From: jinzihao <redacted>
Date: Thu, 13 Aug 2026 09:04:46 +0000 (+0800)
Subject: ggml-cpu/ops: vectorize flash-attention V-cache F16 to F32 conversion (#26947)
X-Git-Tag: upstream/0.0.10438~27
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=eeae28b67e94cbce01f016576803509dbad11d09;p=pkg%2Fggml%2Fsources%2Fllama.cpp

ggml-cpu/ops: vectorize flash-attention V-cache F16 to F32 conversion (#26947)

Co-authored-by: jinzihao <redacted>
---

diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 42ec809ce..25bb74383 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -8941,7 +8941,7 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
             for (int tk = 0; tk < kv_tile; tk++) {
                 const char * v_data = (const char *)v->data + (ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3;
                 if (kv_type == GGML_TYPE_F16) {
-                    ggml_fp16_to_fp32_row((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
+                    ggml_cpu_fp16_to_fp32((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
                 } else {
                     memcpy(V32 + tk * DV, v_data, DV * sizeof(float));
                 }