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

[BugFix]Fix precision issue for LoRA feature (#4141)

Browse Source 
vLLM version: v0.11.0
vLLM main: vllm-project/vllm

### What this PR does / why we need it?
   Fix the precision issue of the LoRA feature in vllm-ascend.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
```bash
pytest tests/lora/test_llama_tp.py::test_llama_lora -s
```
<img width="1319" height="879" alt="lora_test"
src="https://github.com/user-attachments/assets/2a0b2325-5b05-4bbc-ac03-a7c9f0ad9d4c"
/>


- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: hukongyi <hukongyi@cmbchina.com>
This commit is contained in:
hukongyi
2025-12-19 14:22:06 +08:00
committed by GitHub
parent f952de93df
commit ea8f544ce7
6 changed files with 17 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
csrc/kernels/bgmv_expand.cpp
View File

@@ -342,7 +342,7 @@ private:
// declare all dtype kernel
BGMV_EXPAND_TYPE_DECLARE(half)
#if (__CCE_AICORE__ >= 220)
#if !defined(__CCE_AICORE__) || (__CCE_AICORE__ >= 220)
BGMV_EXPAND_TYPE_DECLARE(bfloat16_t)
#endif
@@ -356,8 +356,8 @@ extern void bgmv_expand_impl(AscendType type, void* stream, void* x, void* weigh
bgmv_expand_half<<<blockDim, nullptr, stream>>>(x, weight, indices, indicesSize, yIn, yOut, batchSize, numTokensPerCore,
maxLoRARank, outputHiddenDim, sliceOffset, outputFullDim);
} else if (type == AscendType::BF16) {
#if (__CCE_AICORE__ >= 220)
bgmv_expand_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, indices, indicesSize, yIn, yOut, batchSize,
#if !defined(__CCE_AICORE__) || (__CCE_AICORE__ >= 220)
bgmv_expand_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, indices, indicesSize, yIn, yOut, batchSize,
numTokensPerCore, maxLoRARank, outputHiddenDim,
sliceOffset, outputFullDim);
#endif

6
csrc/kernels/bgmv_shrink.cpp
View File

@@ -226,7 +226,7 @@ private:
// declare all dtype kernel
BGMV_SHRINK_TYPE_DECLARE(half)
#if (__CCE_AICORE__ >= 220)
#if !defined(__CCE_AICORE__) || (__CCE_AICORE__ >= 220)
BGMV_SHRINK_TYPE_DECLARE(bfloat16_t)
#endif
@@ -240,8 +240,8 @@ extern void bgmv_shrink_impl(AscendType type, void* stream, void* x, void* weigh
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, indicesSize, y, batchSize, numTokensPerCore,
#if !defined(__CCE_AICORE__) || (__CCE_AICORE__ >= 220)
bgmv_shrink_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, indices, indicesSize, y, batchSize, numTokensPerCore,
inputHiddenDim, maxLoRARank, scale);
#endif
} else {

6
csrc/kernels/sgmv_expand.cpp
View File

@@ -357,7 +357,7 @@ private:
// declare all dtype kernel
SGMV_EXPAND_TYPE_DECLARE(half)
#if (__CCE_AICORE__ >= 220)
#if !defined(__CCE_AICORE__) || (__CCE_AICORE__ >= 220)
SGMV_EXPAND_TYPE_DECLARE(bfloat16_t)
#endif
@@ -375,8 +375,8 @@ extern void sgmv_expand_impl(AscendType type, void* stream, void* x, void* weigh
numTokensPerCore, maxLoRARank, outputHiddenDim, sliceOffset,
outputFullDim);
} else if (type == AscendType::BF16) {
#if (__CCE_AICORE__ >= 220)
sgmv_expand_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, loraIndices, loraIndicesSize,
#if !defined(__CCE_AICORE__) || (__CCE_AICORE__ >= 220)
sgmv_expand_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, loraIndices, loraIndicesSize,
seqLen, seqLenSize, yIn, yOut, batchSize,
numTokensPerCore, maxLoRARank, outputHiddenDim,
sliceOffset, outputFullDim);

6
csrc/kernels/sgmv_shrink.cpp
View File

@@ -242,7 +242,7 @@ private:
// declare all dtype kernel
SGMV_SHRINK_TYPE_DECLARE(half)
#if (__CCE_AICORE__ >= 220)
#if !defined(__CCE_AICORE__) || (__CCE_AICORE__ >= 220)
SGMV_SHRINK_TYPE_DECLARE(bfloat16_t)
#endif
@@ -260,8 +260,8 @@ extern void sgmv_shrink_impl(AscendType type, void* stream, void* x, void* weigh
numTokensPerCore, inputHiddenDim, maxLoRARank,
scale);
} else if (type == AscendType::BF16) {
#if (__CCE_AICORE__ >= 220)
sgmv_shrink_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, loraIndices, loraIndicesSize,
#if !defined(__CCE_AICORE__) || (__CCE_AICORE__ >= 220)
sgmv_shrink_bfloat16_t<<<blockDim, nullptr, stream>>>(x, weight, loraIndices, loraIndicesSize,
seqLen, seqLenSize,
y, batchSize,
numTokensPerCore, inputHiddenDim, maxLoRARank,