EngineX-Ascend/enginex-ascend-910-vllm
11
0
Fork 1
Code Issues Pull Requests Actions Projects Releases Wiki Activity

init v0.23.0

Browse Source 
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
This commit is contained in:
Sun Ruoxi
2026-08-27 15:11:51 +08:00
parent b582a8e7d1
commit 7f8a1b1f7a
2849 changed files with 712887 additions and 22001 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
csrc/mc2/dispatch_ffn_combine/op_host/op_api/aclnn_dispatch_ffn_combine.cpp Normal file
View File

@@ -0,0 +1,83 @@
/**
* 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>
// 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 aclTensorList* weight1, const aclTensorList* weight2,
const aclTensor* expertId, const aclTensorList* scale1, const aclTensorList* scale2,
const aclTensor* probs, const aclTensor* xActiveMask,
const char* group, int64_t maxOutputSize,
bool transB, bool weightNz, double swigluLimit,
const aclTensor* out, const aclTensor* expertTokenNums,
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 aclTensorList* weight1, const aclTensorList* weight2,
const aclTensor* expertId, const aclTensorList* scale1, const aclTensorList* scale2,
const aclTensor* probs, const aclTensor* xActiveMask,
const char* group, int64_t maxOutputSize, double swigluLimit,
const aclTensor* out, const aclTensor* expertTokenNums,
uint64_t* workspaceSize, aclOpExecutor** executor)
{
bool transB = false;
bool weightNz = true;
aclnnStatus ret = aclnnInnerDispatchFFNCombineGetWorkspaceSize(x, weight1, weight2, expertId, scale1, scale2, probs, xActiveMask, group,
maxOutputSize, transB, weightNz, swigluLimit,
out, expertTokenNums, 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

64
csrc/mc2/dispatch_ffn_combine/op_host/op_api/aclnn_dispatch_ffn_combine.h Normal file
View File

@@ -0,0 +1,64 @@
/**
* 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
/**
* Operator function: fuse all distributed MoE ops from InitRouting through Unpermute.
* @brief First-stage interface of aclnnDispatchFFNCombine that calculates workspace size based on the specific compute flow.
* @domain aclnn_ops_infer
* @param [in] x: The input tensor.
* @param [in] weight1: The first weight tensor.
* @param [in] weight2: The second weight tensor.
* @param [in] expertId: The expert ID tensor.
* @param [in] scale1: The first scale tensor.
* @param [in] scale2: The second scale tensor.
* @param [in] probs: The probabilities tensor.
* @param [in] group: string identifying the communication domain name.
* @param [in] maxOutputSize: The maximum output size.
* @param [out] out: result of computation + communication; same dtype as input.
* @param [out] workspaceSize: workspace size to allocate on the NPU device side.
* @param [out] executor: op executor containing the operator compute flow.
* @return aclnnStatus: status code.
*/
__attribute__((visibility("default"))) aclnnStatus aclnnDispatchFFNCombineGetWorkspaceSize(const aclTensor* x, const aclTensorList* weight1, const aclTensorList* weight2,
const aclTensor* expertId, const aclTensorList* scale1, const aclTensorList* scale2,
const aclTensor* probs, const aclTensor* xActiveMask,
const char* group, int64_t maxOutputSize, double swigluLimit,
const aclTensor* out, const aclTensor* expertTokenNums,
uint64_t* workspaceSize, aclOpExecutor** executor);
/**
* @brief Second-stage interface of aclnnDispatchFFNCombine to execute computation.
* @param [in] workspace: workspace memory address allocated on the NPU device side.
* @param [in] workspace_size: workspace size allocated on the NPU device side, obtained from aclnnDispatchFFNCombineGetWorkspaceSize.
* @param [in] executor: op executor containing the operator compute flow.
* @param [in] stream: acl stream.
* @return aclnnStatus: status code.
*/
__attribute__((visibility("default"))) aclnnStatus aclnnDispatchFFNCombine(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor,
aclrtStream stream);
#ifdef __cplusplus
}
#endif
#endif // OP_API_INC_DISPATCH_FFN_COMBINE_