g = nx.DiGraph()
g.add_node("code_atlas.llm:LLMClient.chat")              # 点を追加
g.add_edge("code_atlas.llm:LLMClient.test_connection",   # 矢印を追加
           "code_atlas.llm:LLMClient.chat")

conda create -n presidio_20260831 python==3.13

pip install "presidio_analyzer[transformers]"

pip install attrs tenacity sniffio requests-toolbelt distro zstandard jsonpatch mypy-extensions --upgrade

pip install  cryptography  marshmallow==3.26.2 referencing

pip install presidio_anonymizer

python -m spacy download en_core_web_sm

pip install presidio_image_redactor

pip install jupyterlab ipywidgets

text = "私は田中太郎。住所は279-0031 千葉県浦安市舞浜１−１。電話番号は03-999-9999。マイナンバーは123456789012。"

from presidio_analyzer import AnalyzerEngine
from presidio_analyzer.nlp_engine import TransformersNlpEngine
from presidio_anonymizer import AnonymizerEngine

model_config = [{"lang_code": "en", "model_name": {
    "spacy": "en_core_web_sm",  # use a small spaCy model for lemmas, tokens etc.
    "transformers": r"H:\BERT_Models\dslim\bert-base-NER"
    }
}]

nlp_engine = TransformersNlpEngine(models=model_config)
analyzer = AnalyzerEngine(nlp_engine=nlp_engine)

results = analyzer.analyze(text=text, language='en')
print(results)

anonymizer = AnonymizerEngine()
anonymized_text = anonymizer.anonymize(text=text, analyzer_results=results)

print(anonymized_text)

Some weights of the model checkpoint at H:\BERT_Models\dslim\bert-base-NER were not used when initializing BertForTokenClassification: ['bert.pooler.dense.bias', 'bert.pooler.dense.weight']
- This IS expected if you are initializing BertForTokenClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing BertForTokenClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
Device set to use cpu
[type: PHONE_NUMBER, start: 36, end: 47, score: 0.4]
text: 私は田中太郎。住所は279-0031 千葉県浦安市舞浜１−１。電話番号は<PHONE_NUMBER>。マイナンバーは123456789012。
items:
[
    {'start': 36, 'end': 50, 'entity_type': 'PHONE_NUMBER', 'text': '<PHONE_NUMBER>', 'operator': 'replace'}
]

python -m spacy download ja_core_news_sm

model_config = [{"lang_code": "ja", "model_name": {
    "spacy": "ja_core_news_sm",
    "transformers": r"H:\BERT_Models\tsmatz\xlm-roberta-ner-japanese"
    }
}]

nlp_engine = TransformersNlpEngine(models=model_config)
analyzer = AnalyzerEngine(nlp_engine=nlp_engine)

results = analyzer.analyze(text=text, language='ja')
print(results)

anonymizer = AnonymizerEngine()
anonymized_text = anonymizer.anonymize(text=text, analyzer_results=results)

print(anonymized_text)

Device set to use cpu
[type: LOCATION, start: 19, end: 21, score: 0.9985260963439941]
text: 私は田中太郎。住所は279-0031 <LOCATION>県浦安市舞浜１−１。電話番号は03-999-9999。マイナンバーは123456789012。
items:
[
    {'start': 19, 'end': 29, 'entity_type': 'LOCATION', 'text': '<LOCATION>', 'operator': 'replace'}
]

c:\ProgramData\anaconda3\envs\presidio_20260831\Lib\site-packages\transformers\pipelines\token_classification.py:490: UserWarning: Tokenizer does not support real words, using fallback heuristic
  warnings.warn(

pip install -U ginza ja_ginza

import spacy
import ginza

# GiNZAのcompound_splitterはsplit_mode未指定だとエラーになるため、
# spacy.load を一時的にラップしてconfigを注入する
_original_spacy_load = spacy.load

def _patched_spacy_load(name, *args, **kwargs):
    if name == "ja_ginza" and "config" not in kwargs:
        kwargs["config"] = {"components": {"compound_splitter": {"split_mode": "A"}}}
    return _original_spacy_load(name, *args, **kwargs)

spacy.load = _patched_spacy_load


from presidio_analyzer import AnalyzerEngine
from presidio_analyzer.nlp_engine import TransformersNlpEngine
from presidio_anonymizer import AnonymizerEngine

model_config = [{"lang_code": "ja", "model_name": {
    "spacy": "ja_ginza",  # ← ja_core_news_sm から変更
    "transformers": r"H:\BERT_Models\tsmatz\xlm-roberta-ner-japanese"
    }
}]

nlp_engine = TransformersNlpEngine(models=model_config)
analyzer = AnalyzerEngine(nlp_engine=nlp_engine)

results = analyzer.analyze(text=text, language='ja')
print(results)

anonymizer = AnonymizerEngine()
anonymized_text = anonymizer.anonymize(text=text, analyzer_results=results)

print(anonymized_text)

Device set to use cpu
[type: LOCATION, start: 19, end: 27, score: 0.9985260963439941]
text: 私は田中太郎。住所は279-0031 <LOCATION>１−１。電話番号は03-999-9999。マイナンバーは123456789012。
items:
[
    {'start': 19, 'end': 29, 'entity_type': 'LOCATION', 'text': '<LOCATION>', 'operator': 'replace'}
]

c:\ProgramData\anaconda3\envs\presidio_20260831\Lib\site-packages\transformers\pipelines\token_classification.py:490: UserWarning: Tokenizer does not support real words, using fallback heuristic
  warnings.warn(

import spacy
import ginza

if not getattr(spacy.load, "_ja_ginza_patched", False):
    _original_spacy_load = spacy.load

    def _make_patched_load(original_load):
        def _patched(name, *args, **kwargs):
            if name == "ja_ginza":
                # "disable" は生成はするが実行しないだけなので、
                # compound_splitter の split_mode 検証エラーを避けられない。
                # "exclude" ならコンポーネント自体を生成しないため、この問題を回避できる。
                exclude = list(kwargs.get("exclude", []))
                for comp in ("compound_splitter", "bunsetu_recognizer"):
                    if comp not in exclude:
                        exclude.append(comp)
                kwargs["exclude"] = exclude
                kwargs.pop("config", None)
            return original_load(name, *args, **kwargs)
        _patched._ja_ginza_patched = True
        return _patched

    spacy.load = _make_patched_load(_original_spacy_load)

from presidio_analyzer import AnalyzerEngine, Pattern, PatternRecognizer, RecognizerRegistry
from presidio_analyzer.nlp_engine import TransformersNlpEngine, NerModelConfiguration
from presidio_anonymizer import AnonymizerEngine


# ============================================================
# NLPエンジン設定(GiNZA + xlm-roberta-ner-japanese)
# ============================================================
model_config = [{"lang_code": "ja", "model_name": {
    "spacy": "ja_ginza",
    "transformers": r"H:\BERT_Models\tsmatz\xlm-roberta-ner-japanese"
    }
}]

ner_model_configuration = NerModelConfiguration(
    model_to_presidio_entity_mapping={
        "PER": "PERSON",
        "PERSON": "PERSON",
        "LOC": "LOCATION",
        "GPE": "LOCATION",
        "ORG": "ORGANIZATION",
        "DAT": "DATE_TIME",
        "DATE": "DATE_TIME",
        "TIM": "DATE_TIME",
        "TIME": "DATE_TIME",
        "MONEY": "MONEY",
        "PERCENT": "PERCENT",
    },
    low_confidence_score_multiplier=0.4,
    low_score_entity_names=["ORGANIZATION"],
    alignment_mode="expand",
    aggregation_strategy="simple",  # 日本語(スペース無し)では"average"等だと、空白ベースのフォールバック判定が誤動作するため"simple"を使う
)

nlp_engine = TransformersNlpEngine(
    models=model_config,
    ner_model_configuration=ner_model_configuration,
)
nlp_engine.load()


# ============================================================
# カスタムレコグナイザー: 郵便番号・マイナンバー
# ============================================================
postal_recognizer = PatternRecognizer(
    supported_entity="JP_POSTAL_CODE",
    patterns=[Pattern(name="jp_postal_code", regex=r"\b\d{3}-\d{4}\b", score=0.85)],
    context=["郵便番号", "〒"],
    supported_language="ja",
)

my_number_recognizer = PatternRecognizer(
    supported_entity="JP_MY_NUMBER",
    patterns=[Pattern(name="jp_my_number", regex=r"(?<!\d)\d{12}(?!\d)", score=0.3)],
    context=["マイナンバー", "個人番号"],
    supported_language="ja",
)


# ============================================================
# レジストリ構築(既定 + カスタムを統合)
# ============================================================
registry = RecognizerRegistry(supported_languages=["ja"])
registry.load_predefined_recognizers(nlp_engine=nlp_engine, languages=["ja"])
registry.add_recognizer(postal_recognizer)
registry.add_recognizer(my_number_recognizer)

analyzer = AnalyzerEngine(
    nlp_engine=nlp_engine,
    registry=registry,
    supported_languages=["ja"],
)


# ============================================================
# 実行
# ============================================================
results = analyzer.analyze(text=text, language='ja')
print(results)

anonymizer = AnonymizerEngine()
anonymized_text = anonymizer.anonymize(text=text, analyzer_results=results)

print(anonymized_text)

Device set to use cpu
[type: PERSON, start: 2, end: 6, score: 0.999563992023468, type: LOCATION, start: 10, end: 30, score: 0.998464047908783, type: JP_POSTAL_CODE, start: 39, end: 47, score: 0.85, type: JP_MY_NUMBER, start: 55, end: 67, score: 0.6499999999999999, type: PHONE_NUMBER, start: 36, end: 47, score: 0.4]
text: 私は<PERSON>。住所は<LOCATION>。電話番号は<PHONE_NUMBER>。マイナンバーは<JP_MY_NUMBER>。
items:
[
    {'start': 52, 'end': 66, 'entity_type': 'JP_MY_NUMBER', 'text': '<JP_MY_NUMBER>', 'operator': 'replace'},
    {'start': 30, 'end': 44, 'entity_type': 'PHONE_NUMBER', 'text': '<PHONE_NUMBER>', 'operator': 'replace'},
    {'start': 14, 'end': 24, 'entity_type': 'LOCATION', 'text': '<LOCATION>', 'operator': 'replace'},
    {'start': 2, 'end': 10, 'entity_type': 'PERSON', 'text': '<PERSON>', 'operator': 'replace'}
]

import pytesseract
from presidio_image_redactor import ImageRedactorEngine
from PIL import Image

pytesseract.pytesseract.tesseract_cmd = r"C:\Program Files\Tesseract-OCR\tesseract.exe"

image = Image.open(r"D:\Workspace\Python\Presidio\pii_test_document.png")

redactor = ImageRedactorEngine()
redacted_image = redactor.redact(image=image, ocr_kwargs={"lang": "jpn"})

# オリジナル表示
display(image)
display(redacted_image)

import spacy
import ginza

if not getattr(spacy.load, "_ja_ginza_patched", False):
    _original_spacy_load = spacy.load

    def _make_patched_load(original_load):
        def _patched(name, *args, **kwargs):
            if name == "ja_ginza":
                exclude = list(kwargs.get("exclude", []))
                for comp in ("compound_splitter", "bunsetu_recognizer"):
                    if comp not in exclude:
                        exclude.append(comp)
                kwargs["exclude"] = exclude
                kwargs.pop("config", None)
            return original_load(name, *args, **kwargs)
        _patched._ja_ginza_patched = True
        return _patched

    spacy.load = _make_patched_load(_original_spacy_load)


from presidio_analyzer import AnalyzerEngine, Pattern, PatternRecognizer, RecognizerRegistry
from presidio_analyzer.nlp_engine import TransformersNlpEngine, NerModelConfiguration
from presidio_image_redactor import ImageRedactorEngine, ImageAnalyzerEngine
import pytesseract
from PIL import Image

pytesseract.pytesseract.tesseract_cmd = r"C:\Program Files\Tesseract-OCR\tesseract.exe"


# ============================================================
# 日本語NLPエンジン(GiNZA + xlm-roberta-ner-japanese)
# ============================================================
model_config = [{"lang_code": "ja", "model_name": {
    "spacy": "ja_ginza",
    "transformers": r"H:\BERT_Models\tsmatz\xlm-roberta-ner-japanese"
    }
}]

ner_model_configuration = NerModelConfiguration(
    model_to_presidio_entity_mapping={
        "PER": "PERSON", "PERSON": "PERSON",
        "LOC": "LOCATION", "GPE": "LOCATION",
        "ORG": "ORGANIZATION",
        "DAT": "DATE_TIME", "DATE": "DATE_TIME",
        "TIM": "DATE_TIME", "TIME": "DATE_TIME",
        "MONEY": "MONEY", "PERCENT": "PERCENT",
    },
    low_confidence_score_multiplier=0.4,
    low_score_entity_names=["ORGANIZATION"],
    alignment_mode="expand",
    aggregation_strategy="simple",
)

nlp_engine = TransformersNlpEngine(models=model_config, ner_model_configuration=ner_model_configuration)
nlp_engine.load()


# ============================================================
# カスタムレコグナイザー(郵便番号・マイナンバー)
# ============================================================
postal_recognizer = PatternRecognizer(
    supported_entity="JP_POSTAL_CODE",
    patterns=[Pattern(name="jp_postal_code", regex=r"\b\d{3}-\d{4}\b", score=0.85)],
    context=["郵便番号", "〒"],
    supported_language="ja",
)
my_number_recognizer = PatternRecognizer(
    supported_entity="JP_MY_NUMBER",
    patterns=[Pattern(name="jp_my_number", regex=r"(?<!\d)\d{12}(?!\d)", score=0.3)],
    context=["マイナンバー", "個人番号"],
    supported_language="ja",
)

registry = RecognizerRegistry(supported_languages=["ja"])
registry.load_predefined_recognizers(nlp_engine=nlp_engine, languages=["ja"])
registry.add_recognizer(postal_recognizer)
registry.add_recognizer(my_number_recognizer)

analyzer = AnalyzerEngine(nlp_engine=nlp_engine, registry=registry, supported_languages=["ja"])


# ============================================================
# 画像レダクター: 日本語analyzerを明示的に組み込む
# ============================================================
image_analyzer_engine = ImageAnalyzerEngine(analyzer_engine=analyzer)
redactor = ImageRedactorEngine(image_analyzer_engine=image_analyzer_engine)

image = Image.open(r"D:\Workspace\Python\Presidio\pii_test_document.png")

redacted_image = redactor.redact(
    image=image,
    ocr_kwargs={"lang": "jpn"},
    language="ja",
)

redacted_image

wsl --install

sudo apt install git git-lfs python3-pip python3-venv

sudo apt install python3.12

python3 -m pip install --user -U uv
uv --version

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc

mkdir -p ~/litert-conversion
cd ~/litert-conversion

python3.12 -m venv litert-lm-20260719

source litert-lm-20260719/bin/activate

python -m pip install --upgrade pip
python -m pip install uv
uv tool install --python 3.12 litert-torch-nightly

litert-torch --help

mkdir -p ~/litert-conversion/output

litert-torch export_hf --model="$HOME/Qwen2.5-Coder-7B-Instruct" --output_dir="$HOME/litert-conversion/output" --externalize_embedder --quantization_recipe=dynamic_wi4_afp32

import os
import psutil
import time
import litert_lm

process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2  # MB

model_path = r"H:\LLM_Models\litert-lm\qwen2.5-coder-7b-it-dynamic_wi4_afp32\qwen2.5-coder-7b-it-dynamic_wi4_afp32.litertlm"

PROMPT = (
    "transformersとFAISSとstreamlitを使った、"
    "QAチャットボットをPythonで構築して。"
)


process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2
start = time.perf_counter()

with litert_lm.Engine(
    str(model_path),
    backend=litert_lm.Backend.CPU(),
) as engine:

    with engine.create_conversation() as conversation:
        response = conversation.send_message(PROMPT)

        answer = response["content"][0]["text"]

        mem_during = process.memory_info().rss / 1024**2

end = time.perf_counter()

mem_after = process.memory_info().rss / 1024**2

print(answer)

print("=" * 60)
print(f"実行前メモリ使用量       : {mem_before:.2f} MB")
print(f"モデル使用中メモリ使用量 : {mem_during:.2f} MB")
print(f"解放後メモリ使用量       : {mem_after:.2f} MB")
print(f"モデル使用中の増加量     : {mem_during - mem_before:.2f} MB")
print(f"実行時間                 : {end - start:.3f} 秒")

import os
import psutil
import time
from optimum.intel import OVModelForCausalLM
from transformers import AutoProcessor
process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2  # MB
start = time.time()
model_path = r"H:\LLM_Models\safetensor\Qwen\Qwen2.5-Coder-7B-Instruct"
processor = AutoProcessor.from_pretrained(model_path)
model = OVModelForCausalLM.from_pretrained(model_path)
messages = [
    {"role": "system", "content": "あなたは有能なAIアシスタントです。"},
    {"role": "user", "content": "面白いジョークを話して。またどこが面白いかも解説してください。"},
]
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Generate output
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
mem_after = process.memory_info().rss / 1024**2  # MB
end = time.time()
print(response)
print("=" * 60)
print(f"実行前メモリ使用量: {mem_before:.2f} MB")
print(f"実行後メモリ使用量: {mem_after:.2f} MB")
print(f"実行時間: {end - start:.3f}")
# OpenVINO IR形式を保存
# 次回以降model_pathに保存済みのOpenVINO IR形式へのpathを渡せば変換が省略される。
open_vino_save_path = r"H:\LLM_Models\open_vino\Qwen\Qwen2.5-Coder-7B-Instruct"
model.save_pretrained(open_vino_save_path)
processor.save_pretrained(open_vino_save_path)

process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2  # MB
start = time.time()
processor = AutoProcessor.from_pretrained(open_vino_save_path)
model = OVModelForCausalLM.from_pretrained(open_vino_save_path)
messages = [
    {"role": "system", "content": "あなたは有能なAIアシスタントです。"},
    {"role": "user", "content": "面白いジョークを話して。またどこが面白いかも解説してください。"},
]
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
mem_after = process.memory_info().rss / 1024**2  # MB
end = time.time()
print(response)
print("=" * 60)
print(f"実行前メモリ使用量: {mem_before:.2f} MB")
print(f"実行後メモリ使用量: {mem_after:.2f} MB")
print(f"実行時間: {end - start:.3f} 秒")

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2  # MB
start = time.time()
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    quantization_config=quantization_config,
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(model_path)
quantization_time = time.time()
messages = [
    {"role": "system", "content": "あなたは有能なAIアシスタントです。"},
    {"role": "user", "content": "面白いジョークを話して。またどこが面白いかも解説してください。"},
]
text = tokenizer.apply_chat_template(
    messages,
    temperature=0.1,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
mem_after = process.memory_info().rss / 1024**2  # MB
end = time.time()
print(response)
print("=" * 60)
print(f"実行前メモリ使用量: {mem_before:.2f} MB")
print(f"実行後メモリ使用量: {mem_after:.2f} MB")
print(f"差引メモリ使用量: {mem_after - mem_before:.2f} MB")
print(f"実行時間: {end - start:.3f} 秒")
print(f"量子化処理を除いた推論時間: {end - quantization_time:.3f} 秒")
del model, tokenizer

import os
import psutil
import time
from optimum.intel import OVModelForCausalLM
from transformers import AutoProcessor
model_path = r"H:\LLM_Models\safetensor\Qwen\Qwen2.5-Coder-7B-Instruct"
# モデル量子化用に追加でロード
from nncf import compress_weights, CompressWeightsMode
process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2  # MB
start = time.time()
model = OVModelForCausalLM.from_pretrained(model_path,
                                           load_in_8bit=False, # これを指定することでデフォルトの8ビット量子化を無効化
                                           compile=False,   # この段階でモデルをコンパイルしても、その後4ビットモデルを作成⇒再コンパイルすることとなり無駄となる。
                                                            # Falseを指定することでコンパイルをキャンセルできる。
                                           )
# modeを指定することで量子化モードを指定できる。指定できるパラメータは以下のマークダウンを参照。
model.model = compress_weights(model.model,
                               mode=CompressWeightsMode.INT4_ASYM,
                              )
# プロセッサも保存する必要があるので読み込み
processor = AutoProcessor.from_pretrained(model_path)
# 量子化したモデルをOpenVINO IR形式で保存
open_vino_save_path_int4a = r"H:\LLM_Models\open_vino\Qwen\Qwen2.5-Coder-7B-Instruct_INT4-ASYM"
model.save_pretrained(open_vino_save_path_int4a)
processor.save_pretrained(open_vino_save_path_int4a)
end = time.time()
print(f"実行時間: {end - start:.3f} 秒")
del model, processor

process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2  # MB
start = time.time()
processor = AutoProcessor.from_pretrained(open_vino_save_path_int4a)
model = OVModelForCausalLM.from_pretrained(open_vino_save_path_int4a)
messages = [
    {"role": "system", "content": "あなたは有能なAIアシスタントです。"},
    {"role": "user", "content": "面白いジョークを話して。またどこが面白いかも解説してください。"},
]
processor = AutoProcessor.from_pretrained(model_path)
inputs = processor.apply_chat_template(
    messages,
    temperature=0.1,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False,
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
mem_after = process.memory_info().rss / 1024**2  # MB
end = time.time()
print(response)
print("=" * 60)
print(f"実行前メモリ使用量: {mem_before:.2f} MB")
print(f"実行後メモリ使用量: {mem_after:.2f} MB")
print(f"差引メモリ使用量: {mem_after - mem_before:.2f} MB")
print(f"実行時間: {end - start:.3f} 秒")
del model, processor

# KVキャッシュの量子化を無効化してみる
process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2  # MB
start = time.time()
processor = AutoProcessor.from_pretrained(open_vino_save_path_int4a)
model = OVModelForCausalLM.from_pretrained(open_vino_save_path_int4a,
                                           ov_config={"KV_CACHE_PRECISION": "f16",
                                                      "PERFORMANCE_HINT": "LATENCY",
                                                      },
                                           )
messages = [
    {"role": "system", "content": "あなたは有能なコーディングAIです。"},
    {"role": "user", "content": "transformersとFAISSとstreamlitを使った、QAチャットボットをpythonで構築して。"},
]
processor = AutoProcessor.from_pretrained(model_path)
inputs = processor.apply_chat_template(
    messages,
    temperature=0.1,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False,
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
outputs = model.generate(**inputs, max_new_tokens=4096)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
mem_after = process.memory_info().rss / 1024**2  # MB
end = time.time()
print(response)
print("=" * 60)
print(f"実行前メモリ使用量: {mem_before:.2f} MB")
print(f"実行後メモリ使用量: {mem_after:.2f} MB")
print(f"差引メモリ使用量: {mem_after - mem_before:.2f} MB")
print(f"実行時間: {end - start:.3f} 秒")
del model, processor

# KVキャッシュの量子化をu4にしてみる
process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2  # MB
start = time.time()
processor = AutoProcessor.from_pretrained(open_vino_save_path_int4a)
model = OVModelForCausalLM.from_pretrained(open_vino_save_path_int4a,
                                           ov_config={"KV_CACHE_PRECISION": "u4",
                                                      "DYNAMIC_QUANTIZATION_GROUP_SIZE": "32",
                                                      "PERFORMANCE_HINT": "LATENCY",
                                                      },
                                           )
messages = [
    {"role": "system", "content": "あなたは有能なコーディングAIです。"},
    {"role": "user", "content": "transformersとFAISSとstreamlitを使った、QAチャットボットをpythonで構築して。"},
]
processor = AutoProcessor.from_pretrained(model_path)
inputs = processor.apply_chat_template(
    messages,
    temperature=0.1,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False,
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
outputs = model.generate(**inputs, max_new_tokens=4096)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
mem_after = process.memory_info().rss / 1024**2  # MB
end = time.time()
print(response)
print("=" * 60)
print(f"実行前メモリ使用量: {mem_before:.2f} MB")
print(f"実行後メモリ使用量: {mem_after:.2f} MB")
print(f"差引メモリ使用量: {mem_after - mem_before:.2f} MB")
print(f"実行時間: {end - start:.3f} 秒")
del model, processor

# NPUが認識されるかチェック
# ['CPU', 'GPU', 'NPU']のようにNPUが表示されればOK
import openvino as ov
core = ov.Core()
print("利用可能なデバイス:", core.available_devices)
for device in core.available_devices:
    print(device, core.get_property(device, "FULL_DEVICE_NAME"))
# 必要なライブラリのインポート
import openvino
import openvino_genai
import transformers
import nncf
print("OpenVINO   :", openvino.__version__)
print("Transformers:", transformers.__version__)
print("NNCF       :", nncf.__version__)

optimum-cli export openvino -m "H:\LLM_Models\safetensor\Qwen\Qwen2.5-Coder-7B-Instruct" --task text-generation-with-past --weight-format int4 --sym --ratio 1.0 --group-size -1 "H:\LLM_Models\open_vino\Qwen\Qwen2.5-Coder-7B-Instruct-int4-npu"

import os
import time
import psutil
import openvino_genai as ov_genai
from transformers import AutoTokenizer
model_path = (
    r"H:\LLM_Models\open_vino\Qwen"
    r"\Qwen2.5-Coder-7B-Instruct-int4-npu"
)
process = psutil.Process(os.getpid())
mem_before = process.memory_info().rss / 1024**2
start = time.time()
tokenizer = AutoTokenizer.from_pretrained(
    model_path,
    trust_remote_code=True,
)
messages = [
    {
        "role": "system",
        "content": "あなたは有能なAIアシスタントです。",
    },
    {
        "role": "user",
        "content": (
            "Pythonで高速な素数判定関数を書いて、"
            "処理内容も説明してください。"
        ),
    },
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
pipeline_config = {
    "MAX_PROMPT_LEN": 1024,
    "MIN_RESPONSE_LEN": 128,
    "GENERATE_HINT": "BEST_PERF",
    "CACHE_DIR": os.path.join(model_path, "npu_cache"),
}
pipe = ov_genai.LLMPipeline(
    model_path,
    "NPU",
    pipeline_config,
)
result = pipe.generate(
    prompt,
    max_new_tokens=512,
    do_sample=False,
)
end = time.time()
mem_after = process.memory_info().rss / 1024**2
print(result)
print("=" * 60)
print(f"実行前メモリ使用量: {mem_before:.2f} MB")
print(f"実行後メモリ使用量: {mem_after:.2f} MB")
print(f"メモリ増加量      : {mem_after - mem_before:+.2f} MB")
print(f"実行時間          : {end - start:.3f} 秒")

vs_BuildTools.exe --layout .\VSLayout --lang en-US ja-JP --add Microsoft.VisualStudio.Workload.VCTools --add Microsoft.VisualStudio.Component.VC.Tools.x86.x64 --add Microsoft.VisualStudio.Component.Windows11SDK.26100 --add Microsoft.VisualStudio.Component.VC.CMake.Project