From ef79941ac9151bd5541e23e4e6328025df06c18e Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 17 Jun 2024 19:20:24 +0300
Subject: [PATCH] llama : disable FA if KV head size do not match

---
 llama.cpp | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/llama.cpp b/llama.cpp
index dd7020dc0..619487515 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -16260,6 +16260,11 @@ struct llama_context * llama_new_context_with_model(
         params.flash_attn = false;
     }
 
+    if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
+        LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
     if (params.type_v != GGML_TYPE_F16 && !params.flash_attn) {
         LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
         return nullptr;