Open faaany opened 1 week ago
Hmm, I always thought that triton comes with torch. I just created a test env and ran
pip install torch
and indeed triton was installed. Under what circumstances can it be missing, is it dependent on the device?
yes, it depends on the device. For the pytorch cuda distribution, triton is installed by default. But for XPU and CPU, triton will not be installed by default (at least for now). So if we think about the pro&cons about adding this marker:
Pros:
Cons:
Just let me know how you decide; I will update my code accordingly. Thanks so much for the review! @BenjaminBossan @SunMarc
Thanks for providing further information on the triton dependency. In that case, I agree that adding an explicit check is fine.
As to whether we should potentially break the tests for other non-CUDA devices: It's hard to say if this would be a good thing or not, as we don't know how that will affect the corresponding maintainers. Probably it's best for Zach to judge when he's back.
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
What does this PR do?
test_dynamo
works on xpu, but fails because it requires extra installation of the triton library. This PR adds therequire_triton
test marker and maketest_dynamo
device-agnostic.