Fix torchchat model not working on demo apps (#3668)

Summary:
As titled fix pytorch/torchchat#810

Pull Request resolved: #3668

Test Plan: Use torchchat to export llama3 pte file, run it in executorch, see it running

Reviewed By: kirklandsign

Differential Revision: D57678514

Pulled By: larryliu0820

fbshipit-source-id: 73437a8b2e4aeeb562e3b1bd53d0fe1411d81456

Author: larryliu0820

.ci/scripts/test_llama.sh

@@ -68,7 +68,7 @@ which "${PYTHON_EXECUTABLE}"
 cmake_install_executorch_libraries() {
     echo "Installing libexecutorch.a, libextension_module.so, libportable_ops_lib.a"
     rm -rf cmake-out
-    retry cmake -DBUCK2="$BUCK" \
+    retry cmake \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
         -DCMAKE_BUILD_TYPE=Debug \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
@@ -85,7 +85,7 @@ cmake_install_executorch_libraries() {
 cmake_build_llama_runner() {
     echo "Building llama runner"
     dir="examples/models/llama2"
-    retry cmake -DBUCK2="$BUCK" \
+    retry cmake \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
         -DCMAKE_BUILD_TYPE=Debug \
         -DEXECUTORCH_BUILD_KERNELS_CUSTOM="$CUSTOM" \

examples/models/llama2/runner/runner.cpp

@@ -76,7 +76,7 @@ Error Runner::load() {
   n_bos_ = getMetadataHelper<int64_t>("get_n_bos", 1);
   n_eos_ = getMetadataHelper<int64_t>("get_n_eos", 1);
   max_seq_len_ = getMetadataHelper<int64_t>("get_max_seq_len", 128);
-  use_kv_cache_ = getMetadataHelper("use_kv_cache", false);
+  use_kv_cache_ = getMetadataHelper("use_kv_cache", true);
   use_sdpa_with_kv_cache_ = getMetadataHelper("use_sdpa_with_kv_cache", false);
   append_eos_ = getMetadataHelper("append_eos_to_prompt", false);