Coverage for src/flag_gems/runtime/backend/_metax/ops/zeros.py: 0%

1import logging 

2 

3import torch 

4import triton 

5import triton.language as tl 

6 

7import flag_gems.runtime as runtime 

8from flag_gems.runtime import device, torch_device_fn 

9from flag_gems.utils import libentry 

10from flag_gems.utils import triton_lang_extension as tle 

11from flag_gems.utils.shape_utils import volume 

12 

13logger = logging.getLogger("flag_gems." + __name__) 

14device_ = device 

15 

16 

17@libentry() 

18@triton.heuristics(runtime.get_heuristic_config("zeros")) 

19@triton.jit 

20def zeros_kernel( 

21 output_ptr, 

22 N, 

23 BLOCK_SIZE: tl.constexpr, 

24): 

25 pid = tle.program_id(axis=0) # We use a 1D launch grid so axis is 0. 

26 block_start = pid * BLOCK_SIZE 

27 offsets = block_start + tl.arange(0, BLOCK_SIZE) 

28 mask = offsets < N 

29 tl.store(output_ptr + offsets, 0.0, mask=mask) 

30 

31 

32def zeros(size, *, dtype=None, layout=None, device=None, pin_memory=None): 

33 logger.debug("METAX GEMS ZEROS") 

34 if dtype is None: 

35 dtype = torch.get_default_dtype() 

36 if device is None: 

37 device = torch.device(device_.name) 

38 

39 out = torch.empty(size, device=device, dtype=dtype) 

40 N = volume(size) 

41 grid_fn = lambda meta: (triton.cdiv(N, meta["BLOCK_SIZE"]),) 

42 with torch_device_fn.device(device): 

43 zeros_kernel[grid_fn](out, N) 

44 return out