fix inf in fused_attention (#41933) (#42032)

paddle/fluid/operators/fused/fmha_ref.h

@@ -16,6 +16,8 @@ limitations under the License. */
 #include "paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h"
 #include "paddle/fluid/operators/transpose_op.cu.h"
 #include "paddle/phi/kernels/funcs/concat_and_split_functor.h"
+#include "paddle/phi/kernels/funcs/elementwise_base.h"
+#include "paddle/phi/kernels/funcs/functors.h"
 #include "paddle/phi/kernels/gpudnn/softmax_gpudnn.h"
 
 namespace paddle {
@@ -117,6 +119,18 @@ class FMHARef {
       v_ptr = k_ptr + k_size;
     }
 
+    {
+      // NOTE(wangxi): We scale Q with 1/sqrt(Dh) before QK^T, because for
+      // float16 calculation, INF may appear in QK^T if we do not scale before.
+      float alpha = 1.0 / sqrt(head_dim_);
+      auto q_tensor = transpose_2_out_tensor->Slice(0, 1);
+      auto functor = phi::funcs::ScaleFunctor<T>(alpha);
+      std::vector<const framework::Tensor*> ins = {&q_tensor};
+      std::vector<framework::Tensor*> outs = {&q_tensor};
+      paddle::operators::LaunchSameDimsElementwiseCudaKernel<T>(dev_ctx_, ins,
+                                                                &outs, functor);
+    }
+
     // q*k^t, batched_gemm
     CBLAS_TRANSPOSE transA = CblasNoTrans;
     CBLAS_TRANSPOSE transB = CblasTrans;
@@ -125,7 +139,7 @@ class FMHARef {
     int gemm_m = seq_len_;
     int gemm_n = out_seq_len;
     int gemm_k = head_dim_;
-    T alpha = static_cast<T>(1.0 / sqrt(head_dim_));
+    T alpha = static_cast<T>(1.0);
     T beta = static_cast<T>(0.0);
     int64_t stride_a = gemm_m * gemm_k;
     int64_t stride_b = gemm_k * gemm_n;
@@ -300,7 +314,9 @@ class FMHARef {
     }
 
     T* qk_out_grad_data = qk_out_grad_tensor->data<T>();
-    alpha = static_cast<T>(1.0 / sqrt(head_dim_));
+    // NOTE(wangxi): For we scale Q with 1/sqrt(Dh) in forward, so we set
+    //   alpha = 1.0 in backward.
+    alpha = static_cast<T>(1.0);
     // recall batchedgemm(nt) fw:  q_ptr * (k_ptr)^t = qk_out
     // bw: dy (seq_len * head_dim) = (dout)^t * x
     transA = CblasTrans;
@@ -314,6 +330,7 @@ class FMHARef {
                      qk_out_grad_data, q_ptr, beta, k_grad_ptr, gemm_batch_size,
                      stride_a, stride_b);
     // dx (seq_len * head_dim) = dout * y
+    alpha = static_cast<T>(1.0 / sqrt(head_dim_));
     transA = CblasNoTrans;
     transB = CblasNoTrans;
     gemm_m = seq_len_;