events
One Trillion Parameters, One Token at a Time
A live Zoom session from the AI Performance Engineering Meetup (UK), presented by Rajat Dandekar of Vizuara, co-author of Build a DeepSeek Model (From Scratch).

Auf Englisch
- Starts
- 2026-11-20 13:30 UTC
- Organizer
- AI Performance Engineering Meetup (UK)
- Location
- Online event
A live Zoom session from the AI Performance Engineering Meetup (UK), presented by Rajat Dandekar of Vizuara, co-author of Build a DeepSeek Model (From Scratch). It traces how a trillion-parameter model is served, from GPU memory and bandwidth constraints to KV-cache methods such as MLA, PagedAttention and prefix caching, then tensor, expert and pipeline parallelism, prefill/decode separation, FP4 and speculative decoding. Examples come from DeepSeek, Kimi and vLLM. It is aimed at engineers building or tuning LLM inference stacks; a recording follows, with registration through Maven.
Event website