Fix vLLM inter-token latency formula

Author: syhya

content/en/posts/2025-05-17-vllm/index.md

@@ -1,7 +1,7 @@
 ---
 title: "vLLM: High-Throughput, Memory-Efficient LLM Serving"
 date: 2025-05-17T10:00:00+08:00
-lastmod: 2025-05-17T10:00:00+08:00
+lastmod: 2026-07-16T10:00:00+08:00
 author: "Yue Shui"
 categories: ["Technical Blog"]
 tags: ["vLLM", "PagedAttention", "LLM Serving", "Inference", "KV Cache", "Memory Optimization", "LLM", "AI Infrastructure", "Deep Learning"]
@@ -47,10 +47,10 @@ $$
 
 ### Inter-Token Latency (ITL)
 
-**Inter-Token Latency (ITL)** represents the average time interval between the generation of two consecutive tokens after the first token has been generated. It reflects the speed at which each subsequent token is generated, calculated as:
+**Inter-Token Latency (ITL)** represents the average time interval between consecutive output tokens after the first token has been generated. For a single request with at least two output tokens, it is calculated as:
 
 $$
-\text{ITL} = \frac{\text{End-to-End Latency} - \text{TTFT}}{\text{Batch Size} \times (\text{Number of Output Tokens} - 1)}
+\text{ITL} = \frac{\text{End-to-End Latency} - \text{TTFT}}{\text{Number of Output Tokens} - 1}
 $$
 
 These metrics reflect the inference engine's response speed, processing efficiency, and concurrency capabilities, serving as important benchmarks for evaluating and optimizing LLM inference performance.

content/zh/posts/2025-05-17-vllm/index.md

@@ -1,7 +1,7 @@
 ---
 title: "vLLM：高吞吐、内存高效的 LLM 服务引擎"
 date: 2025-05-17T10:00:00+08:00
-lastmod: 2025-05-17T10:00:00+08:00
+lastmod: 2026-07-16T10:00:00+08:00
 author: "Yue Shui"
 categories: ["技术博客"]
 tags: ["PagedAttention", "LLM Serving", "Inference", "KV Cache", "Memory Optimization", "LLM", "AI Infrastructure"]
@@ -48,10 +48,10 @@ $$
 
 ### Token 间延迟
 
-**Token 间延迟（Inter Token Latency, ITL）** 表示生成连续 token 时每两个 token 间的平均时间间隔。它体现了模型在生成首个 token 后，每个后续 token 的生成速度，计算公式为：
+**Token 间延迟（Inter Token Latency, ITL）** 表示生成首个 token 后，相邻输出 token 之间的平均时间间隔。对于至少生成两个输出 token 的单个请求，计算公式为：
 
 $$
-\text{ITL} = \frac{\text{End-to-End Latency} - \text{TTFT}}{\text{Batch Size} \times (\text{Number of Output Tokens} - 1)}
+\text{ITL} = \frac{\text{End-to-End Latency} - \text{TTFT}}{\text{Number of Output Tokens} - 1}
 $$
 
 这些指标反映了推理引擎的响应速度、处理效率和并发能力，是评估和优化 LLM 推理性能的重要依据。