Split KV on MHA and Attention ops (#18007)

### Description
Implement Split KV optimization for FlashAttention in MHA and Attention
operators.

### Motivation and Context
Can help further accelerate these ops.

onnxruntime/contrib_ops/cpu/bert/attention_common.h

@@ -55,6 +55,7 @@ struct AttentionParameters {
   int v_hidden_size;          // hidden size of V
   int v_head_size;            // hidden size per head of V
   int num_heads;
+  int num_splits;
   bool is_unidirectional;
   bool past_present_share_buffer;
   bool do_rotary;
@@ -95,9 +96,9 @@ struct GroupQueryAttentionParameters {
   int head_size;
   int kv_hidden_size;
   int kv_num_heads;
+  int num_splits;          // number of splits for splitkv
   bool is_unidirectional;  // causal
   float scale;
-  int num_splits;  // number of splits for splitkv
   AttentionQkvFormat qkv_format;
   AttentionQkvFormat past_kv_format;
 };

onnxruntime/contrib_ops/cuda/bert/attention.cc

@@ -135,8 +135,24 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
   if (use_flash_attention && parameters.sequence_length < min_seq_len_for_flash_attention_packed_qkv_) {
     use_flash_attention = false;
   }
+  // Allocate buffers
+  size_t softmax_lse_accum_bytes = 0;
+  size_t out_accum_bytes = 0;
+  if (use_flash_attention) {
+    using namespace std;
+    auto [num_splits, slse_accum_bytes, o_accum_bytes] = onnxruntime::flash::get_num_splits_and_buffer_sizes(
+        parameters.batch_size, parameters.sequence_length, parameters.kv_sequence_length, parameters.num_heads,
+        parameters.head_size, device_prop.multiProcessorCount);
+    parameters.num_splits = num_splits;
+    softmax_lse_accum_bytes = slse_accum_bytes;
+    out_accum_bytes = o_accum_bytes;
+  }
+  auto softmax_lse_accum_buffer = GetScratchBuffer<void>(softmax_lse_accum_bytes, context->GetComputeStream());
+  auto out_accum_buffer = GetScratchBuffer<void>(out_accum_bytes, context->GetComputeStream());
 #else
   constexpr bool use_flash_attention = false;
+  auto softmax_lse_accum_buffer = GetScratchBuffer<void>(0, context->GetComputeStream());  // nullptr
+  auto out_accum_buffer = GetScratchBuffer<void>(0, context->GetComputeStream());          // nullptr
 #endif
 
   if (!use_flash_attention) {
@@ -273,6 +289,12 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
   data.fused_runner = reinterpret_cast<void*>(fused_runner);
   data.use_flash_attention = use_flash_attention;
   data.use_memory_efficient_attention = use_memory_efficient_attention;
+  if (softmax_lse_accum_buffer != nullptr) {
+    data.softmax_lse_accum = reinterpret_cast<CudaT*>(softmax_lse_accum_buffer.get());
+  }
+  if (out_accum_buffer != nullptr) {
+    data.out_accum = reinterpret_cast<CudaT*>(out_accum_buffer.get());
+  }
 
   return QkvToContext<CudaT>(device_prop, cublas, context->GetComputeStream(), parameters, data);
 }

onnxruntime/contrib_ops/cuda/bert/attention_impl.cu

@@ -316,7 +316,9 @@ Status FlashAttention(
   ORT_RETURN_IF_ERROR(onnxruntime::flash::mha_fwd(
       device_prop, stream, query, key, value, data.output, reinterpret_cast<void*>(data.scratch),
       parameters.batch_size, parameters.num_heads, parameters.num_heads, parameters.head_size,
-      parameters.sequence_length, parameters.total_sequence_length, scale, parameters.is_unidirectional));
+      parameters.sequence_length, parameters.total_sequence_length, scale, parameters.is_unidirectional,
+      parameters.num_splits, reinterpret_cast<void*>(data.softmax_lse_accum), reinterpret_cast<void*>(data.out_accum),
+      true));
 
   DUMP_TENSOR("flash attention output", data.output,
               parameters.batch_size, parameters.sequence_length, parameters.num_heads, parameters.v_head_size);

onnxruntime/contrib_ops/cuda/bert/attention_impl.h

@@ -88,6 +88,11 @@ struct AttentionData {
   T* v = nullptr;
   T* scratch = nullptr;
   AttentionQkvFormat qkv_format = AttentionQkvFormat::Q_K_V_BSNH;
+
+  // Flash buffers
+  T* softmax_lse = nullptr;
+  T* softmax_lse_accum = nullptr;
+  T* out_accum = nullptr;
 };
 
 template <typename T>

onnxruntime/contrib_ops/cuda/bert/flash_attention/flash_api.cc

@@ -140,11 +140,10 @@ void run_mha_fwd(Flash_fwd_params& params, cudaStream_t stream, bool force_split
 // So we find the best efficiency, then find the smallest number of splits that gets 85%
 // of the best efficiency.
 int num_splits_heuristic(int batch_size, int seqlen_q, int seqlen_k, int num_heads, int head_size, int num_SMs,
-                         int max_splits, bool new_kv, bool is_sm8x) {
+                         int max_splits) {
   // This needs to match with run_mha_fwd_splitkv_dispatch
-  const int block_n = is_sm8x ? (head_size <= 64 ? 256 : (head_size <= 128 ? 128 : 64))
-                              : (head_size <= 64 ? 256 : (head_size <= 160 ? 128 : 64));
-  const int num_n_blocks = (seqlen_k + (!new_kv ? 0 : seqlen_q) + block_n - 1) / block_n;
+  const int block_n = head_size <= 64 ? 256 : (head_size <= 128 ? 128 : 64);
+  const int num_n_blocks = (seqlen_k + block_n - 1) / block_n;
   // Technically kBlockM = 64 only for the splitKV kernels, not the standard kernel.
   // In any case we don't expect seqlen_q to be larger than 64 for inference.
   const int num_m_blocks = (seqlen_q + 64 - 1) / 64;
@@ -190,6 +189,26 @@ int num_splits_heuristic(int batch_size, int seqlen_q, int seqlen_k, int num_hea
   return 1;
 }
 
+// Returns (num_splits, softmax_lse_accum bytes, out_accum bytes)
+std::tuple<int, int, int> get_num_splits_and_buffer_sizes(int batch_size, int seqlen_q, int seqlen_k, int num_heads,
+                                                          int head_size, int num_SMs) {
+  int max_splits = 128;
+  // split kv buffers
+  int num_splits = num_splits_heuristic(batch_size, seqlen_q, seqlen_k, num_heads, head_size,
+                                        num_SMs, max_splits);
+  if (num_splits > 1) {
+    // softmax_lse_accum buffer
+    int softmax_lse_accum_bytes = get_softmax_lse_accum_size(num_splits, batch_size, num_heads, seqlen_q);
+    // out_accum buffer
+    auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
+    const int head_size_rounded = round_multiple(head_size, 32);
+    int out_accum_bytes = get_out_accum_size(num_splits, batch_size, num_heads, seqlen_q, head_size_rounded);
+    return {num_splits, softmax_lse_accum_bytes, out_accum_bytes};
+  } else {
+    return {0, 0, 0};
+  }
+}
+
 Status mha_fwd(const cudaDeviceProp& dprops,
                cudaStream_t stream,
                void* q,            // batch_size x seqlen_q x num_heads x head_size

onnxruntime/contrib_ops/cuda/bert/flash_attention/flash_api.h

@@ -31,6 +31,7 @@
 #if USE_FLASH_ATTENTION
 
 #include "core/providers/cuda/cuda_common.h"
+#include <tuple>
 
 namespace onnxruntime {
 namespace flash {
@@ -99,10 +100,9 @@ Status mha_fwd_kvcache(const cudaDeviceProp& dprops,
 );
 
 size_t get_softmax_lse_size(int max_seqlen_q, int batch_size, int num_heads);
-size_t get_softmax_lse_accum_size(int num_splits, int batch_size, int num_heads, int seqlen_q);
-size_t get_out_accum_size(int num_splits, int batch_size, int num_heads, int seqlen_q, int head_size_rounded);
 
-int num_splits_heuristic(int batch_size, int seqlen_q, int seqlen_k, int num_heads, int head_size, int num_SMs, int max_splits, bool new_kv, bool is_sm8x);
+std::tuple<int, int, int> get_num_splits_and_buffer_sizes(int batch_size, int seqlen_q, int seqlen_k, int num_heads,
+                                                          int head_size, int num_SMs);
 
 bool is_supported(const cudaDeviceProp& dprops, int head_size, int num_heads, int num_heads_k);
 

onnxruntime/contrib_ops/cuda/bert/flash_attention/flash_fwd_launch_template.h

@@ -123,17 +123,9 @@ void run_flash_splitkv_fwd(Flash_fwd_params& params, cudaStream_t stream) {
 
 template <typename T, int Headdim>
 void run_mha_fwd_splitkv_dispatch(Flash_fwd_params& params, cudaStream_t stream) {
-  bool is_sm8x = params.dprops->major == 8 && params.dprops->minor > 0;
   constexpr int kBlockM = 64;  // Fixed for all head dimensions
-  if (!is_sm8x) {              // A100, H100
-    // TD [2023-08-28]: nvcc segfaults for headdim 96 with block size 64 x 256,
-    // and for headdim 192 with block size 64 x 128.
-    constexpr int kBlockN = Headdim <= 64 ? 256 : (Headdim <= 160 ? 128 : 64);
-    run_flash_splitkv_fwd<Flash_fwd_kernel_traits<Headdim, kBlockM, kBlockN, 4, false, false, T>>(params, stream);
-  } else {  // Only 99KB of smem, so we have to set kBlockN smaller for Headdim 160 and above
-    constexpr int kBlockN = Headdim <= 64 ? 256 : (Headdim <= 128 ? 128 : 64);
-    run_flash_splitkv_fwd<Flash_fwd_kernel_traits<Headdim, kBlockM, kBlockN, 4, false, false, T>>(params, stream);
-  }
+  constexpr int kBlockN = Headdim <= 64 ? 256 : (Headdim <= 128 ? 128 : 64);
+  run_flash_splitkv_fwd<Flash_fwd_kernel_traits<Headdim, kBlockM, kBlockN, 4, false, false, T>>(params, stream);
 }
 
 template <typename T>

onnxruntime/contrib_ops/cuda/bert/group_query_attention.cc

@@ -116,22 +116,16 @@ Status GroupQueryAttention<T>::ComputeInternal(OpKernelContext* context) const {
   size_t out_accum_bytes = 0;
   size_t seqlens_k_bytes = 0;
   if (use_flash_attention) {
+    // softmax buffer
     softmax_lse_bytes = onnxruntime::flash::get_softmax_lse_size(parameters.sequence_length, parameters.batch_size, parameters.num_heads);
-    // split kv buffers
-    parameters.num_splits = onnxruntime::flash::num_splits_heuristic(
+    // split kv buffer
+    using namespace std;
+    auto [num_splits, slse_accum_bytes, o_accum_bytes] = onnxruntime::flash::get_num_splits_and_buffer_sizes(
         parameters.batch_size, parameters.sequence_length, parameters.kv_sequence_length, parameters.num_heads,
-        parameters.head_size, device_prop.multiProcessorCount, 128, false,
-        device_prop.major == 8 && device_prop.minor > 0);
-    if (parameters.num_splits > 1) {
-      // softmax_lse_accum buffer
-      softmax_lse_accum_bytes = onnxruntime::flash::get_softmax_lse_accum_size(
-          parameters.num_splits, parameters.batch_size, parameters.num_heads, parameters.sequence_length);
-      // out_accum buffer
-      auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
-      const int head_size_rounded = round_multiple(parameters.head_size, 32);
-      out_accum_bytes = onnxruntime::flash::get_out_accum_size(
-          parameters.num_splits, parameters.batch_size, parameters.num_heads, parameters.sequence_length, head_size_rounded);
-    }
+        parameters.head_size, device_prop.multiProcessorCount);
+    parameters.num_splits = num_splits;
+    softmax_lse_accum_bytes = slse_accum_bytes;
+    out_accum_bytes = o_accum_bytes;
     // seqlens_k buffer
     if (past_key != nullptr) {
       seqlens_k_bytes = sizeof(int) * parameters.batch_size;

onnxruntime/contrib_ops/cuda/bert/multihead_attention.cc

@@ -153,8 +153,24 @@ Status MultiHeadAttention<T>::ComputeInternal(OpKernelContext* context) const {
       parameters.sequence_length < min_seq_len_for_flash_attention_packed_qkv_) {
     use_flash_attention = false;
   }
+  // Allocate buffers
+  size_t softmax_lse_accum_bytes = 0;
+  size_t out_accum_bytes = 0;
+  if (use_flash_attention) {
+    using namespace std;
+    auto [num_splits, slse_accum_bytes, o_accum_bytes] = onnxruntime::flash::get_num_splits_and_buffer_sizes(
+        parameters.batch_size, parameters.sequence_length, parameters.kv_sequence_length, parameters.num_heads,
+        parameters.head_size, device_prop.multiProcessorCount);
+    parameters.num_splits = num_splits;
+    softmax_lse_accum_bytes = slse_accum_bytes;
+    out_accum_bytes = o_accum_bytes;
+  }
+  auto softmax_lse_accum_buffer = GetScratchBuffer<void>(softmax_lse_accum_bytes, context->GetComputeStream());
+  auto out_accum_buffer = GetScratchBuffer<void>(out_accum_bytes, context->GetComputeStream());
 #else
   constexpr bool use_flash_attention = false;
+  auto softmax_lse_accum_buffer = GetScratchBuffer<void>(0, context->GetComputeStream());  // nullptr
+  auto out_accum_buffer = GetScratchBuffer<void>(0, context->GetComputeStream());          // nullptr
 #endif
 
   bool use_fused_cross_attention = !use_flash_attention &&
@@ -289,6 +305,12 @@ Status MultiHeadAttention<T>::ComputeInternal(OpKernelContext* context) const {
   data.use_memory_efficient_attention = use_memory_efficient_attention;
   data.cumulated_sequence_length_q_cache = &(this->cumulated_sequence_length_q_cache_);
   data.cumulated_sequence_length_kv_cache = &(this->cumulated_sequence_length_kv_cache_);
+  if (softmax_lse_accum_buffer != nullptr) {
+    data.softmax_lse_accum = reinterpret_cast<CudaT*>(softmax_lse_accum_buffer.get());
+  }
+  if (out_accum_buffer != nullptr) {
+    data.out_accum = reinterpret_cast<CudaT*>(out_accum_buffer.get());
+  }
 
   cublasHandle_t cublas = GetCublasHandle(context);