Coverage for src/flag_gems/ops/addcmul.py: 67%

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(inp, tensor1, tensor2, *, value=1.0, out=None): 

20 logger.debug("GEMS ADDCMUL FORWARD") 

21 if out is not None: 

22 broadcast_shape = torch.broadcast_shapes( 

23 inp.shape, tensor1.shape, tensor2.shape 

24 ) 

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

26 out.resize_(broadcast_shape) 

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

28 return out 

29 else: 

30 return addcmul_forward(inp, tensor1, tensor2, value)