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

[feat] mlapo add bf16 no_quant support (#4852)

Browse Source 
### What this PR does / why we need it?
This PR adds mlapo operation support for bf16 no_quant mode.

### Does this PR introduce _any_ user-facing change?
This PR makes quant related parameters optional. 
### How was this patch tested?
CI passed with new added/existing test.

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

---------

Signed-off-by: chenjunyi <isjunyi.chen@gmail.com>
This commit is contained in:
chenjunyi
2025-12-11 11:06:56 +08:00
committed by GitHub
parent c95c271538
commit c12eb22cbe
12 changed files with 1510 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
csrc/mla_preprocess/op_kernel/mla_preprocess_mix_bf16.hpp
View File

@@ -2386,6 +2386,7 @@ public:
this->num_row = mlaParams_.n;
this->epsilon_ = 1e-6;
this->mlaParams = mlaParams_;
this->hiddenStateDim = mlaParams_.hiddenStateDim;
}
__aicore__ inline void Init(GM_ADDR hiddenStateGm, GM_ADDR quantScale1Gm,
@@ -2692,6 +2693,7 @@ private:
uint32_t blockOffset;
uint32_t perTaskNum;
uint32_t resTaskNum;
uint32_t hiddenStateDim;
MlaTilingData mlaParams;
uint32_t num_core_;
@@ -2795,18 +2797,15 @@ MLAOperation<InDtype, CACHE_MODE, weightFormat1, weightFormat2, weightFormat3, q
uint32_t num_col_align_int8 = (num_col_1 + REPEAT_TIME_256 - 1) / REPEAT_TIME_256 * REPEAT_TIME_256;
uint32_t num_col_align_f16 = (num_col_1 + REPEAT_TIME_128 - 1) / REPEAT_TIME_128 * REPEAT_TIME_128;
uint32_t num_col_align_f32 = (num_col_1 + REPEAT_TIME_64 - 1) / REPEAT_TIME_64 * REPEAT_TIME_64;
const uint32_t base_offset = hiddenStateDim * 6;
AscendC::LocalTensor<InDtype> input_tensor = buf.GetBuffer<BufferType::ASCEND_UB, InDtype>(0);
AscendC::LocalTensor<InDtype> scale_tensor =
buf.GetBuffer<BufferType::ASCEND_UB, InDtype>(HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2);
AscendC::LocalTensor<int8_t> offset_tensor = buf.GetBuffer<BufferType::ASCEND_UB, int8_t>(
HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + 32);
AscendC::LocalTensor<float> res1_tensor =
buf.GetBuffer<BufferType::ASCEND_UB, float>(HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + 64);
AscendC::LocalTensor<InDtype> scale_tensor = buf.GetBuffer<BufferType::ASCEND_UB, InDtype>(base_offset);
AscendC::LocalTensor<int8_t> offset_tensor = buf.GetBuffer<BufferType::ASCEND_UB, int8_t>(base_offset + 32);
AscendC::LocalTensor<float> res1_tensor = buf.GetBuffer<BufferType::ASCEND_UB, float>(base_offset + 64);
AscendC::LocalTensor<float> res3_tensor = buf.GetBuffer<BufferType::ASCEND_UB, float>(
HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + 64 + num_col_align_f32 * 4);
base_offset + 64 + num_col_align_f32 * 4);
AscendC::LocalTensor<int8_t> output_tensor = buf.GetBuffer<BufferType::ASCEND_UB, int8_t>(
HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + HIDDTEN_STATE * 2 + 64 + num_col_align_f32 * 4 +
BUF_FACTOR * num_col_align_f32 * 4 + 64);
base_offset + 64 + num_col_align_f32 * 4 + BUF_FACTOR * num_col_align_f32 * 4 + 64);
Quant1.Launch(output_tensor, input_tensor, scale_tensor, offset_tensor, res1_tensor, res3_tensor);
}
FftsCrossCoreSync<PIPE_MTE3, 0>(QUANT1);