[Common] MXFP8: Add swizzled-scale fast path for cast-only quantization

[WanZzzzzz]

Description

Produce GEMM-ready scales directly in the specialized MXFP8 cast-only kernels. When GEMM-optimized scales are requested, eligible rowwise and bidirectional casts now write scale inverses directly in the GEMM-swizzled layout.

This avoids separate scale-swizzle kernel launches and reduces intermediate scale-memory traffic. Unsupported shapes and configurations, including noop-aware calls, continue to use the generic implementation to preserve existing behavior.

Irregular but supported shapes are handled by zero-initializing padded scale regions before the specialized kernel writes valid scale values.

Fixes: N/A

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

• Extend the specialized rowwise cast-only kernel to write scale inverses directly in GEMM-swizzled layout.
• Add configurable bidirectional cast traits with shared-memory caching and direct swizzled writes for both rowwise and columnwise scales.
• Route eligible MXFP8 cast-only operations through the specialized kernels for both linear and GEMM-swizzled scale layouts.
• Preserve the generic fallback for unsupported types and shapes, partial MXFP8 chunks, grid-size limitations, and noop-aware quantization.
• Zero padded scale regions for irregular GEMM-swizzled tensors to prevent stale values outside the valid scale area.
• Add C++ coverage comparing direct-swizzle results against the existing linear-cast-then-swizzle reference across multiple shapes and data types.
• Add PyTorch regression coverage verifying zeroed row-scale padding for irregular bidirectional shape

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 adds GEMM-swizzled scale output to specialized MXFP8 cast-only kernels, avoiding a separate scale-swizzle launch while retaining generic fallbacks.

• Adds direct rowwise and bidirectional swizzled-scale writes in specialized CUDA kernels.
• Restricts specialized dispatch by shape, grid size, operation type, and noop state.
• Adds C++ and PyTorch coverage for scale layout, padding, data types, tensor ranks, and tail cases.

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	Extends specialized cast-only dispatch to GEMM-swizzled outputs with shape, grid-size, and noop guards while retaining generic fallbacks.
transformer_engine/common/cast/mxfp8/specialized/quantize_mxfp8.cuh	Adds canonical GEMM-swizzled indexing and cached scale flushes for rowwise and bidirectional specialized kernels.
tests/cpp/operator/test_cast_mxfp8.cu	Adds byte-exact CUDA coverage comparing direct swizzled output against linear quantization followed by the established swizzle operation.
tests/pytorch/mxfp8/test_mxfp8_quantize_swizzle_fusion.py	Adds coverage ensuring padded row-scale slots remain zero for irregular bidirectional tensor shapes.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[MXFP8 cast-only quantization] --> B{Supported dtype, shape, grid, and no noop tensor?}
  B -->|Yes: rowwise| C[Specialized rowwise kernel]
  B -->|Yes: bidirectional| D[Specialized bidirectional TMA kernel]
  B -->|No| E[Generic MXFP8 kernel]
  C --> F[Write FP8 data and GEMM-swizzled row scales]
  D --> G[Write rowwise and colwise FP8 data and swizzled scales]
  E --> H[Preserve generic and noop-aware behavior]
  F --> I[GEMM-ready quantized tensor]
  G --> I
  H --> I

Loading

_{Reviews (3): Last reviewed commit: "merge main and fix conflicts" | Re-trigger Greptile}

[Oleg-Goncharov]

/te-ci

[Oleg-Goncharov]

LGTM.

[Oleg-Goncharov]

/te-ci

[Oleg-Goncharov]

LGTM. Pipeline #62101472 passed without errors.