[benchmarks][vllm] Unified Attention benchmark (paged attention)

[Egor-Krivov]

Closes #5257

I also started reporting gbps to the database because many benchmarks are memory bound

[Copilot]

Pull Request Overview

This PR adds a paged attention benchmark for the vLLM library, implementing both 2D and 3D unified attention kernels with tensor descriptor optimizations. The benchmark compares performance against PyTorch reference implementations and reports both throughput (GB/s) and compute (TFlops) metrics.

Key changes:

• Implementation of unified attention benchmark with paged KV cache support
• Enhanced memory bandwidth calculations accounting for actual token usage
• Extended result transformation to report GB/s metrics alongside TFlops
• CI/CD workflow updates to run the new benchmark

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
benchmarks/third_party/vllm/unified_attention_benchmark.py	New comprehensive benchmark for vLLM's unified attention with 2D/3D kernels, supporting various model configurations and attention features
benchmarks/third_party/vllm/transform_results.py	Enhanced to handle non-integer parameter values and report both TFlops and GB/s metrics
benchmarks/third_party/vllm/batched_moe_benchmark.py	Improved memory bandwidth calculation to account for actual activated experts and token usage
.github/workflows/third-party-benchmarks.yml	Added unified attention benchmark to CI workflow and improved command formatting

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.