From d6c757602433a65467e6d8259ec90be6b5fb5980 Mon Sep 17 00:00:00 2001 From: Dani Ferreira Franco Moura Date: Fri, 6 Mar 2026 03:46:43 -0800 Subject: [PATCH] internal change PiperOrigin-RevId: 879546918 --- gemma/flash_attention.cc | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/gemma/flash_attention.cc b/gemma/flash_attention.cc index 0401a1f..7ed1d69 100644 --- a/gemma/flash_attention.cc +++ b/gemma/flash_attention.cc @@ -944,8 +944,8 @@ static HWY_INLINE void QDotKTilexUpTo8TransposedKDoubleWidth( for (size_t i = 0; i < qkv_dim; ++i) { VQ_T k_vec1, k_vec2; if constexpr (HWY_TARGET == HWY_AVX2) { - hwy::Prefetch(k_transposed_span.ptr + (i + 3) * gcpp::KVCache::kTileSize); - hwy::Prefetch(k_transposed_span.ptr + (i + 4) * gcpp::KVCache::kTileSize); + hwy::Prefetch(k_transposed_span.ptr + + (i + 20) * gcpp::KVCache::kTileSize); } Decompress2(df, k_transposed_span, i * gcpp::KVCache::kTileSize, k_vec1, k_vec2);