From 13403d2808d7de6804a8f4195dc37441f067d85b Mon Sep 17 00:00:00 2001 From: biondizzle Date: Fri, 29 May 2026 22:48:50 +0000 Subject: [PATCH] cleanup: remove unused v_head in multi-tile (V via TMA) --- dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh b/dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh index 21938e0c..643f18c6 100644 --- a/dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh +++ b/dsv4/kernels/attention/fmha_6warp_tma_multitile.cuh @@ -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;