OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-07 12:21:27 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity
71,282 Commits 8,304 Branches 150 Tags 1.5 GiB
c77352b5cc
Commit Graph

9 Commits

Author SHA1 Message Date
ydwu4
c77352b5cc Add torch._library.register_fake_class to fakify torchBind class (#122622) 
This PR only adds abstract class registration logic without touching existing tests so they still trace with real script object. The added tests are only for registration APIs and test error messages.

Our design is that the abstract implementation should be in Python. This is much better in terms of usability. But this also has implications for custom op that takes script object as input, which is detailed later in this stack.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/122622
Approved by: https://github.com/zou3519
ghstack dependencies: #122619, #122620, #122621
 2024-04-02 23:52:17 +00:00
ydwu4
46c7235406 add tensor queue example (#122621) 
This PR adds a tensor queue example for later use. It doesn't touch any existing logic. It refactors the tests a little bit to avoid importing the library in unittest setUp.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/122621
Approved by: https://github.com/zou3519
ghstack dependencies: #122619, #122620
 2024-04-02 23:52:17 +00:00
ydwu4
5d6a447357 [torchbind] change to parametrized tests for pre_dispatch (#122620) 
Refactor the tests to make the test more robust.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/122620
Approved by: https://github.com/zou3519
ghstack dependencies: #122619
 2024-04-02 23:52:14 +00:00
Tugsbayasgalan (Tugsuu) Manlaibaatar
443e241cc5 Don't cache predispatch kernels (#121712) 
Summary: Title

Test Plan: CI

Differential Revision: D54791087

Pull Request resolved: https://github.com/pytorch/pytorch/pull/121712
Approved by: https://github.com/ydwu4
 2024-03-12 18:05:59 +00:00
angelayi
e8836759d0 [export] Add effect token to export (#121424) 
Following the creation of effect tokens (https://github.com/pytorch/pytorch/pull/120296), we want to now add support for these tokens in export because the calling/returning convention has changed. The inputs are now `(tokens, params, buffers, constants, user_inputs)` and the outputs are `(tokens, buffer_mutations, user_mutations, user_outputs)`. The graph looks something like:
```
graph():
    %arg0_1 : [num_users=1] = placeholder[target=arg0_1]
    %attr : [num_users=2] = placeholder[target=attr]
    %arg1_1 : [num_users=2] = placeholder[target=arg1_1]
    %with_effects : [num_users=2] = call_function[target=torch._higher_order_ops.effects.with_effects](args = (%arg0_1, _TorchScriptTesting.takes_foo.default, %attr, %arg1_1), kwargs = {})
    %getitem : [num_users=1] = call_function[target=operator.getitem](args = (%with_effects, 0), kwargs = {})
    %getitem_1 : [num_users=1] = call_function[target=operator.getitem](args = (%with_effects, 1), kwargs = {})
    %with_effects_1 : [num_users=2] = call_function[target=torch._higher_order_ops.effects.with_effects](args = (%getitem, _TorchScriptTesting.takes_foo.default, %attr, %getitem_1), kwargs = {})
    %getitem_2 : [num_users=1] = call_function[target=operator.getitem](args = (%with_effects_1, 0), kwargs = {})
    %getitem_3 : [num_users=1] = call_function[target=operator.getitem](args = (%with_effects_1, 1), kwargs = {})
    %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%arg1_1, %getitem_3), kwargs = {})
    return (getitem_2, add)
```

During unlifting, we will first remove the tokens and with_effect calls using the `remove_effect_tokens` pass. (cc @SherlockNoMad on the pass to remove tokens). This is so that this won't change the calling conventions when retracing. The graph after unlifting looks something like:
```
graph():
    %attr_1 : [num_users=2] = get_attr[target=attr]
    %arg1_1 : [num_users=2] = placeholder[target=arg1_1]
    %takes_foo_default_1 : [num_users=1] = call_function[target=torch.ops._TorchScriptTesting.takes_foo.default](args = (%attr_1, %arg1_1), kwargs = {})
    %takes_foo_default : [num_users=1] = call_function[target=torch.ops._TorchScriptTesting.takes_foo.default](args = (%attr_1, %takes_foo_default_1), kwargs = {})
    %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%arg1_1, %takes_foo_default), kwargs = {})
    return (add,)
```

Serialization support will be added in a followup.
Note: tokens only affect custom ops that take in ScriptObjects, not ScriptObject methods yet.

Differential Revision: [D54639390](https://our.internmc.facebook.com/intern/diff/D54639390)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121424
Approved by: https://github.com/tugsbayasgalan
 2024-03-09 02:43:26 +00:00
Angela Yi
413a434846 [export] Convert all export tests to .module() (#118425) 
Test Plan: CI

Differential Revision: D53075379

Pull Request resolved: https://github.com/pytorch/pytorch/pull/118425
Approved by: https://github.com/suo
 2024-01-29 23:06:54 +00:00
suo
d84173c025 [export] fix unlifting of custom class constants (#117979) 
we didn't have a test covering this case, add one.

Aside: we should invest in actually unit testing the lifting/unlifting passes, both separately and also against each other. I have a diff cooking for that.

Differential Revision: [D52962180](https://our.internmc.facebook.com/intern/diff/D52962180/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117979
Approved by: https://github.com/avikchaudhuri
ghstack dependencies: #115222, #117978
 2024-01-23 05:51:00 +00:00
suo
2ae66ddba0 [export] fix test ownership (#117886) 
as title

Differential Revision: [D52924188](https://our.internmc.facebook.com/intern/diff/D52924188/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117886
Approved by: https://github.com/ydwu4
 2024-01-21 01:18:16 +00:00
suo
4057d005ff Initial torchbind support in PT2 (#117697) 
This PR adds the bare minimum functionality to get torchbind working in an e2e testable way on PT2.

It implements:
* ProxyTensor support
* Simple torch.export support (proxytensor-only path, e.g. non-strict).
* add some tests exercising the path.

Because all this is not fully baked, I hide the functionality behind a feature flag (`enable_torchbind_tracing()`) so it does not affect regular users for now.

Still on the agenda:
* Dynamo support
* Actual FakeMode support
* Mutability support

Hoping to get this first bit in as a standalone, as it will unblock some more extensive experimentation/testing going on internally.

Differential Revision: [D51825372](https://our.internmc.facebook.com/intern/diff/D51825372/)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/117697
Approved by: https://github.com/SherlockNoMad
 2024-01-19 06:28:20 +00:00