thundersvm is very slow for large dataset on preparation before train with GPU

[liningbo]

environment:
windows 10
python 3.8
cuda 11.6
thundersvm commit :

smalll dataset train log:
svm fit begin: (404, 13) (404,)
2024-04-06 20:30:50,603 INFO [default] #instances = 404, #features = 13
2024-04-06 20:30:50,825 INFO [default] total memory size is 0.000144489 max mem size is 8
2024-04-06 20:30:50,825 INFO [default] free mem is 7.99986
2024-04-06 20:30:50,825 INFO [default] working set size = 512
2024-04-06 20:30:50,825 INFO [default] training start
2024-04-06 20:30:51,208 INFO [default] global iter = 0, total local iter = 130, diff = 44.8
2024-04-06 20:30:51,219 INFO [default] global iter = 3, total local iter = 207, diff = 0.0999313
2024-04-06 20:30:51,219 INFO [default] training finished
2024-04-06 20:30:51,219 INFO [default] obj = -1203.39
2024-04-06 20:30:51,219 INFO [default] rho = -20.8858
2024-04-06 20:30:51,222 INFO [default] #sv = 395
2024-04-06 20:30:51,223 INFO [default] training finished
train time elapse(s):0.6843371391296387

large dataset train log:
svm fit begin: (1842924, 137) (1842924,)
2024-04-06 19:24:33,588 INFO [default] #instances = 1842924, #features = 137
2024-04-06 20:21:35,682 INFO [default] total memory size is 5.76696 max mem size is 8
2024-04-06 20:21:35,682 INFO [default] free mem is 2.23304
2024-04-06 20:21:35,682 INFO [default] working set size = 128
2024-04-06 20:21:35,935 INFO [default] training start
2024-04-06 20:21:36,376 INFO [default] global iter = 0, total local iter = 64, diff = 22.3704
2024-04-06 20:21:37,309 INFO [default] global iter = 5, total local iter = 224, diff = 18.7142
2024-04-06 20:21:37,309 INFO [default] training finished
2024-04-06 20:21:37,323 INFO [default] obj = -1547.53
2024-04-06 20:21:37,339 INFO [default] rho = 2.02697
2024-04-06 20:21:37,343 INFO [default] #sv = 448
2024-04-06 20:21:45,370 INFO [default] training finished
train time elapse(s):3582.7167501449585

max_mem_size is set to -1, the time between the first log and the second log cost nearly 1 hour before training.
all parameter:
self.params = {
'kernel': 'rbf',
'degree':6,
'gamma':'auto',
'coef0':0.0,
'C':1.0,
'epsilon':0.1,
'tol':0.1,
'probability':False,
'shrinking':False,
'cache_size':None,
'verbose':True,
'max_iter':5,
'n_jobs':-1,
'max_mem_size':-1,
'gpu_id':self.config.gpu_id,
}