Fix qnn runner failure

Differential Revision: D78290436

Pull Request resolved: #12459

Author: larryliu0820

examples/qualcomm/oss_scripts/llama/runner/lhd_token_generator.h

@@ -51,9 +51,9 @@ class LhdTokenGenerator : public TokenGenerator {
                 metadata.use_int64_token},
             stats),
         metadata_(metadata),
-        ngrams_pool_(metadata.vocab_size, metadata.ngram, metadata.gcap),
         lhd_branch_(metadata.ngram - 1, std::vector<int32_t>(metadata.window)),
-        lhd_branch_prev_(metadata.window) {
+        lhd_branch_prev_(metadata.window),
+        ngrams_pool_(metadata.vocab_size, metadata.ngram, metadata.gcap) {
     ET_LOG(
         Info,
         "Use Lookahead decoding: ngram=%d, window=%d, gcap=%d",

examples/qualcomm/oss_scripts/llama/runner/token_generator.cpp

@@ -27,8 +27,8 @@ TokenGenerator::TokenGenerator(
       kv_manager_(kv_manager),
       method_name_(method_name),
       eos_ids_(std::move(eos_ids)),
-      metadata_(metadata),
-      stats_(stats) {
+      stats_(stats),
+      metadata_(metadata) {
   k_cache_in_.resize(metadata_.num_layers);
   v_cache_in_.resize(metadata_.num_layers);
   k_cache_out_.resize(metadata_.num_layers);