ggml-cpu: enable tiled flash attention for non-vector-multiple head dims on x86

[SongXiaoXi]

Overview

The CPU tiled flash attention kernel is gated on the value head size being a multiple of the SIMD vector width. Head sizes like 72 fall back to the per-query kernel, which is an order of magnitude slower and also accumulates P*V in F16.

This PR removes the restriction on 64-bit x86 and handles the leftover columns of the P*V tile GEMM with masked AVX-512 loads/stores/FMA instead of scalar per-column dot products. The behavior is unchanged outside 64-bit x86.

Additional information

FA operator, 16 heads, F16 K/V, Ryzen 9 9950X, 16 threads:

D	Q=KV	master, ms	this PR, ms	speedup
40	784	34.48	1.61	21.4x
40	3136	548.90	19.66	27.9x
72	784	25.13	2.03	12.4x
72	3136	399.84	24.85	16.1x
72	5776	1393.14	84.70	16.4x

Qwen3-VL BF16 mmproj on CPU:

image	master, ms	this PR, ms	speedup
448x448	967.9	338.7	2.86x
896x896	13147.9	1822.8	7.21x
1120x1120	28878.4	3487.1	8.28x

tested on x86_64 only.

Correctness: 225/225 CPU FA backend tests pass, including a new DK=65/DV=67 case with mask, sinks and ALiBi.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - benchmarks and test.

[ggml-gh-bot]

Hi @SongXiaoXi, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 2 open PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[am17an]

/bot review

[ggml-gh-bot]

❌ Code review failed.

Error: failed to fetch diff (503) from https://github.com/ggml-org/llama.cpp/pull/29423.diff

[SongXiaoXi]

/bot review

[ggml-gh-bot]

Sorry @SongXiaoXi, only members of the ggml-org organization can trigger /bot commands.

[SongXiaoXi]

Hi @am17an, is there anything I can do to get the bot review working or help move this PR forward?

[am17an]

/bot review

[ggml-gh-bot]

❌ Code review failed.

Error: failed to fetch diff (503) from https://github.com/ggml-org/llama.cpp/pull/29423.diff

[SongXiaoXi]

Thanks for retrying! It failed with the same failed to fetch diff (503) error again. Is this likely a GitHub issue or a bot issue? I can rebase the branch or try anything else you think might help.

[am17an]

Idk. Time to retire the bot. Generated from one that works. Fix this one and it should be good with me.

---

[P2] Soft-capping gives padded KV entries nonzero attention weight — ops.cpp:9323
Removing the dimension guard exposes an existing tiled-kernel bug to previously correct inputs: padding is set to -INFINITY before soft-capping, which converts it to finite -logit_softcap. Without a mask, those nonexistent keys contribute to softmax.
Reproduced with DK=65, DV=67, Q=75, KV=17, four heads, zero Q/K, unit V, no mask, and logit_softcap=1:

• Before PR: 1.000000
• PR, AVX2 and AVX-512: 0.495767
  Apply the padding exclusion after soft-capping and add an unmasked regression case. The new test uses a mask and disables soft-capping, so it misses this.
  Otherwise, validation passed: 1,011 selected attention tests on each ISA, another 224 head-size-40 tests on AVX-512, and 6,750 GEMM shapes per SSE3/AVX/AVX2/AVX-512 configuration with protected-page boundary checks.