add dispatch_gmm_combine kernel (#3532)

### What this PR does / why we need it?

This PR introduces the Ascend implementation of the
`dispatch_ffn_combine` kernel and wires it into the vLLM-Ascend runtime,
together with follow‑up fixes to ensure the kernel builds and runs
correctly in CI.

- Add full host and device implementation of the `dispatch_ffn_combine`
kernel under `csrc/dispatch_ffn_combine`, including tiling logic, MOE
routing helpers, and kernel utilities for quantized FFN dispatch.
- Integrate the new kernel with the PyTorch binding
(csrc/torch_binding.cpp, csrc/torch_binding_meta.cpp) and the Ascend
runtime (vllm_ascend/ascend_forward_context.py,
vllm_ascend/worker/model_runner_v1.py).
- Extend fused MoE communication and token dispatch support in
`vllm_ascend/ops/fused_moe`, adding methods/utilities needed by the new
dispatch path.
- Update quantization logic in vllm_ascend/quantization/w8a8_dynamic.py
to support the new FFN dispatch flow.
- Fix kernel build issues by adjusting `csrc/build_aclnn.sh`, CMake
configuration, and include/namespace usage in the new kernel files.
- Add an end‑to‑end nightly test
`tests/e2e/nightly/ops/test_dispatch_ffn_combine.py` and helper
utilities in `vllm_ascend/utils.py` to validate the new kernel.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?


- vLLM version: v0.12.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.12.0

---------

Signed-off-by: mojave2 <chenchen145@huawei.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>

csrc/dispatch_ffn_combine/op_host/CMakeLists.txt

@@ -0,0 +1,66 @@
+# Copyright (c) 2025 Huawei Technologies Co., Ltd.
+# This file is a part of the CANN Open Software.
+# Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# ======================================================================================================================
+
+set(_DISPATCH_FFN_INC_OPTS)
+if (EXISTS ${ASCEND_CANN_PACKAGE_PATH}/aarch64-linux/ascendc/include)
+    list(APPEND _DISPATCH_FFN_INC_OPTS -I${ASCEND_CANN_PACKAGE_PATH}/aarch64-linux/ascendc/include)
+elseif (EXISTS ${ASCEND_CANN_PACKAGE_PATH}/arm64-linux/ascendc/include)
+    list(APPEND _DISPATCH_FFN_INC_OPTS -I${ASCEND_CANN_PACKAGE_PATH}/arm64-linux/ascendc/include)
+elseif (EXISTS ${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_SYSTEM_PROCESSOR}-linux/ascendc/include)
+    list(APPEND _DISPATCH_FFN_INC_OPTS -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_SYSTEM_PROCESSOR}-linux/ascendc/include)
+endif()
+if (EXISTS ${CMAKE_SOURCE_DIR}/third_party/catlass/include)
+    list(APPEND _DISPATCH_FFN_INC_OPTS -I${CMAKE_SOURCE_DIR}/third_party/catlass/include)
+endif()
+
+add_ops_compile_options(
+        OP_NAME DispatchFFNCombine
+        OPTIONS --cce-auto-sync=on
+                -Wno-deprecated-declarations
+                -Werror
+                -DHCCL_COMM
+                ${_DISPATCH_FFN_INC_OPTS}
+)
+
+target_sources(op_host_aclnnInner PRIVATE
+        dispatch_ffn_combine_def.cpp
+)
+
+target_sources(opapi PRIVATE
+        aclnn_dispatch_ffn_combine.cpp
+)
+
+if (NOT BUILD_OPEN_PROJECT)
+    target_sources(aclnn_ops_train PRIVATE
+        aclnn_dispatch_ffn_combine.cpp
+    )
+
+    target_sources(aclnn_ops_infer PRIVATE
+        aclnn_dispatch_ffn_combine.cpp
+    )
+endif ()
+
+target_sources(optiling PRIVATE
+        dispatch_ffn_combine_tiling.cpp
+)
+
+target_include_directories(optiling PRIVATE
+        ${CMAKE_CURRENT_SOURCE_DIR}
+        ${CMAKE_CURRENT_SOURCE_DIR}/../op_kernel
+)
+
+target_sources(opsproto PRIVATE
+        dispatch_ffn_combine_proto.cpp
+)
+
+file(GLOB _GMM_Aclnn_header "${CMAKE_CURRENT_SOURCE_DIR}/aclnn_dispatch_ffn_combine.h")
+
+install(FILES ${_GMM_Aclnn_header}
+        DESTINATION ${ACLNN_INC_INSTALL_DIR} OPTIONAL
+)

csrc/dispatch_ffn_combine/op_host/aclnn_dispatch_ffn_combine.cpp

@@ -0,0 +1,84 @@
+/**
+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+#include "aclnn_dispatch_ffn_combine.h"
+#include <algorithm>
+// #include "aclnn_kernels/common/op_error_check.h"
+// #include "opdev/op_log.h"
+// #include "opdev/common_types.h"
+// #include "opdev/platform.h"
+// #include "ophost/matmul_util.h"
+#include <unistd.h>
+#include <vector>
+#include <string>
+#include <iostream>
+#include <fcntl.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/file.h>
+#include <climits>
+#include "../op_host/error_log.h"
+// using namespace op;
+
+// using namespace op;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static constexpr size_t TWO_DIMS = 2;
+static constexpr int64_t KVALUE_MIN = 256;
+static constexpr int64_t KVALUE_MAX = 65535;
+static constexpr size_t HCCL_GROUP_NAME_MAX = 128U;
+enum NnopbaseHcclServerType {
+    NNOPBASE_HCCL_SERVER_TYPE_AICPU = 0,
+    NNOPBASE_HCCL_SERVER_TYPE_MTE,
+    NNOPBASE_HCCL_SERVER_TYPE_END
+};
+
+extern aclnnStatus aclnnInnerDispatchFFNCombineGetWorkspaceSize(const aclTensor* x, const aclTensor* weight1, const aclTensor* weight2,
+                                                         const aclTensor* expertId, const aclTensor* scale1, const aclTensor* scale2,
+                                                         const aclTensor* probs,
+                                                         const char* group, int64_t maxOutputSize,
+                                                         bool transB, bool weightNz,
+                                                         const aclTensor* out,
+                                                         uint64_t* workspaceSize, aclOpExecutor** executor);
+extern aclnnStatus aclnnInnerDispatchFFNCombine(void *workspace, uint64_t workspaceSize,
+                                            aclOpExecutor *executor, aclrtStream stream);
+extern "C" void __attribute__((weak)) NnopbaseSetHcclServerType(void *executor, NnopbaseHcclServerType sType);
+
+
+
+aclnnStatus aclnnDispatchFFNCombineGetWorkspaceSize(const aclTensor* x, const aclTensor* weight1, const aclTensor* weight2,
+                                                    const aclTensor* expertId, const aclTensor* scale1, const aclTensor* scale2,
+                                                    const aclTensor* probs,
+                                                    const char* group, int64_t maxOutputSize,
+                                                    const aclTensor* out,
+                                                    uint64_t* workspaceSize, aclOpExecutor** executor)
+{
+    bool transB = false;
+    bool weightNz = true;
+
+    aclnnStatus ret = aclnnInnerDispatchFFNCombineGetWorkspaceSize(x, weight1, weight2, expertId, scale1, scale2, probs, group, 
+                                                                    maxOutputSize, transB, weightNz,
+                                                                    out, workspaceSize, executor);
+    return ret;
+}
+
+aclnnStatus aclnnDispatchFFNCombine(void* workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)
+{
+    if (NnopbaseSetHcclServerType) {
+        NnopbaseSetHcclServerType(executor, NNOPBASE_HCCL_SERVER_TYPE_MTE);
+    }
+    aclnnStatus ret = aclnnInnerDispatchFFNCombine(workspace, workspaceSize, executor, stream);
+    return ret;
+}
+#ifdef __cplusplus
+}
+#endif

csrc/dispatch_ffn_combine/op_host/aclnn_dispatch_ffn_combine.h

@@ -0,0 +1,61 @@
+/**
+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+#ifndef OP_API_INC_DISPATCH_FFN_COMBINE_
+#define OP_API_INC_DISPATCH_FFN_COMBINE_
+
+#include <string>
+
+#include "aclnn/aclnn_base.h"
+#include "hccl/hccl.h"
+#include "hccl/hccl_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * 算子功能：实现分布式MoE从InitRouting到Unpermute全部算子的融合
+ * @brief aclnnDispatchFFNCombine的第一段接口，根据具体的计算流程，计算workspace大小。
+ * @domain aclnn_ops_infer
+ * @param [in] a: matmul左矩阵，数据类型支持：float16, bf16。
+ * @param [in] b: matmul右矩阵，数据类型支持：float16, bf16。
+ * @param [in] bias: 偏置，数据类型支持：float16, bf16。
+ * @param [in] group: 标识通信域名称的字符串。
+ * @param [in] worldsize: 通信域size，支持2/4/8卡。
+ * @param [in] epRankId: ep本卡Id。取值范围[0, worldSize)，各卡的rankId不能重复
+ * @param [out] c: 计算+通信的结果，数据类型：同输入。
+ * @param [out] workspaceSize: 返回需要在npu device侧申请的workspace大小。
+ * @param [out] executor: 返回op执行器，包含了算子计算流程。
+ * @return aclnnStatus: 返回状态码
+ */
+__attribute__((visibility("default"))) aclnnStatus aclnnDispatchFFNCombineGetWorkspaceSize(const aclTensor* x, const aclTensor* weight1, const aclTensor* weight2,
+                                                                                        const aclTensor* expertId, const aclTensor* scale1, const aclTensor* scale2,
+                                                                                        const aclTensor* probs,
+                                                                                        const char* group, int64_t maxOutputSize,
+                                                                                        const aclTensor* out,
+                                                                                        uint64_t* workspaceSize, aclOpExecutor** executor);
+
+/**
+ * @brief aclnnDispatchGmmCombine的第二段接口，用于执行计算。
+ * @param [in] workspace: 在npu device侧申请的workspace内存起址。
+ * @param [in] workspace_size: 在npu device侧申请的workspace大小，由第一段接口aclnnDispatchFFNCombineGetWorkspaceSize获取。
+ * @param [in] exector: op执行器，包含了算子计算流程。
+ * @param [in] stream: acl stream流。
+ * @return aclnnStatus: 返回状态码
+ */
+__attribute__((visibility("default"))) aclnnStatus aclnnDispatchFFNCombine(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor,
+                                           aclrtStream stream);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // OP_API_INC_GMM_ALLTOALLV_

csrc/dispatch_ffn_combine/op_host/dispatch_ffn_combine_def.cpp

@@ -0,0 +1,88 @@
+/**
+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/*!
+ * \file dispatch_ffn_combine_def.cpp
+ * \brief
+ */
+#include "register/op_def_registry.h"
+
+namespace ops {
+class DispatchFFNCombine : public OpDef {
+ public:
+  explicit DispatchFFNCombine(const char *name) : OpDef(name) {
+    this->Input("a")
+        .ParamType(REQUIRED)
+        .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_BF16})
+        .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
+        .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
+    this->Input("w1")
+        .ParamType(REQUIRED)
+        .DataType({ge::DT_INT8, ge::DT_INT8, ge::DT_INT8})
+        .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_FRACTAL_NZ})
+        .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_FRACTAL_NZ})
+        .IgnoreContiguous();
+    this->Input("w2")
+        .ParamType(REQUIRED)
+        .DataType({ge::DT_INT8, ge::DT_INT8, ge::DT_INT8})
+        .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_FRACTAL_NZ})
+        .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_FRACTAL_NZ})
+        .IgnoreContiguous();
+    this->Input("expertIdx")
+        .ParamType(REQUIRED)
+        .DataType({ge::DT_INT32, ge::DT_INT32, ge::DT_INT32})
+        .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
+        .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
+    this->Input("scale1")
+        .ParamType(REQUIRED)
+        .DataType({ge::DT_INT64, ge::DT_INT64, ge::DT_INT64})
+        .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
+        .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
+    this->Input("scale2")
+        .ParamType(REQUIRED)
+        .DataType({ge::DT_INT64, ge::DT_INT64, ge::DT_INT64})
+        .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
+        .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
+    this->Input("probs")
+        .ParamType(REQUIRED)
+        .DataType({ge::DT_FLOAT, ge::DT_FLOAT, ge::DT_FLOAT})
+        .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
+        .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND});
+
+    // 输出
+    this->Output("out")
+        .ParamType(REQUIRED)
+        .DataType({ge::DT_FLOAT16, ge::DT_BF16, ge::DT_BF16})
+        .Format({ge::FORMAT_ND, ge::FORMAT_ND, ge::FORMAT_ND})
+        .UnknownShapeFormat({ge::FORMAT_ND, ge::FORMAT_ND,ge::FORMAT_ND});
+
+    this->Attr("group").AttrType(REQUIRED).String();
+    this->Attr("M").AttrType(OPTIONAL).Int();
+    this->Attr("transB").AttrType(OPTIONAL).Bool(false);
+    this->Attr("weightNz").AttrType(OPTIONAL).Bool(false);
+
+    OpAICoreConfig aicore_config;
+    aicore_config.DynamicCompileStaticFlag(true)
+        .DynamicFormatFlag(true)
+        .DynamicRankSupportFlag(true)
+        .DynamicShapeSupportFlag(true)
+        .NeedCheckSupportFlag(false)
+        .PrecisionReduceFlag(true)
+        .ExtendCfgInfo("aclnnSupport.value", "support_aclnn")
+        .ExtendCfgInfo("jitCompile.flag", "static_false")
+        .ExtendCfgInfo("multiKernelSupportDynamicGraph.value", "multi_kernel");
+    this->AICore().AddConfig("ascend910_93", aicore_config);
+    this->AICore().AddConfig("ascend910b", aicore_config);
+    this->MC2().HcclGroup("group");
+  }
+};
+
+OP_ADD(DispatchFFNCombine);
+}  // namespace ops

csrc/dispatch_ffn_combine/op_host/dispatch_ffn_combine_proto.cpp

@@ -0,0 +1,40 @@
+/**
+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/*!
+ * \file dispatch_ffn_proto.cpp
+ * \brief
+ */
+#include <graph/utils/type_utils.h>
+#include <register/op_impl_registry.h>
+// #include "../../common/ophost/op_util.h"
+// #include "../../common/ophost/hcom_topo_info.h"
+// #include "log/ops_log.h"
+
+using namespace ge;
+namespace ops {
+const size_t ATTR_GROUP = 0;
+const size_t ATTR_RANK_SIZE = 1;
+const size_t SUPPORT_DIM_SIZE = 2;
+
+static ge::graphStatus InferShapeDispatchFFNCombine(gert::InferShapeContext* context) {
+  return ge::GRAPH_SUCCESS;
+}
+
+static ge::graphStatus InferDataTypeDispatchFFNCombine(gert::InferDataTypeContext* context) {
+  // auto d_type = context->GetInputDataType(0);
+  // context->SetOutputDataType(0, d_type);
+  return ge::GRAPH_SUCCESS;
+}
+
+IMPL_OP_INFERSHAPE(DispatchFFNCombine)
+  .InferShape(InferShapeDispatchFFNCombine)
+  .InferDataType(InferDataTypeDispatchFFNCombine);
+}  // namespace ops

csrc/dispatch_ffn_combine/op_host/dispatch_ffn_combine_tiling.cpp

@@ -0,0 +1,265 @@
+/**
+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+/*!
+ * \file dispatch_ffn_tiling.cpp
+ * \brief
+ */
+#include "vector"
+#include "register/tilingdata_base.h"
+#include "tiling/tiling_api.h"
+#include "error_log.h"
+#include "hcom_topo_info.h"
+#include "register/op_def_registry.h"
+#include "dispatch_ffn_combine_tiling.h"
+#include <vector>
+#include <map>
+#include <algorithm>
+#include "moe_init_routing_quant_v2/moe_init_routing_quant_v2_tiling.h"
+
+using namespace AscendC;
+using namespace ge;
+
+namespace {
+    // 1. 常量定义
+    const char *K_INNER_DEBUG = "DispatchFFNCombine Tiling Debug";
+    constexpr uint32_t ATTR_GROUP_INDEX = 0;
+    constexpr uint32_t ATTR_MAX_OUTPUT_SIZE_INDEX = 1;
+    constexpr uint32_t ATTR_IS_TRANS_B = 2;
+    constexpr uint32_t ATTR_WEIGHT_NZ = 3;
+    constexpr uint64_t INIT_TILINGKEY = 1000000;
+    constexpr uint64_t TILINGKEY_TRANS_B = 1U;
+    constexpr uint64_t TILINGKEY_WEIGHT_NZ = 10;
+    constexpr uint32_t X_INDEX = 0;
+    constexpr uint32_t WEIGHT_INDEX = 1;
+    constexpr uint32_t WEIGHT2_INDEX = 2;
+    constexpr uint32_t EXPERTID_INDEX = 3;
+    constexpr uint32_t BLOCK_NUM = 20;
+    constexpr uint32_t SYSTEM_NEED_WORKSPACE = 16 * 1024 * 1024;
+}
+
+namespace optiling {
+
+static int32_t CeilDev(int32_t num, int32_t div)
+{
+    if (div == 0) {
+        return 0;
+    }
+    return (num + div - 1) / div;
+}
+
+// 解析并校验 rankId, group, worldSize, isTransB 属性值
+static ge::graphStatus DispatchFFNCombineCheckAttrAndSetTiling(gert::TilingContext *context, DispatchFFNCombineInfo& info)
+{
+    auto attrs = context->GetAttrs();
+    OP_TILING_CHECK(attrs == nullptr, OP_LOGE(K_INNER_DEBUG, "attrs is null."), return ge::GRAPH_FAILED);
+
+    // todo：Attr相关tilingdata的设置、校验、打印
+    auto groupPtr = attrs->GetAttrPointer<char>(static_cast<int>(ATTR_GROUP_INDEX));
+    auto maxOutputSizePtr = attrs->GetAttrPointer<int>(ATTR_MAX_OUTPUT_SIZE_INDEX);
+    auto is_trans_b = attrs->GetAttrPointer<bool>(ATTR_IS_TRANS_B);
+    auto weight_nz = attrs->GetAttrPointer<bool>(ATTR_WEIGHT_NZ);
+    OP_TILING_CHECK(groupPtr == nullptr || strlen(groupPtr) == 0,
+    OP_LOGE(K_INNER_DEBUG, "group is invalid."), return GRAPH_FAILED);
+
+    OP_TILING_CHECK(is_trans_b == nullptr,
+        OP_LOGE(K_INNER_DEBUG, "is_trans_b is invalid."), return GRAPH_FAILED);
+    OP_TILING_CHECK(weight_nz == nullptr,
+        OP_LOGE(K_INNER_DEBUG, "weight_nz is invalid."), return GRAPH_FAILED);
+
+    info.maxOutputSize = *maxOutputSizePtr;
+    info.isTransposeB = *is_trans_b;
+    info.isWeightNz = *weight_nz;
+
+    int64_t rankSize;
+    (void)ge::HcomTopoInfo::Instance().GetGroupRankSize(groupPtr, rankSize);
+    info.worldSize = rankSize;
+
+    OP_LOGD(K_INNER_DEBUG, "maxOutputSize=%d ", info.maxOutputSize);
+    OP_LOGD(K_INNER_DEBUG, "rankSize=%d ", info.worldSize);
+
+    return ge::GRAPH_SUCCESS;
+}
+
+// 提取输入张量 A 和 B 的形状，计算出 M、K、N 值
+static ge::graphStatus DispatchFFNCombineCheckShapeAndSetTiling(gert::TilingContext *context, DispatchFFNCombineInfo &info)
+{
+    const char *nodeName = context->GetNodeName();
+    // OPS_LOG_I(nodeName, "DispatchFFnCombine DispatchFFNCombineCheckShapeAndSetTiling.");
+
+    const gert::StorageShape *aStorageShape = context->GetInputShape(X_INDEX);
+    const gert::StorageShape *bStorageShape = context->GetInputShape(WEIGHT_INDEX);
+    const gert::StorageShape *expertIdxShape = context->GetInputShape(EXPERTID_INDEX);
+    uint32_t M = aStorageShape->GetStorageShape().GetDim(0);
+    uint32_t K = aStorageShape->GetStorageShape().GetDim(1);
+    uint32_t expertPerRank = bStorageShape->GetStorageShape().GetDim(0);
+    uint32_t N = bStorageShape->GetStorageShape().GetDim(2);
+    uint32_t topK = expertIdxShape->GetStorageShape().GetDim(1);
+
+    info.M = M;
+    info.N = N;
+    info.K = K;
+    info.expertPerRank = expertPerRank;
+    info.topK = topK;
+    OP_LOGD(K_INNER_DEBUG, "M=%d ", info.M);
+    OP_LOGD(K_INNER_DEBUG, "K=%d ", info.K);
+    OP_LOGD(K_INNER_DEBUG, "N=%d ", info.N);
+    OP_LOGD(K_INNER_DEBUG, "expertPerRank=%d ", info.expertPerRank);
+    OP_LOGD(K_INNER_DEBUG, "topK=%d ", info.topK);
+
+    return ge::GRAPH_SUCCESS;
+}
+
+// 获取当前芯片平台的 AI Core 数目、UB 容量等硬件信息。
+static ge::graphStatus DispatchFFNCombineGetPlatformInfoAndSetTiling(gert::TilingContext *context, DispatchFFNCombineInfo& info)
+{
+    auto ascendcPlatform = platform_ascendc::PlatformAscendC(context->GetPlatformInfo());
+    uint32_t aivNum = ascendcPlatform.GetCoreNumAiv();
+    uint64_t ubSize = 0U;
+    ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSize);
+    info.aivNum = aivNum;
+    info.totalUbSize = ubSize;
+
+    OP_LOGD(K_INNER_DEBUG, "aivNum=%d", info.aivNum);
+    OP_LOGD(K_INNER_DEBUG, "ubSize=%lu", info.totalUbSize);
+
+    return ge::GRAPH_SUCCESS;
+}
+
+void SetTilingData(CoCTiling &cocTilingData, DispatchFFNCombineInfo &info)
+{
+    cocTilingData.m0 = 128;
+    cocTilingData.k0 = 256;
+    cocTilingData.n0 = 256;
+    cocTilingData.swizzleDirect = 1;
+    cocTilingData.swizzleOffset = 7;
+    cocTilingData.ubMoveNum = 16 * 1024;
+    cocTilingData.pValue = 1;
+    cocTilingData.commNpuSplit = info.worldSize;
+    cocTilingData.commDataSplit = 1;
+    cocTilingData.lenPerLoop = cocTilingData.m0 * cocTilingData.n0 / 2;
+}
+
+// 主调度函数：
+// 获取 tilingData ➝ 检查 Attr ➝ 检查 Shape ➝ 获取平台信息 
+// ➝ 调用 SetTilingData（根据rank数目） ➝ 设置 blockDim ➝ 设置 tilingKey ➝ 设置 workspace ➝ 配置通信参数
+
+static ge::graphStatus DispatchFFNCombineTilingFuncImpl(gert::TilingContext *context)
+{
+    const char *nodeName = context->GetNodeName();
+    OP_LOGI(nodeName, "Enter DispatchFFNCombine tiling func.");
+
+    // 1. tilingData
+    DispatchFFNCombineTilingData *tilingData = context->GetTilingData<DispatchFFNCombineTilingData>();
+    OP_TILING_CHECK(tilingData == nullptr, OP_LOGE(nodeName, "tilingData is nullptr."),
+        return ge::GRAPH_FAILED);
+    OP_LOGI(nodeName, "DispatchFFNCombine get tilingData.");
+    DispatchFFNCombineInfo& info = tilingData->dispatchFFNCombineInfo;
+    OP_LOGI(nodeName, "DispatchFFNCombine get tilingData info.");
+
+    OP_TILING_CHECK(DispatchFFNCombineCheckAttrAndSetTiling(context, info) != ge::GRAPH_SUCCESS,
+        OP_LOGE(context->GetNodeName(), "DispatchFFNCombine CheckAttrAndSetTiling Failed"),
+        return ge::GRAPH_FAILED);
+    OP_TILING_CHECK(DispatchFFNCombineCheckShapeAndSetTiling(context, info) != ge::GRAPH_SUCCESS,
+        OP_LOGE(context->GetNodeName(), "DispatchFFNCombine CheckShapeAndSetTiling Failed"),
+        return ge::GRAPH_FAILED);
+    OP_TILING_CHECK(DispatchFFNCombineGetPlatformInfoAndSetTiling(context, info) != ge::GRAPH_SUCCESS,
+        OP_LOGE(context->GetNodeName(), "DispatchFFNCombine GetPlatformInfoAndSetTiling Failed"),
+        return ge::GRAPH_FAILED);
+
+    SetTilingData(tilingData->cocTiling, info);
+
+    // 2. set blockDim
+    uint32_t blockDim = 1U;
+    auto ascendcPlatform = platform_ascendc::PlatformAscendC(context->GetPlatformInfo());
+    auto aicNum = ascendcPlatform.GetCoreNumAic();
+    auto aivNum = ascendcPlatform.GetCoreNumAiv();
+    blockDim = ascendcPlatform.CalcTschBlockDim(aivNum, aicNum, aivNum);
+    context->SetBlockDim(blockDim);
+
+    // 3. set tilingKey
+    uint64_t tilingKey = INIT_TILINGKEY;
+    tilingKey += info.isTransposeB ? TILINGKEY_TRANS_B : 0;
+    tilingKey += info.isWeightNz ? TILINGKEY_WEIGHT_NZ : 0;
+    context->SetTilingKey(tilingKey);
+
+    OP_LOGD(K_INNER_DEBUG, "tilingKey=%d", tilingKey);
+
+    optiling::MoeInitRoutingQuantV2TilingBase moeInitRoutingQuantV2TilingBase;
+    int64_t inuptXDtypeSize = sizeof(int16_t);
+    int64_t scaleDim0 = 0;
+    int64_t ubSize = 196352;
+    int64_t expertCapacity = 0;
+    int64_t expertNum = info.expertPerRank * info.worldSize;
+    int64_t activeNum = 0;
+    int64_t dropPadMode = 0;
+     int64_t expertTokensCountOrCumsumFlag = 2;
+     bool expertTokensBeforeCapacityFlag = false;
+     int64_t quantMode = 1;
+     uint32_t aivNumInitRouting = 2 * BLOCK_NUM;
+    moeInitRoutingQuantV2TilingBase.DoTiling(info.M, info.K, info.topK, expertCapacity, expertNum, activeNum, dropPadMode, 
+        expertTokensCountOrCumsumFlag, expertTokensBeforeCapacityFlag, inuptXDtypeSize, quantMode, scaleDim0, aivNumInitRouting, ubSize);
+    uint64_t initRoutingQuantTilingKey = moeInitRoutingQuantV2TilingBase.tilingKey_;
+    size_t initRoutingWorkspace = moeInitRoutingQuantV2TilingBase.workspaceSize_;
+
+    tilingData->cocTiling.moeInitRoutingQuantV2TilingData = moeInitRoutingQuantV2TilingBase.quantTilingData;
+    tilingData->cocTiling.moeInitRoutingQuantV2TilingData.vbsComputeParamsOp = moeInitRoutingQuantV2TilingBase.quantTilingData.vbsComputeParamsOp;
+    tilingData->cocTiling.moeInitRoutingQuantV2TilingData.vmsMiddleComputeParamsOp = moeInitRoutingQuantV2TilingBase.quantTilingData.vmsMiddleComputeParamsOp;
+    tilingData->cocTiling.moeInitRoutingQuantV2TilingData.sortOutComputeParamsOp = moeInitRoutingQuantV2TilingBase.quantTilingData.sortOutComputeParamsOp;
+    tilingData->cocTiling.moeInitRoutingQuantV2TilingData.srcToDstComputeParamsOp = moeInitRoutingQuantV2TilingBase.quantTilingData.srcToDstComputeParamsOp;
+    tilingData->cocTiling.moeInitRoutingQuantV2TilingData.srcToDstCapacityComputeParamsOp = moeInitRoutingQuantV2TilingBase.quantTilingData.srcToDstCapacityComputeParamsOp;
+    tilingData->cocTiling.moeInitRoutingQuantV2TilingData.gatherOutComputeParamsOp = moeInitRoutingQuantV2TilingBase.quantTilingData.gatherOutComputeParamsOp;
+    tilingData->cocTiling.initRoutingQuantTilingKey = initRoutingQuantTilingKey;
+
+    // 4. workspace
+    size_t *workSpaces = context->GetWorkspaceSizes(1);
+    OP_TILING_CHECK(workSpaces == nullptr, OP_LOGE(nodeName, "workSpaces is nullptr."),
+        return ge::GRAPH_FAILED);
+
+    uint32_t n2 = info.K;
+    uint32_t k2 = info.N / 2;
+
+    uint64_t cocWorkspace = (info.M + 256 - 1) / 256 * 256 * info.topK *sizeof(int32_t) +
+                            info.worldSize * info.worldSize * info.expertPerRank * sizeof(int32_t) * 3 +
+                            info.maxOutputSize * sizeof(float) * 2 +
+                            std::max(info.maxOutputSize * info.N * sizeof(int16_t), info.maxOutputSize * n2 * sizeof(int16_t)) +
+                            std::max(info.maxOutputSize * info.K * sizeof(int8_t), info.maxOutputSize * k2 * sizeof(int8_t));
+
+    workSpaces[0] = SYSTEM_NEED_WORKSPACE + std::max(cocWorkspace, initRoutingWorkspace);
+
+
+    // 5. communication
+    auto attrs = context->GetAttrs();
+    auto group = attrs->GetAttrPointer<char>(static_cast<int>(ATTR_GROUP_INDEX));
+    uint32_t opType = 8U;
+    std::string algConfig = "AlltoAll=level0:fullmesh;level1:pairwise";
+    AscendC::Mc2CcTilingConfig mc2CcTilingConfig(group, opType, algConfig);
+    mc2CcTilingConfig.GetTiling(tilingData->mc2InitTiling);
+    mc2CcTilingConfig.GetTiling(tilingData->mc2CcTiling);
+
+    OP_LOGI(nodeName, "Leave DispatchFFNCombine tiling func.");
+    return ge::GRAPH_SUCCESS;
+}
+
+static ge::graphStatus DispatchFFNCombineTilingFunc(gert::TilingContext* context)
+{
+    return DispatchFFNCombineTilingFuncImpl(context);
+}
+
+struct DispatchFFNCombineCompileInfo {};
+ge::graphStatus TilingParseForDispatchFFNCombine(gert::TilingParseContext *context)
+{
+    (void)context;
+    return ge::GRAPH_SUCCESS;
+}
+
+IMPL_OP_OPTILING(DispatchFFNCombine)
+    .Tiling(DispatchFFNCombineTilingFunc)
+    .TilingParse<DispatchFFNCombineCompileInfo>(TilingParseForDispatchFFNCombine);
+} // namespace optiling

csrc/dispatch_ffn_combine/op_host/error_log.h

@@ -0,0 +1,47 @@
+#ifndef OPS_BUILT_IN_OP_TILING_ERROR_LOG_H_
+#define OPS_BUILT_IN_OP_TILING_ERROR_LOG_H_
+
+#include <string>
+#include "toolchain/slog.h"
+
+#define OP_LOGI(opname, ...)
+#define OP_LOGW(opname, ...)             \
+    do {                                 \
+        printf("[WARN][%s] ", (opname)); \
+        printf(__VA_ARGS__);             \
+        printf("\n");                    \
+    } while (0)
+
+#define OP_LOGE_WITHOUT_REPORT(opname, ...) \
+    do {                                    \
+        printf("[ERRORx][%s] ", (opname));  \
+        printf(__VA_ARGS__);                \
+        printf("\n");                       \
+    } while (0)
+
+#define OP_LOGE(opname, ...)              \
+    do {                                  \
+        printf("[ERROR][%s] ", (opname)); \
+        printf(__VA_ARGS__);              \
+        printf("\n");                     \
+    } while (0)
+
+#define OP_LOGD(opname, ...)
+
+namespace optiling {
+
+#define VECTOR_INNER_ERR_REPORT_TILIING(op_name, err_msg, ...)   \
+    do {                                                         \
+        OP_LOGE_WITHOUT_REPORT(op_name, err_msg, ##__VA_ARGS__); \
+    } while (0)
+
+#define OP_TILING_CHECK(cond, log_func, expr) \
+    do {                                      \
+        if (cond) {                           \
+            log_func;                         \
+            expr;                             \
+        }                                     \
+    } while (0)
+}  // namespace optiling
+
+#endif  // OPS_BUILT_IN_OP_TILING_ERROR_LOG_H_

csrc/dispatch_ffn_combine/op_host/hcom_topo_info.h

@@ -0,0 +1,72 @@
+/* Copyright (c) 2025 Huawei Technologies Co., Ltd.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ * ===================================================================================================================*/
+
+#ifndef METADEF_CXX_INC_EXTERNAL_HCOM_HCOM_TOPO_INFO_H_
+#define METADEF_CXX_INC_EXTERNAL_HCOM_HCOM_TOPO_INFO_H_
+
+#include <unordered_map>
+#include <mutex>
+
+using Status = int32_t;
+
+namespace ge {
+static constexpr uint32_t COMM_MESH = 0b1U;
+static constexpr uint32_t COMM_SWITCH = (COMM_MESH << 1U);
+static constexpr uint32_t COMM_RING = (COMM_MESH << 2U);
+static constexpr uint32_t COMM_PAIRWISE = (COMM_MESH << 3U);
+class HcomTopoInfo {
+ public:
+  enum class TopoLevel {
+    L0 = 0,
+    L1,
+    MAX,
+  };
+  struct TopoLevelDesc {
+    uint32_t comm_sets;
+    uint32_t rank_size;
+  };
+  using TopoDescs = TopoLevelDesc[static_cast<int32_t>(TopoLevel::MAX)];
+  struct TopoInfo {
+    int64_t rank_size;
+    void *notify_handle;
+    TopoDescs topo_level_descs;
+  };
+  static HcomTopoInfo &Instance();
+  bool TopoInfoHasBeenSet(const char_t *group);
+  bool TryGetGroupTopoInfo(const char_t *group, TopoInfo &info);
+  Status SetGroupTopoInfo(const char_t *group, const TopoInfo &info);
+  Status GetGroupRankSize(const char_t *group, int64_t &rank_size);
+  TopoDescs *GetGroupTopoDesc(const char_t *group);
+  Status GetGroupNotifyHandle(const char_t *group, void *&notify_handle);
+  void UnsetGroupTopoInfo(const char_t *group) {
+    const std::lock_guard<std::mutex> lock(mutex_);
+    (void) rank_info_.erase(group);
+  }
+
+  Status SetGroupOrderedStream(const char_t *group, void *stream);
+  Status GetGroupOrderedStream(const char_t *group, void *&stream);
+  void UnsetGroupOrderedStream(const char_t *group) {
+    const std::lock_guard<std::mutex> lock(mutex_);
+    (void) group_to_ordered_stream_.erase(group);
+  };
+
+  Status SetGroupOrderedStream(const int32_t device_id, const char_t *group, void *stream);
+  Status GetGroupOrderedStream(const int32_t device_id, const char_t *group, void *&stream);
+  void UnsetGroupOrderedStream(const int32_t device_id, const char_t *group);
+ private:
+  HcomTopoInfo() = default;
+  ~HcomTopoInfo() = default;
+  std::unordered_map<std::string, TopoInfo> rank_info_;
+  std::mutex mutex_;
+  std::unordered_map<std::string, void*> group_to_ordered_stream_; // 通信域保序流
+  std::unordered_map<int32_t, std::unordered_map<std::string, void*>> device_id_to_group_to_ordered_stream_; // 通信域保序流
+};
+}
+
+#endif // METADEF_CXX_INC_EXTERNAL_HCOM_HCOM_TOPO_INFO_H_

csrc/dispatch_ffn_combine/op_host/tiling_args.h

@@ -0,0 +1,9 @@
+#ifndef TILING_ARGS_H
+#define TILING_ARGS_H
+#include <cstdint>
+
+namespace Moe {
+constexpr uint64_t COMBINE_STATE_WIN_OFFSET = 3U * 1024UL * 1024UL;
+constexpr uint64_t NOTIFY_DISPATCH_WIN_OFFSET = 204U * 1024UL * 1024UL;
+}  // namespace Moe
+#endif  // TILING_ARGS_H