Workaround for https://github.com/pytorch/pytorch/issues/102886
related to: https://github.com/pytorch/pytorch/issues/102476https://github.com/pytorch/pytorch/issues/102475https://github.com/pytorch/pytorch/issues/102474https://github.com/pytorch/pytorch/issues/102473https://github.com/pytorch/pytorch/issues/102473https://github.com/pytorch/pytorch/issues/102472
Since 9aaa12e328 the first inductor (CPU) UT fails until the GPU context is correct initialised and the subsequent UTs pass. CUDA observes the same issue and a workaround was pushed to force initialisation of cuda context by declaring an empty tensor https://github.com/pytorch/pytorch/issues/92627, we have adopted the same approach but have opted for `torch.zeros` which correctly activates the HIP context after the kernel launch.
**Reproducer:**
```
import torch
from torch._subclasses.fake_tensor import FakeTensorMode
import argparse
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Swap between torch.empty and torch.randn operations.')
parser.add_argument('--empty', action='store_true', help='Use torch.empty operation')
parser.add_argument('--rand', action='store_true', help='Use torch.randn operation')
args = parser.parse_args()
torch.cuda.set_device(0)
if args.empty:
torch.empty(1, device="cuda")
elif args.rand:
torch.rand(1, device="cuda")
print(f": hasPrimaryContext: {torch._C._cuda_hasPrimaryContext(0)")
with FakeTensorMode():
p = torch.randn(4, 2, requires_grad=True, device='cuda')
x = torch.randn(8, 4, device='cuda')
y = torch.mm(x, p).square().sum()
y.backward()
```
**ROCm python repro.py --empty**
0: hasPrimaryContext: False
**ROCm python repro.py --rand**
0: hasPrimaryContext: True
**CUDA python repro.py --empty**
0: hasPrimaryContext: True
**CUDA python repro.py --rand**
0: hasPrimaryContext: True
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103149
Approved by: https://github.com/eellison
e936277cc2