Measurement harness for the sliding window attention premium in the vLLM TPU Ragged Paged Attention v3 kernel: per layer decode cost, block size control, throughput, and goodput for Gemma 4 31B on TPU v6e.
google-cloud attention-mechanism gemma pallas gpu-kernels tpu jax inference-optimization tpu-benchmarks tpus tpu-acceleration trillium kernel-optimizers vllm llm-inference paged-attention sliding-window-attention vllm-serve kernel-optimization goodput
-
Updated
Jul 29, 2026 - Python