Add ONNX Runtime GQA-style SDPA benchmark

Pull Request resolved: #18647

Add run_onnx_gqa_sdpa() which faithfully ports the algorithm from
onnxruntime/contrib_ops/cpu/bert/gqa_attention_base.h:
  - Scale baked into GEMM alpha (no separate scaling pass)
  - Scores buffer padded to max_seq_len columns
  - Causal mask: zero out future positions, softmax on valid window only
  - Output always in [B, S, Hq, D] format

Extends validation to verify ONNX GQA output matches custom_sdpa_out
reference. Adds OnnxGQABenchFixture for benchmarking both layouts.
ghstack-source-id: 374666318
@exported-using-ghexport

Differential Revision: [D96044317](https://our.internmc.facebook.com/intern/diff/D96044317/)

Author: kimishpatel

extension/llm/custom_ops/bench_sdpa.cpp

@@ -172,17 +172,145 @@ void run_standard_sdpa(
       });
 }
 
+// ONNX Runtime GQA-style SDPA, faithfully ported from
+// onnxruntime/contrib_ops/cpu/bert/gqa_attention_base.h.
+// Differences from run_standard_sdpa:
+//   1. Scale in GEMM alpha (no separate scaling pass)
+//   2. Scores buffer padded to max_seq_len cols (ONNX's present_buffer_seq_len)
+//   3. Causal mask: zero out future positions, softmax on valid window only
+//   4. Output in [B, S, Hq, D] with stride Hq*D (ONNX's interleaved BNSH->BSNH)
+//
+// When is_transposed=true, inputs are [B,H,S,D]; output is [B,S,Hq,D].
+// When is_transposed=false, inputs are [B,S,H,D]; output is [B,S,Hq,D].
+// Output is always [B, S, Hq, D] to match ONNX's actual output format.
+void run_onnx_gqa_sdpa(
+    const float* q_data,
+    const float* k_data,
+    const float* v_data,
+    float* out_data, // always [B, q_seq_len, Hq, D]
+    float* scores_buf, // must hold batch*Hq*q_seq_len*max_seq_len floats
+    int64_t batch,
+    int64_t Hq,
+    int64_t Hkv,
+    int64_t D,
+    int64_t max_seq_len,
+    int64_t start_pos,
+    int64_t q_seq_len,
+    bool is_transposed) {
+  using executorch::cpublas::TransposeType;
+
+  const int64_t total_seqlen = start_pos + q_seq_len;
+  const float alpha = 1.0f / std::sqrt(static_cast<float>(D));
+  const int64_t heads_per_group = Hq / Hkv;
+  const int64_t hidden_size = Hq * D; // output row stride (ONNX convention)
+
+  // Input strides depend on layout
+  const int64_t ldq = is_transposed ? D : Hq * D;
+  const int64_t ldk = is_transposed ? D : Hkv * D;
+  const int64_t ldv = is_transposed ? D : Hkv * D;
+  // Output is always [B, S, Hq, D] so ldo = Hq * D = hidden_size
+  const int64_t ldo = hidden_size;
+
+  torch::executor::parallel_for(
+      0, batch * Hq, 1, [&](int64_t begin, int64_t end) {
+        for (int64_t idx = begin; idx < end; ++idx) {
+          const int64_t b = idx / Hq;
+          const int64_t h = idx % Hq;
+          const int64_t kv_h = h / heads_per_group;
+
+          const float* q_ptr;
+          const float* k_ptr;
+          const float* v_ptr;
+          if (is_transposed) {
+            q_ptr = q_data + (b * Hq + h) * q_seq_len * D;
+            k_ptr = k_data + (b * Hkv + kv_h) * max_seq_len * D;
+            v_ptr = v_data + (b * Hkv + kv_h) * max_seq_len * D;
+          } else {
+            q_ptr = q_data + b * q_seq_len * Hq * D + h * D;
+            k_ptr = k_data + b * max_seq_len * Hkv * D + kv_h * D;
+            v_ptr = v_data + b * max_seq_len * Hkv * D + kv_h * D;
+          }
+          // Output always [B, S, Hq, D]: head h writes at stride hidden_size
+          float* out_ptr = out_data + b * q_seq_len * hidden_size + h * D;
+
+          // Scores padded to max_seq_len columns (ONNX convention)
+          float* scores = scores_buf + idx * q_seq_len * max_seq_len;
+
+          // GEMM 1: Q @ K^T with scale in alpha
+          executorch::cpublas::gemm(
+              TransposeType::Transpose,
+              TransposeType::NoTranspose,
+              total_seqlen,
+              q_seq_len,
+              D,
+              alpha,
+              k_ptr,
+              ldk,
+              q_ptr,
+              ldq,
+              0.0f,
+              scores,
+              max_seq_len);
+
+          // Causal mask + narrow softmax (ONNX style):
+          // Zero future positions, softmax only on valid [0, causal_len).
+          for (int64_t qi = 0; qi < q_seq_len; ++qi) {
+            float* row = scores + qi * max_seq_len;
+            const int64_t causal_len =
+                std::min(start_pos + qi + 1, total_seqlen);
+
+            for (int64_t j = causal_len; j < total_seqlen; ++j) {
+              row[j] = 0.0f;
+            }
+
+            float max_val = row[0];
+            for (int64_t j = 1; j < causal_len; ++j) {
+              max_val = std::max(max_val, row[j]);
+            }
+            float sum = 0.0f;
+            for (int64_t j = 0; j < causal_len; ++j) {
+              row[j] = std::exp(row[j] - max_val);
+              sum += row[j];
+            }
+            const float inv_sum = 1.0f / sum;
+            for (int64_t j = 0; j < causal_len; ++j) {
+              row[j] *= inv_sum;
+            }
+          }
+
+          // GEMM 2: scores @ V -> output
+          executorch::cpublas::gemm(
+              TransposeType::NoTranspose,
+              TransposeType::NoTranspose,
+              D,
+              q_seq_len,
+              total_seqlen,
+              1.0f,
+              v_ptr,
+              ldv,
+              scores,
+              max_seq_len,
+              0.0f,
+              out_ptr,
+              ldo);
+        }
+      });
+}
+
 // Return max |a - b| across all elements.
-float max_abs_diff(const Tensor& a, const Tensor& b) {
-  const float* a_data = a.const_data_ptr<float>();
-  const float* b_data = b.const_data_ptr<float>();
+float max_abs_diff(const float* a, const float* b, int64_t n) {
   float d = 0.0f;
-  for (int64_t i = 0; i < a.numel(); ++i) {
-    d = std::max(d, std::abs(a_data[i] - b_data[i]));
+  for (int64_t i = 0; i < n; ++i) {
+    d = std::max(d, std::abs(a[i] - b[i]));
   }
   return d;
 }
 
+float max_abs_diff(const Tensor& a, const Tensor& b) {
+  return max_abs_diff(
+      a.const_data_ptr<float>(), b.const_data_ptr<float>(), a.numel());
+}
+
 // Validate a single config: run StandardSDPA and custom_sdpa_out on the same
 // inputs, check outputs match within tolerance. Returns false on mismatch.
 // Only tests standard [B,S,H,D] layout (is_transposed=false).
@@ -268,6 +396,65 @@ bool validate_config(
       (long)q_seq_len,
       diff);
 
+  // Also validate ONNX GQA variant. Output is always [B, S, Hq, D].
+  // Since we only test standard [B,S,H,D] layout, out_ref is already
+  // [B,S,Hq,D] — just copy directly to ref_bshd (no transpose needed).
+  Tensor out_onnx =
+      tf.zeros({(int32_t)batch, (int32_t)q_seq_len, (int32_t)Hq, (int32_t)D});
+  std::vector<float> onnx_scores_buf(batch * Hq * q_seq_len * max_seq_len);
+  run_onnx_gqa_sdpa(
+      q.const_data_ptr<float>(),
+      k.const_data_ptr<float>(),
+      v.const_data_ptr<float>(),
+      out_onnx.mutable_data_ptr<float>(),
+      onnx_scores_buf.data(),
+      batch,
+      Hq,
+      Hkv,
+      D,
+      max_seq_len,
+      start_pos,
+      q_seq_len,
+      false /* is_transposed */);
+
+  // out_ref is already [B, S, Hq, D] (standard layout), compare directly
+  std::vector<float> ref_bshd(batch * q_seq_len * Hq * D);
+  const float* ref_ptr = out_ref.const_data_ptr<float>();
+  std::copy(ref_ptr, ref_ptr + batch * q_seq_len * Hq * D, ref_bshd.data());
+
+  float onnx_diff = max_abs_diff(
+      out_onnx.const_data_ptr<float>(),
+      ref_bshd.data(),
+      batch * q_seq_len * Hq * D);
+  if (onnx_diff > atol) {
+    fprintf(
+        stderr,
+        "FAIL: OnnxGQA standard %s (B=%ld Hq=%ld Hkv=%ld D=%ld sp=%ld sl=%ld) "
+        "max_abs_diff=%.6e > atol=%.6e\n",
+        mode,
+        (long)batch,
+        (long)Hq,
+        (long)Hkv,
+        (long)D,
+        (long)start_pos,
+        (long)q_seq_len,
+        onnx_diff,
+        atol);
+    return false;
+  }
+  fprintf(
+      stderr,
+      "PASS: OnnxGQA standard %s (B=%ld Hq=%ld Hkv=%ld D=%ld sp=%ld sl=%ld) "
+      "max_abs_diff=%.6e\n",
+      mode,
+      (long)batch,
+      (long)Hq,
+      (long)Hkv,
+      (long)D,
+      (long)start_pos,
+      (long)q_seq_len,
+      onnx_diff);
+
   return true;
 }
 
@@ -517,6 +704,132 @@ BENCHMARK_DEFINE_F(StandardSDPABenchFixture, StandardSDPA)
   }
 }
 
+// ONNX Runtime GQA-style benchmark. Faithfully matches the algorithm from
+// gqa_attention_base.h: scale-in-alpha, padded scores buffer, narrow softmax,
+// and output in [B, S, Hq, D] with stride Hq*D.
+class OnnxGQABenchFixture : public benchmark::Fixture {
+ public:
+  // Args: {batch, num_heads_q, num_heads_kv, head_dim, max_seq_len, start_pos,
+  //        query_seq_len, is_transposed}
+  void SetUp(benchmark::State& state) override {
+    int64_t batch = state.range(0);
+    int64_t num_heads_q = state.range(1);
+    int64_t num_heads_kv = state.range(2);
+    int64_t head_dim = state.range(3);
+    int64_t max_seq_len = state.range(4);
+    int64_t start_pos = state.range(5);
+    int64_t q_seq_len = state.range(6);
+    bool is_transposed = state.range(7) != 0;
+
+    std::mt19937 gen(42);
+
+    if (is_transposed) {
+      q_.emplace(tf_.zeros(
+          {(int32_t)batch,
+           (int32_t)num_heads_q,
+           (int32_t)q_seq_len,
+           (int32_t)head_dim}));
+      k_cache_.emplace(tf_.zeros(
+          {(int32_t)batch,
+           (int32_t)num_heads_kv,
+           (int32_t)max_seq_len,
+           (int32_t)head_dim}));
+      v_cache_.emplace(tf_.zeros(
+          {(int32_t)batch,
+           (int32_t)num_heads_kv,
+           (int32_t)max_seq_len,
+           (int32_t)head_dim}));
+    } else {
+      q_.emplace(tf_.zeros(
+          {(int32_t)batch,
+           (int32_t)q_seq_len,
+           (int32_t)num_heads_q,
+           (int32_t)head_dim}));
+      k_cache_.emplace(tf_.zeros(
+          {(int32_t)batch,
+           (int32_t)max_seq_len,
+           (int32_t)num_heads_kv,
+           (int32_t)head_dim}));
+      v_cache_.emplace(tf_.zeros(
+          {(int32_t)batch,
+           (int32_t)max_seq_len,
+           (int32_t)num_heads_kv,
+           (int32_t)head_dim}));
+    }
+    // Output always [B, S, Hq, D] (ONNX convention)
+    output_.emplace(tf_.zeros(
+        {(int32_t)batch,
+         (int32_t)q_seq_len,
+         (int32_t)num_heads_q,
+         (int32_t)head_dim}));
+
+    fill_random(*q_, gen);
+    fill_random(*k_cache_, gen);
+    fill_random(*v_cache_, gen);
+
+    batch_ = batch;
+    num_heads_q_ = num_heads_q;
+    num_heads_kv_ = num_heads_kv;
+    head_dim_ = head_dim;
+    max_seq_len_ = max_seq_len;
+    start_pos_ = start_pos;
+    q_seq_len_ = q_seq_len;
+    is_transposed_ = is_transposed;
+
+    // Scores buffer padded to max_seq_len columns (ONNX convention)
+    int64_t total_units = batch * num_heads_q;
+    scores_buf_.resize(total_units * q_seq_len * max_seq_len);
+  }
+
+  void TearDown(benchmark::State&) override {
+    q_.reset();
+    k_cache_.reset();
+    v_cache_.reset();
+    output_.reset();
+    scores_buf_.clear();
+  }
+
+  TensorFactory<ScalarType::Float> tf_;
+  std::optional<Tensor> q_;
+  std::optional<Tensor> k_cache_;
+  std::optional<Tensor> v_cache_;
+  std::optional<Tensor> output_;
+  std::vector<float> scores_buf_;
+  int64_t batch_ = 0;
+  int64_t num_heads_q_ = 0;
+  int64_t num_heads_kv_ = 0;
+  int64_t head_dim_ = 0;
+  int64_t max_seq_len_ = 0;
+  int64_t start_pos_ = 0;
+  int64_t q_seq_len_ = 0;
+  bool is_transposed_ = false;
+};
+
+BENCHMARK_DEFINE_F(OnnxGQABenchFixture, OnnxGQA)
+(benchmark::State& state) {
+  const float* q_data = q_->const_data_ptr<float>();
+  const float* k_data = k_cache_->const_data_ptr<float>();
+  const float* v_data = v_cache_->const_data_ptr<float>();
+  float* out_data = output_->mutable_data_ptr<float>();
+
+  for (auto _ : state) {
+    run_onnx_gqa_sdpa(
+        q_data,
+        k_data,
+        v_data,
+        out_data,
+        scores_buf_.data(),
+        batch_,
+        num_heads_q_,
+        num_heads_kv_,
+        head_dim_,
+        max_seq_len_,
+        start_pos_,
+        q_seq_len_,
+        is_transposed_);
+  }
+}
+
 /*
  * Benchmark configurations modeled after Llama 3 8B (GQA: 32 q heads, 8 kv
  * heads, head_dim=128). We test decode (seq_len=1) and prefill scenarios at
@@ -565,6 +878,33 @@ BENCHMARK_REGISTER_F(StandardSDPABenchFixture, StandardSDPA)
     ->Args({1, 32, 32, 128, 2048, 256, 1, 1})
     ->ArgNames({"B", "Hq", "Hkv", "D", "MaxS", "StartPos", "SeqLen", "Trans"});
 
+// --- ONNX Runtime GQA-style SDPA ---
+// Same configs as StandardSDPA. Differences: scale-in-alpha, padded scores
+// buffer (ld=MaxS), narrow softmax, output in [B,S,Hq,D] with stride Hq*D.
+BENCHMARK_REGISTER_F(OnnxGQABenchFixture, OnnxGQA)
+    // Standard layout decode at various cache positions
+    ->Args({1, 32, 8, 128, 2048, 0, 1, 0})
+    ->Args({1, 32, 8, 128, 2048, 64, 1, 0})
+    ->Args({1, 32, 8, 128, 2048, 256, 1, 0})
+    ->Args({1, 32, 8, 128, 2048, 512, 1, 0})
+    ->Args({1, 32, 8, 128, 2048, 1024, 1, 0})
+    // Transposed layout decode at same positions
+    ->Args({1, 32, 8, 128, 2048, 0, 1, 1})
+    ->Args({1, 32, 8, 128, 2048, 64, 1, 1})
+    ->Args({1, 32, 8, 128, 2048, 256, 1, 1})
+    ->Args({1, 32, 8, 128, 2048, 512, 1, 1})
+    ->Args({1, 32, 8, 128, 2048, 1024, 1, 1})
+    // Standard layout prefill
+    ->Args({1, 32, 8, 128, 2048, 0, 128, 0})
+    ->Args({1, 32, 8, 128, 2048, 0, 512, 0})
+    // Transposed layout prefill
+    ->Args({1, 32, 8, 128, 2048, 0, 128, 1})
+    ->Args({1, 32, 8, 128, 2048, 0, 512, 1})
+    // Llama 2 style (32 heads, no GQA)
+    ->Args({1, 32, 32, 128, 2048, 256, 1, 0})
+    ->Args({1, 32, 32, 128, 2048, 256, 1, 1})
+    ->ArgNames({"B", "Hq", "Hkv", "D", "MaxS", "StartPos", "SeqLen", "Trans"});
+
 } // namespace
 
 int main(int argc, char** argv) {