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

[Build] Add support for Ascend950 chip (#7151)

Browse Source 
### What this PR does / why we need it?
This PR adds support for the Ascend950 chip. This includes:
- Updating build scripts (`CMakeLists.txt` and `setup.py`) to recognize
the Ascend950 chip and set appropriate compilation flags.
- Disabling a set of custom operators that are not yet supported on the
Ascend950 hardware target.
- Performing a codebase-wide refactoring of `pipe_barrier()` calls to
the namespaced `AscendC::PipeBarrier<>()` for improved code consistency
and adherence to the latest API standards.

Ascend950DT e2e passed (Qwen3-32B-MXFP8) and CI passed
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
This commit is contained in:
linfeng-yuan
2026-03-12 10:25:51 +08:00
committed by GitHub
parent da01a74009
commit 5f3826b093
24 changed files with 246 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
csrc/utils/inc/kernel/pse.h
View File

@@ -336,7 +336,7 @@ __aicore__ inline void PseSlopeCopyIn(LocalTensor<T> &dstTensor, LocalTensor<hal
if (pseInfo.needCast) {
int64_t computeSize = pseInfo.vec1S1RealSize * pseInfo.s2AlignedSize;
Cast(dstTensor, helpTensor, RoundMode::CAST_NONE, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
int64_t s1Offset = pseInfo.s1oIdx * pseInfo.s1BaseSize + pseInfo.vecCoreOffset +
pseInfo.loopIdx * pseInfo.vec1S1BaseSize;
@@ -345,16 +345,16 @@ __aicore__ inline void PseSlopeCopyIn(LocalTensor<T> &dstTensor, LocalTensor<hal
float posShift = float(s2Offset + pseInfo.kvStartIdx - s1Offset - pseInfo.qStartIdx);
Adds(dstTensor, dstTensor, posShift, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
Abs(dstTensor, dstTensor, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
float slopes = ((__gm__ T *)pseSlope)[offset] * -1;
if (pseInfo.pseType == (uint32_t)PseTypeEnum::PSE_INNER_MUL_ADD_SQRT_TYPE) {
Sqrt(dstTensor, dstTensor, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
}
Muls(dstTensor, dstTensor, slopes, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
}
}
}
@@ -373,7 +373,7 @@ __aicore__ inline void PseSlopeCast(LocalTensor<T> &dstTensor, LocalTensor<half>
int64_t offset = bOffset + n2Offset + gOffset;
int64_t computeSize = pseInfo.vec1S1RealSize * pseInfo.s2AlignedSize;
Cast(dstTensor, helpTensor, RoundMode::CAST_NONE, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
int64_t s1Offset = pseInfo.s1oIdx * pseInfo.s1BaseSize + pseInfo.vecCoreOffset +
pseInfo.loopIdx * pseInfo.vec1S1BaseSize;
@@ -382,16 +382,16 @@ __aicore__ inline void PseSlopeCast(LocalTensor<T> &dstTensor, LocalTensor<half>
float posShift = float(s2Offset + pseInfo.kvStartIdx - s1Offset - pseInfo.qStartIdx);
Adds(dstTensor, dstTensor, posShift, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
Abs(dstTensor, dstTensor, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
float slopes = ((__gm__ T *)pseSlope)[offset] * -1;
if (pseInfo.pseType == (uint32_t)PseTypeEnum::PSE_INNER_MUL_ADD_SQRT_TYPE) {
Sqrt(dstTensor, dstTensor, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
}
Muls(dstTensor, dstTensor, slopes, computeSize);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
}
}