merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into npu test=document_fix

Author: swgu98

cmake/external/xpu.cmake

@@ -30,7 +30,7 @@ set(XPU_XFA_LIB_NAME "libxpu_flash_attention.so")
 set(XPU_XPUDNN_LIB_NAME "libxpu_dnn.so")
 
 if(NOT DEFINED XPU_XHPC_BASE_DATE)
-  set(XPU_XHPC_BASE_DATE "dev/20250405")
+  set(XPU_XHPC_BASE_DATE "dev/20250412")
 endif()
 set(XPU_XCCL_BASE_VERSION "3.0.2.5") # For XRE5
 if(NOT DEFINED XPU_XFT_BASE_VERSION)
@@ -46,7 +46,7 @@ if(NOT DEFINED XPU_XRE_BASE_VERSION)
 endif()
 
 if(WITH_XPU_XRE5)
-  set(XPU_XPTI_BASE_VERSION "0.1.0")
+  set(XPU_XPTI_BASE_VERSION "0.2.0")
 else()
   set(XPU_XPTI_BASE_VERSION "0.0.1")
 endif()

paddle/cinn/ir/group_schedule/tactic/tile_broadcast_tactic.cc

@@ -206,6 +206,7 @@ bool IsElementwiseOrBroadcast(const ir::Store& dst, const ir::Load& src) {
 
 bool CheckAllElementwiseOrBroadcast(ir::IRSchedule* sch) {
   for (auto& block : sch->GetAllBlocks()) {
+    if (ir::analyzer::IsReductionSBlock(block)) return false;
     ir::Expr store = ir::analyzer::GetStoreOfSBlock(block);
     auto* store_node = store.As<ir::Store>();
     for (auto& load : CollectLoads(store_node->value)) {
@@ -564,6 +565,7 @@ void TileBroadcastTactic::Apply(ir::IRSchedule* sch,
     block_size = CalcNumWarps(preserved_size_ >> 5);
     if (block_size == -1) {
       applied_layout_ = BroadcastLayout::Invalid;
+      return;
     }
     block_size = std::clamp(block_size << 5, 128, 1024);
   }

paddle/fluid/framework/op_version_proto.cc

@@ -61,7 +61,6 @@ const std::unordered_map<std::string, uint32_t>& GetLegacyOpVersions() {
       {"depthwise_conv2d_transpose", 1},
       {"conv2d", 1},
       {"lamb", 1},
-      {"send_and_recv", 1},
       {"gaussian_random", 1},
       {"unique_consecutive", 1},
       {"conv3d", 1},