Let models provider their own specific special tokens (#4227)

Summary:
Pull Request resolved: #4227

ghstack-source-id: 233588006

Reviewed By: larryliu0820

Differential Revision: D59651199

fbshipit-source-id: dfcdc03434e8b8aac0ded235048e401cdfd327c0

Author: Lunwen He

examples/models/llama2/runner/CMakeLists.txt

@@ -43,7 +43,7 @@ target_include_directories(
 
 if(EXECUTORCH_USE_TIKTOKEN)
   list(APPEND _llama_runner__srcs
-       ${CMAKE_CURRENT_SOURCE_DIR}/../tokenizer/tiktoken.cpp
+       ${CMAKE_CURRENT_SOURCE_DIR}/../tokenizer/llama_tiktoken.cpp
   )
   set(_preprocessor_flag -DET_USE_TIKTOKEN)
 endif()

examples/models/llama2/runner/runner.cpp

@@ -11,7 +11,7 @@
 
 #include <executorch/examples/models/llama2/runner/runner.h>
 #if ET_USE_TIKTOKEN
-#include <executorch/examples/models/llama2/tokenizer/tiktoken.h>
+#include <executorch/examples/models/llama2/tokenizer/llama_tiktoken.h>
 #else /* BPE */
 #include <executorch/examples/models/llama2/tokenizer/bpe_tokenizer.h>
 #endif /* ET_USE_TIKTOKEN*/
@@ -81,7 +81,7 @@ Error Runner::load() {
 
   // Load tokenizer
 #if ET_USE_TIKTOKEN
-  tokenizer_ = std::make_unique<Tiktoken>();
+  tokenizer_ = get_tiktoken_for_llama();
 #else
   tokenizer_ = std::make_unique<BPETokenizer>();
 #endif

examples/models/llama2/tokenizer/llama_tiktoken.cpp

@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/examples/models/llama2/tokenizer/llama_tiktoken.h>
+
+namespace torch {
+namespace executor {
+namespace {
+static constexpr int32_t kSpecialTokensSize = 256;
+static constexpr size_t kBOSTokenIndex = 0;
+static constexpr size_t kEOSTokenIndex = 1;
+
+static inline std::unique_ptr<std::vector<std::string>>
+_get_default_special_tokens() {
+  auto special_tokens =
+      std::make_unique<std::vector<std::string>>(std::vector<std::string>{
+          "<|begin_of_text|>",
+          "<|end_of_text|>",
+          "<|reserved_special_token_0|>",
+          "<|reserved_special_token_1|>",
+          "<|reserved_special_token_2|>",
+          "<|reserved_special_token_3|>",
+          "<|start_header_id|>",
+          "<|end_header_id|>",
+          "<|reserved_special_token_4|>",
+          "<|eot_id|>"});
+
+  // pad the rest of the special tokens with reserved tokens
+  ssize_t reserved_special_token_num = 5;
+  while (special_tokens->size() < kSpecialTokensSize) {
+    special_tokens->emplace_back(
+        "<|reserved_special_token_" +
+        std::to_string(reserved_special_token_num++) + "|>");
+  }
+  return special_tokens;
+}
+
+static inline std::unique_ptr<std::vector<std::string>>
+_get_multimodal_special_tokens() {
+  auto special_tokens =
+      std::make_unique<std::vector<std::string>>(std::vector<std::string>{
+          "<|begin_of_text|>",
+          "<|end_of_text|>",
+          "<|reserved_special_token_0|>",
+          "<|reserved_special_token_1|>",
+          "<|reserved_special_token_2|>",
+          "<|reserved_special_token_3|>",
+          "<|start_header_id|>",
+          "<|end_header_id|>",
+          "<|eom_id|>",
+          "<|eot_id|>",
+          "<|image|>"});
+
+  // pad the rest of the special tokens with reserved tokens except the last
+  // one
+  ssize_t reserved_special_token_num = 4;
+  while (special_tokens->size() < kSpecialTokensSize - 1) {
+    special_tokens->emplace_back(
+        "<|reserved_special_token_" +
+        std::to_string(reserved_special_token_num++) + "|>");
+  }
+
+  special_tokens->emplace_back("<|python_tag|>");
+
+  return special_tokens;
+}
+
+std::unique_ptr<std::vector<std::string>> _get_special_tokens(Version version) {
+  switch (version) {
+    case MULTIMODAL:
+      return _get_multimodal_special_tokens();
+    default:
+      return _get_default_special_tokens();
+  }
+}
+
+} // namespace
+
+std::unique_ptr<Tiktoken> get_tiktoken_for_llama(Version version) {
+  return std::make_unique<Tiktoken>(
+      _get_special_tokens(version), kBOSTokenIndex, kEOSTokenIndex);
+}
+
+} // namespace executor
+} // namespace torch

examples/models/llama2/tokenizer/llama_tiktoken.h

@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <executorch/examples/models/llama2/tokenizer/tiktoken.h>
+
+namespace torch {
+namespace executor {
+
+enum Version {
+  DEFAULT,
+  MULTIMODAL,
+};
+
+std::unique_ptr<Tiktoken> get_tiktoken_for_llama(Version version = DEFAULT);
+
+} // namespace executor
+} // namespace torch

examples/models/llama2/tokenizer/targets.bzl

@@ -23,10 +23,12 @@ def define_common_targets():
         name = "tiktoken",
         srcs = [
             "tiktoken.cpp",
+            "llama_tiktoken.cpp",
         ],
         exported_headers = [
             "tokenizer.h",
             "tiktoken.h",
+            "llama_tiktoken.h",
             "base64.h",
         ],
         exported_deps = [

examples/models/llama2/tokenizer/test/CMakeLists.txt

@@ -26,6 +26,7 @@ set(
   test_tiktoken.cpp
   test_bpe_tokenizer.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/../tiktoken.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/../llama_tiktoken.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/../bpe_tokenizer.cpp
 )
 

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

@@ -6,7 +6,7 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include <executorch/examples/models/llama2/tokenizer/tiktoken.h>
+#include <executorch/examples/models/llama2/tokenizer/llama_tiktoken.h>
 #include <executorch/examples/models/llama2/tokenizer/tokenizer.h>
 #include <executorch/runtime/platform/runtime.h>
 #include <gtest/gtest.h>
@@ -21,7 +21,7 @@ class TiktokenExtensionTest : public Test {
  public:
   void SetUp() override {
     torch::executor::runtime_init();
-    tokenizer_ = std::make_unique<Tiktoken>();
+    tokenizer_ = get_tiktoken_for_llama();
     modelPath_ = std::getenv("RESOURCES_PATH") +
         std::string("/test_tiktoken_tokenizer.model");
   }
@@ -34,7 +34,7 @@ class MultimodalTiktokenV5ExtensionTest : public Test {
  public:
   void SetUp() override {
     torch::executor::runtime_init();
-    tokenizer_ = std::make_unique<Tiktoken>(MULTIMODAL);
+    tokenizer_ = get_tiktoken_for_llama(MULTIMODAL);
     modelPath_ = std::getenv("RESOURCES_PATH") +
         std::string("/test_tiktoken_tokenizer.model");
   }
@@ -144,5 +144,34 @@ TEST_F(TiktokenExtensionTest, TokenizerDecodeOutOfRangeFails) {
   EXPECT_EQ(out.error(), Error::NotSupported);
 }
 
+TEST_F(TiktokenExtensionTest, ConstructionWithInvalidBOSIndex) {
+  // gtest death test doesn't work on iOS:
+  // https://github.com/google/googletest/issues/2834
+#if !GTEST_OS_IOS
+  EXPECT_EXIT(
+      std::make_unique<Tiktoken>(
+          std::make_unique<std::vector<std::string>>(
+              std::vector<std::string>{"<|end_of_text|>"}),
+          1,
+          0),
+      ::testing::KilledBySignal(SIGABRT),
+      "");
+#endif
+}
+
+TEST_F(TiktokenExtensionTest, ConstructionWithInvalidEOSIndex) {
+  // gtest death test doesn't work on iOS:
+  // https://github.com/google/googletest/issues/2834
+#if !GTEST_OS_IOS
+  EXPECT_EXIT(
+      std::make_unique<Tiktoken>(
+          std::make_unique<std::vector<std::string>>(
+              std::vector<std::string>{"<|begin_of_text|>"}),
+          0,
+          1),
+      ::testing::KilledBySignal(SIGABRT),
+      "");
+#endif
+}
 } // namespace executor
 } // namespace torch

examples/models/llama2/tokenizer/tiktoken.cpp

@@ -329,12 +329,38 @@ std::pair<std::vector<uint64_t>, uint64_t> Tiktoken::_encode_with_special_token(
   return std::make_pair(tokens, last_piece_token_len);
 }
 
+Encoder Tiktoken::_build_special_token_encoder(ssize_t num_base_tokens) const {
+  Encoder special_token_encoder;
+  for (ssize_t i = 0; i < _special_tokens->size(); ++i) {
+    special_token_encoder.emplace(_special_tokens->at(i), num_base_tokens + i);
+  }
+  return special_token_encoder;
+}
+
 // -------------------------private method end-------------------------------
 // -------------------------public method start-------------------------------
 
+Tiktoken::Tiktoken(
+    std::unique_ptr<std::vector<std::string>> special_tokens,
+    size_t bos_token_index,
+    size_t eos_token_index)
+    : Tokenizer(),
+      _special_tokens(std::move(special_tokens)),
+      _bos_token_index(bos_token_index),
+      _eos_token_index(eos_token_index) {
+  ET_CHECK_MSG(
+      _bos_token_index < _special_tokens->size(),
+      "invalid bos_token_index %zu",
+      _bos_token_index);
+  ET_CHECK_MSG(
+      _eos_token_index < _special_tokens->size(),
+      "invalid eos_token_index %zu",
+      _eos_token_index);
+}
+
 Error Tiktoken::load(const std::string& path) {
   _encoder = _load_encoder(path);
-  _special_token_encoder = _get_special_tokens(_encoder.size());
+  _special_token_encoder = _build_special_token_encoder(_encoder.size());
 
   _decoder = _build_decoder(_encoder);
   _special_token_decoder = _build_decoder(_special_token_encoder);
@@ -345,8 +371,8 @@ Error Tiktoken::load(const std::string& path) {
 
   // initialize vocab_size, bos_tok, eos_tok
   vocab_size_ = _encoder.size() + _special_token_encoder.size();
-  bos_tok_ = _special_token_encoder.at("<|begin_of_text|>");
-  eos_tok_ = _special_token_encoder.at("<|end_of_text|>");
+  bos_tok_ = _special_token_encoder.at(_special_tokens->at(_bos_token_index));
+  eos_tok_ = _special_token_encoder.at(_special_tokens->at(_eos_token_index));
 
   initialized_ = true;
   return Error::Ok;