Improve PagedAttention dispatch diagnostics

[Baiju Meswani (baijumeswani)]

Summary

Extend the existing CUDA attention debug output with PagedAttention-specific dispatch details.

When ORT_ENABLE_ATTENTION_KERNEL_DEBUG_INFO=1, PagedAttention now reports:

• the selected backend, including XQA versus portable paged decode;
• the number of decode splits;
• the GQA group size;
• the effective KV-length bound after applying the local window;

This makes it easier to verify that an intended optimized path is selected instead of silently falling back. The diagnostics are opt-in and do not change kernel selection or execution when disabled.

[Copilot]

Pull request overview

This PR extends ONNX Runtime’s CUDA attention kernel diagnostics for the PagedAttention contrib op, adding more dispatch-selection details to the opt-in debug output controlled by ORT_ENABLE_ATTENTION_KERNEL_DEBUG_INFO=1.

Changes:

• Add PagedAttention-specific debug fields (decode split count, GQA group size, and effective KV-length bound) to AttentionKernelDebugInfo and its Print() output.
• Populate the new debug fields from PagedAttention::ComputeInternal() during backend selection.
• Add a CUDA unit test that captures stdout and asserts the presence of the new debug markers.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
onnxruntime/test/contrib_ops/paged_attention_op_test.cc	Adds a CUDA test that captures attention debug output and checks for expected PagedAttention dispatch diagnostics.
onnxruntime/contrib_ops/cuda/bert/paged_attention.cc	Populates new debug-info fields for PagedAttention dispatch (including splits/GQA group size/effective KV bound).
onnxruntime/contrib_ops/cuda/bert/attention_kernel_options.h	Extends AttentionKernelDebugInfo with optional fields for additional dispatch diagnostics.
onnxruntime/contrib_ops/cuda/bert/attention_kernel_options.cc	Prints the new optional debug fields when present.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Tianlei Wu (tianleiwu)]

The decode-path fields follow the final fallback state and remain behind the opt-in debug flag. One supported backend is still misclassified; see the inline comment.