[Common] Ensure quantization kernels handle noop properly

[kainzhong]

Description

When quantization kernels are fused with act / dact / dbias, they should ignore the noop flag. Currently we mostly rely on the caller passing the right noop flag but we don't have guards in the kernel.

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:

• Quantization kernels now ignore noop when fused
• MXFP8 now also ignores noop with IS_DBIAS. Also zero_scales_kernel now will ignore noop when the quantization ignores it (otherwise if we skip zeroing but not skip quantization the result could be wrong)
• FP8 kernels now correctly respect noop tensor when it's safe to skip execution
• Add tests

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

This PR corrects noop handling across quantization kernels.

• Fused activation and DBIAS paths now execute regardless of the noop flag.
• Safe non-fused FP8 and MXFP8 paths continue to skip quantization when noop is set.
• MXFP8 scale zeroing now follows the same effective noop behavior as quantization.
• PyTorch CUDA-graph tests verify that skipped quantization preserves output buffers.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/common/cast/mxfp8/quantize_mxfp8.cuh	Aligns scale zeroing and quantization by clearing the effective noop pointer for fused activation and DBIAS variants.
transformer_engine/common/cast/mxfp8/group_quantize_mxfp8.cuh	Prevents grouped DBIAS MXFP8 quantization from incorrectly returning early on noop.
transformer_engine/common/cast/fp8/quantize_fp8.cuh	Propagates noop into safe non-fused FP8 kernels while suppressing it for activation-fused launches.
transformer_engine/common/cast/fp8/group_quantize_fp8.cuh	Restricts grouped FP8 noop early returns to non-activation variants.
transformer_engine/common/cast/fp8_blockwise/group_quantize_fp8_blockwise.cuh	Keeps DBIAS workspace generation active because the subsequent reduction launch is unconditional.
tests/pytorch/test_quantized_tensor.py	Adds CUDA-graph coverage ensuring noop preserves quantized data and scale buffers across formats and layouts.

_{Reviews (2): Last reviewed commit: "Update transformer_engine/common/cast/mx..." | Re-trigger Greptile}