PatchTST is running too slow,GPU utilization is low

[cssuperman]

Description

I have a dataset with only 2 columns ['y','ds'] with 100,000 rows, trained with patchTST model, loading data is very slow, GPU utilization is only 3~15%, no matter how big or small I set batch_size, it's very slow. How to solve this problem？

import pandas as pd
from gluonts.dataset.pandas import PandasDataset
from gluonts.dataset.split import split
from gluonts.torch.model.patch_tst import PatchTSTEstimator

file = 'data.parquet' # dataset has 'y' 'ds' columns and   100000 rows

df = pd.read_parquet(file)
df['y']=df['y'].astype('float32') #avoid problem "mat1 and mat2 must have the same dtype, but got Double and Float"

dataset = PandasDataset(df, freq='500ms',target="y",timestamp="ds")

prediction_length = 16
training_data, test_gen = split(dataset, offset=-1*prediction_length*3)
test_data = test_gen.generate_instances(prediction_length=prediction_length, windows=3)

model = PatchTSTEstimator(
    prediction_length=prediction_length,
    patch_len=prediction_length,
    trainer_kwargs={
                    "max_epochs": 1,
                    },
    batch_size=5000, #batch_size=500,
    num_batches_per_epoch=200, #num_batches_per_epoch=2000, 
).train(training_data)

Environment

Operating system:Win11
Python version: 3.12.11
GluonTS version: 0.16.2
GPU:4090