From 422e7bb3125d87f34078782b322e8bed9e9ac94a Mon Sep 17 00:00:00 2001 From: biondizzle Date: Fri, 29 May 2026 22:54:44 +0000 Subject: [PATCH] cleanup: v_head reference in multi-row (V via TMA now) --- dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh b/dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh index 7d66627b..5f74472d 100644 --- a/dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh +++ b/dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh @@ -77,7 +77,8 @@ fmha_6warp_tma_multirow_kernel(FmhaTmaMultiRowParams 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;