Skip to content

Mohammad UmmairNotes on ML systems, inference, and silicon.

PostsTagsAbout

inference

1 post

  • KV Store: paying memory to avoid recompute

    Without a KV cache, every new token re-attends over the whole prefix. With one, you trade that compute for GPU memory — about 40 GB of it on Llama 3 70B at 128K context, and PagedAttention decides how much of that you waste.

    August 16, 20264 min readDraftinferencelong contextaisoftware engineering
© 2026 Mohammad UmmairCVGitHubLinkedInEmail