Inference planner

Fit the workload, not only the weights.

Import a model configuration or set a planning profile, then make the tradeoff between weight format, context length, concurrent requests, and per-GPU capacity visible.

Inference capacity planner

Plan the workload before choosing the VRAM tier.

Keep weights, KV Cache, runtime reserve, context, and simultaneous requests in one visible capacity model.

How the estimates work →

Model configuration

Import a Hugging Face model profile.

Load the public model configuration to set parameters, attention shape, and available context in this plan.

Loaded weights——
KV Cache at target——
Required per GPU——
Capacity result——

Planning brief

Turn the memory estimate into a runtime decision.

—

What it models

Weight memory, architecture-sensitive KV Cache, a runtime reserve, tensor-parallel GPU shape, and usable memory per GPU.

Serving path

Plan vLLM request capacity

Turn the model baseline into model-length, sequence, TP, and DP starting limits.

Training path

Plan LoRA or QLoRA memory

Separate frozen weights, adapter state, activations, and the distributed training strategy.