openvino-tokenizers 2024.1.0.0

Convert tokenizers into OpenVINO models

OpenVINO Tokenizers

OpenVINO Tokenizers adds text processing operations to OpenVINO.

Features

• Perform tokenization and detokenization without third-party dependencies
• Convert a HuggingFace tokenizer into OpenVINO model tokenizer and detokenizer
• Combine OpenVINO models into a single model
• Add greedy decoding pipeline to text generation model

Installation

(Recommended) Create and activate virtual env:

python3 -m venv venv
source venv/bin/activate
 # or
conda create --name openvino_tokenizers
conda activate openvino_tokenizers

Minimal Installation

Use minimal installation when you have a converted OpenVINO tokenizer:

pip install openvino-tokenizers
 # or
conda install -c conda-forge openvino openvino-tokenizers

Convert Tokenizers Installation

If you want to convert HuggingFace tokenizers into OpenVINO tokenizers:

pip install openvino-tokenizers[transformers]
 # or
conda install -c conda-forge openvino openvino-tokenizers && pip install transformers[sentencepiece] tiktoken

Install Pre-release Version

Use openvino-tokenizers[transformers] to install tokenizers conversion dependencies.

pip install --pre -U openvino openvino-tokenizers --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly

Build and Install from Source

Install OpenVINO archive distribution. Use --no-deps to avoid OpenVINO installation from PyPI.

source path/to/installed/openvino/setupvars.sh
git clone https://github.com/openvinotoolkit/openvino_tokenizers.git
cd openvino_tokenizers
pip install --no-deps .

This command is the equivalent of minimal installation. Install tokenizers conversion dependencies if needed:

pip install transformers[sentencepiece] tiktoken

:warning: Latest commit of OpenVINO Tokenizers might rely on features that are not present in the release OpenVINO version. Use a nightly build of OpenVINO or build OpenVINO Tokenizers from a release branch if you have issues with the build process.

Build and install for development

source path/to/installed/openvino/setupvars.sh
git clone https://github.com/openvinotoolkit/openvino_tokenizers.git
cd openvino_tokenizers
pip install -e .[all]
# verify installation by running tests
cd tests/
pytest .

C++ Installation

You can use converted tokenizers in C++ pipelines with prebuild binaries.

1. Download OpenVINO archive distribution for your OS from here and extract the archive.
2. Download OpenVINO Tokenizers prebuild libraries from here. To ensure compatibility first three numbers of OpenVINO Tokenizers version should match OpenVINO version and OS.
3. Extract OpenVINO Tokenizers archive into OpenVINO installation directory:
   • Windows: <openvino_dir>\runtime\bin\intel64\Release\
   • MacOS_x86: <openvino_dir>/runtime/lib/intel64/Release
   • MacOS_arm64: <openvino_dir>/runtime/lib/arm64/Release/
   • Linux_x86: <openvino_dir>/runtime/lib/intel64/
   • Linux_arm64: <openvino_dir>/runtime/lib/aarch64/

After that you can add binary extension in the code with:

• core.add_extension("openvino_tokenizers.dll") for Windows
• core.add_extension("libopenvino_tokenizers.dylib") for MacOS
• core.add_extension("libopenvino_tokenizers.so") for Linux

and read/compile converted (de)tokenizers models. If you use version 2023.3.0.0, the binary extension file is called (lib)user_ov_extension.(dll/dylib/so).

Usage

:warning: OpenVINO Tokenizers can be inferred on a CPU device only.

Convert HuggingFace tokenizer

OpenVINO Tokenizers ships with CLI tool that can convert tokenizers from Huggingface Hub or Huggingface tokenizers saved on disk:

convert_tokenizer codellama/CodeLlama-7b-hf --with-detokenizer -o output_dir

There is also convert_tokenizer function that can convert tokenizer python object.

import numpy as np
from transformers import AutoTokenizer
from openvino import compile_model, save_model
from openvino_tokenizers import convert_tokenizer

hf_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
ov_tokenizer = convert_tokenizer(hf_tokenizer)

compiled_tokenzier = compile_model(ov_tokenizer)
text_input = ["Test string"]

hf_output = hf_tokenizer(text_input, return_tensors="np")
ov_output = compiled_tokenzier(text_input)

for output_name in hf_output:
    print(f"OpenVINO {output_name} = {ov_output[output_name]}")
    print(f"HuggingFace {output_name} = {hf_output[output_name]}")
# OpenVINO input_ids = [[ 101 3231 5164  102]]
# HuggingFace input_ids = [[ 101 3231 5164  102]]
# OpenVINO token_type_ids = [[0 0 0 0]]
# HuggingFace token_type_ids = [[0 0 0 0]]
# OpenVINO attention_mask = [[1 1 1 1]]
# HuggingFace attention_mask = [[1 1 1 1]]

# save tokenizer for later use
save_model(ov_tokenizer, "openvino_tokenizer.xml")

loaded_tokenizer = compile_model("openvino_tokenizer.xml")
loaded_ov_output = loaded_tokenizer(text_input)
for output_name in hf_output:
    assert np.all(loaded_ov_output[output_name] == ov_output[output_name])

Connect Tokenizer to a Model

To infer and convert the original model, install torch or torch-cpu to the virtual environment.

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from openvino import compile_model, convert_model
from openvino_tokenizers import convert_tokenizer, connect_models

checkpoint = "mrm8488/bert-tiny-finetuned-sms-spam-detection"
hf_tokenizer = AutoTokenizer.from_pretrained(checkpoint)
hf_model = AutoModelForSequenceClassification.from_pretrained(checkpoint)

text_input = ["Free money!!!"]
hf_input = hf_tokenizer(text_input, return_tensors="pt")
hf_output = hf_model(**hf_input)

ov_tokenizer = convert_tokenizer(hf_tokenizer)
ov_model = convert_model(hf_model, example_input=hf_input.data)
combined_model = connect_models(ov_tokenizer, ov_model)
compiled_combined_model = compile_model(combined_model)

openvino_output = compiled_combined_model(text_input)

print(f"OpenVINO logits: {openvino_output['logits']}")
# OpenVINO logits: [[ 1.2007061 -1.4698029]]
print(f"HuggingFace logits {hf_output.logits}")
# HuggingFace logits tensor([[ 1.2007, -1.4698]], grad_fn=<AddmmBackward0>)

Use Extension With Converted (De)Tokenizer or Model With (De)Tokenizer

Import openvino_tokenizers will add all tokenizer-related operations to OpenVINO, after which you can work with saved tokenizers and detokenizers.

import numpy as np
import openvino_tokenizers
from openvino import Core

core = Core()

# detokenizer from codellama sentencepiece model
compiled_detokenizer = core.compile_model("detokenizer.xml")

token_ids = np.random.randint(100, 1000, size=(3, 5))
openvino_output = compiled_detokenizer(token_ids)

print(openvino_output["string_output"])
# ['sc�ouition�', 'intvenord hasient', 'g shouldwer M more']

Text generation pipeline

import numpy as np
from openvino import compile_model, convert_model
from openvino_tokenizers import add_greedy_decoding, convert_tokenizer
from transformers import AutoModelForCausalLM, AutoTokenizer


model_checkpoint = "JackFram/llama-68m"
hf_tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
hf_model = AutoModelForCausalLM.from_pretrained(model_checkpoint, use_cache=False)

# convert hf tokenizer
text_input = ["Quick brown fox jumped "]
ov_tokenizer, ov_detokenizer = convert_tokenizer(hf_tokenizer, with_detokenizer=True, skip_special_tokens=True)
compiled_tokenizer = compile_model(ov_tokenizer)

# transform input text into tokens
ov_input = compiled_tokenizer(text_input)
hf_input = hf_tokenizer(text_input, return_tensors="pt")

# convert Pytorch model to OpenVINO IR and add greedy decoding pipeline to it
ov_model = convert_model(hf_model, example_input=hf_input.data)
ov_model_with_greedy_decoding = add_greedy_decoding(ov_model)
compiled_model = compile_model(ov_model_with_greedy_decoding)

# generate new tokens
new_tokens_size = 10
prompt_size = ov_input["input_ids"].shape[-1]
input_dict = {
    output.any_name: np.hstack([tensor, np.zeros(shape=(1, new_tokens_size), dtype=np.int_)])
    for output, tensor in ov_input.items()
}
for idx in range(prompt_size, prompt_size + new_tokens_size):
    output = compiled_model(input_dict)["token_ids"]
    input_dict["input_ids"][:, idx] = output[:, idx - 1]
    input_dict["attention_mask"][:, idx] = 1
ov_token_ids = input_dict["input_ids"]

hf_token_ids = hf_model.generate(
    **hf_input,
    min_new_tokens=new_tokens_size,
    max_new_tokens=new_tokens_size,
    temperature=0,  # greedy decoding
)

# decode model output
compiled_detokenizer = compile_model(ov_detokenizer)
ov_output = compiled_detokenizer(ov_token_ids)["string_output"]
hf_output = hf_tokenizer.batch_decode(hf_token_ids, skip_special_tokens=True)
print(f"OpenVINO output string: `{ov_output}`")
# OpenVINO output string: `['<s> Quick brown fox was walking through the forest. He was looking for something']`
print(f"HuggingFace output string: `{hf_output}`")
# HuggingFace output string: `['Quick brown fox was walking through the forest. He was looking for something']`

TensorFlow Text Integration

OpenVINO Tokenizers include converters for certain TensorFlow Text operations. Currently, only the MUSE model is supported. Here is an example of model conversion and inference:

import numpy as np
import tensorflow_hub as hub
import tensorflow_text  # register tf text ops
from openvino import convert_model, compile_model
import openvino_tokenizers  # register ov tokenizer ops and translators


sentences = ["dog",  "I cuccioli sono carini.", "私は犬と一緒にビーチを散歩するのが好きです"]
tf_embed = hub.load(
    "https://www.kaggle.com/models/google/universal-sentence-encoder/frameworks/"
    "TensorFlow2/variations/multilingual/versions/2"
)
# convert model that uses Sentencepiece tokenizer op from TF Text
ov_model = convert_model(tf_embed)
ov_embed = compile_model(ov_model, "CPU")

ov_result = ov_embed(sentences)[ov_embed.output()]
tf_result = tf_embed(sentences)

assert np.all(np.isclose(ov_result, tf_result, atol=1e-4))

RWKV Tokenizer

from urllib.request import urlopen

from openvino import compile_model
from openvino_tokenizers import build_rwkv_tokenizer


rwkv_vocab_url = (
    "https://raw.githubusercontent.com/BlinkDL/ChatRWKV/main/tokenizer/rwkv_vocab_v20230424.txt"
)

with urlopen(rwkv_vocab_url) as vocab_file:
    vocab = map(bytes.decode, vocab_file)
    tokenizer, detokenizer = build_rwkv_tokenizer(vocab)

tokenizer, detokenizer = compile_model(tokenizer), compile_model(detokenizer)

print(tokenized := tokenizer(["Test string"])["input_ids"])  # [[24235 47429]]
print(detokenizer(tokenized)["string_output"])  # ['Test string']

Supported Tokenizer Types

Huggingface Tokenizer Type	Tokenizer Model Type	Tokenizer	Detokenizer
Fast	WordPiece	✅	❌
	BPE	✅	✅
	Unigram	❌	❌
Legacy	SentencePiece .model	✅	✅
Custom	tiktoken	✅	✅
RWKV	Trie	✅	✅

Test Results

This report is autogenerated and includes tokenizers and detokenizers tests. The Output Matched, % column shows the percent of test strings for which the results of OpenVINO and Hugingface Tokenizers are the same. To update the report run pytest --update_readme tokenizers_test.py in tests directory.

Output Match by Tokenizer Type

Tokenizer Type	Output Matched, %	Number of Tests
BPE	96.82	3620
SentencePiece	76.33	3620
Tiktoken	97.71	218
WordPiece	90.43	533

Output Match by Model

Tokenizer Type	Model	Output Matched, %	Number of Tests
BPE	EleutherAI/gpt-j-6b	98.90	181
BPE	EleutherAI/gpt-neo-125m	98.90	181
BPE	EleutherAI/gpt-neox-20b	97.79	181
BPE	EleutherAI/pythia-12b-deduped	97.79	181
BPE	KoboldAI/fairseq-dense-13B	98.90	181
BPE	Salesforce/codegen-16B-multi	97.79	181
BPE	ai-forever/rugpt3large_based_on_gpt2	97.79	181
BPE	bigscience/bloom	99.45	181
BPE	databricks/dolly-v2-3b	97.79	181
BPE	facebook/bart-large-mnli	98.90	181
BPE	facebook/galactica-120b	98.34	181
BPE	facebook/opt-66b	98.90	181
BPE	gpt2	98.90	181
BPE	laion/CLIP-ViT-bigG-14-laion2B-39B-b160k	65.19	181
BPE	microsoft/deberta-base	98.90	181
BPE	roberta-base	98.90	181
BPE	sentence-transformers/all-roberta-large-v1	98.90	181
BPE	stabilityai/stablecode-completion-alpha-3b-4k	98.34	181
BPE	stabilityai/stablelm-2-1_6b	98.34	181
BPE	stabilityai/stablelm-tuned-alpha-7b	97.79	181
SentencePiece	NousResearch/Llama-2-13b-hf	100.00	181
SentencePiece	NousResearch/Llama-2-13b-hf_slow	100.00	181
SentencePiece	THUDM/chatglm2-6b	100.00	181
SentencePiece	THUDM/chatglm2-6b_slow	100.00	181
SentencePiece	THUDM/chatglm3-6b	19.34	181
SentencePiece	THUDM/chatglm3-6b_slow	19.34	181
SentencePiece	camembert-base	0.55	181
SentencePiece	camembert-base_slow	74.03	181
SentencePiece	codellama/CodeLlama-7b-hf	100.00	181
SentencePiece	codellama/CodeLlama-7b-hf_slow	100.00	181
SentencePiece	facebook/musicgen-small	80.11	181
SentencePiece	facebook/musicgen-small_slow	74.03	181
SentencePiece	microsoft/deberta-v3-base	93.37	181
SentencePiece	microsoft/deberta-v3-base_slow	100.00	181
SentencePiece	t5-base	81.22	181
SentencePiece	t5-base_slow	75.14	181
SentencePiece	xlm-roberta-base	97.24	181
SentencePiece	xlm-roberta-base_slow	97.24	181
SentencePiece	xlnet-base-cased	61.33	181
SentencePiece	xlnet-base-cased_slow	53.59	181
Tiktoken	Qwen/Qwen-14B-Chat	98.17	109
Tiktoken	Salesforce/xgen-7b-8k-base	97.25	109
WordPiece	ProsusAI/finbert	95.12	41
WordPiece	bert-base-multilingual-cased	95.12	41
WordPiece	bert-base-uncased	95.12	41
WordPiece	cointegrated/rubert-tiny2	80.49	41
WordPiece	distilbert-base-uncased-finetuned-sst-2-english	95.12	41
WordPiece	google/electra-base-discriminator	95.12	41
WordPiece	google/mobilebert-uncased	95.12	41
WordPiece	jhgan/ko-sbert-sts	75.61	41
WordPiece	prajjwal1/bert-mini	95.12	41
WordPiece	rajiv003/ernie-finetuned-qqp	95.12	41
WordPiece	rasa/LaBSE	87.80	41
WordPiece	sentence-transformers/all-MiniLM-L6-v2	75.61	41
WordPiece	squeezebert/squeezebert-uncased	95.12	41

Recreating Tokenizers From Tests

In some tokenizers, you need to select certain settings so that their output is closer to the Huggingface tokenizers:

• THUDM/chatglm2-6b detokenizer always skips special tokens. Use skip_special_tokens=True during conversion
• THUDM/chatglm3-6b detokenizer don't skips special tokens. Use skip_special_tokens=False during conversion
• All tested tiktoken based detokenizers leave extra spaces. Use clean_up_tokenization_spaces=False during conversion