pip:
    - torch-workflow-archiver
    - torch-model-archiver 
    - torchserve 

# 1. 導(dǎo)入huggingface模型
from transformers import AutoModelForCTC, AutoProcessor
import os
modelname = "facebook/wav2vec2-base-960h"
model = AutoModelForCTC.from_pretrained(modelname)
processor = AutoProcessor.from_pretrained(modelname)
modelpath = "model"
os.makedirs(modelpath, exist_ok=True)
model.save_pretrained(modelpath)
processor.save_pretrained(modelpath)
# 2. 自定義handler
import torch
import torchaudio
from transformers import AutoProcessor, AutoModelForCTC
import io
class Wav2VecHandler(object):
    def __init__(self):
        self._context = None
        self.initialized = False
        self.model = None
        self.processor = None
        self.device = None
        # Sampling rate for Wav2Vec model must be 16k
        self.expected_sampling_rate = 16_000
    def initialize(self, context):
        """Initialize properties and load model"""
        self._context = context
        self.initialized = True
        properties = context.system_properties
        # See https://pytorch.org/serve/custom_service.html#handling-model-execution-on-multiple-gpus
        self.device = torch.device("cuda:" + str(properties.get("gpu_id")) if torch.cuda.is_available() else "cpu")
        model_dir = properties.get("model_dir")
        self.processor = AutoProcessor.from_pretrained(model_dir)
        self.model = AutoModelForCTC.from_pretrained(model_dir)
    def handle(self, data, context):
        """Transform input to tensor, resample, run model and return transcribed text."""
        input = data[0].get("data")
        if input is None:
            input = data[0].get("body")
        # torchaudio.load accepts file like object, here `input` is bytes
        model_input, sample_rate = torchaudio.load(io.BytesIO(input), format="WAV")
        # Ensure sampling rate is the same as the trained model
        if sample_rate != self.expected_sampling_rate:
            model_input = torchaudio.functional.resample(model_input, sample_rate, self.expected_sampling_rate)
        model_input = self.processor(model_input, sampling_rate = self.expected_sampling_rate, return_tensors="pt").input_values[0]
        logits = self.model(model_input)[0]
        pred_ids = torch.argmax(logits, axis=-1)[0]
        output = self.processor.decode(pred_ids)
        return [output]

# 打包部署模型文件, 把model部署到torchserve 
torch-model-archiver --model-name Wav2Vec2 --version 1.0 --serialized-file model/pytorch_model.bin --handler ./handler.py --extra-files "model/config.json,model/special_tokens_map.json,model/tokenizer_config.json,model/vocab.json,model/preprocessor_config.json" -f
mv Wav2Vec2.mar model_store
# 啟動model服務(wù), 加載前面打包的model, 并以grpc和http接口向外提供推理服務(wù)
torchserve --start --model-store model_store --models Wav2Vec2=Wav2Vec2.mar --ncs
# Once the server is running, let's try it with:
curl -X POST http://127.0.0.1:8080/predictions/Wav2Vec2 --data-binary '@./sample.wav' -H "Content-Type: audio/basic"
# 暫停torchserve serving
torchserve --stop

torch-model-archiver：用來打包模型
usage: torch-model-archiver [-h] --model-name MODEL_NAME
                            [--serialized-file SERIALIZED_FILE]
                            [--model-file MODEL_FILE] --handler HANDLER
                            [--extra-files EXTRA_FILES]
                            [--runtime {python,python2,python3}]
                            [--export-path EXPORT_PATH]
                            [--archive-format {tgz,no-archive,default}] [-f]
                            -v VERSION [-r REQUIREMENTS_FILE]