[PyTorch] Enable head dim 256 for FA4

[yaox12]

Description

Need FA4 version 4.0.0b11.

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

This PR enables head_dim=256 support for FlashAttention 4 on SM100/103 (Blackwell) by delegating head-dimension validation to FA4's own _validate_head_dims function (available in flash-attn-4 > 4.0.0b10), and adds a properly guarded test_dpa_fa4_hdim256 test that addresses the previous review thread. It also removes stale cuDNN ≥8.9.1 guards from FA4 tests (FA4 is a standalone CUDA kernel independent of cuDNN).

• backends.py: Imports _validate_head_dims from flash_attn.cute.interface and stores it as FlashAttentionUtils.v4_validate_head_dims; the call is now deferred to FA4's own assertion-based validator instead of TE's inline check.
• utils.py: Replaces the static SM architecture table with a live call to v4_validate_head_dims, adds the Callable type annotation, updates the installation hint to 4.0.0b11, and restructures the MLA backward-kernel workaround from elif to a standalone if so it runs independently of the head-dim check.
• test_attention.py: Adds test_dpa_fa4_hdim256 with an explicit device_compute_capability not in ((10, 0), (10, 3)) skip guard; bumps FA QA test suite to 4.0.0b11 on both SM90 and SM100+ branches.

Confidence Score: 4/5

Functional on FA4 4.0.0b11+, but importing _validate_head_dims in the same block as the main FA4 symbols will crash backends.py module load for any environment with an older FA4 installed.

The _validate_head_dims symbol is imported in the same from flash_attn.cute.interface import (...) statement as flash_attn_func and flash_attn_varlen_func. The outer except PackageNotFoundError does not catch ImportError, so any user with FA4 older than 4.0.0b11 will hit an unhandled exception at module load of backends.py, breaking all attention backends — not just FA4. This issue was flagged in the previous review round and remains unaddressed.

transformer_engine/pytorch/attention/dot_product_attention/backends.py — the combined import of _validate_head_dims alongside the two main FA4 symbols needs its own try/except to preserve the graceful fallback.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/attention/dot_product_attention/backends.py	Imports _validate_head_dims in the same from flash_attn.cute.interface import (...) block as the main FA4 functions; an ImportError (not caught by the outer except PackageNotFoundError) crashes backends.py loading for any user with FA4 < 4.0.0b11.
transformer_engine/pytorch/attention/dot_product_attention/utils.py	Adds v4_validate_head_dims class attribute and delegates head-dim validation to FA4's own _validate_head_dims; installation steps correctly updated to 4.0.0b11; MLA workaround restructured with if (was elif) to run independently after the head-dim check.
tests/pytorch/attention/test_attention.py	Adds test_dpa_fa4_hdim256 with explicit SM100/103 skipif guard (addressing the previous thread); removes stale cuDNN ≥8.9.1 guards from all FA4 tests (FA4 is a standalone CUDA kernel, not cuDNN-dependent).
qa/L3_pytorch_FA_versions_test/test.sh	Bumps FA4 test version from 4.0.0b8 to 4.0.0b11 for both SM90 and SM100+ branches, matching the minimum version required for _validate_head_dims and the new hdim=256 dedicated kernel.

_{Reviews (4): Last reviewed commit: "Merge branch 'main' into xiny/headdim256..." | Re-trigger Greptile}

[yaox12]

/te-ci pytorch L3

[yaox12]

@vcherepanov-nv @KshitijLakhani Please review.

[sudhakarsingh27]

Should this still be checked when FlashAttentionUtils.v4_validate_head_dims == None?

[yaox12]

I double checked that this is a bug of FA4. Kernels produce wrong results on these shapes but they're allowed by v4_validate_head_dims, so we have to filter them out manually.
Raise an issue to FA4. Dao-AILab/flash-attention#2552

[vcherepanov-nv]

LGTM

[yaox12]

/te-ci pytorch L3