Update throughput chart and improve README

README.md

@@ -49,7 +49,14 @@ In terms of the evaluation of reasoning ability， Ring-lite-linear-preview achi
 
 ## Inference Speed
 
-To evaluate the generation throughput, we deploy Ring-lite-linear and the softmax-attention-based Ring-lite based on vLLM on a single NVIDIA A100 GPU. Specifically, the input sequence length is fixed to 1. The end-to-end (E2E) generation time required for generating output sequences of varying lengths is illustrated below. It is shown in the figure that at 32k output length, Ring-lite-linear-preview achieves 2.2× throughput of Ring-lite.
+To evaluate the generation throughput, we deploy Ring-lite-linear and the softmax-attention-based Ring-lite based on vLLM on a single NVIDIA A100 GPU. We conduct two sets of experiments:
+
+1. **Long Input Evaluation**: We measure the time-to-first-token (TTFT) with varying input sequence lengths (from 512 to 384k tokens) using batch size 1 and TP=1. As shown in the top figure, at 384k input length, Ring-lite-linear achieves 3.5× faster TTFT compared to the softmax-attention-based model.
+
+2. **Long Output Evaluation**: We fix the input sequence length to 1 and measure the end-to-end (E2E) generation time required for generating output sequences of varying lengths (from 512 to 32k tokens) with batch size 64 and TP=1. As illustrated in the bottom figure, at 32k output length, Ring-lite-linear achieves 2.2× throughput of the softmax-attention-based Ring-lite.
+
+These results demonstrate that our hybrid linear attention mechanism significantly improves both input processing efficiency and generation throughput, especially for long context scenarios.
+
 
 <p align="center">
     <img src="https://huggingface.co/inclusionAI/Ring-lite-linear-preview/resolve/main/throughput.png" width="600"/>