In the default pull flow, the decode instance waits for routing metadata, allocates destination blocks, prepares the transfer, and initiates a read after prefill completes. The new flow registers decode-side destination blocks in advance; once prefill and registration are both ready, the prefill side issues a NIXL write, with setup handled by a background thread.
The vLLM post reports results from a 1P1D Qwen3-32B benchmark on Amazon EC2 P5en instances connected with EFA: up to 3× better time to first token (TTFT) and up to 30% better time per output token (TPOT) than pull mode. The authors note that results depend on workload and configuration. They say the feature was built and validated on AWS Trainium clusters and contributed to vLLM as accelerator-agnostic. https://vllm.ai/blog/2026-10-14-nixl-push-connector