biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

cleanup: remove unused v_head in multi-tile (V via TMA)

Browse Source
This commit is contained in:
biondizzle
2026-05-29 22:48:50 +00:00
parent 74145a31cc
commit 13403d2808
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh
View File

@@ -78,7 +78,8 @@ fmha_6warp_tma_multitile_kernel(FmhaTmaMultiTileParams params) {
const float scale = params.scale;
const bf16_t* __restrict__ q_head = params.q + head_idx * params.q_head_stride + batch_idx * params.q_batch_stride;
const bf16_t* __restrict__ v_head = params.v + head_idx * params.v_head_stride + batch_idx * params.v_batch_stride;
// v_head kept for reference; V loaded via TMA
// const bf16_t* __restrict__ v_head = params.v + head_idx * params.v_head_stride + batch_idx * params.v_batch_stride;
bf16_t* __restrict__ o_head = params.o + head_idx * params.o_head_stride + batch_idx * params.o_batch_stride;
float* __restrict__ lse_head = params.lse ? params.lse + head_idx * params.lse_head_stride + batch_idx * params.lse_batch_stride : nullptr;
CUtensorMap* __restrict__ my_tma_k = params.tma_k + batch_idx * params.n_h + head_idx;