* 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 advance_step.h
* \brief
*/
#ifndef ADVANCE_STEP_H
#define ADVANCE_STEP_H
#include <type_traits>
#include "kernel_operator.h"
#include "lib/matmul_intf.h"
namespace AdvanceStepNs {
using namespace AscendC;
template <typename T>
class KernelAdvanceStep {
public:
TPipe pipe;
__aicore__ inline KernelAdvanceStep(){};
__aicore__ inline void Init(GM_ADDR input_tokens, GM_ADDR sampled_token_ids, GM_ADDR input_positions,
GM_ADDR seq_lens, GM_ADDR slot_mapping, GM_ADDR block_tables, GM_ADDR workspace,
const AdvanceStepTilingData* tilingData);
__aicore__ inline void Process();
private:
GlobalTensor<T> input_tokens_ptr;
GlobalTensor<T> sampled_token_ids_ptr;
GlobalTensor<T> input_positions_ptr;
GlobalTensor<T> seq_lens_ptr;
GlobalTensor<T> slot_mapping_ptr;
GlobalTensor<T> block_tables_ptr;
int64_t blockIdx;
int64_t block_tables_stride;
int64_t num_seqs;
int64_t num_queries;
int64_t block_size;
int64_t total_core_num;
};
template <typename T>
__aicore__ inline void KernelAdvanceStep<T>::Init(GM_ADDR input_tokens, GM_ADDR sampled_token_ids,
GM_ADDR input_positions, GM_ADDR seq_lens, GM_ADDR slot_mapping,
GM_ADDR block_tables, GM_ADDR workspace,
const AdvanceStepTilingData* tilingData)
{
input_tokens_ptr.SetGlobalBuffer((__gm__ T*)input_tokens);
sampled_token_ids_ptr.SetGlobalBuffer((__gm__ T*)sampled_token_ids);
input_positions_ptr.SetGlobalBuffer((__gm__ T*)input_positions);
seq_lens_ptr.SetGlobalBuffer((__gm__ T*)seq_lens);
slot_mapping_ptr.SetGlobalBuffer((__gm__ T*)slot_mapping);
block_tables_ptr.SetGlobalBuffer((__gm__ T*)block_tables);
blockIdx = GetBlockIdx();
total_core_num = tilingData->needCoreNum;
block_tables_stride = tilingData->blockTablesStride;
num_seqs = tilingData->numSeqs;
num_queries = tilingData->numQueries;
block_size = tilingData->blockSize;
}
template <typename T>
__aicore__ inline void KernelAdvanceStep<T>::Process()
{
int64_t n_pad = num_seqs - num_queries;
if (n_pad && blockIdx == 0) {
int64_t offset = num_queries;
for (int i = 0; i < n_pad; i += total_core_num) {
input_tokens_ptr.SetValue(offset + i, 0);
input_positions_ptr.SetValue(offset + i, 0);
slot_mapping_ptr.SetValue(offset + i, -1);
}
}
for (int index = 0; index < total_core_num; index++) {
if (index >= num_queries) {
return;
}
input_tokens_ptr.SetValue(index, sampled_token_ids_ptr.GetValue(index));
int64_t seq_len = seq_lens_ptr.GetValue(index);
int64_t next_seq_len = seq_len + 1;
int64_t next_input_pos = next_seq_len - 1;
seq_lens_ptr.SetValue(index, next_seq_len);
input_positions_ptr.SetValue(index, next_input_pos);
int64_t block_index = next_input_pos / block_size;
int64_t block_offset = next_input_pos % block_size;
int64_t slot_num = (block_tables_ptr.GetValue(block_index) + block_tables_stride * index) * block_size +
block_offset;
slot_mapping_ptr.SetValue(index, slot_num);
}
}
}
#endif