[matthiasdiener]

Description

Triton 3.8 miscompiles MoE permutation on gfx1250 for multi-warp kernels that apply both tl.cumsum and a reduction to the same block. Triton 3.9 fixes that miscompilation, but the code is slow on both versions (as well as on gfx950).

• Token-major row_id_map (all AMD GPUs): three new kernels count each expert's tokens per block of tokens, prefix-sum the counts over blocks, then write each block's rows. They read whole token rows, need no sort, and none of them both scans and reduces a block. The row_id_map format is unchanged.
• gfx1250 only: unpermute_with_mask_map unpermutes the probs in a kernel of their own (_unpermute_probs_kernel) instead of in the first hidden blocks of _unpermute_kernel. Other GPUs keep the fused kernel.

Performance results: https://github.com/AMD-ROCm-Internal/frameworks-internal/issues/395#issue-6169144290

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

[ipanfilo]

Maybe add Triton ticket so it is easier to track