Masaki Kozuki

ref: - https://docs.python.org/3/library/typing.html#typing.Generator - https://docs.python.org/3/library/collections.abc.html#collections.abc.Generator

following #102

It feels like 2ea1933e2a33c81f3b347dea0493e1064ee3734a has some effects on the test case in question. The change of `all` -> `any` is something I'm not quite confident though.

this would occur when `dtypesIfCUDA` or `dtypesIfROCM` are not set: https://github.com/pytorch/pytorch/blob/c3c4465f504282f61f962027adf2bb9ea196b976/torch/testing/_internal/opinfo/core.py#L1342-L1348 while https://github.com/pytorch/pytorch/blob/c3c4465f504282f61f962027adf2bb9ea196b976/torch/testing/_internal/opinfo/core.py#L717-L727 seems to me that `OpInfo`'s post init or dunder init would use and set `dtypes` as `dtypesIfCUDA`...

Convert `ForeachFuncInfo` to `dataclass`

11

- `ForeachFuncInfo` to `dataclass` for smaller diff from `OpInfo` - `skips` to `decorators` and `skip` to `xfail`

Test foreach functions with all dtypes except qints

1

Set `dtypes` and the others to all dtypes except qints, with some required xfails Related to #124726.

- [ ] prepare Dockerfile - [ ] push the build docker image to dockerhub - [ ] write tests - [ ] run `pytest tests` in the docker

Hopefully with - [ ] some tests - [ ] jit option - [ ] amp

features

1

- [ ] SAGAN

[distributed][Tensor Parallelism] Implement early transform for Column-wise Parallel

1

this implements a trace transform that converts one or more linear layers into column-wise tensor parallel ones by (1) sharding their weight and bias and (2) inserting needed communication after...