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

[kernel] add AscendC op: lightning_indexer and sparse_flash_attention (#4625)

Browse Source 
### What this PR does / why we need it?
Provide high-performance AscendC operators lightning_indexer and
sparse_flash_attention to boost the execution performance of the
DeepSeek v3.2 model. Meanwhile, adapt the two AscendC operators to
vllm-ascend framework.

### Does this PR introduce _any_ user-facing change?
No (only underlying operator optimizations, with no user-facing changes)

### How was this patch tested?

- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2

Signed-off-by: MingYang119 <songmingyang@huawei.com>
This commit is contained in:
Song Mingyang
2025-12-03 09:53:10 +08:00
committed by GitHub
parent 7f2673ea2d
commit 18b90b501d
28 changed files with 9772 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

72
csrc/lightning_indexer/op_host/lightning_indexer_def.cpp Normal file
View File

@@ -0,0 +1,72 @@
/**
* This program is free software, you can redistribute it and/or modify it.
* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This file is a part of the CANN Open Software.
* Licensed under 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 lightning_indexer_def.cpp
* \brief
*/
#include <cstdint>
#include "register/op_def_registry.h"
namespace ops {
class LightningIndexer : public OpDef {
public:
explicit LightningIndexer(const char *name) : OpDef(name)
{
this->Input("query")
.ParamType(REQUIRED)
.DataType({ge::DT_BF16, ge::DT_FLOAT16})
.FormatList({ge::FORMAT_ND})
.AutoContiguous();
this->Input("key")
.ParamType(REQUIRED)
.DataType({ge::DT_BF16, ge::DT_FLOAT16})
.FormatList({ge::FORMAT_ND})
.AutoContiguous();
this->Input("weights")
.ParamType(REQUIRED)
.DataType({ge::DT_BF16, ge::DT_FLOAT16})
.FormatList({ge::FORMAT_ND})
.AutoContiguous();
this->Input("actual_seq_lengths_query")
.ParamType(OPTIONAL)
.DataType({ge::DT_INT32, ge::DT_INT32})
.FormatList({ge::FORMAT_ND})
.AutoContiguous();
this->Input("actual_seq_lengths_key")
.ParamType(OPTIONAL)
.DataType({ge::DT_INT32, ge::DT_INT32})
.FormatList({ge::FORMAT_ND})
.AutoContiguous();
this->Input("block_table")
.ParamType(OPTIONAL)
.DataTypeList({ge::DT_INT32})
.FormatList({ge::FORMAT_ND})
.AutoContiguous();
this->Output("sparse_indices").ParamType(REQUIRED).DataTypeList({ge::DT_INT32}).FormatList({ge::FORMAT_ND});
this->Attr("layout_query").AttrType(OPTIONAL).String("BSND");
this->Attr("layout_key").AttrType(OPTIONAL).String("PA_BSND");
this->Attr("sparse_count").AttrType(OPTIONAL).Int(2048); // 2048: Default value, filter the top 2048
this->Attr("sparse_mode").AttrType(OPTIONAL).Int(3); // 3: Default value, only calculate the lower triangular matrix
OpAICoreConfig aicore_config;
aicore_config.DynamicCompileStaticFlag(true)
.DynamicFormatFlag(true)
.DynamicRankSupportFlag(true)
.DynamicShapeSupportFlag(true)
.NeedCheckSupportFlag(false)
.PrecisionReduceFlag(true)
.ExtendCfgInfo("aclnnSupport.value", "support_aclnn")
.ExtendCfgInfo("jitCompile.flag", "static_false,dynamic_false");
this->AICore().AddConfig("ascend910b", aicore_config);
this->AICore().AddConfig("ascend910_93", aicore_config);
}
};
OP_ADD(LightningIndexer);
} // namespace ops