* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (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 mla_preprocess_infershape.cpp
* \brief
*/
#include <register/op_impl_registry.h>
#include "log/log.h"
using namespace ge;
namespace ops {
constexpr size_t ATTR_CACHE_MODE = 9;
constexpr size_t INPUT_INPUT = 0;
constexpr size_t INPUT_WUK = 18;
constexpr size_t INPUT_KV_CACHE = 19;
constexpr size_t INPUT_KV_CACHE_ROPE = 20;
constexpr size_t OUTPUT_Q_OUT = 0;
constexpr size_t OUTPUT_KV_CACHE_OUT = 1;
constexpr size_t OUTPUT_Q_ROPE_OUT = 2;
constexpr size_t OUTPUT_KR_CACHE_OUT = 3;
constexpr size_t DIM_TWO = 2;
constexpr size_t DIM_THREE = 3;
constexpr size_t DIM_SHAPE_Q = 512;
constexpr size_t DIM_SHAPE_QR = 64;
constexpr size_t DIM_SHAPE_Q_CACHE = 576;
static ge::graphStatus InferShapeMlaPreprocess(gert::InferShapeContext* context) {
OP_LOGD(context->GetNodeName(), "Begin to do InferShapeMlaPreprocess.");
auto attrs = context->GetAttrs();
OP_CHECK_NULL_WITH_CONTEXT(context, attrs);
const int64_t *cacheModePtr = attrs->GetAttrPointer<int64_t>(ATTR_CACHE_MODE);
if (cacheModePtr == nullptr) {
return ge::GRAPH_FAILED;
}
const int64_t cacheMode = *cacheModePtr;
const gert::Shape* kvCacheShape = context->GetInputShape(INPUT_KV_CACHE);
const gert::Shape* kvCacheRopeShape = context->GetInputShape(INPUT_KV_CACHE_ROPE);
const gert::Shape* inputShape = context->GetInputShape(INPUT_INPUT);
const gert::Shape* wukShape = context->GetInputShape(INPUT_WUK);
const int64_t tokenNum = inputShape->GetDim(0);
const int64_t headNum = wukShape->GetDim(0);
gert::Shape* qOutShape = context->GetOutputShape(OUTPUT_Q_OUT);
gert::Shape* kvCacheOutShape = context->GetOutputShape(OUTPUT_KV_CACHE_OUT);
*kvCacheOutShape = *kvCacheShape;
qOutShape->SetDimNum(DIM_THREE);
qOutShape->SetDim(0, tokenNum);
qOutShape->SetDim(1, headNum);
if (cacheMode != 0) {
gert::Shape* qRopeOutShape = context->GetOutputShape(OUTPUT_Q_ROPE_OUT);
gert::Shape* krCacheOutShape = context->GetOutputShape(OUTPUT_KR_CACHE_OUT);
*krCacheOutShape = *kvCacheRopeShape;
qOutShape->SetDim(DIM_TWO, DIM_SHAPE_Q);
qRopeOutShape->SetDimNum(DIM_THREE);
qRopeOutShape->SetDim(0, tokenNum);
qRopeOutShape->SetDim(1, headNum);
qRopeOutShape->SetDim(DIM_TWO, DIM_SHAPE_QR);
} else {
qOutShape->SetDim(DIM_TWO, DIM_SHAPE_Q_CACHE);
}
OP_LOGD(context->GetNodeName(), "End to do InferShapeMlaPreprocess.");
return ge::GRAPH_SUCCESS;
}
static ge::graphStatus InferDataTypeMlaPreprocess(gert::InferDataTypeContext* context) {
OP_LOGD(context->GetNodeName(), "Begin to do InferDataTypeMlaPreprocess.");
auto attrs = context->GetAttrs();
OP_CHECK_NULL_WITH_CONTEXT(context, attrs);
const int64_t *cacheModePtr = attrs->GetAttrPointer<int64_t>(ATTR_CACHE_MODE);
if (cacheModePtr == nullptr) {
return ge::GRAPH_FAILED;
}
const int64_t cacheMode = *cacheModePtr;
auto inputDtype = context->GetInputDataType(INPUT_INPUT);
auto kvCacheDtype = context->GetInputDataType(INPUT_KV_CACHE);
context->SetOutputDataType(OUTPUT_Q_OUT, kvCacheDtype);
context->SetOutputDataType(OUTPUT_KV_CACHE_OUT, kvCacheDtype);
if (cacheMode != 0) {
context->SetOutputDataType(OUTPUT_Q_ROPE_OUT, inputDtype);
context->SetOutputDataType(OUTPUT_KR_CACHE_OUT, inputDtype);
}
OP_LOGD(context->GetNodeName(), "End to do InferDataTypeMlaPreprocess.");
return ge::GRAPH_SUCCESS;
}
IMPL_OP_INFERSHAPE(MlaPreprocess).InferShape(InferShapeMlaPreprocess).InferDataType(InferDataTypeMlaPreprocess);
}