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

12
csrc/kernels/sgmv_shrink.cpp
View File

@@ -94,7 +94,7 @@ private:
AscendC::LocalTensor<float> xTmpTensor = tmpBufferX_.Get<float>();
AscendC::LocalTensor<X_T> xLocal = inQueueX_.DeQue<X_T>();
Cast(xTmpTensor, xLocal, AscendC::RoundMode::CAST_NONE, inputHiddenDim_);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
inQueueX_.FreeTensor(xLocal);
}
@@ -153,20 +153,20 @@ private:
AscendC::LocalTensor<X_T> xLocal = inQueueX_.DeQue<X_T>();
Cast(xTmpTensor, xLocal, AscendC::RoundMode::CAST_NONE, numElements);
Cast(wTmpTensor, wLocal, AscendC::RoundMode::CAST_NONE, numElements);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
inQueueX_.FreeTensor(xLocal);
inQueueW_.FreeTensor(wLocal);
} else {
Cast(wTmpTensor, wLocal, AscendC::RoundMode::CAST_NONE, numElements);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
inQueueW_.FreeTensor(wLocal);
}
// dot product of the one tile of X and W
Mul(wTmpTensor, xTmpTensor, wTmpTensor, numElements);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
// reduce sum generate one number, which is the summation of all the dot product
ReduceSum<float>(wTmpTensor, wTmpTensor, wTmpTensor, numElements);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
acc += wTmpTensor.GetValue(0);
}
@@ -192,7 +192,7 @@ private:
AscendC::LocalTensor<Y_T> yOutLocal = outQueueY_.AllocTensor<Y_T>();
Muls(yOutLocal, yLocal, scale_, maxLoRARank_);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
outQueueY_.EnQue<Y_T>(yOutLocal);
}