#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/custom_functions/atb/AtbCommon.h"
#include <acl/acl.h>
using namespace std;
namespace atb {
using PagedAttentionParam = atb::infer::PagedAttentionParam;
void _npu_paged_attention_quant(const at::Tensor &query, const at::Tensor &key_cache, const at::Tensor &value_cache, const int64_t num_kv_heads, const int64_t num_heads, const double scale_value, const at::Tensor &block_table, const at::Tensor &context_lens,
const int64_t quant_type, const int64_t outdata_type, const at::Tensor &k_descale, const at::Tensor &v_descale, at::Tensor &out)
{
const c10::OptionalDeviceGuard device_guard(device_of(query));
OpParamCache<PagedAttentionParam>& pagedAttentionParamCache = OpParamCache<PagedAttentionParam>::getInstance();
PagedAttentionParam pagedparam;
pagedparam.headNum = num_heads;
pagedparam.qkScale = scale_value;
pagedparam.kvHeadNum = num_kv_heads;
pagedparam.maskType = PagedAttentionParam::UNDEFINED;
pagedparam.batchRunStatusEnable = false;
auto quanttype = static_cast<PagedAttentionParam::QuantType>(quant_type);
pagedparam.quantType = quanttype;
auto outdataType = static_cast<aclDataType>(outdata_type);
pagedparam.outDataType = outdataType;
pagedparam.hasQuantOffset = false;
pagedparam.compressType = PagedAttentionParam::COMPRESS_TYPE_UNDEFINED;
pagedparam.calcType = PagedAttentionParam::CALC_TYPE_UNDEFINED;
pagedparam.scaleType = PagedAttentionParam::SCALE_TYPE_TOR;
pagedparam.inputLayout = atb::infer::TYPE_BSND;
pagedparam.mlaVHeadSize = 0;
ParamSetter paramsetter;
paramsetter.Input(query, true)
.Input(key_cache)
.Input(value_cache)
.Input(block_table, true)
.Input(context_lens, true)
.Input(k_descale, true)
.Input(v_descale, true)
.Output(out);
auto opPaged = pagedAttentionParamCache.getOperation(pagedparam, "PagedAttentionOperation");
RunAtbCmd(opPaged, paramsetter, "PagedAttentionOperation");
return;
}
namespace {
TORCH_LIBRARY_FRAGMENT(atb, m)
{
m.def("_npu_paged_attention_quant(Tensor query, Tensor key_cache, Tensor value_cache, int num_kv_heads, int num_heads, float scale_value, Tensor block_table, Tensor context_lens, int quant_type, int outdata_type, Tensor k_descale, Tensor v_descale, Tensor(a!) out) -> ()");
}
}
namespace {
TORCH_LIBRARY_IMPL(atb, PrivateUse1, m)
{
m.impl("_npu_paged_attention_quant", TORCH_FN(atb::_npu_paged_attention_quant));
}
}
}