torch.nn.functional.scaled_addmm_

torch.nn.functional.scaled_addmm_(input, mat1, mat2, scale_a, scale_recipe_a, scale_b, scale_recipe_b, swizzle_a=None, swizzle_b=None, contraction_dim=(), use_fast_accum=False, *, beta=1.0, alpha=1.0)

In-place version of scaled_addmm().

This function accumulates in input.dtype (float16, bfloat16, or float32), preserves the storage of input, and returns input. A serialized WGRAD loop can create the first contribution with scaled_mm(), then use scaled_addmm_ for later contributions.

Warning

In-place accumulation is not safe for concurrent writers. Callers must serialize writes or provide external coordination.

Return type:

Tensor