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

[Bugfix][LoRA][Operator] Fix LoRA custom operators accuracy issue (#2672)

Browse Source 
### What this PR does / why we need it?
Fix the LoRA accuracy issue that introduced by custom AscendC operator
"bgmv_shrink, sgmv_shrink, bgmv_expand, sgmv_epand".

The bug details are: 
- In the kernel function, if you want to call GlobalTensor.GetSize
method, you have to pass the second parameter of bufferSize when you
call GlobalTensor.SetGlobalBuffer first.
- Or GlobalTensor.GetSize method will return a random value.
- You can refer to [this
doc](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/81RC1alpha002/apiref/ascendcopapi/atlasascendc_api_07_00024.html).

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
pytest -sv tests/e2e/singlecard/test_ilama_lora.py
pytest -sv tests/e2e/multicard/test_ilama_lora_tp2.py

- vLLM version: v0.10.1.1
- vLLM main:
a344a5aa0a

---------

Signed-off-by: paulyu12 <paulyu0307@gmail.com>
Signed-off-by: paulyu12 <507435917@qq.com>
Co-authored-by: paulyu12 <paulyu0307@gmail.com>
This commit is contained in:
yupeng
2025-09-02 11:46:59 +08:00
committed by GitHub
parent 214b32a346
commit 9f1e054fe3
9 changed files with 99 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
csrc/kernels/bgmv_shrink.cpp
View File

@@ -29,7 +29,7 @@ public:
public:
__aicore__ inline BGMVShrink(AscendC::TPipe *pipe) : pipe_(pipe) {}
__aicore__ inline void Init(__gm__ void *x, __gm__ void *weight, __gm__ void *indices, __gm__ void *y,
__aicore__ inline void Init(__gm__ void *x, __gm__ void *weight, __gm__ void *indices, uint32_t indicesSize, __gm__ void *y,
uint32_t batchSize, uint32_t numTokensPerCore, uint32_t inputHiddenDim,
uint32_t maxLoRARank, float scale)
{
@@ -44,7 +44,7 @@ public:
xGm_.SetGlobalBuffer((__gm__ X_T *)x);
yOutGm_.SetGlobalBuffer((__gm__ Y_T *)y);
wGm_.SetGlobalBuffer((__gm__ W_T *)weight);
indicesGm_.SetGlobalBuffer((__gm__ int64_t *)indices);
indicesGm_.SetGlobalBuffer((__gm__ int64_t *)indices, indicesSize);
pipe_->InitBuffer(inQueueX_, BUFFER_NUM, TILE_LENGTH * sizeof(X_T));
pipe_->InitBuffer(inQueueW_, BUFFER_NUM, TILE_LENGTH * sizeof(W_T));
@@ -214,13 +214,13 @@ private:
#define BGMV_SHRINK_TYPE_DECLARE(TYPE) \
extern "C" __global__ __aicore__ void bgmv_shrink_##TYPE(__gm__ void* x, __gm__ void* weight, __gm__ void* indices,\
__gm__ void* y, uint32_t batchSize, \
uint32_t indicesSize, __gm__ void* y, uint32_t batchSize, \
uint32_t numTokensPerCore, uint32_t inputHiddenDim, \
uint32_t maxLoRARank, float scale) \
{ \
AscendC::TPipe pipe; \
BGMVShrink<TYPE> op(&pipe); \
op.Init(x, weight, indices, y, batchSize, numTokensPerCore, inputHiddenDim, maxLoRARank, scale); \
op.Init(x, weight, indices, indicesSize, y, batchSize, numTokensPerCore, inputHiddenDim, maxLoRARank, scale); \
op.Process(); \
}
@@ -231,17 +231,17 @@ BGMV_SHRINK_TYPE_DECLARE(half)
#endif
namespace vllm_ascend {
extern void bgmv_shrink_impl(AscendType type, void* stream, void* x, void* weight, void* indices,
extern void bgmv_shrink_impl(AscendType type, void* stream, void* x, void* weight, void* indices, uint32_t indicesSize,
void* y, uint32_t batchSize, uint32_t numTokensPerCore, uint32_t inputHiddenDim,
uint32_t maxLoRARank, float scale)
{
uint32_t blockDim = (batchSize + numTokensPerCore - 1) / numTokensPerCore;
if (type == AscendType::FP16) {
bgmv_shrink_half<<<blockDim, nullptr, stream>>>(x, weight, indices, y, batchSize, numTokensPerCore,
bgmv_shrink_half<<<blockDim, nullptr, stream>>>(x, weight, indices, indicesSize, y, batchSize, numTokensPerCore,
inputHiddenDim, maxLoRARank, scale);
} else if (type == AscendType::BF16) {
#if (__CCE_AICORE__ >= 220)
bgmv_shrink_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, indices, y, batchSize, numTokensPerCore,
bgmv_shrink_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, indices, indicesSize, y, batchSize, numTokensPerCore,
inputHiddenDim, maxLoRARank, scale);
#endif
} else {