Summary
Groundwork PR
- Adds the hop
- Adds the quack impl
- Only supports pointwise
- no aux buffers
- All this to make follow up prs easier to review
sanity check
Notes for reviewer
API;
# Do be made public later
from torch._higher_order_ops import flex_gemm
out = flex_gemm(
gemm_op, # torch.mm / torch.addmm / aten op overload expanded to a wider set scaled_mm, grouped friends
gemm_args, # tuple of tensor/scalar operands orignal args ot he base func
epilogue_fn, # Python fn over accumulator/result
*,
gemm_kwargs=None, # non-tensor kwargs only
kernel_options={}, # dict of options, will be backend, fastmath, tune
)Pull Request resolved: #186483
Approved by: https://github.com/mlazos, https://github.com/eellison
ghstack dependencies: #186944
SOCIAL SHARE CARD GENERATOR