Skip to content

Details

Friday, November 20 · 5:30–6:30 a.m. Pacific / 13:30–14:30 UTC · Live Zoom session hosted by Rajat Dandekar on Maven. A recording will be available afterward.
You click Send. A cluster of GPUs has to turn a trillion parameters into your next token. Where do the weights fit, what has to move, and what happens when thousands of people are waiting for an answer at once?
Dr. Rajat Dandekar, co-founder of Vizuara and co-author of Build a DeepSeek Model (From Scratch), will use examples from DeepSeek, Kimi, and vLLM to build up the serving stack from one token to a full cluster.

  • Do the GPU math. Work out why a trillion parameters at 16-bit precision take about 2 TB just for weights, why bandwidth can bottleneck decoding, and how batching lets requests share the cost of reading weights.
  • Make the KV cache fit. See what multi-head latent attention (MLA), PagedAttention, and prefix caching each do to reduce memory use or reuse work, including what Kimi K3 does differently.
  • Split the work across the cluster. Compare tensor, expert, and pipeline parallelism; see why you might run prefill and decode on separate GPUs; and learn where FP4 and speculative decoding fit.

For engineers building or tuning LLM inference stacks, or working out where to start when latency or GPU memory becomes a bottleneck. Bring your team.
Register free on Maven for the Zoom session
A Meetup RSVP isn’t a Maven registration.

You may also like