title: the four choices when the pool holds too few users, in order of cost
source: manuals/vllm-101/sections/2-1-count-the-memory-before-you-pick-a-gpu.md
lines: 83-90
lang: text
note: The first choice is the one that this ticket is about.
---
## Decide what to change

If the count is too small for your traffic, you have four choices. They are in order of cost:

1. Lower `--max-model-len` to the longest context that your users send. The pool stays the same size, and one long request can no longer take all of it.
2. Use an FP8 KV cache. It halves the bytes per token. [The FP8 KV page](#s-fit-more-users-with-an-fp8-kv-cache) shows the cost.
3. Use FP8 weights. They free memory for the cache. [The FP8 weights page](#s-fit-a-bigger-model-with-fp8-weights) shows when this pays.
4. Use a GPU with more memory, or split the model across GPUs. [The parallelism page](#s-choose-tensor-pipeline-or-data-parallelism) covers this.
