[Common] Fix Grouped MXFP8 tensor-map validation and synchronization

[Oleg-Goncharov]

Description

This PR fixes two correctness issues in the Grouped MXFP8 quantization kernels introduced by the optimized work mapper.

First, the direct VARYING_FIRST_DIM path assumed that every tensor's first dimension was aligned to the 128-row kernel tile, but only the combined logical dimension was validated. This could allow a CTA to cross an expert boundary and use incorrect per-expert metadata.

Second, device-modified TMA descriptors were copied from shared to global memory without the release/acquire protocol required by the TMA memory model. The descriptor updater now follows the recommended warp-collective publication sequence.

Addresses #3474

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:

• Restore validation that every VARYING_FIRST_DIM tensor has a first dimension divisible by the 128-row kernel tile.
• Align shared-memory CUtensorMap storage to 128 bytes.
• Publish modified TMA descriptors with a warp-collective tensormap.cp_fenceproxy operation using GPU-scope release semantics.
• Use the matching GPU-scope acquire fence before consuming descriptors from another CTA.
• Launch the descriptor updater with a full warp and update the grouped quantize and dequantize paths accordingly.

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 fixes Grouped MXFP8 tensor-map validation and safely publishes device-modified TMA descriptors across CTAs.

• Restores per-tensor row-alignment validation for the direct varying-first-dimension mapper.
• Uses aligned shared descriptor storage with warp-collective GPU-scope release and matching acquire semantics.
• Launches descriptor updaters with a full warp and updates dynamic shared-memory alignment call sites.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the descriptor updater now launches a complete warp, reaches the warp-collective publication instruction convergently, and pairs GPU-scope release and acquire operations across updater and consumer CTAs.

Important Files Changed

Filename	Overview
transformer_engine/common/cast/core/grouped_tma.cuh	Implements aligned, warp-collective tensor-map publication and a matching GPU-scope acquire; the previously reported publication issue is fixed.
transformer_engine/common/cast/mxfp8/group_quantize_mxfp8.cuh	Restores per-tensor direct-mapper validation and launches the descriptor updater with the required full warp.
transformer_engine/common/cast/mxfp8/group_dequantize_mxfp8.cuh	Makes the updater body warp-convergent and preserves a full-warp launch for tensor-map publication.
transformer_engine/common/cast/fp8_blockwise/group_quantize_fp8_blockwise.cuh	Replaces the removed shared-memory alignment helper with the common aligned-pointer utility.
transformer_engine/common/cast/mxfp8/group_scaled_swiglu_mxfp8.cuh	Migrates dynamic shared-memory alignment to the common utility without changing the buffer layout.
transformer_engine/common/cast/nvfp4/specialized/quantize_transpose_nvfp4_tuned_1D.cuh	Migrates dynamic shared-memory alignment to the common utility without changing kernel behavior.

_{Reviews (3): Last reviewed commit: "[pre-commit.ci] auto fixes from pre-comm..." | Re-trigger Greptile}