Skip to content

Fix SparseFlashMla small q_head handling on A5 #3

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
yjyang62 wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

Fix SparseFlashMla small q_head handling on A5 #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ffdf7ad
Fix SparseFlashMla small head sinks expansion
cursoragent Jul 7, 2026
1c4e15d
Fix SparseFlashMla A5 small q_head vector split
cursoragent Jul 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion attention/sparse_flash_mla/op_kernel/arch22/sparse_flash_mla_csa_block_vector.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -333,7 +333,7 @@ __aicore__ inline void SMLAVectorBlock<SMLAT>::CopySinksIn()
repeatParams.blockLen = constInfo.qHeadNum;
repeatParams.srcStride = 0U;
repeatParams.dstStride = 0U;
for (uint32_t i = 1U; i <= 256U / constInfo.qHeadNum; i++) {
for (uint32_t i = 1U; i < 256U / constInfo.qHeadNum; i++) {
DataCopy(sinksBrcbUb[constInfo.qHeadNum * BLOCK_ELEMENT_NUM * i], sinksBrcbUb, repeatParams);
}
PipeBarrier<PIPE_V>();
Expand Down
2 changes: 1 addition & 1 deletion attention/sparse_flash_mla/op_kernel/arch22/sparse_flash_mla_swa_block_vector.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -289,7 +289,7 @@ __aicore__ inline void SWAVectorBlock<SMLAT>::CopySinksIn()
repeatParams.blockLen = constInfo.qHeadNum;
repeatParams.srcStride = 0U;
repeatParams.dstStride = 0U;
for (uint32_t i = 1U; i <= 256U / constInfo.qHeadNum; i++) {
for (uint32_t i = 1U; i < 256U / constInfo.qHeadNum; i++) {
DataCopy(sinksBrcbUb[constInfo.qHeadNum * BLOCK_ELEMENT_NUM * i], sinksBrcbUb, repeatParams);
}
PipeBarrier<PIPE_V>();
Expand Down
5 changes: 5 additions & 0 deletions attention/sparse_flash_mla/op_kernel/arch35/sparse_flash_mla_csa_block_vector.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -487,6 +487,11 @@ __aicore__ inline void CSABlockVec<TEMPLATE_ARGS>::ProcessVec1(
ConstInfo &constInfo)
{
bmm1ResBuf.WaitCrossCore();
if (unlikely(runInfo.halfMRealSize == 0)) {
bmm1ResBuf.SetCrossCore();
outputBuf.SetCrossCore();
return;
}

LocalTensor<float> sumUb = this->softmaxSumBuf[runInfo.multiCoreIdxMod2].template Get<float>();
LocalTensor<float> maxUb = this->softmaxMaxBuf[runInfo.multiCoreIdxMod2].template Get<float>();
Expand Down
6 changes: 5 additions & 1 deletion attention/sparse_flash_mla/op_kernel/arch35/sparse_flash_mla_kvcache.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,7 +162,11 @@ __aicore__ inline void ComputeSouterParam(RunParamStr& runParam, const ConstInfo
}

runParam.cubeMOuterOffset = cubeSOuterOffset * constInfo.gSize;
runParam.halfMRealSize = (runParam.mRealSize + 1) >> 1;
if (!IS_SPLIT_G && runParam.mRealSize <= 32U) {
runParam.halfMRealSize = runParam.mRealSize;
} else {
runParam.halfMRealSize = (runParam.mRealSize + 1) >> 1;
}
runParam.firstHalfMRealSize = runParam.halfMRealSize;
if (constInfo.subBlockIdx == 1) {
runParam.halfMRealSize = runParam.mRealSize - runParam.halfMRealSize;
Expand Down
88 changes: 88 additions & 0 deletions attention/sparse_flash_mla/tests/pytest/sparse_flash_mla_paramset.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -275,6 +275,94 @@
"ori_win_left": [127],
"ori_win_right": [0]
},
"csa_g32_test_prefill":{
"testcase_name": ["csa_g32_test_prefill"],
"layout_q": ["TND"],
"layout_kv": ["PA_BBND"],
"q_type": [torch.bfloat16],
"ori_kv_type": [torch.bfloat16],
"cmp_kv_type": [torch.bfloat16],
"B": [1],
"S1": [2],
"T1": [2],
"T2": [256],
"N1": [32],
"N2": [1],
"D": [512],
"K": [512],
"block_num1": [2],
"block_num2": [1],
"block_size1": [128],
"block_size2": [128],
"cu_seqlens_q": [[0, 2]],
"cu_seqlens_ori_kv": [[0, 256]],
"cu_seqlens_cmp_kv": [[0, 64]],
"seqused_ori_kv": [[256]],
"seqused_cmp_kv": [[64]],
"cmp_residual_kv": [[0]],
"softmax_scale": [0.04419417],
"cmp_ratio": [4],
"return_softmax_lse": [False],
"ori_mask_mode": [4],
"cmp_mask_mode": [3],
"ori_win_left": [127],
"ori_win_right": [0]
},
"csa_g16_test_prefill":{
"testcase_name": ["csa_g16_test_prefill"],
"layout_q": ["TND"],
"layout_kv": ["PA_BBND"],
"q_type": [torch.bfloat16],
"ori_kv_type": [torch.bfloat16],
"cmp_kv_type": [torch.bfloat16],
"B": [1],
"S1": [2],
"T1": [2],
"T2": [256],
"N1": [16],
"N2": [1],
"D": [512],
"K": [512],
"block_num1": [2],
"block_num2": [1],
"block_size1": [128],
"block_size2": [128],
"cu_seqlens_q": [[0, 2]],
"cu_seqlens_ori_kv": [[0, 256]],
"cu_seqlens_cmp_kv": [[0, 64]],
"seqused_ori_kv": [[256]],
"seqused_cmp_kv": [[64]],
"cmp_residual_kv": [[0]],
"softmax_scale": [0.04419417],
"cmp_ratio": [4],
"return_softmax_lse": [False],
"ori_mask_mode": [4],
"cmp_mask_mode": [3],
"ori_win_left": [127],
"ori_win_right": [0]
},
"swa_g16_test_prefill":{
"testcase_name": ["swa_g16_test_prefill"],
"layout_q": ["TND"],
"layout_kv": ["PA_BBND"],
"q_type": [torch.bfloat16],
"ori_kv_type": [torch.bfloat16],
"B": [1],
"S1": [2],
"T1": [2],
"N1": [16],
"N2": [1],
"D": [512],
"block_num1": [2],
"block_size1": [128],
"cu_seqlens_q": [[0, 2]],
"seqused_ori_kv": [[256]],
"softmax_scale": [0.04419417],
"return_softmax_lse": [False],
"ori_mask_mode": [4],
"ori_win_left": [127],
"ori_win_right": [0]
},
"csa_g128_test_prefill":{
"testcase_name": ["csa_g128_test_prefill"],
"layout_q": ["TND"],
Expand Down