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

[Ops][Misc] Refactor and optimize CausalConv1d for Ascend (#7495)

Browse Source 
### What this PR does / why we need it?
During the prefill phase of Qwen3-Next and Qwen3.5, the
`torch.ops._C_ascend.causal_conv1d_fn` operator exhibits significant
performance bottlenecks. To address this, we have re-implemented the
optimization using `torch.ops._C_ascend.npu_causal_conv1d_custom`.

### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
1 accuracy test
```
[2026-03-20 16:44:22,961] [ais_bench] [INFO] Start launch task state board ...
+-----------------------------+-----------+------------+-------------+----------+-------------------------------------------+---------------------+
| Task Name                   |   Process | Progress   | Time Cost   | Status   | Log Path                                  | Extend Parameters   |
+=============================+===========+============+=============+==========+===========================================+=====================+
| vllm-api-general-chat/gsm8k |   2918978 | NA         | 0:00:01     | finish   | logs/eval/vllm-api-general-chat/gsm8k.out | None                |
+-----------------------------+-----------+------------+-------------+----------+-------------------------------------------+---------------------+
[2026-03-20 16:44:34,284] [ais_bench] [INFO] Evaluation tasks completed.
[2026-03-20 16:44:34,287] [ais_bench] [INFO] Summarizing evaluation results...
dataset    version    metric    mode      vllm-api-general-chat
---------  ---------  --------  ------  -----------------------
gsm8k      271d0b     accuracy  gen                       96.21
```
2 ut modify test
`pytest -sv
/home/c30006096/vllm-ascend/tests/e2e/nightly/single_node/ops/singlecard_ops/triton/test_causal_conv1d.py::test_ascend_causal_conv1d`

- vLLM version: v0.17.0
- vLLM main:
8b6325758c

Signed-off-by: wenba0 <3054239545@qq.com>
Signed-off-by: jiaojiao <56385650+wenba0@users.noreply.github.com>
This commit is contained in:
jiaojiao
2026-03-24 00:07:12 +08:00
committed by GitHub
parent e942b62d74
commit 1de805ce0a
16 changed files with 907 additions and 554 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
csrc/causal_conv1d/op_kernel/causal_conv1d.cpp
View File

@@ -18,13 +18,16 @@
namespace {
template <typename T>
// NOTE:
// Dtype is provided via AscendC compile macros (e.g. DTYPE_X / ORIG_DTYPE_X), so tiling key does not need to carry dtype.
template <typename T>
__aicore__ inline void RunCausalConv1d(GM_ADDR x, GM_ADDR weight, GM_ADDR bias, GM_ADDR convStates,
GM_ADDR queryStartLoc, GM_ADDR cacheIndices, GM_ADDR hasInitialState,
GM_ADDR y, const NsCausalConv1d::CausalConv1dTilingData* tilingData)
GM_ADDR queryStartLoc, GM_ADDR cacheIndices, GM_ADDR initialStateMode,
GM_ADDR numAcceptedTokens, GM_ADDR y, const CausalConv1dTilingData* tilingData)
{
NsCausalConv1d::CausalConv1d<T> op;
op.Init(x, weight, bias, convStates, queryStartLoc, cacheIndices, hasInitialState, y, tilingData);
op.Init(x, weight, bias, convStates, queryStartLoc, cacheIndices, initialStateMode, numAcceptedTokens, y, tilingData);
op.Process();
}
@@ -32,27 +35,24 @@ __aicore__ inline void RunCausalConv1d(GM_ADDR x, GM_ADDR weight, GM_ADDR bias,
template <uint32_t schMode>
__global__ __aicore__ void causal_conv1d(GM_ADDR x, GM_ADDR weight, GM_ADDR bias, GM_ADDR convStates,
GM_ADDR queryStartLoc, GM_ADDR cacheIndices, GM_ADDR hasInitialState,
GM_ADDR y, GM_ADDR workspace, GM_ADDR tiling)
GM_ADDR queryStartLoc, GM_ADDR cacheIndices, GM_ADDR initialStateMode,
GM_ADDR numAcceptedTokens, GM_ADDR y, GM_ADDR workspace, GM_ADDR tiling)
{
REGISTER_TILING_DEFAULT( NsCausalConv1d::CausalConv1dTilingData);
// GET_TILING_DATA_WITH_STRUCT( NsCausalConv1d::CausalConv1dTilingData, tilingData, tiling);
GET_TILING_DATA(tilingData, tiling);
REGISTER_TILING_DEFAULT(CausalConv1dTilingData);
GET_TILING_DATA_WITH_STRUCT(CausalConv1dTilingData, tilingData, tiling);
#if defined(ORIG_DTYPE_X)
#if (ORIG_DTYPE_X == DT_FLOAT16)
RunCausalConv1d<half>(x, weight, bias, convStates, queryStartLoc, cacheIndices, hasInitialState, y, &tilingData);
RunCausalConv1d<half>(x, weight, bias, convStates, queryStartLoc, cacheIndices, initialStateMode, numAcceptedTokens, y, &tilingData);
#elif (ORIG_DTYPE_X == DT_BF16)
RunCausalConv1d<bfloat16_t>(x, weight, bias, convStates, queryStartLoc, cacheIndices, hasInitialState, y, &tilingData);
#elif (ORIG_DTYPE_X == DT_FLOAT)
RunCausalConv1d<float>(x, weight, bias, convStates, queryStartLoc, cacheIndices, hasInitialState, y, &tilingData);
RunCausalConv1d<bfloat16_t>(x, weight, bias, convStates, queryStartLoc, cacheIndices, initialStateMode, numAcceptedTokens, y, &tilingData);
#endif
#else
#if (DTYPE_X == DT_FLOAT16)
RunCausalConv1d<half>(x, weight, bias, convStates, queryStartLoc, cacheIndices, hasInitialState, y, &tilingData);
RunCausalConv1d<half>(x, weight, bias, convStates, queryStartLoc, cacheIndices, initialStateMode, numAcceptedTokens, y, &tilingData);
#elif (DTYPE_X == DT_BF16)
RunCausalConv1d<bfloat16_t>(x, weight, bias, convStates, queryStartLoc, cacheIndices, hasInitialState, y, &tilingData);
#elif (DTYPE_X == DT_FLOAT)
RunCausalConv1d<float>(x, weight, bias, convStates, queryStartLoc, cacheIndices, hasInitialState, y, &tilingData);
RunCausalConv1d<bfloat16_t>(x, weight, bias, convStates, queryStartLoc, cacheIndices, initialStateMode, numAcceptedTokens, y, &tilingData);
#endif
#endif
}
}