[Kernel] LoRA - Enable CUDAGraphs for V1

[varun-sundar-rabindranath]

Enable CUDAGraphs support for V1

[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.

🚀

[varun-sundar-rabindranath]

^ changes are to avoid errors such as,

  raise ConstraintViolationError(
torch.fx.experimental.symbolic_shapes.ConstraintViolationError: Constraints violated (L['input_ids'].size()[0], L['positions'].size()[0])! For more information, run with TORCH_LOGS="+dynamic".
  - Not all values of RelaxedUnspecConstraint(L['input_ids'].size()[0]) are valid because L['input_ids'].size()[0] was inferred to be a constant (8192).
  - Not all values of RelaxedUnspecConstraint(L['positions'].size()[0]) are valid because L['positions'].size()[0] was inferred to be a constant (8192).

[varun-sundar-rabindranath]

x here is the input_ids. In V1, we don't zero out the cuda graph pad region.
Avoid the in-place update here to prevent accumulating garbage into the input buffer.

[varun-sundar-rabindranath]

During torch.compile, LoRA static buffers like in

vllm/vllm/lora/punica_wrapper/punica_base.py

Line 133 in 5305673

self._token_lora_indices = torch.empty(max_num_batched_tokens,

and

vllm/vllm/lora/ops/triton_ops/v1/v1_kernel_metadata.py

Line 24 in 5305673

token_lora_mapping = torch.empty(max_num_tokens,

get captured along with their sizes and strides (they aren't dynamic)

When max_num_batched_tokens changes, and when the captured graph is executed, we hit assert_size_stride asserts on these tensors. As a solution, we simply recompile when max_num_batched_tokens change.