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

28
csrc/kernels/sgmv_expand.cpp
View File

@@ -198,7 +198,7 @@ private:
}
} else {
Cast(xDup, xLocal, AscendC::RoundMode::CAST_NONE, maxLoRARank_);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
for (int32_t i = maxLoRARank_; i < NUM_ELEMENTS_PER_REPEAT; i += maxLoRARank_) {
for (int32_t j = 0; j < maxLoRARank_; j++) {
@@ -230,15 +230,15 @@ private:
AscendC::LocalTensor<Y_T> yInLocal = inQueueY_.DeQue<Y_T>();
AscendC::LocalTensor<float> yInLocalFP32 = inBufferY_.Get<float>();
Cast(yInLocalFP32, yInLocal, AscendC::RoundMode::CAST_NONE, numElements);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
inQueueY_.FreeTensor(yInLocal);
Add(yLocal, yLocal, yInLocalFP32, numElements);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
AscendC::LocalTensor<Y_T> yOutLocal = outQueueY_.AllocTensor<Y_T>();
Cast(yOutLocal, yLocal, AscendC::RoundMode::CAST_RINT, numElements);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
outQueueY_.EnQue<Y_T>(yOutLocal);
}
@@ -254,40 +254,40 @@ private:
AscendC::LocalTensor<float> wTmpTensor = tmpBufferW_.Get<float>();
Cast(wTmpTensor, wLocal, AscendC::RoundMode::CAST_NONE, MASK_COUNT, blockReduceRepeatCount, castParams_);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
inQueueW_.FreeTensor(wLocal);
Mul(wTmpTensor, xDup, wTmpTensor, MASK_COUNT, blockReduceRepeatCount, dotProductParams_);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
if (maxLoRARank_ == LORA_RANK_8) {
BlockReduceSum(yLocal[progress], wTmpTensor, blockReduceRepeatCount, MASK_COUNT,
reduceSumParams_.dstRepStride, reduceSumParams_.srcBlkStride, reduceSumParams_.srcRepStride);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
} else if (maxLoRARank_ == LORA_RANK_16) {
BlockReduceSum(wTmpTensor, wTmpTensor, blockReduceRepeatCount, MASK_COUNT,
reduceSumParams_.dstRepStride, reduceSumParams_.srcBlkStride, reduceSumParams_.srcRepStride);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
PairReduceSum(yLocal[progress], wTmpTensor, pairReduceRepeat16, MASK_COUNT,
reduceSumParams_.dstRepStride, reduceSumParams_.srcBlkStride, reduceSumParams_.srcRepStride);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
} else if (maxLoRARank_ == LORA_RANK_32) {
BlockReduceSum(wTmpTensor, wTmpTensor, blockReduceRepeatCount, MASK_COUNT,
reduceSumParams_.dstRepStride, reduceSumParams_.srcBlkStride, reduceSumParams_.srcRepStride);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
PairReduceSum(wTmpTensor, wTmpTensor, pairReduceRepeat16, MASK_COUNT,
reduceSumParams_.dstRepStride, reduceSumParams_.srcBlkStride, reduceSumParams_.srcRepStride);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
PairReduceSum(yLocal[progress], wTmpTensor, pairReduceRepeat32, MASK_COUNT,
reduceSumParams_.dstRepStride, reduceSumParams_.srcBlkStride, reduceSumParams_.srcRepStride);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
} else if (maxLoRARank_ == LORA_RANK_64) {
BlockReduceSum(wTmpTensor, wTmpTensor, blockReduceRepeatCount, MASK_COUNT,
reduceSumParams_.dstRepStride, reduceSumParams_.srcBlkStride, reduceSumParams_.srcRepStride);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
BlockReduceSum(yLocal[progress], wTmpTensor, pairReduceRepeat16, MASK_COUNT,
reduceSumParams_.dstRepStride, reduceSumParams_.srcBlkStride, reduceSumParams_.srcRepStride);
pipe_barrier(PIPE_V);
AscendC::PipeBarrier<PIPE_V>();
}
}