https://sandeep007734.github.io/tags/systems/Recent content in Systems on Sandeep KumarHugoen-usFri, 07 Aug 2026 00:00:00 +0530https://sandeep007734.github.io/posts/flashattention/Sun, 08 Mar 2026 00:00:00 +0530https://sandeep007734.github.io/posts/flashattention/How FlashAttention avoids materializing quadratic attention matrices by combining tiled execution, online softmax, and kernel fusion.https://sandeep007734.github.io/posts/pagedattention/Tue, 03 Mar 2026 00:00:00 +0530https://sandeep007734.github.io/posts/pagedattention/How PagedAttention uses virtual-memory-style KV-cache paging to reduce fragmentation and increase LLM serving throughput.https://sandeep007734.github.io/posts/flashattention-evolution/Fri, 07 Aug 2026 00:00:00 +0530https://sandeep007734.github.io/posts/flashattention-evolution/How FlashAttention-2, -3, and -4 rescheduled the same exact-attention algorithm as GPU bottlenecks shifted from occupancy to asynchronous pipelines and specialized memory.https://sandeep007734.github.io/posts/flashinfer/Fri, 07 Aug 2026 00:00:00 +0530https://sandeep007734.github.io/posts/flashinfer/How FlashInfer unifies heterogeneous KV-cache layouts, specialized attention kernels, and dynamic serving schedules in one inference-oriented engine.