set default value for use_int32_token

Author: shewu-quic

examples/models/llama/runner/runner.cpp

@@ -129,17 +129,17 @@ Error Runner::load() {
       temperature_);
   text_prefiller_ = std::make_unique<llm::TextPrefiller>(
       text_decoder_runner_.get(),
-      metadata_.at(kUseInt32Token),
       metadata_.at(kUseKVCache),
-      metadata_.at(kEnableDynamicShape));
+      metadata_.at(kEnableDynamicShape),
+      metadata_.at(kUseInt32Token));
 
   text_token_generator_ = std::make_unique<llm::TextTokenGenerator>(
       tokenizer_.get(),
       text_decoder_runner_.get(),
-      metadata_.at(kUseInt32Token),
       metadata_.at(kUseKVCache),
       std::move(eos_ids),
-      &stats_);
+      &stats_,
+      metadata_.at(kUseInt32Token));
 
   return Error::Ok;
 }

extension/llm/runner/text_prefiller.cpp

@@ -17,9 +17,9 @@ namespace llm {
 
 TextPrefiller::TextPrefiller(
     TextDecoderRunner* text_decoder_runner,
-    bool use_int32_token,
     bool use_kv_cache,
-    bool enable_parallel_prefill)
+    bool enable_parallel_prefill,
+    bool use_int32_token)
     : text_decoder_runner_(text_decoder_runner),
       use_int32_token_(use_int32_token),
       use_kv_cache_(use_kv_cache),

extension/llm/runner/text_prefiller.h

@@ -24,9 +24,9 @@ class ET_EXPERIMENTAL TextPrefiller {
  public:
   TextPrefiller(
       TextDecoderRunner* text_decoder_runner,
-      bool use_int32_token,
       bool use_kv_cache_,
-      bool enable_parallel_prefill);
+      bool enable_parallel_prefill,
+      bool use_int32_token = false);
   /**
    * Prefill an LLM Module with the given text input.
    * @param prompt_tokens The text prompt tokens to the LLM Module. Encoded by

extension/llm/runner/text_token_generator.h

@@ -23,10 +23,10 @@ class ET_EXPERIMENTAL TextTokenGenerator {
   TextTokenGenerator(
       Tokenizer* tokenizer,
       TextDecoderRunner* text_decoder_runner,
-      bool use_int32_token,
       bool use_kv_cache,
       std::unique_ptr<std::unordered_set<uint64_t>>&& eos_ids,
-      Stats* stats)
+      Stats* stats,
+      bool use_int32_token = false)
       : tokenizer_(tokenizer),
         text_decoder_runner_(text_decoder_runner),
         eos_ids_(std::move(eos_ids)),