aclnnSwigluGroupQuantGrad【免费下载链接】ops-nn本项目是CANN提供的神经网络类计算算子库实现网络在NPU上加速计算。项目地址: https://gitcode.com/cann/ops-nn 查看源码产品支持情况产品是否支持Ascend 950PR/Ascend 950DT√Atlas A3 训练系列产品/Atlas A3 推理系列产品×Atlas A2 训练系列产品/Atlas A2 推理系列产品×Atlas 200I/500 A2 推理产品×Atlas 推理系列产品×Atlas 训练系列产品×功能说明接口功能SwigluGroupQuantGrad算子实现SwiGLU激活函数分组量化的反向梯度计算。用于计算输入梯度grad_x和权重梯度grad_weight。算子支持范围支持MoE场景传入groupIndex和非MoE场景groupIndex传空支持可选的Clamp反向传播掩码支持可选的Weight梯度计算。计算流程步骤〇GroupIndex处理可选→ 计算trunc步骤一输入切分将x切分为x0和x1步骤二Clamp处理可选步骤三SwiGLU反向传播计算步骤四Weight梯度计算可选步骤五梯度拼接输出MoE场景GroupIndex处理公式$$\text{trunc} \sum_{g0}^{G-1} \text{groupIndex}[g]$$其中$G$ 为MoE专家分组数后续所有步骤仅处理前 $\text{trunc}$ 行数据。输入切分公式$$\mathbf{x}_0[t, h] \mathbf{x}[t, h], \quad h \in [0, H)$$$$\mathbf{x}_1[t, h] \mathbf{x}[t, h H], \quad h \in [0, H)$$Clamp处理公式当clamp_limit 0时$$\mathbf{x}_0[t, h] \min(\mathbf{x}_0[t, h], c)$$$$\mathbf{x}_1[t, h] \min(\max(\mathbf{x}_1[t, h], -c), c)$$其中 $c$ 为clamp_limit。SiLU梯度公式$$\frac{d\text{SiLU}}{d\mathbf{x}_0} \sigma(\mathbf{x}_0) \cdot \left(1 \mathbf{x}_0 \cdot (1 - \sigma(\mathbf{x}_0))\right)$$其中$\sigma(\mathbf{x}_0) \frac{1}{1 e^{-\mathbf{x}_0}}$输入梯度计算公式$$\mathbf{grad}{x_0}[t, h] \mathbf{grad}{y_0}[t, h] \cdot \mathbf{x}_1[t, h] \cdot \frac{d\text{SiLU}}{d\mathbf{x}_0}[t, h]$$$$\mathbf{grad}{x_1}[t, h] \mathbf{grad}{y_0}[t, h] \cdot \text{SiLU}(\mathbf{x}0[t, h])$$其中如果提供了weight则 $\mathbf{grad}{y_0} \mathbf{grad}{\text{output}} \cdot \mathbf{weight}$如果未提供weight则 $\mathbf{grad}{y_0} \mathbf{grad}_{\text{output}}$Weight梯度计算公式可选$$\mathbf{grad}{\text{weight}}[t] \sum{h0}^{H-1} \mathbf{grad}{\text{output}}[t, h] \cdot \mathbf{y}{\text{origin}}[t, h]$$其中$\mathbf{y}_{\text{origin}}$ 为SwiGLU前向传播的原始激活值输出沿最后一维H维度求和。Clamp反向传播掩码公式当clamp_limit 0时$$\mathbf{grad}{x_0}[t, h] \mathbf{grad}{x_0}[t, h] \cdot \mathbb{I}(\mathbf{x}_0[t, h] c)$$$$\mathbf{grad}{x_1}[t, h] \mathbf{grad}{x_1}[t, h] \cdot \mathbb{I}(-c \mathbf{x}_1[t, h] c)$$其中 $\mathbb{I}$ 为指示函数。梯度拼接与GroupIndex处理公式$$\mathbf{grad}x[t, h] \begin{cases}\mathbf{grad}{x_0}[t, h] h \in [0, H) \\mathbf{grad}_{x_1}[t, h-H] h \in [H, 2H)\end{cases}$$$$\mathbf{grad}_x[t, :] \mathbf{grad}_x[t, :] \cdot \mathbb{I}(t \text{trunc})$$函数原型每个算子分为两段式接口必须先调用aclnnSwigluGroupQuantGradGetWorkspaceSize接口获取计算所需workspace大小以及包含了算子计算流程的执行器再调用aclnnSwigluGroupQuantGrad接口执行计算。aclnnStatus aclnnSwigluGroupQuantGradGetWorkspaceSize( const aclTensor *gradY, const aclTensor *x, const aclTensor *weightOptional, const aclTensor *yOriginOptional, const aclIntArray *groupIndexOptional, double clampLimit, const aclTensor *gradXOut, const aclTensor *gradWeightOutOptional, uint64_t *workspaceSize, aclOpExecutor **executor)aclnnStatus aclnnSwigluGroupQuantGrad( void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)aclnnSwigluGroupQuantGradGetWorkspaceSize参数说明参数名输入/输出描述使用说明数据类型数据格式维度(shape)非连续TensorgradYaclTensor*输入梯度输出张量来自下游层的梯度。shape[T, H]或[B, S, H]。T为token数量B为batch sizeS为sequence lengthH为hidden size。BFLOAT16、FLOAT16、FLOATND2-3√xaclTensor*输入前向传播的输入张量。shape[T, 2H]或[B, S, 2H]最后一维必须为偶数。最后一维的H与gradY的H对应。BFLOAT16、FLOAT16、FLOATND2-3√weightOptionalaclTensor*输入可选MoE权重张量。shape[T, 1]或[B, S, 1]需与gradY的第一维一致。当提供weight时必须同时提供yOrigin才能计算gradWeight。FLOATND2-3√yOriginOptionalaclTensor*输入可选SwiGLU前向传播的原始激活值输出。shape[T, H]或[B, S, H]需与gradY的shape一致。当提供weight时必须同时提供yOrigin才能计算gradWeight。BFLOAT16、FLOAT16、FLOATND2-3√groupIndexOptionalaclTensor*输入可选GroupIndex张量动态核分配。shape[G]dtypeINT64。G为MoE专家分组数。groupIndex内元素要求为非递减。INT64ND1√clampLimitfloat输入Clamp阈值。取值范围≥0.0。clampLimit0表示不启用Clamp反向传播掩码。FLOAT---gradXOutaclTensor*输出输入梯度张量。shape[T, 2H]或[B, S, 2H]与x一致。数据类型与gradY/x保持一致。BFLOAT16、FLOAT16、FLOATND2-3√gradWeightOutOptionalaclTensor*输出可选权重梯度张量。当提供weight时输出shape[T, 1]或[B, S, 1]。数据类型为FLOAT。FLOATND2-3√workspaceSizeuint64_t*输出返回需要在Device侧申请的workspace大小。-----executoraclOpExecutor**输出返回op执行器包含了算子计算流程。-----返回值aclnnStatus返回状态码具体参见aclnn返回码。第一段接口会完成入参校验出现以下场景时报错返回码错误码描述ACLNN_ERR_PARAM_NULLPTR161001必选参数gradY/x/gradXOut为nullptr。ACLNN_ERR_INNER_TILING_ERROR161002gradY、x、weight等输入变量的数据类型和数据格式不在支持的范围内。ACLNN_ERR_INNER_TILING_ERROR561002多个输入tensor之间的shape信息不匹配、输入属性不在取值范围详见参数说明。aclnnSwigluGroupQuantGrad参数说明参数名输入/输出描述workspace输入在Device侧申请的workspace内存地址。workspaceSize输入在Device侧申请的workspace大小由第一段接口aclnnSwigluGroupQuantGradGetWorkspaceSize获取。executor输入op执行器包含了算子计算流程。stream输入指定执行任务的Stream。返回值aclnnStatus返回状态码具体参见aclnn返回码。约束说明确定性计算aclnnSwigluGroupQuantGrad默认确定性实现。输入shape约束x最后一维必须为偶数$2H$gradY最后一维为 $H$与x最后一维的一半对应gradY与x的前n-1维shape必须一致可选参数约束weight提供时必须同时提供yOrigin才能计算gradWeightweight的shape需与gradY的第一维一致yOrigin的shape需与gradY一致数据类型约束gradY、x、yOrigin、gradX数据类型必须一致FLOAT、FLOAT16或BFLOAT16weight、gradWeight必须为FLOAT类型groupIndex必须为INT64类型Clamp约束clampLimit必须 ≥ 0.0clampLimit0表示不启用Clamp反向传播掩码调用示例示例代码如下仅供参考具体编译和执行过程请参考编译与运行样例。#include iostream #include vector #include acl/acl.h #include aclnnop/aclnn_swiglu_group_quant_grad.h #define CHECK_RET(cond, return_expr) \ do { \ if (!(cond)) { \ return_expr; \ } \ } while (0) #define LOG_PRINT(message, ...) \ do { \ printf(message, ##__VA_ARGS__); \ } while (0) int64_t GetShapeSize(const std::vectorint64_t shape) { int64_t shapeSize 1; for (auto i : shape) { shapeSize * i; } return shapeSize; } int Init(int32_t deviceId, aclrtStream* stream) { auto ret aclInit(nullptr); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclInit failed. ERROR: %d\n, ret); return ret); ret aclrtSetDevice(deviceId); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtSetDevice failed. ERROR: %d\n, ret); return ret); ret aclrtCreateStream(stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtCreateStream failed. ERROR: %d\n, ret); return ret); return 0; } template typename T int CreateAclTensor(const std::vectorT hostData, const std::vectorint64_t shape, void** deviceAddr, aclDataType dataType, aclTensor** tensor) { auto size GetShapeSize(shape) * sizeof(T); auto ret aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtMalloc failed. ERROR: %d\n, ret); return ret); ret aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtMemcpy failed. ERROR: %d\n, ret); return ret); std::vectorint64_t strides(shape.size(), 1); for (int64_t i shape.size() - 2; i 0; i--) { strides[i] shape[i 1] * strides[i 1]; } *tensor aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), *deviceAddr); return 0; } template typename T int CreateAclTensorWithValue(const std::vectorint64_t shape, void** deviceAddr, aclDataType dataType, aclTensor** tensor, T value) { int64_t shapeSize GetShapeSize(shape); std::vectorT hostData(shapeSize, value); return CreateAclTensor(hostData, shape, deviceAddr, dataType, tensor); } int main() { int32_t deviceId 0; aclrtStream stream; auto ret Init(deviceId, stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(Init acl failed. ERROR: %d\n, ret); return ret); std::vectorint64_t gradYShape {512, 512}; std::vectorint64_t xShape {512, 1024}; std::vectorint64_t weightShape {512, 1}; std::vectorint64_t yOriginShape {512, 512}; std::vectorint64_t groupIndexShape {256}; std::vectorint64_t gradXShape {512, 1024}; std::vectorint64_t gradWeightShape {512, 1}; void* gradYDeviceAddr nullptr; void* xDeviceAddr nullptr; void* weightDeviceAddr nullptr; void* yOriginDeviceAddr nullptr; void* groupIndexDeviceAddr nullptr; void* gradXDeviceAddr nullptr; void* gradWeightDeviceAddr nullptr; aclTensor* gradYTensor nullptr; aclTensor* xTensor nullptr; aclTensor* weightTensor nullptr; aclTensor* yOriginTensor nullptr; aclIntArray* groupIndexArray nullptr; aclTensor* gradXTensor nullptr; aclTensor* gradWeightTensor nullptr; int64_t gradYSize GetShapeSize(gradYShape); std::vectorfloat gradYHostData(gradYSize, 1.0f); for (int64_t i 0; i gradYSize; i) { gradYHostData[i] static_castfloat(i % 10) * 0.1f; } int64_t xSize GetShapeSize(xShape); std::vectorfloat xHostData(xSize, 1.0f); for (int64_t i 0; i xSize; i) { xHostData[i] static_castfloat((i % 20) - 10) * 0.5f; } int64_t weightSize GetShapeSize(weightShape); std::vectorfloat weightHostData(weightSize, 1.0f); for (int64_t i 0; i weightSize; i) { weightHostData[i] static_castfloat((i % 5) 1) * 0.2f; } int64_t yOriginSize GetShapeSize(yOriginShape); std::vectorfloat yOriginHostData(yOriginSize, 1.0f); for (int64_t i 0; i yOriginSize; i) { yOriginHostData[i] static_castfloat((i % 8) 1) * 0.3f; } int64_t groupIndexSize GetShapeSize(groupIndexShape); std::vectorint64_t groupIndexHostData(groupIndexSize, 0); int64_t groupStride 512 / 256; for (int64_t i 0; i groupIndexSize; i) { groupIndexHostData[i] i * groupStride; } ret CreateAclTensor(gradYHostData, gradYShape, gradYDeviceAddr, aclDataType::ACL_FLOAT16, gradYTensor); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensor(xHostData, xShape, xDeviceAddr, aclDataType::ACL_FLOAT16, xTensor); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensor(weightHostData, weightShape, weightDeviceAddr, aclDataType::ACL_FLOAT, weightTensor); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensor(yOriginHostData, yOriginShape, yOriginDeviceAddr, aclDataType::ACL_FLOAT16, yOriginTensor); CHECK_RET(ret ACL_SUCCESS, return ret); std::vectorint64_t groupArray {256, 256}; groupIndexArray aclCreateIntArray(groupArray.data(), groupArray.size()); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorWithValuefloat(gradXShape, gradXDeviceAddr, aclDataType::ACL_FLOAT16, gradXTensor, 0.0f); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorWithValuefloat(gradWeightShape, gradWeightDeviceAddr, aclDataType::ACL_FLOAT, gradWeightTensor, 0.0f); CHECK_RET(ret ACL_SUCCESS, return ret); float clampLimit 1.0f; uint64_t workspaceSize 0; aclOpExecutor* executor; ret aclnnSwigluGroupQuantGradGetWorkspaceSize(gradYTensor, xTensor, weightTensor, yOriginTensor, groupIndexArray, clampLimit, gradXTensor, gradWeightTensor, workspaceSize, executor); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclnnSwigluGroupQuantGradGetWorkspaceSize failed. ERROR: %d\n, ret); return ret); void* workspaceAddr nullptr; if (workspaceSize 0) { ret aclrtMalloc(workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(allocate workspace failed. ERROR: %d\n, ret); return ret); } ret aclnnSwigluGroupQuantGrad(workspaceAddr, workspaceSize, executor, stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclnnSwigluGroupQuantGrad failed. ERROR: %d\n, ret); return ret); ret aclrtSynchronizeStream(stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtSynchronizeStream failed. ERROR: %d\n, ret); return ret); auto gradXResultSize GetShapeSize(gradXShape); std::vectorfloat gradXResultData(gradXResultSize, 0); ret aclrtMemcpy(gradXResultData.data(), gradXResultData.size() * sizeof(float), gradXDeviceAddr, gradXResultSize * sizeof(float), ACL_MEMCPY_DEVICE_TO_HOST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(copy gradX result from device to host failed. ERROR: %d\n, ret); return ret); LOG_PRINT(gradX output (first 10 elements):\n); for (int64_t i 0; i 10 i gradXResultSize; i) { LOG_PRINT(gradX[%ld] %f\n, i, gradXResultData[i]); } auto gradWeightResultSize GetShapeSize(gradWeightShape); std::vectorfloat gradWeightResultData(gradWeightResultSize, 0); ret aclrtMemcpy(gradWeightResultData.data(), gradWeightResultData.size() * sizeof(float), gradWeightDeviceAddr, gradWeightResultSize * sizeof(float), ACL_MEMCPY_DEVICE_TO_HOST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(copy gradWeight result from device to host failed. ERROR: %d\n, ret); return ret); LOG_PRINT(gradWeight output (first 10 elements):\n); for (int64_t i 0; i 10 i gradWeightResultSize; i) { LOG_PRINT(gradWeight[%ld] %f\n, i, gradWeightResultData[i]); } aclDestroyTensor(gradYTensor); aclDestroyTensor(xTensor); aclDestroyTensor(weightTensor); aclDestroyTensor(yOriginTensor); aclDestroyTensor(gradXTensor); aclDestroyTensor(gradWeightTensor); aclrtFree(gradYDeviceAddr); aclrtFree(xDeviceAddr); aclrtFree(weightDeviceAddr); aclrtFree(yOriginDeviceAddr); aclrtFree(groupIndexDeviceAddr); aclrtFree(gradXDeviceAddr); aclrtFree(gradWeightDeviceAddr); if (workspaceSize 0) { aclrtFree(workspaceAddr); } aclrtDestroyStream(stream); aclrtResetDevice(deviceId); aclFinalize(); return 0; }【免费下载链接】ops-nn本项目是CANN提供的神经网络类计算算子库实现网络在NPU上加速计算。项目地址: https://gitcode.com/cann/ops-nn创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考
CANN/ops-nn SwiGLU分组量化梯度算子
发布时间：2026/6/17 15:47:07
aclnnSwigluGroupQuantGrad【免费下载链接】ops-nn本项目是CANN提供的神经网络类计算算子库实现网络在NPU上加速计算。项目地址: https://gitcode.com/cann/ops-nn 查看源码产品支持情况产品是否支持Ascend 950PR/Ascend 950DT√Atlas A3 训练系列产品/Atlas A3 推理系列产品×Atlas A2 训练系列产品/Atlas A2 推理系列产品×Atlas 200I/500 A2 推理产品×Atlas 推理系列产品×Atlas 训练系列产品×功能说明接口功能SwigluGroupQuantGrad算子实现SwiGLU激活函数分组量化的反向梯度计算。用于计算输入梯度grad_x和权重梯度grad_weight。算子支持范围支持MoE场景传入groupIndex和非MoE场景groupIndex传空支持可选的Clamp反向传播掩码支持可选的Weight梯度计算。计算流程步骤〇GroupIndex处理可选→ 计算trunc步骤一输入切分将x切分为x0和x1步骤二Clamp处理可选步骤三SwiGLU反向传播计算步骤四Weight梯度计算可选步骤五梯度拼接输出MoE场景GroupIndex处理公式$$\text{trunc} \sum_{g0}^{G-1} \text{groupIndex}[g]$$其中$G$ 为MoE专家分组数后续所有步骤仅处理前 $\text{trunc}$ 行数据。输入切分公式$$\mathbf{x}_0[t, h] \mathbf{x}[t, h], \quad h \in [0, H)$$$$\mathbf{x}_1[t, h] \mathbf{x}[t, h H], \quad h \in [0, H)$$Clamp处理公式当clamp_limit 0时$$\mathbf{x}_0[t, h] \min(\mathbf{x}_0[t, h], c)$$$$\mathbf{x}_1[t, h] \min(\max(\mathbf{x}_1[t, h], -c), c)$$其中 $c$ 为clamp_limit。SiLU梯度公式$$\frac{d\text{SiLU}}{d\mathbf{x}_0} \sigma(\mathbf{x}_0) \cdot \left(1 \mathbf{x}_0 \cdot (1 - \sigma(\mathbf{x}_0))\right)$$其中$\sigma(\mathbf{x}_0) \frac{1}{1 e^{-\mathbf{x}_0}}$输入梯度计算公式$$\mathbf{grad}{x_0}[t, h] \mathbf{grad}{y_0}[t, h] \cdot \mathbf{x}_1[t, h] \cdot \frac{d\text{SiLU}}{d\mathbf{x}_0}[t, h]$$$$\mathbf{grad}{x_1}[t, h] \mathbf{grad}{y_0}[t, h] \cdot \text{SiLU}(\mathbf{x}0[t, h])$$其中如果提供了weight则 $\mathbf{grad}{y_0} \mathbf{grad}{\text{output}} \cdot \mathbf{weight}$如果未提供weight则 $\mathbf{grad}{y_0} \mathbf{grad}_{\text{output}}$Weight梯度计算公式可选$$\mathbf{grad}{\text{weight}}[t] \sum{h0}^{H-1} \mathbf{grad}{\text{output}}[t, h] \cdot \mathbf{y}{\text{origin}}[t, h]$$其中$\mathbf{y}_{\text{origin}}$ 为SwiGLU前向传播的原始激活值输出沿最后一维H维度求和。Clamp反向传播掩码公式当clamp_limit 0时$$\mathbf{grad}{x_0}[t, h] \mathbf{grad}{x_0}[t, h] \cdot \mathbb{I}(\mathbf{x}_0[t, h] c)$$$$\mathbf{grad}{x_1}[t, h] \mathbf{grad}{x_1}[t, h] \cdot \mathbb{I}(-c \mathbf{x}_1[t, h] c)$$其中 $\mathbb{I}$ 为指示函数。梯度拼接与GroupIndex处理公式$$\mathbf{grad}x[t, h] \begin{cases}\mathbf{grad}{x_0}[t, h] h \in [0, H) \\mathbf{grad}_{x_1}[t, h-H] h \in [H, 2H)\end{cases}$$$$\mathbf{grad}_x[t, :] \mathbf{grad}_x[t, :] \cdot \mathbb{I}(t \text{trunc})$$函数原型每个算子分为两段式接口必须先调用aclnnSwigluGroupQuantGradGetWorkspaceSize接口获取计算所需workspace大小以及包含了算子计算流程的执行器再调用aclnnSwigluGroupQuantGrad接口执行计算。aclnnStatus aclnnSwigluGroupQuantGradGetWorkspaceSize( const aclTensor *gradY, const aclTensor *x, const aclTensor *weightOptional, const aclTensor *yOriginOptional, const aclIntArray *groupIndexOptional, double clampLimit, const aclTensor *gradXOut, const aclTensor *gradWeightOutOptional, uint64_t *workspaceSize, aclOpExecutor **executor)aclnnStatus aclnnSwigluGroupQuantGrad( void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)aclnnSwigluGroupQuantGradGetWorkspaceSize参数说明参数名输入/输出描述使用说明数据类型数据格式维度(shape)非连续TensorgradYaclTensor*输入梯度输出张量来自下游层的梯度。shape[T, H]或[B, S, H]。T为token数量B为batch sizeS为sequence lengthH为hidden size。BFLOAT16、FLOAT16、FLOATND2-3√xaclTensor*输入前向传播的输入张量。shape[T, 2H]或[B, S, 2H]最后一维必须为偶数。最后一维的H与gradY的H对应。BFLOAT16、FLOAT16、FLOATND2-3√weightOptionalaclTensor*输入可选MoE权重张量。shape[T, 1]或[B, S, 1]需与gradY的第一维一致。当提供weight时必须同时提供yOrigin才能计算gradWeight。FLOATND2-3√yOriginOptionalaclTensor*输入可选SwiGLU前向传播的原始激活值输出。shape[T, H]或[B, S, H]需与gradY的shape一致。当提供weight时必须同时提供yOrigin才能计算gradWeight。BFLOAT16、FLOAT16、FLOATND2-3√groupIndexOptionalaclTensor*输入可选GroupIndex张量动态核分配。shape[G]dtypeINT64。G为MoE专家分组数。groupIndex内元素要求为非递减。INT64ND1√clampLimitfloat输入Clamp阈值。取值范围≥0.0。clampLimit0表示不启用Clamp反向传播掩码。FLOAT---gradXOutaclTensor*输出输入梯度张量。shape[T, 2H]或[B, S, 2H]与x一致。数据类型与gradY/x保持一致。BFLOAT16、FLOAT16、FLOATND2-3√gradWeightOutOptionalaclTensor*输出可选权重梯度张量。当提供weight时输出shape[T, 1]或[B, S, 1]。数据类型为FLOAT。FLOATND2-3√workspaceSizeuint64_t*输出返回需要在Device侧申请的workspace大小。-----executoraclOpExecutor**输出返回op执行器包含了算子计算流程。-----返回值aclnnStatus返回状态码具体参见aclnn返回码。第一段接口会完成入参校验出现以下场景时报错返回码错误码描述ACLNN_ERR_PARAM_NULLPTR161001必选参数gradY/x/gradXOut为nullptr。ACLNN_ERR_INNER_TILING_ERROR161002gradY、x、weight等输入变量的数据类型和数据格式不在支持的范围内。ACLNN_ERR_INNER_TILING_ERROR561002多个输入tensor之间的shape信息不匹配、输入属性不在取值范围详见参数说明。aclnnSwigluGroupQuantGrad参数说明参数名输入/输出描述workspace输入在Device侧申请的workspace内存地址。workspaceSize输入在Device侧申请的workspace大小由第一段接口aclnnSwigluGroupQuantGradGetWorkspaceSize获取。executor输入op执行器包含了算子计算流程。stream输入指定执行任务的Stream。返回值aclnnStatus返回状态码具体参见aclnn返回码。约束说明确定性计算aclnnSwigluGroupQuantGrad默认确定性实现。输入shape约束x最后一维必须为偶数$2H$gradY最后一维为 $H$与x最后一维的一半对应gradY与x的前n-1维shape必须一致可选参数约束weight提供时必须同时提供yOrigin才能计算gradWeightweight的shape需与gradY的第一维一致yOrigin的shape需与gradY一致数据类型约束gradY、x、yOrigin、gradX数据类型必须一致FLOAT、FLOAT16或BFLOAT16weight、gradWeight必须为FLOAT类型groupIndex必须为INT64类型Clamp约束clampLimit必须 ≥ 0.0clampLimit0表示不启用Clamp反向传播掩码调用示例示例代码如下仅供参考具体编译和执行过程请参考编译与运行样例。#include iostream #include vector #include acl/acl.h #include aclnnop/aclnn_swiglu_group_quant_grad.h #define CHECK_RET(cond, return_expr) \ do { \ if (!(cond)) { \ return_expr; \ } \ } while (0) #define LOG_PRINT(message, ...) \ do { \ printf(message, ##__VA_ARGS__); \ } while (0) int64_t GetShapeSize(const std::vectorint64_t shape) { int64_t shapeSize 1; for (auto i : shape) { shapeSize * i; } return shapeSize; } int Init(int32_t deviceId, aclrtStream* stream) { auto ret aclInit(nullptr); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclInit failed. ERROR: %d\n, ret); return ret); ret aclrtSetDevice(deviceId); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtSetDevice failed. ERROR: %d\n, ret); return ret); ret aclrtCreateStream(stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtCreateStream failed. ERROR: %d\n, ret); return ret); return 0; } template typename T int CreateAclTensor(const std::vectorT hostData, const std::vectorint64_t shape, void** deviceAddr, aclDataType dataType, aclTensor** tensor) { auto size GetShapeSize(shape) * sizeof(T); auto ret aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtMalloc failed. ERROR: %d\n, ret); return ret); ret aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtMemcpy failed. ERROR: %d\n, ret); return ret); std::vectorint64_t strides(shape.size(), 1); for (int64_t i shape.size() - 2; i 0; i--) { strides[i] shape[i 1] * strides[i 1]; } *tensor aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), *deviceAddr); return 0; } template typename T int CreateAclTensorWithValue(const std::vectorint64_t shape, void** deviceAddr, aclDataType dataType, aclTensor** tensor, T value) { int64_t shapeSize GetShapeSize(shape); std::vectorT hostData(shapeSize, value); return CreateAclTensor(hostData, shape, deviceAddr, dataType, tensor); } int main() { int32_t deviceId 0; aclrtStream stream; auto ret Init(deviceId, stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(Init acl failed. ERROR: %d\n, ret); return ret); std::vectorint64_t gradYShape {512, 512}; std::vectorint64_t xShape {512, 1024}; std::vectorint64_t weightShape {512, 1}; std::vectorint64_t yOriginShape {512, 512}; std::vectorint64_t groupIndexShape {256}; std::vectorint64_t gradXShape {512, 1024}; std::vectorint64_t gradWeightShape {512, 1}; void* gradYDeviceAddr nullptr; void* xDeviceAddr nullptr; void* weightDeviceAddr nullptr; void* yOriginDeviceAddr nullptr; void* groupIndexDeviceAddr nullptr; void* gradXDeviceAddr nullptr; void* gradWeightDeviceAddr nullptr; aclTensor* gradYTensor nullptr; aclTensor* xTensor nullptr; aclTensor* weightTensor nullptr; aclTensor* yOriginTensor nullptr; aclIntArray* groupIndexArray nullptr; aclTensor* gradXTensor nullptr; aclTensor* gradWeightTensor nullptr; int64_t gradYSize GetShapeSize(gradYShape); std::vectorfloat gradYHostData(gradYSize, 1.0f); for (int64_t i 0; i gradYSize; i) { gradYHostData[i] static_castfloat(i % 10) * 0.1f; } int64_t xSize GetShapeSize(xShape); std::vectorfloat xHostData(xSize, 1.0f); for (int64_t i 0; i xSize; i) { xHostData[i] static_castfloat((i % 20) - 10) * 0.5f; } int64_t weightSize GetShapeSize(weightShape); std::vectorfloat weightHostData(weightSize, 1.0f); for (int64_t i 0; i weightSize; i) { weightHostData[i] static_castfloat((i % 5) 1) * 0.2f; } int64_t yOriginSize GetShapeSize(yOriginShape); std::vectorfloat yOriginHostData(yOriginSize, 1.0f); for (int64_t i 0; i yOriginSize; i) { yOriginHostData[i] static_castfloat((i % 8) 1) * 0.3f; } int64_t groupIndexSize GetShapeSize(groupIndexShape); std::vectorint64_t groupIndexHostData(groupIndexSize, 0); int64_t groupStride 512 / 256; for (int64_t i 0; i groupIndexSize; i) { groupIndexHostData[i] i * groupStride; } ret CreateAclTensor(gradYHostData, gradYShape, gradYDeviceAddr, aclDataType::ACL_FLOAT16, gradYTensor); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensor(xHostData, xShape, xDeviceAddr, aclDataType::ACL_FLOAT16, xTensor); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensor(weightHostData, weightShape, weightDeviceAddr, aclDataType::ACL_FLOAT, weightTensor); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensor(yOriginHostData, yOriginShape, yOriginDeviceAddr, aclDataType::ACL_FLOAT16, yOriginTensor); CHECK_RET(ret ACL_SUCCESS, return ret); std::vectorint64_t groupArray {256, 256}; groupIndexArray aclCreateIntArray(groupArray.data(), groupArray.size()); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorWithValuefloat(gradXShape, gradXDeviceAddr, aclDataType::ACL_FLOAT16, gradXTensor, 0.0f); CHECK_RET(ret ACL_SUCCESS, return ret); ret CreateAclTensorWithValuefloat(gradWeightShape, gradWeightDeviceAddr, aclDataType::ACL_FLOAT, gradWeightTensor, 0.0f); CHECK_RET(ret ACL_SUCCESS, return ret); float clampLimit 1.0f; uint64_t workspaceSize 0; aclOpExecutor* executor; ret aclnnSwigluGroupQuantGradGetWorkspaceSize(gradYTensor, xTensor, weightTensor, yOriginTensor, groupIndexArray, clampLimit, gradXTensor, gradWeightTensor, workspaceSize, executor); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclnnSwigluGroupQuantGradGetWorkspaceSize failed. ERROR: %d\n, ret); return ret); void* workspaceAddr nullptr; if (workspaceSize 0) { ret aclrtMalloc(workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(allocate workspace failed. ERROR: %d\n, ret); return ret); } ret aclnnSwigluGroupQuantGrad(workspaceAddr, workspaceSize, executor, stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclnnSwigluGroupQuantGrad failed. ERROR: %d\n, ret); return ret); ret aclrtSynchronizeStream(stream); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(aclrtSynchronizeStream failed. ERROR: %d\n, ret); return ret); auto gradXResultSize GetShapeSize(gradXShape); std::vectorfloat gradXResultData(gradXResultSize, 0); ret aclrtMemcpy(gradXResultData.data(), gradXResultData.size() * sizeof(float), gradXDeviceAddr, gradXResultSize * sizeof(float), ACL_MEMCPY_DEVICE_TO_HOST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(copy gradX result from device to host failed. ERROR: %d\n, ret); return ret); LOG_PRINT(gradX output (first 10 elements):\n); for (int64_t i 0; i 10 i gradXResultSize; i) { LOG_PRINT(gradX[%ld] %f\n, i, gradXResultData[i]); } auto gradWeightResultSize GetShapeSize(gradWeightShape); std::vectorfloat gradWeightResultData(gradWeightResultSize, 0); ret aclrtMemcpy(gradWeightResultData.data(), gradWeightResultData.size() * sizeof(float), gradWeightDeviceAddr, gradWeightResultSize * sizeof(float), ACL_MEMCPY_DEVICE_TO_HOST); CHECK_RET(ret ACL_SUCCESS, LOG_PRINT(copy gradWeight result from device to host failed. ERROR: %d\n, ret); return ret); LOG_PRINT(gradWeight output (first 10 elements):\n); for (int64_t i 0; i 10 i gradWeightResultSize; i) { LOG_PRINT(gradWeight[%ld] %f\n, i, gradWeightResultData[i]); } aclDestroyTensor(gradYTensor); aclDestroyTensor(xTensor); aclDestroyTensor(weightTensor); aclDestroyTensor(yOriginTensor); aclDestroyTensor(gradXTensor); aclDestroyTensor(gradWeightTensor); aclrtFree(gradYDeviceAddr); aclrtFree(xDeviceAddr); aclrtFree(weightDeviceAddr); aclrtFree(yOriginDeviceAddr); aclrtFree(groupIndexDeviceAddr); aclrtFree(gradXDeviceAddr); aclrtFree(gradWeightDeviceAddr); if (workspaceSize 0) { aclrtFree(workspaceAddr); } aclrtDestroyStream(stream); aclrtResetDevice(deviceId); aclFinalize(); return 0; }【免费下载链接】ops-nn本项目是CANN提供的神经网络类计算算子库实现网络在NPU上加速计算。项目地址: https://gitcode.com/cann/ops-nn创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考
