Fix paddle.dist & paddle.nn.functional.normalize (#74448)

fxyfxy777 · web-flow · commit 239dff3e5977 · 2025-08-07T18:54:20.000+08:00
* fix: fix normalization logic in p_norm kernels

* Code Formatting

* zancun

* fix_dist_normalize

* back

* huifu zhushi

* pre-commit

* Fix code style

* Standardize code comments
diff --git a/paddle/phi/kernels/funcs/reduce_grad_functions.h b/paddle/phi/kernels/funcs/reduce_grad_functions.h
@@ -38,10 +38,10 @@ void ReduceGradFunctor(const Context& dev_ctx,
   auto x_dims = input0.dims();
   auto reduced_dims_v = common::vectorize(x_dims);
   std::vector<int> dims_ref = dims;
-  Eigen::array<int, D> broadcast_dim;
+  Eigen::array<int64_t, D> broadcast_dim;
   for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1;
 
-  int broad_cast_times = 1;
+  int64_t broad_cast_times = 1;
   for (size_t i = 0; i < dims_ref.size(); ++i) {
     if (dims_ref[i] < 0) {
       dims_ref[i] = x_rank + dims_ref[i];
@@ -142,7 +142,7 @@ void LaunchReduceGradKernel(const Context& dev_ctx,
     auto& place = *dev_ctx.eigen_device();
     // *dev_ctx.eigen_device();
     auto broadcast_dim =
-        Eigen::array<int, 1>({{static_cast<int>(input0->numel())}});
+        Eigen::array<int64_t, 1>({{static_cast<int64_t>(input0->numel())}});
     functor(place,
             &x,
             &x_reduce,
diff --git a/paddle/phi/kernels/gpu/dist_kernel.cu b/paddle/phi/kernels/gpu/dist_kernel.cu
@@ -63,6 +63,18 @@ struct PowFunctor {
   Ty p_order_;
 };
 
+template <typename Tx,
+          typename Ty,
+          typename Tout>  // Tx is high precision, Tout is low/out precision
+struct PowFunctorHighPrecision {
+  HOSTDEVICE explicit inline PowFunctorHighPrecision(const Ty& p_order)
+      : p_order_(p_order) {}
+  HOSTDEVICE inline Tx operator()(const Tx x) const {
+    return static_cast<Tout>(pow(static_cast<Ty>(x), p_order_));
+  }
+  Ty p_order_;
+};
+
 template <typename T, typename Functor>
 __global__ void ReduceSumWithSubtract(
     const T* x, const T* y, T* out, int64_t N, Functor func) {
@@ -126,16 +138,17 @@ void DistKernel(const Context& dev_ctx,
   DenseTensor intermediate;
   const T* x_ptr = x.data<T>();
   const T* y_ptr = y.data<T>();
+
   T* o_ptr = dev_ctx.template Alloc<T>(out);
   auto stream = dev_ctx.stream();
 
   auto xdim = x.dims();
   if (xdim == y.dims()) {  // same shape
-    auto n = x.numel();
+    int64_t n = x.numel();
+
     auto config = phi::backends::gpu::GetGpuLaunchConfig1D(dev_ctx, n);
     intermediate.Resize(common::make_ddim({config.block_per_grid.x}));
     T* i_ptr = dev_ctx.template Alloc<T>(&intermediate);
-
     std::vector<int64_t> axis_dims = {static_cast<int64_t>(-1)};
     std::vector<int> reduce_axis =
         funcs::details::GetReduceDim(axis_dims, xdim.size(), true);
@@ -166,15 +179,23 @@ void DistKernel(const Context& dev_ctx,
       ReduceSumWithSubtract<T>
           <<<config.block_per_grid.x, config.thread_per_block.x, 0, stream>>>(
               x_ptr, y_ptr, i_ptr, n, OtherOrderFunctor<T, MT>(p_order));
-      phi::funcs::ReduceKernel<T, T, kps::AddFunctor, kps::IdentityFunctor<MT>>(
-          dev_ctx, intermediate, out, kps::IdentityFunctor<MT>(), reduce_axis);
-
-      const DenseTensor* tmp_norm = out;
-      std::vector<const DenseTensor*> ins = {tmp_norm};
+      DenseTensor out_other;
+      out_other.Resize(out->dims());
+      dev_ctx.template Alloc<MT>(&out_other);
+
+      phi::funcs::
+          ReduceKernel<T, MT, kps::AddFunctor, kps::IdentityFunctor<MT>>(
+              dev_ctx,
+              intermediate,
+              &out_other,
+              kps::IdentityFunctor<MT>(),
+              reduce_axis);
+      std::vector<const DenseTensor*> ins = {&out_other};
       std::vector<DenseTensor*> outs = {out};
-      MT p_order_ = static_cast<MT>(static_cast<MT>(1.) / p_order);
+
+      MT p_order_ = static_cast<MT>(1.f / p_order);
       phi::funcs::ElementwiseKernel<T>(
-          dev_ctx, ins, &outs, PowFunctor<T, MT>(p_order_));
+          dev_ctx, ins, &outs, PowFunctorHighPrecision<MT, MT, T>(p_order_));
     }
 
   } else {
diff --git a/paddle/phi/kernels/gpu/p_norm_grad_kernel.cu b/paddle/phi/kernels/gpu/p_norm_grad_kernel.cu
@@ -42,10 +42,12 @@ struct AbsMaxAndMinGradFunctor {
 
 template <typename T>
 struct PNormGradFunctor {
+  using MT = typename phi::dtype::MPTypeTrait<T>::Type;
   HOSTDEVICE explicit inline PNormGradFunctor(float porder, float eps) {
-    this->porder = static_cast<T>(porder - 1.);
-    this->eps = static_cast<T>(eps);
+    this->porder = static_cast<MT>(porder - 1.);
+    this->eps = static_cast<MT>(eps);
   }
+
   template <typename Context,
             typename X,
             typename Y,
@@ -59,12 +61,33 @@ struct PNormGradFunctor {
                   DY* dy,
                   const Dim& dim,
                   int size) {
+    auto x_mt = x->template cast<MT>();
+    auto y_mt = y->template cast<MT>();
+    auto dy_mt = dy->template cast<MT>();
+
+    auto norm_pow = y_mt.pow(-this->porder);
+    auto mask_norm_nonzero = (y_mt != static_cast<MT>(0)).template cast<MT>();
+
+    // Set to 0 where porder < 0 and x == 0
+    MT zero = static_cast<MT>(0);
+    auto mask_x_zero = (x_mt == zero).template cast<MT>();
+
+    MT is_porder_negative =
+        this->porder < zero ? static_cast<MT>(1) : static_cast<MT>(0);
+    auto invalid_mask = (mask_x_zero * is_porder_negative);
+    auto safe_pow =
+        x_mt.abs().pow(this->porder) * (static_cast<MT>(1) - invalid_mask);
+
     dx->device(place) =
-        (*x).abs().pow(this->porder) * (*x).sign() * dy->broadcast(dim) *
-        (*y + y->constant(eps)).pow(-this->porder).broadcast(dim);
+        (safe_pow * x_mt.sign() * dy_mt.broadcast(dim) *
+         norm_pow.broadcast(dim) *
+         mask_norm_nonzero.broadcast(dim)  // Mask out positions where norm == 0
+         )
+            .template cast<T>();
   }
-  T porder;
-  T eps;
+
+  MT porder;
+  MT eps;
 };
 
 template <typename T, typename Context>
diff --git a/paddle/phi/kernels/gpu/p_norm_kernel.cu b/paddle/phi/kernels/gpu/p_norm_kernel.cu
@@ -124,31 +124,38 @@ void PNormKernel(const Context& dev_ctx,
     phi::funcs::ElementwiseKernel<T>(
         dev_ctx, ins, &outs, UnsignedPowFunctor<T>(1. / porder));
 #else
+    DenseTensor out_temp;
+    out_temp.Resize(out_norm->dims());
+    dev_ctx.template Alloc<MT>(&out_temp);
+
     if (porder == 1.0) {
       // fast 1-norm
       phi::funcs::ReduceKernel<T, T, kps::AddFunctor, FabsFunctor<T>>(
           dev_ctx, *in_x, out_norm, FabsFunctor<T>(), reduce_axis);
     } else if (porder == 2.0) {
       // fast 2-norm
-      phi::funcs::ReduceKernel<T, T, kps::AddFunctor, SquareFunctor<T>>(
-          dev_ctx, *in_x, out_norm, SquareFunctor<T>(), reduce_axis);
+      phi::funcs::ReduceKernel<T, MT, kps::AddFunctor, SquareFunctor<MT>>(
+          dev_ctx, *in_x, &out_temp, SquareFunctor<MT>(), reduce_axis);
     } else if (porder == 3.0) {
       // fast 3-norm
-      phi::funcs::ReduceKernel<T, T, kps::AddFunctor, FabsCubicFunctor<T>>(
-          dev_ctx, *in_x, out_norm, FabsCubicFunctor<T>(), reduce_axis);
+      phi::funcs::ReduceKernel<T, MT, kps::AddFunctor, FabsCubicFunctor<MT>>(
+          dev_ctx, *in_x, &out_temp, FabsCubicFunctor<MT>(), reduce_axis);
     } else {
       // vanilla norm
-      phi::funcs::ReduceKernel<T, T, kps::AddFunctor, UnsignedPowFunctor<T>>(
-          dev_ctx, *in_x, out_norm, UnsignedPowFunctor<T>(porder), reduce_axis);
+      phi::funcs::ReduceKernel<T, MT, kps::AddFunctor, UnsignedPowFunctor<MT>>(
+          dev_ctx,
+          *in_x,
+          &out_temp,
+          UnsignedPowFunctor<MT>(porder),
+          reduce_axis);
     }
 
     if (porder != 1.0) {
-      // save computation when porder is 1.0
-      const DenseTensor* tmp_norm = out_norm;
-      std::vector<const DenseTensor*> ins = {tmp_norm};
+      std::vector<const DenseTensor*> ins = {&out_temp};
       std::vector<DenseTensor*> outs = {out_norm};
+      MT p_order_ = static_cast<MT>(1.f / porder);
       phi::funcs::ElementwiseKernel<T>(
-          dev_ctx, ins, &outs, UnsignedPowFunctor<T>(1. / porder));
+          dev_ctx, ins, &outs, UnsignedPowFunctor<MT>(p_order_));
     }
 #endif
   }
