Why is tslearn.metrics.dtw(gpu) much slower than dtadistance.dtw(cpu)?

[Meoooww]

device = torch.device("cuda")  
s_t = torch.tensor(metric_today[:], dtype=torch.float32).reshape(-1, 1).to(device)     # length == 1440
s_y = torch.tensor(metric_yesterday[:], dtype=torch.float32).reshape(-1, 1).to(device)
s_l = torch.tensor(metric_lastweek[:], dtype=torch.float32).reshape(-1, 1).to(device)
t1 = time.time() - start_t
slice_yesterday = dtw(s_t, s_y, be="pytorch")  
slice_week = dtw(s_t, s_l, be="pytorch")
slice = min(slice_yesterday, slice_week)
t2 = time.time() - t1

tslearn.metrics.dtw use 20s, t2 is much larger than t1

from dtaidistance import dtw
slice_yesterday = dtw.distance(metric_today[start:], metric_yesterday[start:])
slice_week = dtw.distance(metric_today[start:], metric_lastweek[start:])
slice = min(slice_yesterday, slice_week)

dtaidistance.dtw only use 0.68s

why? Shouldn't GPU computing be faster?