snnn
diff --git a/‎cmake/CMakeLists.txt
Lines changed: 17 additions & 0 deletions b/‎cmake/CMakeLists.txt
Lines changed: 17 additions & 0 deletions
diff --git a/‎onnxruntime/contrib_ops/cpu/bert/attention_common.h
Lines changed: 8 additions & 2 deletions b/‎onnxruntime/contrib_ops/cpu/bert/attention_common.h
Lines changed: 8 additions & 2 deletions
diff --git a/‎onnxruntime/contrib_ops/cuda/bert/attention.cc
Lines changed: 15 additions & 3 deletions b/‎onnxruntime/contrib_ops/cuda/bert/attention.cc
Lines changed: 15 additions & 3 deletions
diff --git a/‎onnxruntime/contrib_ops/cuda/bert/attention_impl.cu
Lines changed: 93 additions & 3 deletions b/‎onnxruntime/contrib_ops/cuda/bert/attention_impl.cu
Lines changed: 93 additions & 3 deletions
diff --git a/‎onnxruntime/contrib_ops/cuda/bert/attention_impl.h
Lines changed: 15 additions & 0 deletions b/‎onnxruntime/contrib_ops/cuda/bert/attention_impl.h
Lines changed: 15 additions & 0 deletions
diff --git a/‎onnxruntime/contrib_ops/cuda/bert/attention_kernel_options.cc
Lines changed: 17 additions & 0 deletions b/‎onnxruntime/contrib_ops/cuda/bert/attention_kernel_options.cc
Lines changed: 17 additions & 0 deletions
@@ -106,6 +106,7 @@ option(onnxruntime_USE_LLVM "Build TVM with LLVM" OFF)
 option(onnxruntime_USE_VSINPU "Build with VSINPU support" OFF)
 
 cmake_dependent_option(onnxruntime_USE_FLASH_ATTENTION "Build flash attention kernel for scaled dot product attention" ON "onnxruntime_USE_CUDA" OFF)
+cmake_dependent_option(onnxruntime_USE_LEAN_ATTENTION "Build lean attention kernel for scaled dot product attention" ON "onnxruntime_USE_CUDA; NOT WIN32" OFF)
 option(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION "Build memory efficient attention kernel for scaled dot product attention" ON)
 
 option(onnxruntime_BUILD_FOR_NATIVE_MACHINE "Enable this option for turning on optimization specific to this machine" OFF)
@@ -751,21 +752,30 @@ if (onnxruntime_USE_CUDA)
 
   if (onnxruntime_DISABLE_CONTRIB_OPS)
     set(onnxruntime_USE_FLASH_ATTENTION OFF)
+    set(onnxruntime_USE_LEAN_ATTENTION OFF)
     set(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION OFF)
   endif()
+
   if (CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 11.6)
     message( STATUS "Turn off flash attention since CUDA compiler version < 11.6")
     set(onnxruntime_USE_FLASH_ATTENTION OFF)
+    set(onnxruntime_USE_LEAN_ATTENTION OFF)
     set(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION OFF)
   elseif(WIN32 AND CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 12)
     message( STATUS "Flash-Attention unsupported in Windows with CUDA compiler version < 12.0")
     set(onnxruntime_USE_FLASH_ATTENTION OFF)
   endif()
+
   if (CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 11.4)
     message( FATAL_ERROR "Failed build due to CUDA compiler version < 11.4")
   endif()
+  if (WIN32)
+    message( STATUS "Lean Attention unsupported in Windows")
+    set(onnxruntime_USE_LEAN_ATTENTION OFF)
+  endif()
 else()
   set(onnxruntime_USE_FLASH_ATTENTION OFF)
+  set(onnxruntime_USE_LEAN_ATTENTION OFF)
   set(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION OFF)
 endif()
 
@@ -779,6 +789,13 @@ if (onnxruntime_USE_CUDA)
       list(APPEND ORT_PROVIDER_FLAGS -DUSE_FLASH_ATTENTION=1)
       list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_FLASH_ATTENTION=1)
     endif()
+
+    if (onnxruntime_USE_LEAN_ATTENTION)
+      message( STATUS "Enable lean attention for CUDA EP")
+      list(APPEND ORT_PROVIDER_FLAGS -DUSE_LEAN_ATTENTION=1)
+      list(APPEND ORT_PROVIDER_CMAKE_FLAGS -Donnxruntime_USE_LEAN_ATTENTION=1)
+    endif()
+
     if (onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION)
       message( STATUS "Enable memory efficient attention for CUDA EP")
       list(APPEND ORT_PROVIDER_FLAGS -DUSE_MEMORY_EFFICIENT_ATTENTION=1)
 
@@ -48,6 +48,7 @@ enum AttentionKernelType {
   AttentionKernel_CutlassMemoryEfficientAttention,
   AttentionKernel_FlashAttention,
   AttentionKernel_CudnnFlashAttention,
+  AttentionKernel_LeanAttention,
   AttentionKernel_Default
 };
 
@@ -65,7 +66,6 @@ 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;
   int rotary_embedding;
   bool is_unidirectional;
   bool past_present_share_buffer;
@@ -208,10 +208,13 @@ enum class AttentionBackend : int {
   CUDNN_FLASH_ATTENTION = 8,  // reserved for cuDNN flash attention.
   MATH = 16,                  // unfused kernel cannot be disabled right now.
 
-  // The following kernels might be deprecated in the future.
+  // The following TRT kernels might be deprecated in the future.
   TRT_FLASH_ATTENTION = 32,
   TRT_CROSS_ATTENTION = 64,
   TRT_CAUSAL_ATTENTION = 128,
+
+  // Experimental kernels
+  LEAN_ATTENTION = 256,
 };
 
 // Environment variable to enable debug information of attention kernel to be printed. Default is 0 (disabled).
@@ -239,6 +242,9 @@ constexpr const char* kDisableMemoryEfficientAttention = "ORT_DISABLE_MEMORY_EFF
 // Environment variable to enable or disable flash attention. Default is 0 (enabled).
 constexpr const char* kDisableFlashAttention = "ORT_DISABLE_FLASH_ATTENTION";
 
+// Environment variable to enable or disable lean attention. Default is 0 (disabled).
+constexpr const char* kEnableLeanAttention = "ORT_ENABLE_LEAN_ATTENTION";
+
 // Minimum sequence length to perfer memory efficient attention when data type is float32
 constexpr const char* kMinSeqLenForEfficientAttentionFp32 = "ORT_MIN_SEQ_LEN_EFFICIENT_ATTENTION_FP32";
 
 
@@ -102,6 +102,9 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
   const int sm = device_prop.major * 10 + device_prop.minor;
   const bool is_mask_1d_seq_len = parameters.mask_type == AttentionMaskType::MASK_1D_KEY_SEQ_LEN;
 
+  typedef typename ToCudaType<T>::MappedType CudaT;
+  AttentionData<CudaT> data;
+
 #if USE_FLASH_ATTENTION
   bool use_flash_attention = !disable_flash_attention_ &&
                              (nullptr == attention_bias) &&
@@ -118,21 +121,26 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
     use_flash_attention = false;
   }
   // Allocate buffers
+  size_t softmax_lse_bytes = 0;
   size_t softmax_lse_accum_bytes = 0;
   size_t out_accum_bytes = 0;
   if (use_flash_attention) {
+    softmax_lse_bytes = onnxruntime::flash::get_softmax_lse_size(sequence_length, batch_size, parameters.num_heads);
+
     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.total_sequence_length, parameters.num_heads,
         parameters.head_size, device_prop.multiProcessorCount);
-    parameters.num_splits = static_cast<int>(num_splits);
+    data.num_splits = static_cast<int>(num_splits);
     softmax_lse_accum_bytes = slse_accum_bytes;
     out_accum_bytes = o_accum_bytes;
   }
+  auto softmax_lse_buffer = GetScratchBuffer<void>(softmax_lse_bytes, context->GetComputeStream());
   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_buffer = GetScratchBuffer<void>(0, context->GetComputeStream());
   auto softmax_lse_accum_buffer = GetScratchBuffer<void>(0, context->GetComputeStream());  // nullptr
   auto out_accum_buffer = GetScratchBuffer<void>(0, context->GetComputeStream());          // nullptr
 #endif
@@ -247,6 +255,7 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
   constexpr size_t element_size = sizeof(T);
   constexpr bool use_fused_cross_attention = false;
   constexpr bool use_cudnn_flash_attention = false;
+  constexpr bool use_lean_attention = false;
   size_t workSpaceSize = GetAttentionWorkspaceSize(element_size,
                                                    parameters.batch_size,
                                                    parameters.num_heads,
@@ -257,14 +266,13 @@ Status Attention<T>::ComputeInternal(OpKernelContext* context) const {
                                                    parameters.total_sequence_length,
                                                    fused_runner,
                                                    use_flash_attention,
+                                                   use_lean_attention,
                                                    use_fused_cross_attention,
                                                    use_memory_efficient_attention,
                                                    use_cudnn_flash_attention,
                                                    false);
   IAllocatorUniquePtr<void> work_space = IAllocator::MakeUniquePtr<void>(allocator, workSpaceSize, false, context->GetComputeStream());
 
-  typedef typename ToCudaType<T>::MappedType CudaT;
-  AttentionData<CudaT> data;
   data.gemm_buffer = reinterpret_cast<CudaT*>(gemm_buffer.get());
   if (nullptr != bias) {
     data.bias = reinterpret_cast<const CudaT*>(bias->Data<T>());
@@ -289,6 +297,10 @@ 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_buffer != nullptr) {
+    data.softmax_lse = reinterpret_cast<CudaT*>(softmax_lse_buffer.get());
+  }
+
   if (softmax_lse_accum_buffer != nullptr) {
     data.softmax_lse_accum = reinterpret_cast<CudaT*>(softmax_lse_accum_buffer.get());
   }
 
@@ -39,6 +39,7 @@ limitations under the License.
 #include "contrib_ops/cuda/bert/cutlass_fmha/memory_efficient_attention.h"
 #include "contrib_ops/cuda/bert/cudnn_fmha/cudnn_flash_attention.h"
 #include "contrib_ops/cuda/bert/flash_attention/flash_api.h"
+#include "contrib_ops/cuda/bert/lean_attention/lean_api.h"
 #include "contrib_ops/cuda/bert/attention_impl.h"
 
 using namespace onnxruntime::cuda;
@@ -108,6 +109,7 @@ size_t GetAttentionWorkspaceSize(
     size_t total_sequence_length,
     void* fused_runner,
     bool use_flash_attention,
+    bool use_lean_attention,
     bool use_fused_cross_attention,
     bool use_memory_efficient_attention,
     bool use_cudnn_flash_attention,
@@ -119,12 +121,20 @@ size_t GetAttentionWorkspaceSize(
 
 #if USE_FLASH_ATTENTION
   if (use_flash_attention) {
-    return qkv_bytes + onnxruntime::flash::get_softmax_lse_size(sequence_length, batch_size, num_heads);
+    return qkv_bytes;
   }
 #else
   ORT_UNUSED_PARAMETER(use_flash_attention);
 #endif
 
+#if USE_LEAN_ATTENTION
+  if (use_lean_attention) {
+    return qkv_bytes;
+  }
+#else
+  ORT_UNUSED_PARAMETER(use_lean_attention);
+#endif
+
 #if USE_MEMORY_EFFICIENT_ATTENTION
   if (use_memory_efficient_attention) {
     size_t fmha_buffer_bytes = 0;
@@ -301,10 +311,10 @@ Status FlashAttention(
 
   constexpr bool is_bf16 = false;
   ORT_RETURN_IF_ERROR(onnxruntime::flash::mha_fwd(
-      device_prop, stream, data.q, data.k, data.v, data.output, reinterpret_cast<void*>(data.scratch),
+      device_prop, stream, data.q, data.k, data.v, data.output, reinterpret_cast<void*>(data.softmax_lse),
       parameters.batch_size, parameters.num_heads, parameters.num_heads, parameters.head_size,
       parameters.sequence_length, parameters.total_sequence_length, scale, 0.0, parameters.is_unidirectional, is_bf16,
-      false, parameters.num_splits, reinterpret_cast<void*>(data.softmax_lse_accum),
+      false, data.num_splits, reinterpret_cast<void*>(data.softmax_lse_accum),
       reinterpret_cast<void*>(data.out_accum), data.qkv_format == AttentionQkvFormat::Q_K_V_BSNH));
 
   return Status::OK();
@@ -326,6 +336,81 @@ Status FlashAttention(
 }
 #endif
 
+#if USE_LEAN_ATTENTION
+template <typename T>
+Status LeanAttention(
+    const cudaDeviceProp& device_prop,
+    cudaStream_t stream,
+    contrib::AttentionParameters& parameters,
+    AttentionData<T>& data,
+    float scale) {
+  assert(data.qkv_format == AttentionQkvFormat::Q_K_V_BSNH ||
+         data.qkv_format == AttentionQkvFormat::Q_K_V_BSNH_BNSH_BNSH);
+  assert(nullptr == data.mask_index);
+  assert(nullptr == data.attention_bias);
+  assert(parameters.head_size == parameters.v_head_size);
+
+  constexpr bool is_bf16 = false;
+
+  ORT_RETURN_IF_ERROR(onnxruntime::lean::mha_fwd_kvcache(
+    device_prop, stream,
+    data.q,
+    data.k, // k_cache
+    data.v, // v_cache
+    nullptr,  // new_k (we have appended new_k to k_cache)
+    nullptr,  // new_v (we have appended new_v to k_cache)
+    data.output,
+    reinterpret_cast<void*>(data.softmax_lse),
+    nullptr, // seqlens_k
+    nullptr, // cos_cache
+    nullptr, // sin_cache
+    nullptr, // block_table
+    parameters.batch_size,
+    parameters.num_heads,
+    parameters.num_heads, // num_heads_k
+    parameters.head_size,
+    parameters.sequence_length, // seqlen_q
+    parameters.total_sequence_length, // seqlen_k
+    0, // seqlen_k_new
+    0, // rotary_dim
+    scale, // softmax_scale
+    parameters.is_unidirectional,
+    is_bf16,
+    false, // past_bsnh
+    data.num_splits,
+    data.grid_dim_z,
+    data.max_tiles_per_tb,
+    data.high_load_tbs,
+    data.tiles_per_head,
+    reinterpret_cast<void*>(data.softmax_lse_accum),
+    reinterpret_cast<void*>(data.out_accum),
+    data.lean_sync_flag,
+    -1, // local_window_size
+    false, // is_rotary_interleaved
+    false // is_packed_qkv
+    ));
+
+  return Status::OK();
+}
+
+template <>
+Status LeanAttention(
+    const cudaDeviceProp& device_prop,
+    cudaStream_t stream,
+    contrib::AttentionParameters& parameters,
+    AttentionData<float>& data,
+    float scale) {
+  ORT_UNUSED_PARAMETER(device_prop);
+  ORT_UNUSED_PARAMETER(stream);
+  ORT_UNUSED_PARAMETER(parameters);
+  ORT_UNUSED_PARAMETER(data);
+  ORT_UNUSED_PARAMETER(scale);
+  return ORT_MAKE_STATUS(ONNXRUNTIME, StatusCode::NOT_IMPLEMENTED, "lean attention does not support float tensor");
+}
+#endif
+
+
+
 template <typename T>
 Status CudnnFlashAttention(
     cudnnHandle_t cudnn_handle,
@@ -641,6 +726,11 @@ Status QkvToContext(
   // For raw attention mask, the scalar 1/sqrt(H) is moved to combine with softmax computation.
   const float scale = parameters.scale == 0.0f ? 1.f / sqrt(static_cast<float>(qk_head_size))
                                                : parameters.scale;
+#if USE_LEAN_ATTENTION
+  if (data.use_lean_attention) {
+    return LeanAttention(device_prop, stream, parameters, data, scale);
+  }
+#endif
 
 #if USE_FLASH_ATTENTION
   if (data.use_flash_attention) {
 
@@ -53,6 +53,7 @@ size_t GetAttentionWorkspaceSize(
     size_t total_sequence_length,
     void* fused_runner,
     bool use_flash_attention,
+    bool use_lean_attention,
     bool use_fused_cross_attention,
     bool use_memory_efficient_attention,
     bool use_cudnn_flash_attention,
@@ -102,6 +103,19 @@ struct AttentionData {
   T* softmax_lse_accum = nullptr;
   T* out_accum = nullptr;
 
+  // Flash Atttention and Lean Attention
+  int num_splits;
+
+  // Lean Attention
+  bool use_lean_attention = false;
+#if USE_LEAN_ATTENTION
+  int grid_dim_z = 0;
+  int max_tiles_per_tb = 0;
+  int high_load_tbs = 0;
+  int tiles_per_head = 0;
+  int* lean_sync_flag = nullptr;
+#endif
+
   // For Debugging
   size_t workspace_bytes = 0;
   bool allow_debug_info = false;
@@ -115,6 +129,7 @@ struct AttentionData {
 
   void PrintDebugInfo() const {
     std::cout << "flash=" << use_flash_attention
+              << ", lean=" << use_lean_attention
               << ", efficient=" << use_memory_efficient_attention
               << ", fused_runner=" << (fused_runner != nullptr)
               << ", fused_cross=" << (fused_cross_attention_kernel != nullptr)
 
@@ -17,6 +17,9 @@ namespace onnxruntime {
 void AttentionKernelOptions::Initialize(int value, bool use_build_flag, bool check_cudnn_version) {
   if (value > 0) {
     use_flash_attention_ = (value & static_cast<int>(AttentionBackend::FLASH_ATTENTION)) > 0;
+#if USE_LEAN_ATTENTION
+    use_lean_attention_ = (value & static_cast<int>(AttentionBackend::LEAN_ATTENTION)) > 0;
+#endif
     use_efficient_attention_ = (value & static_cast<int>(AttentionBackend::EFFICIENT_ATTENTION)) > 0;
     use_trt_fused_attention_ = (value & static_cast<int>(AttentionBackend::TRT_FUSED_ATTENTION)) > 0;
     use_cudnn_flash_attention_ = (value & static_cast<int>(AttentionBackend::CUDNN_FLASH_ATTENTION)) > 0;
@@ -26,6 +29,9 @@ void AttentionKernelOptions::Initialize(int value, bool use_build_flag, bool che
     use_trt_causal_attention_ = (value & static_cast<int>(AttentionBackend::TRT_CAUSAL_ATTENTION)) > 0;
   } else {
     use_flash_attention_ = !ParseEnvironmentVariableWithDefault<bool>(kDisableFlashAttention, false);
+#if USE_LEAN_ATTENTION
+    use_lean_attention_ = ParseEnvironmentVariableWithDefault<bool>(kEnableLeanAttention, false);
+#endif
     use_efficient_attention_ = !ParseEnvironmentVariableWithDefault<bool>(kDisableMemoryEfficientAttention, false);
     use_trt_fused_attention_ = !ParseEnvironmentVariableWithDefault<bool>(kDisableFusedSelfAttention, false);
     use_cudnn_flash_attention_ = ParseEnvironmentVariableWithDefault<bool>(kEnableCudnnFlashAttention, false);
@@ -61,6 +67,10 @@ void AttentionKernelOptions::Initialize(int value, bool use_build_flag, bool che
     use_flash_attention_ = false;
 #endif
 
+#ifndef USE_LEAN_ATTENTION
+    use_lean_attention_ = false;
+#endif
+
 #ifndef USE_MEMORY_EFFICIENT_ATTENTION
     use_efficient_attention_ = false;
 #endif
@@ -81,6 +91,9 @@ void AttentionKernelOptions::Print() const {
   std::stringstream sstream;
   sstream << "AttentionKernelOptions:";
   sstream << " FLASH_ATTENTION=" << int(use_flash_attention_);
+#if USE_LEAN_ATTENTION
+  sstream << " LEAN_ATTENTION=" << int(use_lean_attention_);
+#endif
   sstream << " EFFICIENT_ATTENTION=" << int(use_efficient_attention_);
   sstream << " TRT_FUSED_ATTENTION=" << int(use_trt_fused_attention_);
   sstream << " CUDNN_FLASH_ATTENTION=" << int(use_cudnn_flash_attention_);
@@ -131,6 +144,10 @@ void AttentionKernelDebugInfo::Print(const char* operator_name,
   sstream << " SdpaKernel=";
   if (use_flash_attention.has_value() && use_flash_attention.value()) {
     sstream << "FLASH_ATTENTION";
+#if USE_LEAN_ATTENTION
+  } else if (use_lean_attention.has_value() && use_lean_attention.value()) {
+    sstream << "LEAN_ATTENTION";
+#endif
   } else if (use_efficient_attention.has_value() && use_efficient_attention.value()) {
     sstream << "EFFICIENT_ATTENTION";
   } else if (use_trt_fused_attention.has_value() && use_trt_fused_attention.value()) {