Add Python API HF Embedded JSON tokenizer support (#860)

* add python api hf embdedded json tokenizer support

* remove xlmrobertatokenizer test as it is not on HF

---------

Co-authored-by: Sayan Shaw <[email protected]>

onnxruntime_extensions/_cuops.py

@@ -491,6 +491,16 @@ def get_outputs(cls):
         ]
 
 
+class HfJsonTokenizer(CustomOp):
+    @classmethod
+    def get_inputs(cls):
+        return [cls.io_def('str', onnx_proto.TensorProto.STRING, ['N'])]
+
+    @classmethod
+    def get_outputs(cls):
+        return [cls.io_def("ids", onnx.TensorProto.INT64, ['N', None])]
+
+
 # TODO: have a C++ impl.
 def _argsort_op(x, dim):
     d = numpy.argsort(x, dim)
@@ -544,4 +554,4 @@ def build_graph(cls, op_class, *args, **kwargs):
 
     @staticmethod
     def get_op_class(op_type):
-        return globals()[op_type]
+        return globals()[op_type]

onnxruntime_extensions/cvt.py

@@ -12,6 +12,24 @@
 from ._hf_cvt import HFTokenizerConverter, HFTokenizerOnnxGraph  # noqa
 from ._ortapi2 import make_onnx_model, SingleOpGraph
 
+import os
+import numpy as np
+import tempfile
+import shutil
+
+# edit environment variables to avoid protobuf version mismatch
+os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+
+from transformers.convert_slow_tokenizer import SpmConverter  # noqa: E402
+from transformers import AutoTokenizer  # noqa: E402
+from tokenizers import decoders, normalizers, pre_tokenizers, Regex  # noqa: E402
+
+
+OrtxTokenizer = None
+try:
+    from onnxruntime_extensions.pp_api import Tokenizer as OrtxTokenizer
+except ImportError:
+    pass
 
 _is_torch_available = False
 try:
@@ -24,11 +42,150 @@
 
 _PRE_POST_PAIR = {'TrieTokenizer': "TrieDetokenizer"}
 
+def _get_prepend_scheme(add_prefix_space: bool, original_tokenizer) -> str:
+    if add_prefix_space:
+        prepend_scheme = "always"
+        if not getattr(original_tokenizer, "legacy", True):
+            prepend_scheme = "first"
+    else:
+        prepend_scheme = "never"
+    return prepend_scheme
+
+
+class Baichuan2Converter(SpmConverter):
+    handle_byte_fallback = True
+
+    def __init__(self, original_tokenizer):
+        super().__init__(original_tokenizer)
+        original_tokenizer.add_prefix_space = False
+
+    def vocab(self, proto):
+        vocab = [
+            (self.original_tokenizer.convert_ids_to_tokens(0), 0.0),
+            (self.original_tokenizer.convert_ids_to_tokens(1), 0.0),
+            (self.original_tokenizer.convert_ids_to_tokens(2), 0.0),
+        ]
+        vocab += [(piece.piece, piece.score) for piece in proto.pieces[3:]]
+        return vocab
+
+    def unk_id(self, proto):
+        unk_id = 0
+        return unk_id
+
+    def decoder(self, replacement, add_prefix_space):
+        sequence = [
+            decoders.Replace("▁", " "),
+            decoders.ByteFallback(),
+            decoders.Fuse(),
+        ]
+        if add_prefix_space:
+            sequence += [decoders.Strip(content=" ", left=1)]
+        return decoders.Sequence(sequence)
+
+    def normalizer(self, proto):
+        if getattr(self.original_tokenizer, "legacy", True):
+            sequence = []
+            if getattr(self.original_tokenizer, "add_prefix_space", True):
+                sequence += [normalizers.Prepend(prepend="▁")]
+            sequence += [normalizers.Replace(pattern=" ", content="▁")]
+            return normalizers.Sequence(sequence)
+        return None  # non-legacy, no normalizer
+
+    def pre_tokenizer(self, replacement, add_prefix_space):
+        if not getattr(self.original_tokenizer, "legacy", True):  # non-legacy, we need a replace
+            prepend_scheme = _get_prepend_scheme(add_prefix_space, self.original_tokenizer)
+            return pre_tokenizers.Metaspace(replacement=replacement, prepend_scheme=prepend_scheme, split=False)
+        else:
+            return super().pre_tokenizer(replacement, add_prefix_space)
+
+
+class ChatGlmConverter(SpmConverter):
+    def normalizer(self, proto):
+        precompiled_charsmap = proto.normalizer_spec.precompiled_charsmap
+        _normalizers = [
+            normalizers.Strip(left=False, right=True),  # stripping is important
+            normalizers.Replace(Regex(" {2,}"), "▁"),
+        ]
+        return normalizers.Sequence([normalizers.Precompiled(precompiled_charsmap)] + _normalizers)
+
+    def pre_tokenizer(self, replacement, add_prefix_space):
+        prepend_scheme = "always"
+        if hasattr(self.original_tokenizer, "legacy") and not self.original_tokenizer.legacy:
+            prepend_scheme = "first"
+        return pre_tokenizers.Metaspace(
+            replacement=replacement, add_prefix_space=add_prefix_space, prepend_scheme=prepend_scheme
+        )
+
+
+JSON_TOKEN_CONVERTERS = {
+    "BaichuanTokenizer": Baichuan2Converter,
+    "ChatGLMTokenizer": ChatGlmConverter,
+}
+
+# Save tokenizer JSON files using HuggingFace AutoTokenizer
+def convert_tokenizer(model_path, output_dir):
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    if output_dir is None:
+        if os.path.isdir(model_path):
+            output_dir = model_path
+        else:
+            # create a temporary directory
+            output_dir = tempfile.mkdtemp()
+            tokenizer.save_pretrained(output_dir)
+        json_path = os.path.join(output_dir, "tokenizer.json")
+
+    if type(tokenizer).__name__ in JSON_TOKEN_CONVERTERS:
+        GenericSpmConverter = JSON_TOKEN_CONVERTERS[type(tokenizer).__name__]
+
+    converted = GenericSpmConverter(tokenizer).converted()
+    converted.save(json_path)
+    print(f"**Tokenizer saved to {json_path}")
+    return output_dir
+
+# Validate tokenizer files downloaded from memory
+def validate_tokenizer(model_path, output_dir):
+    test_sentence = "I like walking my cute dog\n and\x17 then, 生活的真谛是   \t\t\t\t \n\n61"
+    if OrtxTokenizer is None:
+        print("onnxruntime_extensions package was built with C API enabled, skipping tokenization test")
+    ortx_tokenizer = OrtxTokenizer(output_dir)
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, use_fast=False)
+    expected_ids = tokenizer(test_sentence, return_tensors="np")["input_ids"]
+    ortx_ids = np.asarray(ortx_tokenizer.tokenize(test_sentence))
+    assert np.array_equal(expected_ids[0], ortx_ids), f"Tokenization mismatch: {expected_ids[0]} != {ortx_ids}"
+    print("Tokenization test passed")
+
+# Download tokenizer JSON files from memory
+def download_tokenizer(tokenizer_dir, output_dir):
+    try:
+        from transformers.utils import cached_file
+
+        resolved_full_file = cached_file(tokenizer_dir, "tokenizer.json")
+        resolved_config_file = cached_file(tokenizer_dir, "tokenizer_config.json")
+    except ImportError:
+        raise ValueError(f"Directory '{tokenizer_dir}' not found and transformers is not available")
+    if not os.path.exists(resolved_full_file):
+        raise FileNotFoundError(f"Downloaded HF file '{resolved_full_file}' cannot be found")
+    if os.path.dirname(resolved_full_file) != os.path.dirname(resolved_config_file):
+        raise FileNotFoundError(
+            f"Downloaded HF files '{resolved_full_file}' " f"and '{resolved_config_file}' are not in the same directory"
+        )
+
+    if output_dir is None or len(output_dir) == 0:
+        output_dir = os.path.dirname(resolved_full_file)
+        print(f"Using {output_dir} as output directory")
+        return output_dir
+    else:
+        # copy the files to the output directory
+        shutil.copy(resolved_full_file, output_dir)
+        shutil.copy(resolved_config_file, output_dir)
+        return output_dir
+
 
 def gen_processing_models(processor: Union[str, object],
                           pre_kwargs: dict = None,
                           post_kwargs: dict = None,
                           opset: int = None,
+                          schema_v2: bool = False,
                           **kwargs):
     """
     Generate the pre- and post-processing ONNX model, basing on the name or HF class.
@@ -47,6 +204,9 @@ def gen_processing_models(processor: Union[str, object],
         Keyword arguments for generating the post-processing model
     opset: int
         the target opset version of the model
+    schema_v2: bool
+        the flag for using embedded tokenizer files; this option leverages the blob-loading functionality
+        which loads HF tokenizers from memory rather than using the tokenizer files in HF JSON format.
     kwargs:
         The additional arguments for generating models
 
@@ -58,39 +218,89 @@ def gen_processing_models(processor: Union[str, object],
     if pre_kwargs is None and post_kwargs is None:
         raise ValueError(
             "Either pre_kwargs or post_kwargs should be provided. None means no processing graph output.")
-    if isinstance(processor, str):
+
+    # If true, we get the tokenizer JSON files by either downloading from cache or using HuggingFace AutoTokenizer
+    # to convert them, and then create an ONNX model with the JSON files as strings in the model attributes (attrs).
+    if schema_v2:
+        model_name = processor if isinstance(processor, str) else type(processor).__name__
+
+        converted_tokenizer = {"Baichuan2", "chatglm"}
+        need_convert = False
+        for token in converted_tokenizer:
+            if model_name.find(token) != -1:
+                need_convert = True
+                break
+
+        if need_convert:
+            model_dir = convert_tokenizer(model_name)
+            validate_tokenizer(model_name, None)
+        else:
+            model_dir = download_tokenizer(model_name, None)
+
+        # Load the content of tokenizer.json into a string
+        with open(f"{model_dir}/tokenizer.json", "r", encoding="utf-8") as f:
+            tokenizer_vocab = f.read()
+
+        # Load the content of tokenizer_config.json into a string
+        with open(f"{model_dir}/tokenizer_config.json", "r", encoding="utf-8") as f:
+            tokenizer_config = f.read()
+
+        # Create an ONNX model with these JSON file strings in attrs
         g_pre, g_post = (None, None)
-        if pre_kwargs:
-            g_pre = SingleOpGraph.build_graph(processor, **pre_kwargs)
-        if post_kwargs:
+        if pre_kwargs is not None:
+            # Add tokenizer_vocab and tokenizer_config to the kwargs
+            # so they are added to attrs in build_graph
+            pre_kwargs['tokenizer_vocab'] = tokenizer_vocab
+            pre_kwargs['tokenizer_config'] = tokenizer_config
+            g_pre = SingleOpGraph.build_graph("HfJsonTokenizer", **pre_kwargs)
+        if post_kwargs is not None:
             if pre_kwargs is None:
                 cls_name = processor
             else:
                 if processor not in _PRE_POST_PAIR:
                     raise RuntimeError(
                         f"Cannot locate the post processing operator name from {processor}")
                 cls_name = _PRE_POST_PAIR[processor]
+            # Add tokenizer_vocab and tokenizer_config to the kwargs
+            # so they are added to attrs in build_graph
+            post_kwargs['tokenizer_vocab'] = tokenizer_vocab
+            post_kwargs['tokenizer_config'] = tokenizer_config
             g_post = SingleOpGraph.build_graph(cls_name, **post_kwargs)
         return make_onnx_model(g_pre) if g_pre else None, make_onnx_model(g_post) if g_post else None
-
-    cls_name = type(processor).__name__
-    if cls_name == "WhisperProcessor":
-        if WhisperDataProcGraph is None:
-            raise ValueError(
-                "The Whisper processor needs torch.onnx support, please install pytorch 2.0 and above")
-        _converter = WhisperDataProcGraph(processor, opset=opset, **kwargs)
-        pre_m = _converter.pre_processing(
-            **pre_kwargs) if pre_kwargs is not None else None
-        post_m = _converter.post_processing(
-            **post_kwargs) if post_kwargs is not None else None
-        return pre_m, post_m
-    elif HFTokenizerOnnxGraph.is_supported(processor):
-        _converter = HFTokenizerOnnxGraph(processor)
-        pre_g = _converter.pre_processing(
-            **pre_kwargs) if pre_kwargs is not None else None
-        post_g = _converter.post_processing(
-            **post_kwargs) if post_kwargs is not None else None
-        return make_onnx_model(pre_g) if pre_g else None, \
-            make_onnx_model(post_g) if post_g else None
     else:
-        raise ValueError(f"Unsupported processor/tokenizer: {cls_name}")
+        if isinstance(processor, str):
+            g_pre, g_post = (None, None)
+            if pre_kwargs:
+                g_pre = SingleOpGraph.build_graph(processor, **pre_kwargs)
+            if post_kwargs:
+                if pre_kwargs is None:
+                    cls_name = processor
+                else:
+                    if processor not in _PRE_POST_PAIR:
+                        raise RuntimeError(
+                            f"Cannot locate the post processing operator name from {processor}")
+                    cls_name = _PRE_POST_PAIR[processor]
+                g_post = SingleOpGraph.build_graph(cls_name, **post_kwargs)
+            return make_onnx_model(g_pre) if g_pre else None, make_onnx_model(g_post) if g_post else None
+
+        cls_name = type(processor).__name__
+        if cls_name == "WhisperProcessor":
+            if WhisperDataProcGraph is None:
+                raise ValueError(
+                    "The Whisper processor needs torch.onnx support, please install pytorch 2.0 and above")
+            _converter = WhisperDataProcGraph(processor, opset=opset, **kwargs)
+            pre_m = _converter.pre_processing(
+                **pre_kwargs) if pre_kwargs is not None else None
+            post_m = _converter.post_processing(
+                **post_kwargs) if post_kwargs is not None else None
+            return pre_m, post_m
+        elif HFTokenizerOnnxGraph.is_supported(processor):
+            _converter = HFTokenizerOnnxGraph(processor)
+            pre_g = _converter.pre_processing(
+                **pre_kwargs) if pre_kwargs is not None else None
+            post_g = _converter.post_processing(
+                **post_kwargs) if post_kwargs is not None else None
+            return make_onnx_model(pre_g) if pre_g else None, \
+                make_onnx_model(post_g) if post_g else None
+        else:
+            raise ValueError(f"Unsupported processor/tokenizer: {cls_name}")

test/test_embedded_tokenizer.py

@@ -0,0 +1,43 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+import unittest
+
+import numpy as np
+from transformers import AutoTokenizer, GPT2Tokenizer
+from onnxruntime_extensions import OrtPyFunction, gen_processing_models, ort_inference
+
+
+class TestEmbeddedTokenizer(unittest.TestCase):
+    def test_clip_tokenizer(self):
+        tokenizer = AutoTokenizer.from_pretrained(
+            "openai/clip-vit-base-patch32", use_fast=False)
+        text = """
+               1. Testing long text with multiple lines to check newline handling
+               2. As well as words with apostrophes such as you're, i'm, don't, etc.
+               3. And weird characters such as . , ~ ? ( ) " [ ] ! : - .
+               """
+        ids = tokenizer.encode(text, return_tensors="np")
+
+        ort_tok = OrtPyFunction.from_model(gen_processing_models(
+            tokenizer,
+            pre_kwargs={"WITH_DEFAULT_INPUTS": True})[0],
+            schema_v2=True)
+        actual_ids = ort_tok([text])[0]
+        np.testing.assert_array_equal(ids, actual_ids)
+
+    def test_gpt2_tokenizer(self):
+        tokenizer = GPT2Tokenizer.from_pretrained(
+            "Xenova/gpt-4", use_fast=False)
+        text = "Testing words with apostrophes such as you're, i'm, don't, etc."
+        ids = tokenizer.encode(text, return_tensors="np")
+
+        ort_tok = OrtPyFunction.from_model(gen_processing_models(
+            tokenizer,
+            pre_kwargs={"WITH_DEFAULT_INPUTS": True})[0],
+            schema_v2=True)
+        actual_ids = ort_tok([text])[0]
+        np.testing.assert_array_equal(ids, actual_ids)
+
+
+if __name__ == '__main__':
+    unittest.main()