pip install openvino-dev[pytorch]==2022.3.0
pip install --upgrade --upgrade-strategy eager "optimum[neural-compressor]"
pip install --upgrade --upgrade-strategy eager "optimum[openvino]"
pip install --upgrade --upgrade-strategy eager "optimum[ipex]"

Export OpenVINO Model

Python

from transformers import AutoTokenizer
from optimum.intel import OVWeightQuantizationConfig
from optimum.intel.openvino import OVModelForCausalLM
from optimum.exporters.openvino.convert import export_tokenizer
from pathlib import Path
import os
#fp16 int8 int4
precision="int8"
#导出模型的路径
ir_model_path = Path("./qwen0.5b-ov")
if ir_model_path.exists() == False:
    os.mkdir(ir_model_path)
compression_configs = {
    "sym": False,
    "group_size": 128,
    "ratio": 0.8,
}
#加载模型
model_path = "Qwen/Qwen1.5-0.5B-Chat"

print("====Exporting IR=====")
if precision == "int4":
    ov_model = OVModelForCausalLM.from_pretrained(model_path, export=True,
                                                    compile=False, quantization_config=OVWeightQuantizationConfig(
                                                        bits=4, **compression_configs))
elif precision == "int8":
    ov_model = OVModelForCausalLM.from_pretrained(model_path, export=True,
                                                    compile=True, load_in_8bit=True)
else:
    ov_model = OVModelForCausalLM.from_pretrained(model_path, export=True,
                                                    compile=False, load_in_8bit=False)

ov_model.save_pretrained(ir_model_path)

tokenizer = AutoTokenizer.from_pretrained(
    model_path)
tokenizer.save_pretrained(ir_model_path)

print("====Exporting IR tokenizer=====")
export_tokenizer(tokenizer, ir_model_path)

Usage

python

from optimum.intel.openvino import OVModelForCausalLM
from transformers import (AutoTokenizer, AutoConfig,
                          TextIteratorStreamer)
#导出模型的路径
model_dir = "./qwen0.5b-ov"
ov_config = {"PERFORMANCE_HINT": "LATENCY",
             "NUM_STREAMS": "1", "CACHE_DIR": ""}
tokenizer = AutoTokenizer.from_pretrained(
    model_dir)
ov_model = OVModelForCausalLM.from_pretrained(
    model_dir,
    device="cpu",
    ov_config=ov_config,
    config=AutoConfig.from_pretrained(model_dir),
    trust_remote_code=True,
)
streamer = TextIteratorStreamer(
    tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True
)
prompt = "今天天气如何？"
length=len(prompt)
messages = [
{"role": "user", "content": prompt}
]
model_inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
)
generate_kwargs = dict(
        input_ids=model_inputs,
        max_new_tokens=length,
        temperature=0.1,
        max_length=500,
        do_sample=True,
        top_p=1.0,
        top_k=50,
        repetition_penalty=1.1,
        streamer=streamer,
        pad_token_id=151645,
    )
generated_ids = ov_model.generate(**generate_kwargs)
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Anthonyfhd/Qwen1.5-0.5B-Chat-IntelOptimum

Model Details

Requirements

Export OpenVINO Model

Usage

Anthonyfhd/
Qwen1.5-0.5B-Chat-IntelOptimum