Fix spm converted FastTokenizer issue on non-ascii char (#778)

* Fix spm converted tokenizer issue on non-ascii char

* remove pkg_resource in python

operators/tokenizer/bpe_decoder.hpp

@@ -42,10 +42,10 @@ struct KernelBpeDecoder {
       return status;
     } else {
       auto um = ParseId2String(byte_decoder);
-      std::transform(um.begin(), um.end(),
-                     std::inserter(byte_decoder_, byte_decoder_.end()),
-                     [](const auto& p) { return std::make_pair(static_cast<char32_t>(p.first),
-                                                               ort_extensions::narrow<unsigned char>(std::stoul(p.second))); });
+      std::transform(um.begin(), um.end(), std::inserter(byte_decoder_, byte_decoder_.end()), [](const auto& p) {
+        return std::make_pair(static_cast<char32_t>(p.first),
+                              ort_extensions::narrow<unsigned char>(std::stoul(p.second)));
+      });
     }
 
     std::string added_tokens;
@@ -59,8 +59,7 @@ struct KernelBpeDecoder {
     ORTX_RETURN_IF_ERROR(OrtW::GetOpAttribute(info, "all_special_ids", all_special_ids));
     if (!all_special_ids.empty()) {
       auto um = ParseId2String(all_special_ids);
-      std::transform(um.begin(), um.end(),
-                     std::inserter(all_special_ids_, all_special_ids_.end()),
+      std::transform(um.begin(), um.end(), std::inserter(all_special_ids_, all_special_ids_.end()),
                      [](const auto& p) { return p.first; });
     }
 
@@ -116,8 +115,29 @@ struct KernelBpeDecoder {
     arr_vocab_.shrink_to_fit();
   }
 
-  OrtxStatus Compute(const ortc::Tensor<int64_t>& ids,
-                     ortc::Tensor<std::string>& output) const {
+  const std::string spm_underscore{"\xe2\x96\x81"};
+
+  static bool IsSpmByteWord(std::string_view word) {
+    return word.size() == 6 && word[0] == '<' && word[1] == '0' && word[2] == 'x' && word[5] == '>';
+  }
+
+  static std::string ReplaceAll(std::string_view s, const std::string& search, const std::string& replace) {
+    std::string result;
+    for (size_t pos = 0;; pos += search.length()) {
+      auto new_pos = s.find(search, pos);
+      if (new_pos == std::string::npos) {
+        result += s.substr(pos, s.size() - pos);
+        break;
+      }
+      result += s.substr(pos, new_pos - pos);
+      result += replace;
+      pos = new_pos;
+    }
+
+    return result;
+  }
+
+  OrtxStatus Compute(const ortc::Tensor<int64_t>& ids, ortc::Tensor<std::string>& output) const {
     const int64_t* p_ids = ids.Data();
     const auto& ids_dim = ids.Shape();
     std::vector<int64_t> output_dim = {1};
@@ -126,6 +146,8 @@ struct KernelBpeDecoder {
       std::copy(ids_dim.begin(), ids_dim.begin() + ids_dim.size() - 1, output_dim.begin());
     }
 
+    bool spm_mode = byte_decoder_.count(ustring(spm_underscore)[0]) > 0;
+
     size_t seq_len = ids_dim.back();
     size_t string_batch = ids.NumberOfElement() / seq_len;
     std::vector<std::string> decoded_strings;
@@ -148,24 +170,37 @@ struct KernelBpeDecoder {
 
         if (added_tokens_.count(token)) {
           const std::string& ws = added_tokens_.at(token);
-          decoded_token = (std::string)ws;
+          decoded_token.assign(ws);
         } else if (static_cast<size_t>(token) < arr_vocab_.size()) {
-          const auto str = ustring(arr_vocab_[token]);
-          for (auto wchr : str) {
-            if (byte_decoder_.count(wchr) == 0) {
-              if (wchr <= char32_t(0xFF)) {
-                decoded_token.push_back(static_cast<char>(wchr));
-                continue;
-              }
-              if (skip_special_tokens_) {
-                continue;
-              } else {
-                decoded_token = unk_token_;
-                break;
+          const auto piece = arr_vocab_[token];
+          if (spm_mode) {
+            // sentencepiece case, which doesn't really have a byte decoder
+            if ((IsSpmByteWord(piece))) {
+              char buf[3] = {piece[3], piece[4], 0};  // something like <0x20>
+              char token = {static_cast<char>(strtol(buf, NULL, 16))};
+              decoded_token.push_back(token);
+            } else {
+              decoded_token.append(ReplaceAll(piece, spm_underscore, " "));
+            }
+          } else {
+            // the common bpe case
+            const auto str = ustring(piece);
+            for (auto wchr : str) {
+              if (byte_decoder_.count(wchr) == 0) {
+                if (wchr <= char32_t(0xFF)) {
+                  decoded_token.push_back(static_cast<char>(wchr));
+                  continue;
+                }
+                if (skip_special_tokens_) {
+                  continue;
+                } else {
+                  decoded_token = unk_token_;
+                  break;
+                }
               }
+              char uchr = byte_decoder_.at(wchr);
+              decoded_token.push_back(uchr);
             }
-            char uchr = byte_decoder_.at(wchr);
-            decoded_token.push_back(uchr);
           }
         } else {
           if (skip_special_tokens_) {
@@ -183,15 +218,13 @@ struct KernelBpeDecoder {
           }
         }
 
-        if (whitespace_token_ &&
-            f_special && (tok_idx > 0 && !f_special_last)) {
+        if (whitespace_token_ && f_special && (tok_idx > 0 && !f_special_last)) {
           text.push_back(' ');
         }
 
         text.append(decoded_token);
 
-        if (whitespace_token_ &&
-            f_special && tok_idx != count - 1) {
+        if (whitespace_token_ && f_special && tok_idx != count - 1) {
           text.push_back(' ');
         }
 

operators/tokenizer/bpe_streaming.hpp

@@ -53,25 +53,7 @@ class BpeStreamingDecoder : public KernelBpeDecoder {
     return {};
   }
 
-  static std::string ReplaceAll(std::string_view s, const std::string& search, const std::string& replace) {
-    std::string result;
-    for (size_t pos = 0;; pos += search.length()) {
-      auto new_pos = s.find(search, pos);
-      if (new_pos == std::string::npos) {
-        result += s.substr(pos, s.size() - pos);
-        break;
-      }
-      result += s.substr(pos, new_pos - pos);
-      result += replace;
-      pos = new_pos;
-    }
 
-    return result;
-  }
-
-  static bool IsSpmByteWord(std::string_view word) {
-    return word.size() == 6 && word[0] == '<' && word[1] == '0' && word[2] == 'x' && word[5] == '>';
-  }
 
   OrtxStatus Id2Token(extTokenId_t id,
                       std::string& token,
@@ -119,7 +101,6 @@ class BpeStreamingDecoder : public KernelBpeDecoder {
   }
 
   OrtxStatus SpmId2Token(extTokenId_t id, std::string& token, bool& f_special_last) const {
-    const char spm_underscore[] = "\xe2\x96\x81";
 
     std::string piece = id < arr_vocab_.size() ? arr_vocab_[id] : "";
     bool f_special = false;

test/test_cliptok.py

@@ -1,7 +1,6 @@
 import unittest
 import numpy as np
 import onnxruntime as _ort
-import pkg_resources
 
 from pathlib import Path
 from onnx import helper, onnx_pb as onnx_proto
@@ -150,8 +149,4 @@ def test_optional_outputs(self):
 
 
 if __name__ == "__main__":
-    try:
-        dist = pkg_resources.get_distribution('ftfy')
-    except pkg_resources.DistributionNotFound:
-        raise Exception("WARNING: ftfy is not installed - it is required for parity between CLIPTokenizer and CLIPTokenizerFast.")
     unittest.main()

test/test_fast_tokenizer.py

@@ -21,14 +21,25 @@ def test_llama_tokenizer(self):
         np.testing.assert_array_equal(ids[0], actual_ids[0])
 
     def test_mistral(self):
-        tokenizer = AutoTokenizer.from_pretrained("mistral-community/Mistral-7B-v0.2", use_fast=True)
+        tokenizer = AutoTokenizer.from_pretrained(
+            "mistral-community/Mistral-7B-v0.2", use_fast=True)
         text = "\nOnce upon a time, I was really into monochromatic makeup looks. I have a lot of coppery and bronze "
         ids = tokenizer.encode(text, return_tensors="np")
 
         ort_tok, _ = gen_processing_models(tokenizer, pre_kwargs={})
         actual_ids, *_ = ort_inference(ort_tok, [text])
         np.testing.assert_array_equal(ids[0], actual_ids[0])
 
+    def test_phi_3_mini(self):
+        tokenizer = AutoTokenizer.from_pretrained(
+            "microsoft/Phi-3-mini-128k-instruct", use_fast=True)
+        text = "what are you? \n 给 weiss ich, über was los ist \n"
+        ids = tokenizer.encode(text, return_tensors="np")
+
+        ort_tok, _ = gen_processing_models(tokenizer, pre_kwargs={})
+        actual_ids, *_ = ort_inference(ort_tok, [text])
+        np.testing.assert_array_equal(ids[0], actual_ids[0][1:])
+
 
 if __name__ == '__main__':
     unittest.main()