies, custom generation validation, and built-in validation bypass.
tests/pytorch/test_custom_recipe.py	Strengthens GroupedLinear CustomRecipe coverage with exact forward and backward comparisons against current scaling.
tests/pytorch/test_identity_quantizer.py	Updates identity and CPU-offload tests for the unified split API and its optional bias-gradient return.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  GL[GroupedLinear forward or backward] --> SQ[_grouped_quantization._split_quantize]
  SQ --> N{Reference quantizer}
  N -->|None| P[Cast and split tensors]
  N -->|DebugQuantizer| D[Debug multi-tensor quantization]
  N -->|HybridQuantizer| H[Rowwise and columnwise split quantization]
  N -->|Native quantizer| B[Native split or bgrad quantization]
  N -->|Custom or Identity| F[Python fallback or identity views]
  B --> O[Per-expert outputs and optional dbiases]
  D --> O
  H --> O
  F --> O
  P --> O

Loading

_{Reviews (1): Last reviewed commit: "Refactor GroupedLinear quantization disp..." | Re-trigger Greptile}

[negvet]

/te-ci pytorch L1

[zhongbozhu]

Can we rename this file to something like split quantization? It's not really a grouped quantization

[zhongbozhu]

Same comment for renaming

[zhongbozhu]

same comment

[zhongbozhu]

overall LGTM, just have some questions

[zhongbozhu]

why do we need this cast(Sequence[Quantizer], quantizers)? feels like some extra CPU overhead?