OpenVINO トークナイザー: OpenVINO パイプラインにテキスト処理を組み込み#

%pip install -Uq pip 
%pip install --pre -Uq openvino-tokenizers[transformers] --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly 
%pip install "torch>=2.1" --extra-index-url https://download.pytorch.org/whl/cpu

from pathlib import Path 

tokenizer_dir = Path("tokenizer/") 
model_id = "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"

model_dir = Path(Path(model_id).name) 

if not model_dir.exists():
    # 元のモデルを変換 
    # %pip install -U "git+https://github.com/huggingface/optimum-intel.git" "nncf>=2.8.0" onnx 
    # %optimum-cli export openvino -m $model_id --task text-generation-with-past $model_dir 

    # すでに変換されたモデルをロード 
    from huggingface_hub import hf_hub_download 

    hf_hub_download( 
        "chgk13/TinyLlama-1.1B-intermediate-step-1431k-3T", 
        filename="openvino_model.xml", 
        local_dir=model_dir, 
    ) 
    hf_hub_download( 
        "chgk13/TinyLlama-1.1B-intermediate-step-1431k-3T", 
        filename="openvino_model.bin", 
        local_dir=model_dir, 
    )

import numpy as np 
from tqdm.notebook import trange 
from pathlib import Path 
from openvino_tokenizers import add_greedy_decoding 
from openvino_tokenizers.constants import EOS_TOKEN_ID_NAME 
from openvino import Core 

core = Core() 

# LLM 上に貪欲復号サブグラフを追加して、最も可能性の高いトークンを出力として取得 
ov_model = add_greedy_decoding(core.read_model(model_dir / "openvino_model.xml")) 
compiled_model = core.compile_model(ov_model) 
infer_request = compiled_model.create_infer_request()

text_input = ["Quick brown fox jumped"] 

model_input = {name.any_name: output for name, output in tokenizer(text_input).items()} 

if "position_ids" in (input.any_name for input in infer_request.model_inputs): 
    model_input["position_ids"] = np.arange(model_input["input_ids"].shape[1], dtype=np.int64)[np.newaxis, :]
# ビームサーチなし、idx を 0 に設定 
model_input["beam_idx"] = np.array([0], dtype=np.int32) 
# 文末トークンは、モデルがテキスト生成の終了を意味する 
# トークナイザー/デトークナイザー ov.Model オブジェクトの rt_info から EOS トークン ID を読み取り 
eos_token = ov_tokenizer.get_rt_info(EOS_TOKEN_ID_NAME).value 

tokens_result = np.array([[]], dtype=np.int64) 

# 推論前にモデル内の KV キャッシュをリセット 
infer_request.reset_state() 
max_infer = 10 

for _ in trange(max_infer): 
    infer_request.start_async(model_input) 
    infer_request.wait() 

    # 最初の推論で最後のトークンの予測を取得 
    output_token = infer_request.get_output_tensor().data[:, -1:] 
    tokens_result = np.hstack((tokens_result, output_token)) 
    if output_token[0, 0] == eos_token: 
        break 

    # 新しい推論のための入力を準備 
    model_input["input_ids"] = output_token 
    model_input["attention_mask"] = np.hstack((model_input["attention_mask"].data, [[1]])) 
    model_input["position_ids"] = np.hstack( 
        ( 
            model_input["position_ids"].data, 
            [[model_input["position_ids"].data.shape[-1]]], 
        ) 
    ) 

text_result = detokenizer(tokens_result)["string_output"] 
print(f"Prompt:\n{text_input[0]}") 
print(f"Generated:\n{text_result[0]}")

model_id = "mrm8488/bert-tiny-finetuned-sms-spam-detection" 
model_dir = Path(Path(model_id).name) 

if not model_dir.exists():     %pip install -qU git+https://github.com/huggingface/optimum-intel.git onnx 
    !optimum-cli export openvino --model $model_id --task text-classification $model_dir 
    !convert_tokenizer $model_id -o $model_dir

from openvino import Core, save_model 
from openvino_tokenizers import connect_models 

core = Core() 
text_input = ["Free money!!!"] 

ov_tokenizer = core.read_model(model_dir / "openvino_tokenizer.xml") 
ov_model = core.read_model(model_dir / "openvino_model.xml") 
combined_model = connect_models(ov_tokenizer, ov_model) 
save_model(combined_model, model_dir / "combined_openvino_model.xml") 

compiled_combined_model = core.compile_model(combined_model) 
openvino_output = compiled_combined_model(text_input) 

print(f"Logits: {openvino_output['logits']}")