[Kernel] Add needs_fixed_stride_order tag to most GEMMs

[tlrmchlsmth]

PyTorch 2.6.0 changed the default behavior for custom operators to be "requires_contiguous". In GEMM kernels, we often transpose or otherwise transform our weights offline, which would be in conflict with this change. This PR changes adds {at::Tag::needs_fixed_stride_order} to most GEMMs in csrc/torch_bindings.cpp

See #12721 (comment)

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[zou3519]

PyTorch 2.6.0 changed the default behavior for custom operators to be "requires_contiguous"

For some more context, this was a mistake and we're going to flip the default back in PyTorch 2.7. Is it possible to only set the tag for PyTorch 2.6? The reason why I ask is in PyTorch 2.8 we're going to flip the default again to something better ("match exact eager-mode strides") and setting it to at::Tag::needs_fixed_stride_order now means that you won't pick up that change

[tlrmchlsmth]

PyTorch 2.6.0 changed the default behavior for custom operators to be "requires_contiguous"

For some more context, this was a mistake and we're going to flip the default back in PyTorch 2.7. Is it possible to only set the tag for PyTorch 2.6? The reason why I ask is in PyTorch 2.8 we're going to flip the default again to something better ("match exact eager-mode strides") and setting it to at::Tag::needs_fixed_stride_order now means that you won't pick up that change

Ah, thanks for that context. I'll try it

[zou3519]

(that being said, I don't know how to actually switch on the PyTorch version like that in C++, so if it's not possible we can just leave a mental note here)

[mgoin]

Thanks LGTM! Maybe we could override ops to add a wrapper function, but still it wouldn't be clear which functions require it.