GPU Memory Behavior in Long-Context Inference长上下文推理中的 GPU 内存行为
Questions and instrumentation for understanding memory access patterns and bandwidth utilization in long-context serving.用于理解长上下文服务中内存访问模式与带宽利用率的问题和测量方法。
Working outline研究提纲 · Hypotheses and measurement plans; verified results are still being attached.包含假设与测量计划，经验证的结果仍在补充中。
Questions
Which allocations scale with sequence length, which are retained across requests, and which accesses constrain the measured path? TODO: attach validated traces.
问题
哪些内存分配会随序列长度增长，哪些会跨请求保留，又有哪些访问限制了被测路径？TODO：附上经过验证的性能轨迹。