Closed leonardozcm closed 9 months ago
It seems to be related to torch.bmm, pls try the following code on igpu:
import torch
import intel_extension_for_pytorch
import torch.nn as nn
torch_bmm = torch.bmm
device = 'xpu'
def test_func():
key_states = torch.ones((12, 1500, 64)).to(device)
query_states = torch.ones((12, 1500, 64)).to(device)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
torch.xpu.synchronize()
return attn_weights
for iter in range(100):
print(f"iter {iter}")
attn_weights = test_func()
Describe the bug
When running Whisper-Medium on an iGPU, serious unknown errors occur, leading to system crashes. And I think the memory usage of the iGPU (about 2GB) is far from reaching the system's upper limit.
To reproduce:
For the audio file you may refer to https://github.com/intel-analytics/BigDL/issues/8793#issuecomment-1690927181