Home/Guides
2 deployment guidesUnderstand
Understand
the estimate.
Short answers are useful, but capacity decisions need the reasoning behind them. These guides explain the inputs that move LLM memory and GPU fit.
Review the calculation methodology →Weights + overhead + headroom
Read guide ↗02How much VRAM does an LLM need?
A complete calculation method for model weights, quantization overhead, KV cache, runtime workspace, and a safe deployment margin.
- VRAM calculator
- quantization
- local LLM
Context + concurrency + cache precision
Read guide ↗How context windows use GPU memory
See why KV cache grows with tokens and concurrent sequences—and why the model’s maximum context is not always the practical serving target.
- KV cache
- context window
- concurrency
A practical sequenceFrom a model name
From a model name
to a GPU request.
- 01Estimate the weight floor
Use parameter count and the actual storage precision.
- 02Add workload memory
Set context, concurrency, cache precision, and runtime overhead.
- 03Leave usable headroom
A fit at 100% of advertised VRAM is not a stable operating plan.
- 04Validate and source
Test the real runtime, then request matching capacity.
Skip to your workload
Use the interactive worksheet.
Choose a model and change the exact inputs that affect its memory envelope.