Upgrade to DLPack 1.0. #145000

ysiraichi · 2025-01-16T21:38:45Z

Stack from ghstack (oldest at bottom):

This PR makes the necessary changes in order to upgrade PyTorch DLPack
support to version 1.0. In summary, we add support for the following:

Support both DLManagedTensor and DLManagedTensorVersioned when
producing and consuming DLPack capsules
New parameter for __dlpack__ method: max_version
Version checks:
- Fallback to old implementation if no max_version or if version
  lower than 1.0
- Check that the to-be-consumed capsule is of version up to 1.X

In order to accommodate these new specifications, this PR adds the
following main changes:

torch._C._to_dlpack_versioned Python API (Module.cpp): new Python
API for creating a versioned DLPack capsule (called by __dlpack__
method)
DLPackTraits<T> class (DLConvertor.h): select the correct
traits (e.g. capsule name, conversion functions) depending on which
DLPack tensor class is being used
toDLPackImpl<T> function (DLConvertor.cpp): populates the
common fields of both classes
fromDLPackImpl<T> function (DLConvertor.cpp): constructs a tensor
from a DLPAck capsule
fillVersion<T> function (DLConvertor.cpp): populates the version
field for DLManagedTensorVersioned (no-op for DLManagedTensor)
tensor_fromDLPackImpl<T> function (tensor_new.cpp): outer function
for constructing a tensor out of a DLPack capsule that also marks the
capsule as used

[ghstack-poisoned]

pytorch-bot · 2025-01-16T21:38:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/145000

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 2f1782e with merge base 85bfaf8 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

This PR makes the necessary changes in order to upgrade PyTorch DLPack support to version 1.0. In summary, we add support for the following: - Support both `DLManagedTensor` and `DLManagedTensorVersioned` when producing and consuming DLPack capsules - New parameter for `__dlpack__` method: `max_version` - Version checks: - Fallback to old implementation if no `max_version` or if version lower than 1.0 - Check that the to-be-consumed capsule is of version up to 1.X In order to accommodate these new specifications, this PR adds the following main changes: - `torch._C._to_dlpack_versioned` Python API (Module.cpp): new Python API for creating a versioned DLPack capsule (called by `__dlpack__` method) - `DLPackTraits<T>` class (DLConvertor.h): select the correct capsule name depending on which DLPack tensor class is being used - `toDLPackImpl<T>` function (DLConvertor.cpp): populates the common fields of both classes - `fillVersion<T>` function (DLConvertor.cpp): populates the version field for `DLManagedTensorVersioned` (no-op for `DLManagedTensor`) - `fromDLPackImpl<T>` function (tensor_new.cpp): common function for creating an `at::Tensor` for both classes, leaving the possible version check for its caller ghstack-source-id: 3ca1169 Pull Request resolved: #145000

[ghstack-poisoned]

This PR makes the necessary changes in order to upgrade PyTorch DLPack support to version 1.0. In summary, we add support for the following: - Support both `DLManagedTensor` and `DLManagedTensorVersioned` when producing and consuming DLPack capsules - New parameter for `__dlpack__` method: `max_version` - Version checks: - Fallback to old implementation if no `max_version` or if version lower than 1.0 - Check that the to-be-consumed capsule is of version up to 1.X In order to accommodate these new specifications, this PR adds the following main changes: - `torch._C._to_dlpack_versioned` Python API (Module.cpp): new Python API for creating a versioned DLPack capsule (called by `__dlpack__` method) - `DLPackTraits<T>` class (DLConvertor.h): select the correct capsule name depending on which DLPack tensor class is being used - `toDLPackImpl<T>` function (DLConvertor.cpp): populates the common fields of both classes - `fromDLPackImpl<T>` function (DLConvertor.cpp): constructs a tensor from a DLPAck capsule - `fillVersion<T>` function (DLConvertor.cpp): populates the version field for `DLManagedTensorVersioned` (no-op for `DLManagedTensor`) - `tensor_fromDLPackImpl<T>` function (tensor_new.cpp): outer function for constructing a tensor out of a DLPack capsule that also marks the capsule as used ghstack-source-id: e2d39d9 Pull Request resolved: #145000

[ghstack-poisoned]

This PR makes the necessary changes in order to upgrade PyTorch DLPack support to version 1.0. In summary, we add support for the following: - Support both `DLManagedTensor` and `DLManagedTensorVersioned` when producing and consuming DLPack capsules - New parameter for `__dlpack__` method: `max_version` - Version checks: - Fallback to old implementation if no `max_version` or if version lower than 1.0 - Check that the to-be-consumed capsule is of version up to 1.X In order to accommodate these new specifications, this PR adds the following main changes: - `torch._C._to_dlpack_versioned` Python API (Module.cpp): new Python API for creating a versioned DLPack capsule (called by `__dlpack__` method) - `DLPackTraits<T>` class (DLConvertor.h): select the correct capsule name depending on which DLPack tensor class is being used - `toDLPackImpl<T>` function (DLConvertor.cpp): populates the common fields of both classes - `fromDLPackImpl<T>` function (DLConvertor.cpp): constructs a tensor from a DLPAck capsule - `fillVersion<T>` function (DLConvertor.cpp): populates the version field for `DLManagedTensorVersioned` (no-op for `DLManagedTensor`) - `tensor_fromDLPackImpl<T>` function (tensor_new.cpp): outer function for constructing a tensor out of a DLPack capsule that also marks the capsule as used ghstack-source-id: e58ba67 Pull Request resolved: #145000

[ghstack-poisoned]

This PR makes the necessary changes in order to upgrade PyTorch DLPack support to version 1.0. In summary, we add support for the following: - Support both `DLManagedTensor` and `DLManagedTensorVersioned` when producing and consuming DLPack capsules - New parameter for `__dlpack__` method: `max_version` - Version checks: - Fallback to old implementation if no `max_version` or if version lower than 1.0 - Check that the to-be-consumed capsule is of version up to 1.X In order to accommodate these new specifications, this PR adds the following main changes: - `torch._C._to_dlpack_versioned` Python API (Module.cpp): new Python API for creating a versioned DLPack capsule (called by `__dlpack__` method) - `DLPackTraits<T>` class (DLConvertor.h): select the correct capsule name depending on which DLPack tensor class is being used - `toDLPackImpl<T>` function (DLConvertor.cpp): populates the common fields of both classes - `fromDLPackImpl<T>` function (DLConvertor.cpp): constructs a tensor from a DLPAck capsule - `fillVersion<T>` function (DLConvertor.cpp): populates the version field for `DLManagedTensorVersioned` (no-op for `DLManagedTensor`) - `tensor_fromDLPackImpl<T>` function (tensor_new.cpp): outer function for constructing a tensor out of a DLPack capsule that also marks the capsule as used ghstack-source-id: 1f9150c Pull Request resolved: #145000

[ghstack-poisoned]

This PR makes the necessary changes in order to upgrade PyTorch DLPack support to version 1.0. In summary, we add support for the following: - Support both `DLManagedTensor` and `DLManagedTensorVersioned` when producing and consuming DLPack capsules - New parameter for `__dlpack__` method: `max_version` - Version checks: - Fallback to old implementation if no `max_version` or if version lower than 1.0 - Check that the to-be-consumed capsule is of version up to 1.X In order to accommodate these new specifications, this PR adds the following main changes: - `torch._C._to_dlpack_versioned` Python API (Module.cpp): new Python API for creating a versioned DLPack capsule (called by `__dlpack__` method) - `DLPackTraits<T>` class (DLConvertor.h): select the correct capsule name depending on which DLPack tensor class is being used - `toDLPackImpl<T>` function (DLConvertor.cpp): populates the common fields of both classes - `fromDLPackImpl<T>` function (DLConvertor.cpp): constructs a tensor from a DLPAck capsule - `fillVersion<T>` function (DLConvertor.cpp): populates the version field for `DLManagedTensorVersioned` (no-op for `DLManagedTensor`) - `tensor_fromDLPackImpl<T>` function (tensor_new.cpp): outer function for constructing a tensor out of a DLPack capsule that also marks the capsule as used ghstack-source-id: 063107b Pull Request resolved: #145000

albanD

Change sounds ok but what is the expectations in terms of BC when interracting with libraries that didn't upgrade to latest dlpack yet? Is that ok for all Tensors to be of the new version?

albanD · 2025-02-03T19:32:11Z

aten/src/ATen/DLConvertor.h

-TORCH_API Tensor fromDLPack(DLManagedTensor* src);
-TORCH_API Tensor
-fromDLPack(DLManagedTensor* src, std::function<void(void*)> deleter);
+TORCH_API DLManagedTensorVersioned* toDLPack(const Tensor& src);


Is this API used by C++ libraries? This would be a BC-breaking change for these users right?

I think it's being used by PyTorch/XLA. Yes, it's definitely BC-breaking. I was thinking that, since DLPack 1.0 should be the new default, the old version should have the name with a suffix. However, now that you brought this up, not being BC-breaking sounds more important.

In summary, I will change the names so that we are not BC-breaking.

albanD · 2025-02-03T19:32:39Z

aten/src/ATen/DLConvertor.h

+TORCH_API DLManagedTensor* toDLPackUnversioned(const Tensor& src);
+TORCH_API Tensor fromDLPack(
+    DLManagedTensorVersioned* src,
+    std::optional<std::function<void(void*)>> deleter = std::nullopt);


std::function is optional by default and doesn't require std::optional right?

aten/src/ATen/dlpack.h

albanD · 2025-02-03T19:34:21Z

torch/__init__.py

@@ -2270,7 +2270,7 @@ def compiled_with_cxx11_abi() -> builtins.bool:
    matrix_rank,
    solve,
 )
-from torch.utils.dlpack import from_dlpack, to_dlpack
+from torch.utils.dlpack import from_dlpack, to_dlpack, to_dlpack_unversioned


While I understand this matches current code, I don't think we want to have this as a new top level API? Being part of torch.utils.dlpack is enough?

That comment sounds right - I don't think it should be needed. For the introduction strategy to a versioned protocol, see the explanation and prototype code (if max_version is None: ....) at https://data-apis.org/array-api/latest/API_specification/generated/array_api.array.__dlpack__.html

If it's useful/needed to have a Python function to use here so that can be called from within Tensor.__dlpack__, then making it a private function by prepending an underscore should be fine I think.

albanD · 2025-02-03T19:35:59Z

torch/csrc/Module.cpp

@@ -1599,6 +1616,7 @@ static std::initializer_list<PyMethodDef> TorchMethods = {
     METH_NOARGS,
     nullptr},
    {"_to_dlpack", THPModule_toDLPack, METH_O, nullptr},
+    {"_to_dlpack_unversioned", THPModule_toDLPackUnversioned, METH_O, nullptr},


Why do we still need this one?

This does look necessary to me. For DLPack to change the ABI once, there's a dance that needs doing to be not-super-disruptive: continue returning the old (0.8) version, unless the consumer indicates it can handle the new (1.X) version by passing in max_version=(1, 0) (or (1, x) in the future).

albanD · 2025-02-03T19:37:02Z

torch/utils/dlpack.py

+        try:
+            # Try running __dlpack__ while specifying `max_version` argument.
+            dlpack = ext_tensor.__dlpack__(**kwargs)
+        except TypeError:


Is it guaranteed that they will fail with TypeError?

That should be the case - it's the standard error in Python for using a non-existing keyword.

rgommers

Thanks for tackling this @ysiraichi! This looks like a good start. A few high-level comments:

As noted inline, __dlpack__ gained 3 new keywords, so far this adds only max_version. I suspect it's safer to add all at once, because other libraries are probably going to assume that if max_version is present, the 1.0 support is complete.
It would be good to have new Python-level tests in test/test_dlpack.py. That will also make the changes in logic easier to review.
- There's one testing TODO for a very old numpy version there that may be nice to take along:
  
  pytorch/test/test_dlpack.py
  
  Line 239 in e9f6e27
  
  # TODO: add interchange tests once NumPy 1.22 (dlpack support) is required
I think this is inactionable right now, but adding for completeness: DLPack gained a new DLPACK_FLAG_BITMASK_READ_ONLY field, which in the future can feed into PyTorch's copy-on-write (COW) feature.

aten/src/ATen/dlpack.h

rgommers · 2025-02-04T15:46:01Z

torch/utils/dlpack.py

+        try:
+            # Try running __dlpack__ while specifying `max_version` argument.
+            dlpack = ext_tensor.__dlpack__(**kwargs)
+        except TypeError:


That should be the case - it's the standard error in Python for using a non-existing keyword.

rgommers · 2025-02-04T15:50:51Z

torch/__init__.py

@@ -2270,7 +2270,7 @@ def compiled_with_cxx11_abi() -> builtins.bool:
    matrix_rank,
    solve,
 )
-from torch.utils.dlpack import from_dlpack, to_dlpack
+from torch.utils.dlpack import from_dlpack, to_dlpack, to_dlpack_unversioned


That comment sounds right - I don't think it should be needed. For the introduction strategy to a versioned protocol, see the explanation and prototype code (if max_version is None: ....) at https://data-apis.org/array-api/latest/API_specification/generated/array_api.array.__dlpack__.html

rgommers · 2025-02-04T15:52:01Z

torch/_tensor.py

+
+            max_version (tuple[int, int] or None): An optional Python tuple with
+            2 integers, representing the maximum version the caller supports. If
+            None is passed, then PyTorch will fallback to DLPack 0.X, where versions


I'd rephrase as "If None (default), PyTorch will use DLPack 0.8".

rgommers · 2025-02-04T15:55:19Z

torch/__init__.py

@@ -2270,7 +2270,7 @@ def compiled_with_cxx11_abi() -> builtins.bool:
    matrix_rank,
    solve,
 )
-              from torch.utils.dlpack import from_dlpack, to_dlpack

            
+from torch.utils.dlpack import from_dlpack, to_dlpack, to_dlpack_unversioned


If it's useful/needed to have a Python function to use here so that can be called from within Tensor.__dlpack__, then making it a private function by prepending an underscore should be fine I think.

rgommers · 2025-02-04T16:01:30Z

torch/_tensor.py

@@ -1661,7 +1661,7 @@ def __torch_function__(cls, func, types, args=(), kwargs=None):

    __torch_dispatch__ = _C._disabled_torch_dispatch_impl

-    def __dlpack__(self, stream=None):
+    def __dlpack__(self, stream=None, max_version=None):


Note that there are also new dl_device and copy keywords (API docs).

rgommers · 2025-02-04T16:05:23Z

torch/csrc/Module.cpp

@@ -1599,6 +1616,7 @@ static std::initializer_list<PyMethodDef> TorchMethods = {
     METH_NOARGS,
     nullptr},
    {"_to_dlpack", THPModule_toDLPack, METH_O, nullptr},
+    {"_to_dlpack_unversioned", THPModule_toDLPackUnversioned, METH_O, nullptr},


This does look necessary to me. For DLPack to change the ABI once, there's a dance that needs doing to be not-super-disruptive: continue returning the old (0.8) version, unless the consumer indicates it can handle the new (1.X) version by passing in max_version=(1, 0) (or (1, x) in the future).

ysiraichi · 2025-02-07T14:16:05Z

@albanD @rgommers Since you have already reviewed this PR, I was thinking about adding the new __dlpack__ keywords + tests in a new PR stacked on top of this one. There are 2 reasons for that:

It would be easier to review the new changes
We could merge the whole stack at once when that PR is ready

Let me know if that works for you.

rgommers · 2025-02-07T15:14:54Z

@ysiraichi a stack with two separate PRs, with the dl_device and copy keywords in the second PR, sounds perfectly fine to me. Merging the changes in one go will be nice, but functionally those two keywords don't interact with max_version, so it'll be a clean split.

albanD · 2025-02-07T15:46:12Z

Separate PR on the stack sounds ok.
Note that you won't be able to land both together though if the bottom PR's CI is red though, they must all pass CI

[ghstack-poisoned]

This PR makes the necessary changes in order to upgrade PyTorch DLPack support to version 1.0. In summary, we add support for the following: - Support both `DLManagedTensor` and `DLManagedTensorVersioned` when producing and consuming DLPack capsules - New parameter for `__dlpack__` method: `max_version` - Version checks: - Fallback to old implementation if no `max_version` or if version lower than 1.0 - Check that the to-be-consumed capsule is of version up to 1.X In order to accommodate these new specifications, this PR adds the following main changes: - `torch._C._to_dlpack_versioned` Python API (Module.cpp): new Python API for creating a versioned DLPack capsule (called by `__dlpack__` method) - `DLPackTraits<T>` class (DLConvertor.h): select the correct capsule name depending on which DLPack tensor class is being used - `toDLPackImpl<T>` function (DLConvertor.cpp): populates the common fields of both classes - `fromDLPackImpl<T>` function (DLConvertor.cpp): constructs a tensor from a DLPAck capsule - `fillVersion<T>` function (DLConvertor.cpp): populates the version field for `DLManagedTensorVersioned` (no-op for `DLManagedTensor`) - `tensor_fromDLPackImpl<T>` function (tensor_new.cpp): outer function for constructing a tensor out of a DLPack capsule that also marks the capsule as used ghstack-source-id: 2b774ce Pull Request resolved: #145000

leofang · 2025-03-11T16:50:00Z

Gentle ping @ysiraichi, any chance we can get this work wrapped up in the near future? Note that we just tagged DLPack v1.1: https://github.com/dmlc/dlpack/releases/tag/v1.1, which added a few more dtype enums.

ysiraichi · 2025-03-11T17:03:54Z

Yeah. I'm still working on this. Have 2 more PRs to be opened.
Haven't had the time to get back to them, though. Will do so soon.

[ghstack-poisoned]

ysiraichi · 2025-04-04T18:16:49Z

Here is a summary of the latest updates after the last reviews:

Renamed the DLPack API, removing the Unversioned suffix
Renamed the DLPack API, adding Versioned to the newly added functions
Removed the import of to_dlpack_unversioned
- Using torch._C._to_dlpack() and torch._C._to_dlpack_versioned() directly
Added test_max_version.py, which tests both the versioned and unversioned execution paths

@albanD @rgommers
Sorry for the delay on this PR.
Let me know what you think of this new set of changes.

[ghstack-poisoned]

ysiraichi · 2025-04-25T23:35:41Z

Hi, @albanD @rgommers. Sorry for taking too much time on this. I believe this stack should be good for another round of reviews. Let me know what you think.

[ghstack-poisoned]

msaroufim · 2025-05-14T05:28:10Z

torch/csrc/utils/tensor_new.cpp

+
+template <class T>
+at::Tensor tensor_fromDLPackImpl(PyObject* data, T* tensor) {
+  // HACK: Ensure that we hold the GIL here just in case the


n00b q: what's the context here?

torch/utils/dlpack.py

msaroufim · 2025-05-14T05:51:57Z

torch/_tensor.py

@@ -1690,9 +1690,14 @@ def __dlpack__(self, stream=None):
            both streams.  If None or -1 is passed then no synchronization is performed.
            If 1 (on CUDA) or 0 (on ROCM) then the default stream is used for
            synchronization.
+
+            max_version (tuple[int, int] or None): An optional Python tuple with


whats the tradeoff with being conservative on the default value?

I guess we follow the default from https://data-apis.org/array-api/latest/API_specification/generated/array_api.array.__dlpack__.html ?

msaroufim · 2025-05-14T05:54:25Z

torch/utils/dlpack.py

 class DLDeviceType(enum.IntEnum):
    # Enums as in DLPack specification (aten/src/ATen/dlpack.h)
    kDLCPU = 1,
-    kDLGPU = 2,
-    kDLCPUPinned = 3,


why change the name?

msaroufim · 2025-05-14T05:58:10Z

Will defer to @albanD for merge but this seems like a fairly safe PR to land considering BC guarantees are there and despite an upgrade it doesn't change the way dlpack works in pytorch all that much and if anyhing adds more layers of tests in every new PR

albanD

Sounds good!
Mostly small comments.
The main thing would be to add a BC-breaking note on exactly which user facing APIs are changed.

albanD · 2025-05-14T20:46:19Z

aten/src/ATen/dlpack.h

+ *
+ * \note This is the current standard DLPack exchange data structure.
+ */
+struct DLManagedTensorVersioned {


I assume this struct is defined by the dlpack standard?
I'm curious why we define all of these by hand vs using a standard include from dlpack? I guess it doesn't exist?

albanD · 2025-05-14T20:47:54Z

test/test_dlpack.py

+        def test(device, **kwargs):
+            inp = make_tensor((5,), dtype=torch.float32, device=device)
+            out = torch.from_dlpack(inp.__dlpack__(**kwargs))
+            self.assertEqual(inp, out)


Can we query the version field on the dlpack Tensor to ensure we get the object we want?

albanD · 2025-05-14T20:48:35Z

torch/_tensor.py

@@ -1673,7 +1673,7 @@ def __torch_function__(cls, func, types, args=(), kwargs=None):

    __torch_dispatch__ = _C._disabled_torch_dispatch_impl

-    def __dlpack__(self, stream=None):
+    def __dlpack__(self, stream=None, max_version=None):


These are kwargs only as well.
This will help with adding more kwargs later without any issue

albanD · 2025-05-14T20:50:12Z

torch/_tensor.py

@@ -1690,9 +1690,14 @@ def __dlpack__(self, stream=None):
            both streams.  If None or -1 is passed then no synchronization is performed.
            If 1 (on CUDA) or 0 (on ROCM) then the default stream is used for
            synchronization.
+
+            max_version (tuple[int, int] or None): An optional Python tuple with


I guess we follow the default from https://data-apis.org/array-api/latest/API_specification/generated/array_api.array.__dlpack__.html ?

albanD · 2025-05-14T20:50:46Z

torch/_tensor.py

        """
        if has_torch_function_unary(self):
-            return handle_torch_function(Tensor.__dlpack__, (self,), self, stream)
+            args = (self, stream, max_version)


Move to kwargs

albanD · 2025-05-14T20:51:59Z

torch/_tensor.py

        """
        if has_torch_function_unary(self):
-            return handle_torch_function(Tensor.__dlpack__, (self,), self, stream)
+            args = (self, stream, max_version)
+            return handle_torch_function(Tensor.__dlpack__, (self,), *args)


bug below not from this PR: we should check both is_conj() and is_neg() as they will fail the same way!

albanD · 2025-05-14T21:00:13Z

torch/utils/dlpack.py

 class DLDeviceType(enum.IntEnum):
    # Enums as in DLPack specification (aten/src/ATen/dlpack.h)
    kDLCPU = 1,
-    kDLGPU = 2,
-    kDLCPUPinned = 3,
+    kDLCUDA = 2,


These are BC-breaking right?

albanD · 2025-05-14T21:01:20Z

torch/utils/dlpack.py

+        try:
+            # Try running __dlpack__ while specifying `max_version` argument.
+            dlpack = ext_tensor.__dlpack__(**kwargs)
+        except TypeError:


According to the standard, it says BufferError should be raised if you don't get a good max_version. Should we handle that here?

Update

b74e768

[ghstack-poisoned]

pytorchbot added the open source label Jan 16, 2025

Update

69b71c1

[ghstack-poisoned]

ysiraichi added module: dlpack release notes: python_frontend python frontend release notes category labels Feb 1, 2025

Update

4c8f732

[ghstack-poisoned]

Update

b3728b4

[ghstack-poisoned]

ysiraichi requested review from rgommers and albanD February 2, 2025 16:12

Update

04e69ec

[ghstack-poisoned]

albanD reviewed Feb 3, 2025

View reviewed changes

rgommers reviewed Feb 4, 2025

View reviewed changes

rgommers mentioned this pull request Feb 4, 2025

Tag an official v1.0 release dmlc/dlpack#149

Open

Update

0104701

[ghstack-poisoned]

Update

6b5e62e

[ghstack-poisoned]

This was referenced Mar 28, 2025

[DLPack] add NumPy exchange tests. #150216

Open

Fix DLPack stream logic. #150217

Open

[DLPack] Add support for missing keyword-arguments. #150218

Open

ysiraichi added 4 commits March 28, 2025 17:33

Rebased.

d899557

[ghstack-poisoned]

Update

8bf0ee2

[ghstack-poisoned]

Update

69ad77c

[ghstack-poisoned]

Update

8231c62

[ghstack-poisoned]

ysiraichi mentioned this pull request Apr 4, 2025

Raise BufferError for DLPack buffer-related errors. #150691

Open

Update

37daf8a

[ghstack-poisoned]

Update

2f1782e

[ghstack-poisoned]

msaroufim reviewed May 14, 2025

View reviewed changes

torch/utils/dlpack.py Show resolved Hide resolved

msaroufim reviewed May 14, 2025

View reviewed changes

albanD reviewed May 14, 2025

View reviewed changes

ogrisel mentioned this pull request May 16, 2025

Clarification of output array type when metrics accept multiclass/multioutput scikit-learn/scikit-learn#31286

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Upgrade to DLPack 1.0. #145000

Upgrade to DLPack 1.0. #145000

Upgrade to DLPack 1.0. #145000

Are you sure you want to change the base?

Upgrade to DLPack 1.0. #145000

Conversation

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/145000

✅ No Failures

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment