Coverage for src/flag_gems/ops/normal.py: 93%

1import logging 

2 

3import torch 

4import triton 

5 

6from flag_gems.ops.randn import randn_kernel 

7from flag_gems.runtime import torch_device_fn 

8from flag_gems.utils import pointwise_dynamic 

9from flag_gems.utils.random_utils import philox_backend_seed_offset 

10from flag_gems.utils.shape_utils import broadcast_shapes, volume 

11 

12logger = logging.getLogger(__name__) 

13UNROLL = 4 

14 

15 

16@pointwise_dynamic( 

17 is_tensor=[True, True, True], promotion_methods=[(0, 1, 2, "DEFAULT")] 

18) 

19@triton.jit 

20def transform_func_tensor_tensor(val, std, mean): 

21 return val * std + mean 

22 

23 

24@pointwise_dynamic( 

25 is_tensor=[True, False, True], promotion_methods=[(0, 1, 2, "DEFAULT")] 

26) 

27@triton.jit 

28def transform_func_tensor_float(val, std, mean): 

29 return val * std + mean 

30 

31 

32@pointwise_dynamic( 

33 is_tensor=[True, True, False], promotion_methods=[(0, 1, 2, "DEFAULT")] 

34) 

35@triton.jit 

36def transform_func_float_tensor(val, std, mean): 

37 return val * std + mean 

38 

39 

40@pointwise_dynamic( 

41 is_tensor=[True, False, False], promotion_methods=[(0, 1, 2, "DEFAULT")] 

42) 

43@triton.jit 

44def transform_func_float_float(val, std, mean): 

45 return val * std + mean 

46 

47 

48def normal_distribution(shape, device, *, generator=None, out=None): 

49 if out is None: 

50 out = torch.empty(shape, device=device, dtype=torch.float32) 

51 N = volume(shape) 

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

53 

54 increment = triton.cdiv(N, UNROLL) 

55 philox_seed, philox_offset = philox_backend_seed_offset( 

56 increment, generator=generator 

57 ) 

58 with torch_device_fn.device(device): 

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

60 return out 

61 

62 

63def normal_tensor_tensor(mean, std, *, generator=None): 

64 logger.debug("GEMS NORMAL_TENSOR_TENSOR") 

65 shape = broadcast_shapes([mean.shape, std.shape]) 

66 device = mean.device 

67 out = normal_distribution(shape, device) 

68 return transform_func_tensor_tensor(out, std, mean) 

69 

70 

71def normal_tensor_float(mean, std, *, generator=None): 

72 logger.debug("GEMS NORMAL_TENSOR_FLOAT") 

73 shape = mean.shape 

74 device = mean.device 

75 out = normal_distribution(shape, device) 

76 return transform_func_tensor_float(out, std, mean) 

77 

78 

79def normal_float_tensor(mean, std, *, generator=None): 

80 logger.debug("GEMS NORMAL_FLOAT_TENSOR") 

81 shape = std.shape 

82 device = std.device 

83 out = normal_distribution(shape, device) 

84 return transform_func_float_tensor(out, std, mean) 

85 

86 

87def normal_(self, mean=0, std=1, *, generator=None): 

88 logger.debug("GEMS NORMAL_") 

89 shape = self.shape 

90 device = self.device 

91 self = normal_distribution(shape, device, generator=None, out=self) 

92 transform_func_float_float(self, std, mean, out0=self) 

93 return self