Add quantized input support to cpu_sdpa

[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

cpu_sdpa (unfused SDPA) previously only supported float inputs.
When the model uses quantized Q/K/V (int8 with per-channel scales
and zero_points), decode fell back to cpu_flash_attention, missing
the ~25-30% throughput improvement from unfused SDPA.

This adds quantized support to cpu_sdpa by:

• Accepting optional quantization params (zero_points, scales for Q/K/V)
• Using _q_at_k_gemm for Q@K^T (handles both int8 and float)
• Using _qk_at_v_gemm for scores@V (handles both int8 and float)
• Applying scaling factor separately (fused with mask add or max reduction)
• Allocating a dequantization buffer for V when quantized

The dispatch in op_sdpa.cpp is updated to route quantized decode
(seq_len==1) through cpu_sdpa instead of cpu_flash_attention.

Differential Revision: D96044310

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18649

• 📄 Preview Python docs built from this PR

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

❌ 2 New Failures, 3 Cancelled Jobs

As of commit 01ab6c4 with merge base fb1618e ():

NEW FAILURES - The following jobs have failed:

• Build Presets / apple (llm) / build (gh)
• 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-buck / macos / macos-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.