ensure 'torch' CUDA wheels are installed in CI, test that 'torch' is an optional dependency by jameslamb · Pull Request #425 · rapidsai/cugraph-gnn

jameslamb · 2026-03-09T14:19:26Z

Contributes to rapidsai/build-planning#256 and #410

Ensures that `torch` CUDA wheels are always installed in CI

"unintentionally installed a CPU-only torch from PyPI" is a failure mode I've seen a few times for CI in this project, and each time it's happened it has take a bit of time to figure out that that was the root cause.

This PR tries to fix that by:

using a stricter install pattern to guarantee a CUDA torch wheel is installed if a compatible one exists
using local versions like +cu130 to help prevent pulling in packages from PyPI
adding dependencies.yaml items for "oldest" dependencies so CI covers a range of supported versions

I tested similar patterns in rapidsai/rmm#2279 and saw them work well there.

Makes `torch` truly optional

We want these packages to be installable and importable without torch, for use in RAPIDS DLFW builds (where we don't install torch alongside RAPIDS because it's build in other processes).

I'd started relying on the assumption that they worked that way in this PR, but quickly realized that that isn't true... torch is used unconditionally in many ways in these libraries.

This PR fixes that. It makes torch optional and adds testing to ensure it stays that way:

copying the import_optional machinery from cugraph-pyg into pylibwholegraph
using import_optional() for all torch imports in the libraries
using pytest.importorskip("torch") in tests
deferring some torch inputs from import time to run time
adding the flake8-tidy-imports:banned-api check from ruff to enforce that import torch isn't used anywhere in library code or test code
explicitly testing that the libraries are still installable and at least 1 unit test can run successfully after a pip uninstall torch
adding a check in ci/validate_wheel.sh confirming that torch doesn't make it into any wheel metadata (which could happen due to mistakes in dependencies.yaml / pyproject.toml)

Notes for Reviewers

Pulling these changes out of #413 , so CI in this repo can immediately benefit from them and so #419 can be reverted.

When this is merged, #413 will have a smaller diff and just be focused on testing against a range of CTKs.

copy-pr-bot · 2026-03-09T14:19:30Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

jameslamb · 2026-03-09T14:20:57Z

dependencies.yaml

+          - matrix:
+            packages:
+              - ogb
+              - sentence-transformers


Moving all these (torchdata, pydantic, ogb, sentence-transformers) so that depends_on_pytorch can be limited to only PyTorch packages. That makes its use in wheel installs a bit easier (we can pip download --no-deps the output of the list and know the only package will be torch).

Contributes to rapidsai/build-planning#256, by splitting the conda-only parts of #425 out to hoepfully simplify it. * limits `depends_on_pytorch` in `dependencies.yaml` to only `torch` / `pytorch` - *(this will be helpful in #425 to generate a requirement for `pip download --no-deps`)* * declares `pylibwholegraph`'s dependency on `packaging` * adds some `.gitignore` rules for Python package file types (was creating these in local testing) Authors: - James Lamb (https://github.com/jameslamb) Approvers: - Bradley Dice (https://github.com/bdice) URL: #427

…#429) Splitting some changes off of #425 * re-enable arm64 conda tests * adding a small test with `torch.cuda.is_available()` in conda CI tests * updates `ci/test_*` scripts to re-use `ci/run_*` scripts ## Notes for Reviewers ### How I tested this Ran the full PR and nightly matrices of `conda-python-tests` here. https://github.com/rapidsai/cugraph-gnn/actions/runs/22935844055/job/66568277622?pr=429 All except arm64 + CUDA 12.2 passed (that combination doesn't have `pytorch` conda packages), so added matrix filters skipping that one combination: #429 (comment) ### How can we enable arm tests now? Since they were originally filtered out, `pytorch` packages with GPU support have made it to `conda-forge`. So the reasoning that there aren't available packages (#167 (comment)) no longer holds. We should now be able to test on arm64 😁 ### Why add this import test? `torch` is an optional dependency of both `cugraph-pyg` and `pylibwholegraph`. The things done to keep it optional introduce risk of it accidentally not being installed in CI, or a CPU-only version of it being installed. Right now, when that happens you're confronted with a hard-to-understand error like ```text E AttributeError: module 'torch._C' has no attribute '_cuda_changeCurrentAllocator' ``` Hopefully we never accidentally install CPU-only `torch` in CI again, but if it does happen the test introduced in this PR should save some debugging time. ### Why re-use the `run_*` scripts from `test_*` scripts? Reduces duplication. And more importantly, ensures this project's tests are runnable in DLFW builds (where only the `run_*` scripts are used because they are not tightly coupled to RAPIDS CI setup). Authors: - James Lamb (https://github.com/jameslamb) Approvers: - Bradley Dice (https://github.com/bdice) - Alex Barghi (https://github.com/alexbarghi-nv) URL: #429

alexbarghi-nv

Just one minor comment; otherwise should be ok.

alexbarghi-nv · 2026-03-13T18:40:15Z

python/pylibwholegraph/pylibwholegraph/torch/wholegraph_env.py

@@ -171,8 +173,6 @@ def get_cpp_extension_src_path():


 def compile_cpp_extension():


Should we be doing an import_optional for torch.utils?

I was thinking we wouldn't need to because this usage will just try to hit the torch we already imported with torch = import_optional("torch") at the top of the module.

cugraph-gnn/python/pylibwholegraph/pylibwholegraph/torch/wholegraph_env.py

Line 195 in 4dcf7eb

torch.utils.cpp_extension.load(

However it could be useful to do that more specific import for those annoying "uninstall didn't fully work" situations like this: https://github.com/rapidsai/cugraph-gnn/pull/425/changes#r2921258039

Because then you'd get a more informative error than like "module 'torch' has no attribute 'utils".

I'll add that here.

Did this in eed447c

alexbarghi-nv · 2026-03-13T18:42:24Z

pyproject.toml

+]
+select = [
+    # (flake8-tidy-imports) banned-api
+    "TID251"


Thanks for adding this check.

python/pylibwholegraph/pylibwholegraph/torch/wholegraph_env.py

python/pylibwholegraph/pylibwholegraph/test_utils/test_comm.py

ci/download-torch-wheels.sh

…ncies (#5453) Contributes to rapidsai/build-planning#257 * ensures that when `torch` is installed in wheels CI, it's always a CUDA variant * removes unused test dependencies `ogb`, `pydantic`, `torchdata`, and `torchmetrics` - *I suspect these are left over from before GNN packages moved out to https://github.com/rapidsai/cugraph-gnn* ## Notes for Reviewers ### How does this help with testing against a mix of CTKs? `torch` CUDA wheels tightly pin to `nvidia-{thing}` wheels and soon, `cuda-toolkit` (see rapidsai/build-planning#255). Forcing the use of CUDA `torch` wheels ensures we'll catch dependency conflicts in CI. Without this, `pip` could silently fall back to CPU-only `torch` from pypi.org. ### How I tested this This PR uses some patterns I've tested elsewhere too: * rapidsai/rmm#2270 (comment) * rapidsai/cugraph-gnn#425 I also tested a commit here with both the PR and nightly test matrices, to be sure we covered everything: https://github.com/rapidsai/cugraph/actions/runs/23062364778/job/66996357196?pr=5453 Authors: - James Lamb (https://github.com/jameslamb) Approvers: - Alex Barghi (https://github.com/alexbarghi-nv) - Gil Forsyth (https://github.com/gforsyth) URL: #5453

jameslamb · 2026-03-16T14:23:34Z

python/pylibwholegraph/pylibwholegraph/tests/pylibwholegraph/test_wholememory_binding.py



-def test_dlpack():
+def test_dlpack(torch):


Just picking a place on the diff so this can be threaded.

pylibwholegraph tests are failing like this:

_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ /pyenv/versions/3.11.15/lib/python3.11/site-packages/torch/cuda/__init__.py:569: in set_device torch._C._cuda_setDevice(device) _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ def _lazy_init(): global _initialized, _queued_calls if is_initialized() or hasattr(_tls, "is_initializing"): return with _initialization_lock: # We be double-checked locking, boys! This is OK because # the above test was GIL protected anyway. The inner test # is for when a thread blocked on some other thread which was # doing the initialization; when they get the lock, they will # find there is nothing left to do. if is_initialized(): return # It is important to prevent other threads from entering _lazy_init # immediately, while we are still guaranteed to have the GIL, because some # of the C calls we make below will release the GIL if _is_in_bad_fork(): > raise RuntimeError( "Cannot re-initialize CUDA in forked subprocess. To use CUDA with " "multiprocessing, you must use the 'spawn' start method" ) E RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method /pyenv/versions/3.11.15/lib/python3.11/site-packages/torch/cuda/__init__.py:398: RuntimeError

and this:

def test_wholegraph_unweighted_sample( graph_node_count, ... torch, ): gpu_count = wmb.fork_get_gpu_count() > assert gpu_count > 0 E assert -1 > 0 tests/wholegraph_torch/ops/test_wholegraph_unweighted_sample_without_replacement.py:376: AssertionError

(build link)

I'm not sure what the cause is. Nightly tests did not fail this way (https://github.com/rapidsai/cugraph-gnn/actions/runs/23132494004/job/67188677016), so it's likely this is some issue introduced by this PR :/

If I run just this test using the same pytest invocation as in ci/run_pylibwholegraph_pytests.sh, it fails

$ pytest -rs --cache-clear --forked --import-mode=append tests/pylibwholegraph/test_wholememory_binding.py torch = <module 'torch' from '/pyenv/versions/3.11.15/lib/python3.11/site-packages/torch/__init__.py'> def test_dlpack(torch): gpu_count = wmb.fork_get_gpu_count() > assert gpu_count > 0 E assert -1 > 0 tests/pylibwholegraph/test_wholememory_binding.py:110: AssertionError --------------------------------------------------------------------------- Captured stdout call ---------------------------------------------------------------------------- fork_get_device_count failed.

But running the same code in a Python process? No problem.

import pylibwholegraph.binding.wholememory_binding as wmb wmb.fork_get_gpu_count() # 1

🤔

I've replicated the failure locally. My initial thought was to run the tests more like cuGraph-PyG does, but I think what's going on in pylibwholegraph should be more or less functionally equivalent. I'm pretty busy today but I will take another look as soon as I get a chance.

Thank you! I have a local repro too and will try to look more when I have time.

I can say I tried just removing the use of pytest-forked (unsure why that's used here), and unfortunately it wasn't that simple 😂

Yeah, I forgot to say I also just used regular pytest and it failed with the same error.

Ok, with some help I managed to figure out what's going wrong, I think. It looks like the tests in pylibwholegraph don't like it when torch constructs the CUDA context, resulting in this error when the forked processes try to create their own contexts. The problem is that importing torch creates a CUDA context...

I think I might have a solution. Give me a minute.

alexbarghi-nv · 2026-03-16T17:32:59Z

@jameslamb let me see if I can replicate this locally. I think I might have some ideas on how to fix it.

alexbarghi-nv · 2026-03-16T22:31:52Z

python/pylibwholegraph/pylibwholegraph/utils/imports.py

+        raise RuntimeError(f"This feature requires the '{self.name}' package/module")
+
+
+class FoundModule:


@jameslamb this is the solution I came up with - AI threw a bunch of different ideas at me, all of which failed miserably, so I guess human ingenuity still isn't obsolete. Basically, this makes all optional imports lazy so the actual import only occurs upon use. This prevents creating a CUDA context, and the tests pass, hopefully without any other weird implications...

Ok thanks! Interesting, makes sense to me that that could work.

This doesn't work in the case where torch is not installed though:

ImportError while importing test module '/__w/cugraph-gnn/cugraph-gnn/python/pylibwholegraph/pylibwholegraph/tests/pylibwholegraph/test_wholememory_tensor.py'. Hint: make sure your test modules/packages have valid Python names. Traceback: /pyenv/versions/3.11.15/lib/python3.11/importlib/__init__.py:126: in import_module return _bootstrap._gcd_import(name[level:], package, level) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ tests/pylibwholegraph/test_wholememory_tensor.py:6: in <module> from pylibwholegraph.torch.initialize import init_torch_env_and_create_wm_comm /pyenv/versions/3.11.15/lib/python3.11/site-packages/pylibwholegraph/torch/__init__.py:17: in <module> from .embedding import ( /pyenv/versions/3.11.15/lib/python3.11/site-packages/pylibwholegraph/torch/embedding.py:19: in <module> from .tensor import WholeMemoryTensor /pyenv/versions/3.11.15/lib/python3.11/site-packages/pylibwholegraph/torch/tensor.py:16: in <module> from .wholegraph_env import wrap_torch_tensor, get_wholegraph_env_fns, get_stream /pyenv/versions/3.11.15/lib/python3.11/site-packages/pylibwholegraph/torch/wholegraph_env.py:13: in <module> torch_utils_cpp_ext = import_optional("torch.utils.cpp_extension") ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ /pyenv/versions/3.11.15/lib/python3.11/site-packages/pylibwholegraph/utils/imports.py:56: in import_optional if find_spec(mod) is not None: ^^^^^^^^^^^^^^ E ModuleNotFoundError: No module named 'torch'

(build link)

There might be clue in that... all of the unit tests were passing in CI on this PR, with and without torch installed, before I added that import_optional("torch.utils.cpp_extension") in response to 79a6efe#r2933127641

i.e. before these commits:

eed447c

eed447c

Maybe going 2 levels deep in a single import_optional() just does not work well with this machinery?

ahhh yeah ok ... importlib.util.find_spec() cannot process dotted names.

from importlib.util import find_spec find_spec("torch") find_spec("torch.autograd") # Traceback (most recent call last): # File "<stdin>", line 1, in <module> # File "<frozen importlib.util>", line 95, in find_spec # ModuleNotFoundError: No module named 'torch'

I'll look for a different approach here.

I think I have a fix.

Ok let's see if 456857a fixes this

There, I'm catching the ModuleNotFoundError / ImportError that could happen with dotted imports and treating that as another example of "thing not installed".

Ok that fixed pylibwholegraph. There's one more issue in cugraph-pyg:

Traceback (most recent call last): File "<string>", line 1, in <module> import cugraph_pyg; print(f'cugraph-pyg version: {cugraph_pyg.__version__}') ^^^^^^^^^^^^^^^^^^ File "/pyenv/versions/3.13.12/lib/python3.13/site-packages/cugraph_pyg/__init__.py", line 6, in <module> import cugraph_pyg.data File "/pyenv/versions/3.13.12/lib/python3.13/site-packages/cugraph_pyg/data/__init__.py", line 6, in <module> from cugraph_pyg.data.graph_store import ( GraphStore, ) File "/pyenv/versions/3.13.12/lib/python3.13/site-packages/cugraph_pyg/data/graph_store.py", line 36, in <module> else torch_geometric.data.GraphStore ^^^^^^^^^^^^^^^^^^^^ File "/pyenv/versions/3.13.12/lib/python3.13/site-packages/cugraph_pyg/utils/imports.py", line 50, in __getattr__ self.mod = import_module(self.mod) ~~~~~~~~~~~~~^^^^^^^^^^ File "/pyenv/versions/3.13.12/lib/python3.13/importlib/__init__.py", line 88, in import_module return _bootstrap._gcd_import(name[level:], package, level) ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/pyenv/versions/3.13.12/lib/python3.13/site-packages/torch_geometric/__init__.py", line 3, in <module> import torch ModuleNotFoundError: No module named 'torch'

(build link)

When torch_geometric is available but torch is not, this check falls back to trying to import torch_geometric and fails:

cugraph-gnn/python/cugraph-pyg/cugraph_pyg/data/graph_store.py

Lines 33 to 37 in 4dcf7eb

class GraphStore(

object

if isinstance(torch_geometric, MissingModule)

else torch_geometric.data.GraphStore

):

I think because find_spec("torch_geometric") works, but import_module("torch_geometric") fails (because torch_geometric unconditionally imports torch).

I've pushed 0afcdd7 hopefully resolving that.

python/pylibwholegraph/pylibwholegraph/utils/imports.py

ci/test_wheel_pylibwholegraph.sh

python/pylibwholegraph/pylibwholegraph/utils/imports.py

pyproject.toml

…torch-testing

python/cugraph-pyg/cugraph_pyg/utils/imports.py

KyleFromNVIDIA

Approved with a couple of small comments

ci/test_wheel_cugraph-pyg.sh

KyleFromNVIDIA · 2026-03-17T18:19:20Z

ci/validate_wheel.sh

+# Use '|| true' so grep not finding any matches (exit 1) does not kill the script under set -e
+unzip -p "${WHEEL_FILE}" '*.dist-info/METADATA' \
+| grep -E '^Requires-Dist:.*torch[><=!~ ]+.*' \
+| tee matches.txt || true


Why the tee? You could pipe the output into a variable and then echo it instead.

tee specifically to avoid the complexity of storing the output in a variable just to echo it.

KyleFromNVIDIA · 2026-03-17T18:34:47Z

ci/test_wheel_cugraph-pyg.sh

+fi
+
+rapids-logger "import cugraph-pyg (no 'torch')"
+./ci/uninstall-torch-wheels.sh


Is there any reason you couldn't do the no-torch tests before the torch tests? That would save you the trouble of uninstalling torch.

Because to run the tests without torch, we still need to install cugraph-pyg and its dependencies, any of which might pull in torch as a requirement.

Imperative code that force-uninstalls torch right here afterwards is easier and less error-prone than trying to be careful about not letting torch into the environment at the beginning, in my opinion.

alexbarghi-nv

👍

alexbarghi-nv · 2026-03-18T18:51:47Z

/merge

ensure 'torch' CUDA wheels are installed in CI

13825b5

jameslamb added breaking Introduces a breaking change improvement Improves an existing functionality labels Mar 9, 2026

jameslamb commented Mar 9, 2026

View reviewed changes

help git understand the diff

698f115

jameslamb mentioned this pull request Mar 9, 2026

wheels CI: test mix of cuda-toolkit version in CI rapidsai/build-planning#256

Open

30 tasks

jameslamb added 5 commits March 9, 2026 10:24

use rapids-generate-pip-constraints, fix typo

1c457b8

handle the fallback case better, other fixes

066d5c4

echo wheel name

6f73e44

more pin fiddling

271eb7e

fix validation script

5a4064e

jameslamb mentioned this pull request Mar 10, 2026

make PyTorch installation in conda test jobs stricter #427

Merged

jameslamb added 8 commits March 10, 2026 10:19

just wheels changes

0d7215e

even fewer changes

fc30204

revert gitignore changes

97e2c02

add 'torch_only'

355d5aa

testing

4aad5b4

more updates

426c5ff

make 'torch' optional everywhere

7ac88d3

more changes

104b8bf

jameslamb added 4 commits March 10, 2026 15:45

more torch fixes... unconditional references in argument defaults

4b479f7

merge main

d055f9d

handle more unconditional 'torch' references (this time in type hints)

7bbf218

revert

4cacebf

This was referenced Mar 10, 2026

CI: restore arm64 conda tests, re-use run_* scripts in test_* scripts #429

Merged

[BUG] pylibwholegraph: nightly wheel tests failing: "driver on your system is too old" #410

Open

check in debugging code temporarily

36843b6

This was referenced Mar 13, 2026

ensure 'torch' CUDA wheels are installed in CI, remove unused dependencies rapidsai/cugraph#5453

Merged

wheels: build with oldest CUDA 13 CTK (13.0.x) rapidsai/build-planning#257

Open

alexbarghi-nv approved these changes Mar 13, 2026

View reviewed changes

one more import

eed447c

greptile-apps bot reviewed Mar 13, 2026

View reviewed changes

python/pylibwholegraph/pylibwholegraph/torch/wholegraph_env.py Outdated Show resolved Hide resolved

fix

79a6efe

greptile-apps bot reviewed Mar 13, 2026

View reviewed changes

python/pylibwholegraph/pylibwholegraph/test_utils/test_comm.py Show resolved Hide resolved

ci/download-torch-wheels.sh Show resolved Hide resolved

jameslamb commented Mar 16, 2026

View reviewed changes

make optional imports lazy

a61a427

alexbarghi-nv reviewed Mar 16, 2026

View reviewed changes

greptile-apps bot reviewed Mar 16, 2026

View reviewed changes

python/pylibwholegraph/pylibwholegraph/utils/imports.py Show resolved Hide resolved

ci/test_wheel_pylibwholegraph.sh Show resolved Hide resolved

fix module check - meant to change to use find_spec

96201b6

alexbarghi-nv reviewed Mar 16, 2026

View reviewed changes

python/pylibwholegraph/pylibwholegraph/utils/imports.py Show resolved Hide resolved

greptile-apps bot reviewed Mar 16, 2026

View reviewed changes

pyproject.toml Show resolved Hide resolved

Merge branch 'release/26.04' of github.com:rapidsai/cugraph-gnn into …

b8276e2

…torch-testing

greptile-apps bot reviewed Mar 17, 2026

View reviewed changes

python/cugraph-pyg/cugraph_pyg/utils/imports.py Show resolved Hide resolved

handle dotted imports, make ruff selections explicit

456857a

jameslamb force-pushed the torch-testing branch from d03a9bf to 456857a Compare March 17, 2026 15:54

more import-time patching

0afcdd7

KyleFromNVIDIA approved these changes Mar 17, 2026

View reviewed changes

jameslamb added 2 commits March 17, 2026 14:10

remove unnecessary CUDA_MAJOR

27f8fdd

Merge branch 'release/26.04' into torch-testing

abf7313

alexbarghi-nv approved these changes Mar 18, 2026

View reviewed changes

rapids-bot bot merged commit 330b135 into rapidsai:release/26.04 Mar 18, 2026
62 checks passed

jameslamb deleted the torch-testing branch March 18, 2026 19:30

jameslamb mentioned this pull request Mar 19, 2026

wheels: build with CUDA 13.0, test against mix of CTK versions, make 'torch-geometric' fully optional for 'cugraph-pyg' #434

Open

jakirkham mentioned this pull request Mar 20, 2026

Use PyBuffer_FillInfo for simple buffers & simplify Python buffer cleanup #436

Open

		@@ -171,8 +173,6 @@ def get_cpp_extension_src_path():


		def compile_cpp_extension():

		raise RuntimeError(f"This feature requires the '{self.name}' package/module")


		class FoundModule:

	class GraphStore(
	object
	if isinstance(torch_geometric, MissingModule)
	else torch_geometric.data.GraphStore
	):

Conversation

jameslamb commented Mar 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Ensures that torch CUDA wheels are always installed in CI

Makes torch truly optional

Notes for Reviewers

Uh oh!

copy-pr-bot bot commented Mar 9, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alexbarghi-nv left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alexbarghi-nv commented Mar 16, 2026

Uh oh!

alexbarghi-nv Mar 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jameslamb Mar 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

KyleFromNVIDIA left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alexbarghi-nv left a comment

Choose a reason for hiding this comment

Uh oh!

alexbarghi-nv commented Mar 18, 2026

Uh oh!

Uh oh!

jameslamb commented Mar 9, 2026 •

edited

Loading

Ensures that `torch` CUDA wheels are always installed in CI

Makes `torch` truly optional

alexbarghi-nv Mar 16, 2026 •

edited

Loading

jameslamb Mar 17, 2026 •

edited

Loading