EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Refactor the ops PyTorch adapter，cleanup for csrc/torch_binding.cpp (#6732)

Browse Source 
### What this PR does / why we need it?
Refactor the ops PyTorch adapter，cleanup for csrc/torch_binding.cpp,
more details see
https://github.com/vllm-project/vllm-ascend/issues/6486

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
install the new package to test the new modification, here is the
result:


- vLLM version: v0.15.0
- vLLM main:
9562912cea

---------

Signed-off-by: liziyu <liziyu16@huawei.com>
Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Signed-off-by: luomin2005 <luomin2005@huawei.com>
Co-authored-by: liziyu <56102866+liziyu179@users.noreply.github.com>
Co-authored-by: wangxiaoteng <wangxiaoteng@huawei.com>
This commit is contained in:
luomin2005
2026-02-24 09:12:43 +08:00
committed by GitHub
parent f0caeeadcb
commit f41eeeb11e
15 changed files with 1037 additions and 735 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
csrc/lightning_indexer_vllm/lightning_indexer_vllm_torch_adpt.h Normal file
View File

@@ -0,0 +1,73 @@
/*
* Copyright (c) Huawei Technologies Co., Ltd. 2026. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef LIGHTING_INDEXER_VLLM_TORCH_ADPT_H
#define LIGHTING_INDEXER_VLLM_TORCH_ADPT_H
namespace vllm_ascend {
at::Tensor npu_lightning_indexer(
const at::Tensor &query, const at::Tensor &key, const at::Tensor &weights,
const c10::optional<at::Tensor> &actual_seq_lengths_query,
const c10::optional<at::Tensor> &actual_seq_lengths_key,
const c10::optional<at::Tensor> &block_table, c10::string_view layout_query,
c10::string_view layout_key, int64_t sparse_count, int64_t sparse_mode)
{
// npu tensor max size
constexpr int32_t SIZE = 8;
constexpr int32_t DIM_0 = 0;
constexpr int32_t DIM_1 = 1;
constexpr int32_t DIM_2 = 2;
constexpr int32_t DIM_3 = 3;
TORCH_CHECK(query.numel() > 0, "Query is empty.");
TORCH_CHECK(key.numel() > 0, "Key is empty.");
TORCH_CHECK(weights.numel() > 0, "Weights is empty.");
for (size_t i = 0; i < query.sizes().size(); i++) {
TORCH_CHECK(query.size(i) > 0, "All values within query's shape should be greater "
"than 0, but shape[", i, "] is ", query.size(i));
}
TORCH_CHECK(sparse_count > 0, "sparse count should be greater than 0, but now is ", sparse_count);
at::SmallVector<int64_t, SIZE> output_size;
std::string query_layout_str = std::string(layout_query);
std::string key_layout_str = std::string(layout_key);
if (query_layout_str == "BSND") {
output_size = {query.size(DIM_0), query.size(DIM_1), key.size(DIM_2), sparse_count};
} else {
int n_dim_index = 0;
n_dim_index = (key_layout_str == "TND") ? DIM_1 : DIM_2;
output_size = {query.size(DIM_0), key.size(n_dim_index), sparse_count};
}
at::Tensor lightning_indexer_output = at::empty(output_size, query.options().dtype(at::kInt));
// convert str
char *query_layout_ptr = const_cast<char *>(query_layout_str.c_str());
char *key_layout_ptr = const_cast<char *>(key_layout_str.c_str());
EXEC_NPU_CMD(
aclnnLightningIndexerVllm,
query,
key,
weights,
actual_seq_lengths_query,
actual_seq_lengths_key,
block_table,
query_layout_ptr,
key_layout_ptr,
sparse_count,
sparse_mode,
lightning_indexer_output);
return lightning_indexer_output;
}
}
#endif