feat: add xlite as third npu backend (support qwen3-dense/qwen3-moe/glm-4.7/glm-5.1/glm-5.2)

Author: uuzWY

CMakeLists.txt

@@ -9,6 +9,7 @@ option(USE_MUSA "Enable MUSA support" OFF)
 option(USE_DCU "Enable DCU support" OFF)
 option(USE_MACA "Enable MACA support" OFF)
 option(ENABLE_HA "Enable Mooncake etcd-based high availability support" OFF)
+option(USE_XLITE "Enable xlite backend (NPU only)" OFF)
 add_compile_definitions(YLT_ENABLE_IBV)
 add_definitions(-DYLT_ENABLE_IBV)
 set(YLT_ENABLE_IBV ON)
@@ -477,6 +478,12 @@ if(USE_NPU)
   set(CMAKE_VERBOSE_MAKEFILE ON)
   add_definitions(-DTORCH_HIGHER_THAN_PTA6)
 
+  if (USE_XLITE)
+    add_definitions(-DUSE_XLITE)
+    include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/xlite.cmake)
+    message(STATUS "USE_XLITE is ON")
+  endif()
+
   # Use vcpkg header files as the first priority search directory,
   #-> because the scope of third-party software managed by vcpkg is used throughout the entire xllm.
   message(STATUS "VCPKG_INCLUDE_DIR = ${CMAKE_BINARY_DIR}/vcpkg_installed/${VCPKG_TARGET_TRIPLET}/include")

cmake/xlite.cmake

@@ -0,0 +1,23 @@
+# xlite link helper: find_package(xlite) when USE_NPU AND USE_XLITE are ON.
+
+function(xllm_link_xlite target)
+  if(NOT USE_NPU OR NOT USE_XLITE)
+    return()
+  endif()
+
+  if(NOT TARGET xlite::xlite)
+    execute_process(
+      COMMAND ${Python_EXECUTABLE} -c "import xlite; print(xlite.cmake_prefix_path)"
+      OUTPUT_VARIABLE _XLITE_CMAKE_PREFIX
+      OUTPUT_STRIP_TRAILING_WHITESPACE
+      RESULT_VARIABLE _XLITE_IMPORT_RESULT)
+    if(NOT _XLITE_IMPORT_RESULT EQUAL 0 OR _XLITE_CMAKE_PREFIX STREQUAL "")
+      message(FATAL_ERROR "USE_XLITE is ON but xlite not found. Install xlite or pass -DUSE_XLITE=OFF.")
+    endif()
+
+    find_package(xlite REQUIRED CONFIG PATHS "${_XLITE_CMAKE_PREFIX}" NO_DEFAULT_PATH)
+    message(STATUS "xlite::xlite found via find_package (${_XLITE_CMAKE_PREFIX})")
+  endif()
+
+  target_link_libraries(${target} PRIVATE xlite::xlite)
+endfunction()

xllm/CMakeLists.txt

@@ -127,6 +127,9 @@ if(USE_MUSA)
   )
 endif()
 
+if (USE_NPU AND USE_XLITE)
+  xllm_link_xlite(xllm)
+endif()
 # install xllm
 install(TARGETS xllm RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
 

xllm/core/framework/model/causal_lm.h

@@ -49,6 +49,10 @@ namespace layer {
 struct AttentionMetadata;
 }
 
+namespace xlite {
+class XliteModelHolder;
+}
+
 struct ModelGraphMetadataState {
   virtual ~ModelGraphMetadataState() = default;
 };
@@ -151,6 +155,9 @@ class CausalLM : public torch::nn::Module {
     NOT_IMPLEMENTED();
     return false;
   }
+
+  // xlite runtime access, nullptr for non-xlite models.
+  virtual xlite::XliteModelHolder* get_xlite_holder() { return nullptr; }
 #endif
 
   virtual layer::LmHead get_lm_head() {
@@ -444,6 +451,14 @@ class CausalLMImpl : public CausalLM {
                                                      requested_rolling_slots,
                                                      model_id);
   }
+
+  // Forward to inner Model.
+  xlite::XliteModelHolder* get_xlite_holder() override {
+    if constexpr (detail::has_get_xlite_holder<Model>::value) {
+      return model_->get_xlite_holder();
+    }
+    return CausalLM::get_xlite_holder();
+  }
 #endif
 
   layer::LmHead get_lm_head() override {

xllm/core/framework/model/model_traits.h

@@ -246,6 +246,15 @@ struct has_init_or_refresh_rolling_runtime<
         std::declval<int32_t>(),
         std::declval<const std::string&>()))>> : std::true_type {};
 
+// SFINAE: xlite-backend models expose get_xlite_holder().
+template <typename T, typename = void>
+struct has_get_xlite_holder : std::false_type {};
+
+template <typename T>
+struct has_get_xlite_holder<
+    T,
+    std::void_t<decltype(std::declval<T>()->get_xlite_holder())>>
+    : std::true_type {};
 #endif
 
 template <typename T, typename = void>

xllm/core/layers/xlite/xlite_attn_meta_builder.h

@@ -0,0 +1,84 @@
+/* Copyright 2025-2026 The xLLM Authors.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    https://github.com/jd-opensource/xllm/blob/main/LICENSE
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// Build xlite XModelAttnMeta from xllm ModelInputParams.
+
+#pragma once
+
+#include <xlite/xlite.h>
+
+#include <algorithm>
+#include <cstdint>
+#include <vector>
+
+#include "core/framework/model/model_input_params.h"
+#include "core/layers/xlite/xlite_init_utils.h"
+
+namespace xllm::xlite {
+
+class XliteAttnMetaBuilder {
+ public:
+  static void Build(const ModelInputParams& params,
+                    const torch::Tensor& positions,
+                    uint32_t block_size,
+                    XModelAttnMeta& m,
+                    int64_t pad_count = 0) {
+    // version=0: xlite recomputes position from cachedLens (framework's
+    // positions tensor is off-by-one on decode).
+    m.version = 0;
+    m.lens.clear();
+    m.cachedLens.clear();
+    m.blockTables.clear();
+
+    const auto& host = params.attention.host;
+    int n = params.meta.num_sequences;
+    uint32_t bs = block_size;
+
+    // block_tables may be undefined in edge cases (DP empty shard).
+    const bool has_real_seqs =
+        n > 0 && host.block_tables.defined() && host.block_tables.dim() >= 2;
+    if (has_real_seqs) {
+      auto block_acc = host.block_tables.accessor<int32_t, 2>();
+      for (int s = 0; s < n; ++s) {
+        int32_t q_len = host.q_seq_lens[s];
+        int32_t kv_len = host.kv_seq_lens[s];
+        m.lens.push_back(static_cast<uint32_t>(q_len));
+        m.cachedLens.push_back(
+            static_cast<uint32_t>(std::max(0, kv_len - q_len)));  // clamp >= 0
+        int32_t nblocks =
+            (kv_len + static_cast<int32_t>(bs) - 1) / static_cast<int32_t>(bs);
+        std::vector<uint32_t> row(nblocks);
+        for (int32_t b = 0; b < nblocks; ++b) {
+          row[b] = static_cast<uint32_t>(block_acc[s][b]);
+        }
+        m.blockTables.push_back(std::move(row));
+      }
+    }
+
+    // DP padding: append dummy seq so sum(lens) aligns across DP groups.
+    if (pad_count > 0) {
+      m.lens.push_back(static_cast<uint32_t>(pad_count));
+      m.cachedLens.push_back(0);
+      int32_t nblocks =
+          (static_cast<int32_t>(pad_count) + static_cast<int32_t>(bs) - 1) /
+          static_cast<int32_t>(bs);
+      std::vector<uint32_t> row(nblocks, 0);
+      m.blockTables.push_back(std::move(row));
+    }
+    InitXTensor(m.vllmPosition, positions);
+  }
+};
+
+}  // namespace xllm::xlite