#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/custom_functions/atb/AtbCommon.h"
#include <acl/acl.h>
namespace atb {
using PagedAttentionParam = atb::infer::PagedAttentionParam;
void _npu_paged_attention(const at::Tensor &query, const at::Tensor &key_cache, const at::Tensor &value_cache, int64_t num_kv_heads, int64_t num_heads, double scale_value,
const at::Tensor &block_table, const at::Tensor &context_lens, at::Tensor &out, const c10::optional<at::Tensor> &workspace)
{
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;
pagedparam.quantType = PagedAttentionParam::TYPE_QUANT_UNDEFINED;
pagedparam.outDataType = ACL_DT_UNDEFINED;
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)
.Output(out);
auto opPaged = pagedAttentionParamCache.getOperation(pagedparam, "PagedAttentionOperation");
if (workspace.has_value() && workspace.value().defined()) {
RunAtbCmdWithWorkspace(opPaged, paramsetter, "PagedAttentionOperation", workspace.value());
} else {
RunAtbCmd(opPaged, paramsetter, "PagedAttentionOperation");
}
return;
}
at::Tensor _npu_paged_attention_get_workspace(const at::Tensor &query, const at::Tensor &key_cache, const at::Tensor &value_cache, int64_t num_kv_heads, int64_t num_heads, double scale_value,
const at::Tensor &block_table, const at::Tensor &context_lens, 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;
pagedparam.quantType = PagedAttentionParam::TYPE_QUANT_UNDEFINED;
pagedparam.outDataType = ACL_DT_UNDEFINED;
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)
.Output(out);
auto opPaged = pagedAttentionParamCache.getOperation(pagedparam, "PagedAttentionOperation");
uint64_t workspace_size = GetAtbWorkspaceSizeCmd(opPaged, paramsetter, "PagedAttentionOperation");
at::Tensor workspace_tensor = at::empty({workspace_size}, query.options().dtype(at::kByte));
return workspace_tensor;
}
namespace {
TORCH_LIBRARY_FRAGMENT(atb, m)
{
m.def("_npu_paged_attention(Tensor query, Tensor key_cache, Tensor value_cache, int num_kv_heads, int num_heads, float scale_value, Tensor block_table, Tensor context_lens, Tensor(a!) out, *,Tensor? workspace=None) -> ()");
m.def("_npu_paged_attention_get_workspace(Tensor query, Tensor key_cache, Tensor value_cache, int num_kv_heads, int num_heads, float scale_value, Tensor block_table, Tensor context_lens, Tensor(a!) out) -> Tensor");
}
}
namespace {
TORCH_LIBRARY_IMPL(atb, PrivateUse1, m)
{
m.impl("_npu_paged_attention", TORCH_FN(atb::_npu_paged_attention));
m.impl("_npu_paged_attention_get_workspace", TORCH_FN(atb::_npu_paged_attention_get_workspace));
}
}
}