[Executorch] Add non-flash SDPA for decode

[kimishpatel]

Stack from ghstack (oldest at bottom):

• temp changes to enable benchmarking on device via adb #18719
• Add xplat cxx_binary target for bench_transposed_cache benchmark #18718
• Add non-flash SDPA path gated by ET_USE_UNFUSED_SDPA #18717
• Add ONNX Runtime GQA-style SDPA benchmark #18716
• Add GEMM-based standard SDPA benchmark and fix custom_sdpa_out signature #18715
• Support separate K/V seq dim in custom_sdpa op #18714
• Add --no_transposed_cache CLI flag for export pipeline #18713
• Plumb transposed cache config through export pipeline #18712
• Add benchmarks for transposed vs standard KV cache layout #18711
• temp hack #18710
• Add support for transposed kv cache in all variants of custom kv cache #18709
• Add contig requirement to aot binding for llm custom ops #18708
• Add support for seq at dim=2 #18707
• Make kv cache update op support transposed cache #18706
• Add xplat cxx_binary target for bench_sdpa #18705
• Use unfused SDPA for short sequences (q_len <= 128 or kv_len <= 128) #18651
• Limit softmax to causally-valid elements in cpu_sdpa #18650
• Add quantized input support to cpu_sdpa #18649
• -> [Executorch] Add non-flash SDPA for decode #18648
• Add ONNX Runtime GQA-style SDPA benchmark #18647
• Add GEMM-based standard SDPA benchmark #18646
• Fix SLEEF preprocessor macro name to match ATen vec headers #18645

Add cpu_sdpa template function in op_sdpa_impl.h that provides a
simpler SDPA implementation using standard GEMM (no tiling). This is
useful as a baseline and for cases where flash attention is not optimal.

The implementation uses a single SeqDim parameter for all tensors and
supports causal masking, attention masks, GQA, and multi-threading.

During decode (seq_len == 1), the tiled flash attention implementation
has unnecessary overhead from its blocking/tiling logic. The simpler
unfused SDPA path using direct GEMM is more efficient for single-query
attention, yielding ~25-30% decode throughput improvement on S25
(41 -> 53 tok/s for 1.4B parameter model).

This makes cpu_sdpa always available (previously gated behind
ET_USE_UNFUSED_SDPA) and dispatches to it when seq_len == 1 and
inputs are not quantized. Prefill continues to use flash attention.

Differential Revision: D96044318

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18648

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure, 2 Cancelled Jobs

As of commit beb7b11 with merge base fb1618e ():

NEW FAILURE - The following job has failed:

• Lint / lintrunner (gh)
  >>> Lint for extension/llm/custom_ops/op_sdpa_impl.h:

CANCELLED JOBS - The following jobs were cancelled. Please retry:

• pull / unittest / windows / windows-job (gh)
• pull / unittest-editable / windows / windows-job (gh)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[digantdesai]

Review automatically exported from Phabricator review in Meta.