The three components
- Weights: 22 GB at bfloat16, 11-12 GB at 8-bit, 7-8 GB at 4-bit.
- KV cache: grows with the context window you reserve and the number of concurrent requests.
- Runtime overhead: the inference engine, tokenizer and activation buffers need their own room.
Why the context window matters
Falcon 2 11B is configured at 8,192 tokens. A server that reserves the full window for every request allocates far more cache than one that reserves 2,048, and the difference can be the margin between fitting and not fitting. Set the limit you actually need.
A simple planning rule
Take the weight size, add the cache for your context length multiplied by how many users you expect at once, then add about ten percent. If that total exceeds your VRAM, change one of the three inputs: quantise, shorten the context, or limit concurrency. There is no fourth option at a fixed memory size.
Related pages
Sources
falcon2.lol is an independent third-party site. It is not affiliated with, endorsed by, sponsored by or operated by the Technology Innovation Institute, and it does not speak for TII. Every number on this page is attributed to the source listed above.