-
Notifications
You must be signed in to change notification settings - Fork 21.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[DTensor] implement dist_cat as a sharding prop rule #92677
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/92677
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 3ea6b8c: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
ghstack-source-id: a27c5a3dffcb7b29dc8d22de45013a199f2ffb00 Pull Request resolved: #92677
[ghstack-poisoned]
ghstack-source-id: d4cbf117b4e630d80fe39229251d73724fae55cc Pull Request resolved: #92677
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
See comments inlined, we should make sure the xfail
of cat to be removed so that it passes all the possible cases.
return output_sharding | ||
|
||
|
||
def _update_schema_suggestion_for_cat( |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
can you tell me what exactly this function is doing? it looks like a lot of duplicate logic with the rule itself and I am not quite sure what this function is used for.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
einop_rule
expects the op_schema
argument to have its args_schema
in form [DTensorSpec, DTensorSpec, ...]
but when it's passed into cat_rule
the schema is actually [List[DTensorSpec]]
. That's why I convert the args_schema
at the beginning of cat_rule
(https://github.com/pytorch/pytorch/pull/92677/files#diff-ebc7be1151cf411ce7edf46c4ca1cabb74cd953a2bdf47e04b4cc733c31f6085R492) before feeding it into einop_rule
. Thus, we need to convert it back if a schema_suggestion
is present here.
[ghstack-poisoned]
ghstack-source-id: 8ec685e5998e48321a16a752d2b1a7c5a6c84ed4 Pull Request resolved: #92677
[ghstack-poisoned]
ghstack-source-id: f536f348cc4ed8e049eb9bdf1462415c18a89839 Pull Request resolved: #92677
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
lgtm, thanks for working on it! left a couple of suggestions and some question.
dim_word = free_dim[:dim] + alphabet[i] + free_dim[dim:] | ||
einop_notation_list.append(dim_word) | ||
else: | ||
einop_notation_list.append(alphabet[i]) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
is this the empty tensor annotation where it have a single char?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Not entirely for empty tensor but empty tensor whose ndim
is smaller than other tensors. This is for case like concatenating Tensor([], shape=torch.Size([0]))
with Tensor([[1, 2], [3, 4]], shape=torch.Size([2, 2]))
.
In this case, an empty annotation may still work but we want to ensure that the dim char for cat_dim
in output tensor annotation must appear in input as well. Adding each input tensor's cat_dim
dim char into annotation guarantees that.
[ghstack-poisoned]
ghstack-source-id: 4c2f6291dbcdad2b1674b3db36dc46a21a9159ce Pull Request resolved: #92677
@pytorchmergebot merge -g |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Stack from ghstack (oldest at bottom):