Coverage for src/flag_gems/ops/addcmul.py: 95%

1import logging 

2 

3import torch 

4import triton 

5 

6from flag_gems.utils import pointwise_dynamic 

7 

8logger = logging.getLogger(__name__) 

9 

10 

11@pointwise_dynamic( 

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

13) 

14@triton.jit 

15def addcmul_forward(x, t1, t2, value): 

16 return x + value * t1 * t2 

17 

18 

19def addcmul_out(inp, tensor1, tensor2, *, value=1.0, out): 

20 logger.debug("GEMS ADDCMUL_OUT") 

21 broadcast_shape = torch.broadcast_shapes(inp.shape, tensor1.shape, tensor2.shape) 

22 if list(out.shape) != list(broadcast_shape): 

23 out.resize_(broadcast_shape) 

24 addcmul_forward(inp, tensor1, tensor2, value, out0=out) 

25 return out 

26 

27 

28def addcmul(inp, tensor1, tensor2, *, value=1.0): 

29 """Functional entry; keep alongside ``addcmul.out`` for dispatch coverage.""" 

30 logger.debug("GEMS ADDCMUL") 

31 broadcast_shape = torch.broadcast_shapes(inp.shape, tensor1.shape, tensor2.shape) 

32 dtype = torch.promote_types( 

33 inp.dtype, torch.promote_types(tensor1.dtype, tensor2.dtype) 

34 ) 

35 out = torch.empty(broadcast_shape, device=inp.device, dtype=dtype) 

36 return addcmul_out(inp, tensor1, tensor2, value=value, out=out)