LLM capacity planner
Turn a model and workload into a hardware boundary.
Choose what the model must do, then keep loaded weights, KV Cache or training activations, runtime reserve, and per-GPU capacity in the same decision path.
Start with the real intent
Choose the work that will pressure the memory budget.
Interactive inference
Local chat
Plan one interactive session from model weights, context length, KV Cache, and the available local GPU.
Start this path →Long working context
Coding assistant
Keep repository or file context and the active request count visible before choosing the hardware boundary.
Start this path →Retrieved context
RAG workflow
Include retrieved passages in the context budget and keep simultaneous request pressure explicit.
Start this path →Shared endpoint
API serving
Establish a model fit, then continue into per-replica context, sequence, KV Cache, and vLLM limits.
Start this path →Adapter training
LoRA training
Carry the model and GPU capacity into a BF16 base-weight, adapter, optimizer, and activation plan.
Start this path →Quantized adapter training
QLoRA training
Carry the model and GPU capacity into a 4-bit base-weight, adapter, optimizer, and activation plan.
Start this path →Reviewed model profile
Start from source-bound model facts.
Learn the moving parts
Read only what changes the capacity plan.
Memory model
Separate model weights, KV Cache, and runtime reserve
A parameter count alone cannot tell you whether a model will fit at the context length and concurrency you need.
Read guide →Serving capacity
Plan context length and concurrency together
A model that fits for one chat can run out of capacity once long requests or parallel users share the same GPU.
Read guide →vLLM model length
Set vLLM max model length from the KV Cache budget
Choose a useful maximum request length that leaves enough cache for the number of active requests you want to serve.
Read guide →vLLM concurrency
Plan vLLM concurrency before raising max_num_seqs
Translate a request target into cache pressure before exposing it as an endpoint concurrency limit.
Read guide →LoRA mode
Choose LoRA or QLoRA from the base-weight budget
Choose the adapter path only after you know how much of the GPU is already occupied by the frozen base model.
Read guide →Training memory
Set sequence length and micro-batch before increasing VRAM
Control activation memory with the two workload inputs that move it most: sequence length and micro-batch size.
Read guide →