Coverage for src/flag_gems/runtime/backend/_tsingmicro/ops/randn_like.py: 0%

1import logging 

2 

3import torch 

4import triton 

5 

6from flag_gems.runtime import torch_device_fn 

7from flag_gems.utils.random_utils import philox_backend_seed_offset 

8 

9from .randn import randn_kernel 

10 

11logger = logging.getLogger(__name__) 

12UNROLL = 4 

13 

14 

15def randn_like( 

16 x, *, dtype=None, layout=None, device=None, pin_memory=None, memory_format=None 

17): 

18 logger.debug("GEMS_TSINGMICRO RANDN_LIKE") 

19 if device is None: 

20 device = x.device.index 

21 if dtype is None: 

22 dtype = x.dtype 

23 out = torch.empty_like(x, device=device, dtype=dtype) 

24 N = x.numel() 

25 grid_fn = lambda meta: (triton.cdiv(N, meta["BLOCK"] * UNROLL),) 

26 # (TODO) Using Triton autotuner makes kernel parameters opaque to the caller, 

27 # hence we cannot obtain the per thread offset as in Pytorch. 

28 increment = triton.cdiv(N, UNROLL) 

29 philox_seed, philox_offset = philox_backend_seed_offset(increment) 

30 with torch_device_fn.device(x.device): 

31 randn_kernel[grid_fn](out, N, philox_seed, philox_offset) 

32 return out