wconstab/ltc core hash #3148
wconstab/ltc core hash #3148
Conversation
|
yep mistake. Btw, this isn't ready yet. I was just seeing if I could push. I just got the right bazel version and trying to build locally, but I am sure some things are missing/wrong. |
|
Sounds good, let me know when it is ready 😄 |
wconstab
force-pushed
the
wconstab/ltc_core_hash
branch
from
dda0426 to
ad9e44c
Compare
As part of migration of core lazy tensor functionality to PyTorch core, this uses the newly added torch::lazy::Hash functions from PyTorch. Note: while the Hash* functions are largely identical to the original XLA ones, the underlying uint128 class is from protobuf instead of absl, since it was a slightly smaller dependency to ingest and get building on multiple OS/platform combinations for PyTorch.
wconstab
force-pushed
the
wconstab/ltc_core_hash
branch
from
ad9e44c to
0c8bd01
Compare
|
@JackCaoG OK- this is ready for review now. I am in the process of landing the pytorch/pytorch PR. |
torch_xla/csrc/ops/generic_slice.cpp
Outdated
| [&]() { return NodeOutputShape(input, base_indices, sizes); }, | ||
| /*num_outputs=*/1, xla::util::MHash(base_indices, sizes)), | ||
| /*num_outputs=*/1, | ||
| torch::lazy::MHash(torch::lazy::Hash(base_indices), |
There was a problem hiding this comment.
why couldn't it just be
torch::lazy::MHash(base_indices, sizes);
here?
There was a problem hiding this comment.
yea, ill take another look here. I think I had trouble with the getting torch/core's MHash to work natively with torch_xla/util's Hash(absl::Span) impl.
I might have to define or declare the absl one in a different way so that the torch one can find it?
There was a problem hiding this comment.
I found a solution here, which is to define a new MHash template in torch_util.h which specializes on absl::span.
|
@wconstab Could you also include your pytorch pr link in the pr description? This will make it easier for us to track this pr in the future. |
|
Sure. got reverted already though, some extra windows builds that didn't run on the PR. I'll tag the new pytorch PR here once it's ready. |
- remove from places that only used it for util::Hash
wconstab
force-pushed
the
wconstab/ltc_core_hash
branch
from
622c095 to
b67be8e
Compare
|
I'm landing the pytorch changes again now. I'm not going to rush to merge the xla side in case the pytorch side gets reverted. I'll wait a few days then if stable press the button. Btw what do I do about the .torch_pin file? I guess I can delete it once pytorch/pytorch is merged and then CI will still pass by running against master. |
Yea, you can delete the |
…_data" Looks like these files are getting used by downstream xla (as of pytorch/xla#3148) so we need to include them in our package_data so that it gets included when we build `bdist_wheel` like in our GHA workflows Signed-off-by: Eli Uriegas <eliuriegasfb.com> Differential Revision: [D32622241](https://our.internmc.facebook.com/intern/diff/D32622241) [ghstack-poisoned]
Looks like these files are getting used by downstream xla (as of pytorch/xla#3148) so we need to include them in our package_data so that it gets included when we build `bdist_wheel` like in our GHA workflows Signed-off-by: Eli Uriegas <eliuriegasfb.com> Differential Revision: [D32622241](https://our.internmc.facebook.com/intern/diff/D32622241) [ghstack-poisoned]
…_data" Looks like these files are getting used by downstream xla (as of pytorch/xla#3148) so we need to include them in our package_data so that it gets included when we build `bdist_wheel` like in our GHA workflows Signed-off-by: Eli Uriegas <eliuriegasfb.com> Differential Revision: [D32622241](https://our.internmc.facebook.com/intern/diff/D32622241) [ghstack-poisoned]
Looks like these files are getting used by downstream xla (as of pytorch/xla#3148) so we need to include them in our package_data so that it gets included when we build `bdist_wheel` like in our GHA workflows Signed-off-by: Eli Uriegas <eliuriegasfb.com> Differential Revision: [D32622241](https://our.internmc.facebook.com/intern/diff/D32622241) [ghstack-poisoned]
2 participants
Depends on pytorch PR pytorch/pytorch#66181