Add float8 datatype to XPU OPs #1393
Conversation
There was a problem hiding this comment.
kernels related: _local_scalar_dense foreach_tensor_copy cat_xpu where_xpu
@yucai-intel : what this comment tries to say? Can you, please, update PR description highlighting which XPU ops will be available for FP8 datatype after merging this PR? and whether there are any ops which won't be available with FP8 and we should expect over PRs to bring support for them?
| AT_DISPATCH_SWITCH( \ | ||
| TYPE, \ | ||
| NAME, \ | ||
| AT_PRIVATE_CASE_TYPE_USING_HINT( \ |
There was a problem hiding this comment.
Why so strange indent below? If I find definition correctly it expands to the regular case <id>: {}. So each item can be placed just below the previous one.
Did you just copy from https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/cuda/ForeachBinaryOpList.cu#L288? I suggest to fix the indent here for sycl kernel.
#define AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, HINT, ...) \
case enum_type: { \
AT_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type); \
using HINT [[maybe_unused]] = c10::impl::ScalarTypeToCPPTypeT<enum_type>; \
return __VA_ARGS__(); \
}
| }); | ||
|
|
||
| // AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND3( |
There was a problem hiding this comment.
what's the reason to have this commented out? Either remove or add in-source comment explaining why this might be needed.
kernels related:
_local_scalar_dense
foreach_tensor_copy
cat_xpu
where_xpu