From c2f79c31c1a5c9f719eb7b1650cfc9997454b4c3 Mon Sep 17 00:00:00 2001
From: yirongjie <yirj0809@gmail.com>
Date: Sat, 9 Nov 2024 12:44:43 +0000
Subject: [PATCH] =?UTF-8?q?fix=EF=BC=9A=20NPU=20affect=20CPU?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/Layer.hpp                   | 2 +-
 src/backends/cpu/CPUKVCache.cpp | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/Layer.hpp b/src/Layer.hpp
index 752e272c..7d729b84 100644
--- a/src/Layer.hpp
+++ b/src/Layer.hpp
@@ -125,7 +125,7 @@ class Layer {
             do_init = !inited_loaded;
             if (op_ == nullptr) {
 #ifdef USE_QNN
-                if (param_["type"] == KVCACHE || param_["type"] == KVCACHENPU) {
+                if ((param_["type"] == KVCACHE || param_["type"] == KVCACHENPU) && (Backend::global_backends.find(MLLM_QNN) != Backend::global_backends.end())) {
                     if (kv_cache_map.find(name_) == kv_cache_map.end()) {
                         // for the prefill part, we need to create a new op
                         param_["type"] = KVCACHENPU;
diff --git a/src/backends/cpu/CPUKVCache.cpp b/src/backends/cpu/CPUKVCache.cpp
index 204648ab..c487a04a 100644
--- a/src/backends/cpu/CPUKVCache.cpp
+++ b/src/backends/cpu/CPUKVCache.cpp
@@ -5,9 +5,9 @@
 #include "Types.hpp"
 
 int n_pack = 16;
-#ifndef USE_QNN
+// #ifndef USE_QNN
 #define KVCache_TYPE_16
-#endif
+// #endif
 namespace mllm {
 CPUKVCache::CPUKVCache(Backend *bn, string opName, int n_rep, int cache_max, int threadCount) :
     thread_count(threadCount), Op(bn, opName) {