Remove llama related stuff out of bpe_tokenizer

Pull Request resolved: #4235

We don't need to initialize `vocab_`, `vocab_scores_`, etc. They will be initialized anyway while loading the tokenizer binary. A benefit of removing them is that we can remove these llama related default values and make `bpe_tokenizer` agnostic to models.
ghstack-source-id: 233769845

Differential Revision: [D59664556](https://our.internmc.facebook.com/intern/diff/D59664556/)

Author: helunwencser

examples/models/llama2/tokenizer/bpe_tokenizer.cpp

@@ -24,12 +24,6 @@ static int compare_tokens(const void* a, const void* b) {
 }
 
 BPETokenizer::BPETokenizer() : Tokenizer() {
-  vocab_size_ = kDefaultVocabSize;
-  vocab_ = std::make_unique<char*[]>(kDefaultVocabSize);
-  vocab_scores_ = std::make_unique<float[]>(kDefaultVocabSize);
-  sorted_vocab_ = std::make_unique<TokenIndex[]>(kDefaultVocabSize);
-  bos_tok_ = kDefaultBosTokenId;
-  eos_tok_ = kDefaultEosTokenId;
   for (int i = 0; i < 256; i++) {
     byte_pieces_[i * 2] = (unsigned char)i;
     byte_pieces_[i * 2 + 1] = '\0';

examples/models/llama2/tokenizer/bpe_tokenizer.h

@@ -14,11 +14,6 @@
 namespace torch {
 namespace executor {
 
-// Default values for llama2
-constexpr int32_t kDefaultVocabSize = 32000;
-constexpr uint64_t kDefaultBosTokenId = 1;
-constexpr uint64_t kDefaultEosTokenId = 2;
-
 struct TokenIndex {
   const char* str;
   int32_t id;
@@ -38,10 +33,10 @@ class BPETokenizer : public Tokenizer {
       const override;
 
  private:
-  std::unique_ptr<char*[]> vocab_;
-  std::unique_ptr<float[]> vocab_scores_;
-  std::unique_ptr<TokenIndex[]> sorted_vocab_;
-  unsigned int max_token_length_;
+  std::unique_ptr<char*[]> vocab_ = nullptr;
+  std::unique_ptr<float[]> vocab_scores_ = nullptr;
+  std::unique_ptr<TokenIndex[]> sorted_vocab_ = nullptr;
+  unsigned int max_token_length_ = 0;
   unsigned char byte_pieces_[512]; // stores all single-byte strings
 };
 } // namespace executor

examples/models/llama2/tokenizer/test/test_bpe_tokenizer.cpp

@@ -57,5 +57,15 @@ TEST_F(TokenizerExtensionTest, TokenizerMetadataIsExpected) {
   EXPECT_EQ(tokenizer_->eos_tok(), 0);
 }
 
+TEST_F(TokenizerExtensionTest, SafeToDestruct) {
+  // Safe to destruct initialized tokenizer.
+  tokenizer_->load(modelPath_);
+  tokenizer_.reset();
+
+  // Safe to destruct uninitialized tokenizer.
+  tokenizer_ = std::make_unique<BPETokenizer>();
+  tokenizer_.reset();
+}
+
 } // namespace executor
 } // namespace torch

examples/models/llama2/tokenizer/tokenizer.h

@@ -28,7 +28,7 @@ namespace executor {
 
 class Tokenizer {
  public:
-  explicit Tokenizer() : initialized_(false) {}
+  explicit Tokenizer() {}
   virtual ~Tokenizer() {}
 
   virtual Error load(const std::string& tokenizer_path) = 0;
@@ -69,9 +69,10 @@ class Tokenizer {
   }
 
  protected:
-  bool initialized_;
-  int32_t vocab_size_;
-  uint64_t bos_tok_, eos_tok_;
+  bool initialized_ = false;
+  int32_t vocab_size_ = 0;
+  uint64_t bos_tok_ = 0;
+  uint64_t eos_tok_ = 0;
 };
 
 } // namespace executor