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 centralizes GroupedLinear split-quantization and bias-gradient dispatch in a dedicated helper module while limiting grouped-quantizer validation to CustomRecipe generations.

• Replaces separate plain, hybrid, and bias-gradient split helpers with a unified _split_quantize API.
• Preserves native, hybrid, identity, debug, and Python fallback paths behind one dispatch boundary.
• Adds coverage for CustomRecipe validation caching, built-in recipe behavior, native bias-gradient dispatch, hybrid storage, identity quantization, and CPU offload.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/module/_split_quantization.py	Introduces the unified split-quantization, hybrid-storage, native bias-gradient, validation, identity, and fallback dispatch helpers.
transformer_engine/pytorch/module/grouped_linear.py	Delegates split quantization to the new module and restricts generation compatibility validation and caching to CustomRecipe quantizers.
tests/pytorch/test_hybrid_quantization.py	Expands coverage for dispatch selection, exact quantized storage, validation caching, hybrid policies, and built-in recipe behavior.
tests/pytorch/test_custom_recipe.py	Strengthens CustomRecipe GroupedLinear coverage by comparing forward and backward results exactly with current scaling.
tests/pytorch/test_identity_quantizer.py	Updates identity and hybrid split tests for the unified API, including dbias and CPU-offload behavior.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A["GroupedLinear forward/backward"] --> B["_split_quantization._split_quantize"]
  B --> C{"Quantized output requested?"}
  C -- No --> D["Cast and torch.split"]
  C -- Yes --> E{"Compute dbias with supported native quantizer?"}
  E -- Yes --> F["tex.bgrad_quantize per split"]
  E -- No --> G{"Expert 0 quantizer type"}
  G -- Debug --> H["DebugQuantizer.multi_tensor_quantize"]
  G -- Hybrid --> I["Split rowwise and columnwise sub-quantizers"]
  G -- Native --> J["tex.split_quantize"]
  G -- Identity or custom --> K["Views or Python quantizer fallback"]

Loading

_{Reviews (4): Last reviewed commit: "fix lint" | Re-trigger Greptile}

[negvet]

/te-ci pytorch L1

[zhongbozhu]

overall LGTM, just have some questions

[zhongbozhu]

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

[negvet]

This is only for the type checker, like assume there are no None elements. At runtime, it returns the same object unchanged, it does not iterate through.

[zhongbozhu]

/te-ci pytorch

[negvet]

/te-ci pytorch

[vthumbe1503]

LGTM. We can rerun CI after fixing merge conflicts

[negvet]

/te-ci pytorch

[negvet]

@vthumbe1503 ci is green

[vthumbe1503]

@negvet Had missed this in earlier review. But could we please move the is_grad_enabled inside the validate function to reduce CPU overheads in cases of non-custom recipes? Can be a seperate PR as well

[negvet]

Sure, will do in a separate PR

[negvet]

#3405