llama.cpp v1.0.0 VRAM Budget Calculator for GGUF Models
calibrated on llama.cpp v1.0.0

llama.cpp only, GGUF models only — every dial is a real launch flag or a GGUF fact, and only GGUF repos can actually run here. One equation prices any model at granular level; the dense and MoE panels are just two architecture-set price sheets, not fixed demo models. Every slider's wall is computed live from every other slider in its panel; when a dial hits its wall the red note names the dials to move — click one to apply it. Load a GGUF repo id for a real config; safetensors ids load as shape-only what-if. Panels are independent experiments and shared hardware applies to both.
per card:  weights/TP + slots · ctx · KV B/token/TP + ubatch · compute B/token + draft  ≤  GPU − driver − slack
HARDWARE · shared by both panels GiB GiB GiB/cardVRAM each card spends on the CUDA context before any model loads %safety margin kept unused on each card - covers spikes the equation cannot price
all hardware values and costs are GiB — the binary units llama.cpp itself uses · model card file sizes are shown in decimal GB as HF lists them (8.42 GB = 7.84 GiB) · * GPUs assumes identical cards — mixed-size rigs (e.g. 24+12) are not modeled; bracket them as N × smallest card or the big card alone — details in “How this works”
LOAD FROM HUGGINGFACE metadata only — GGUF headers and model cards are read, never model weights; fetched ids are cached 30 days
examples:
llama.cpp VRAM Budget Calculator — llama.cpp memory model, flags & prices calibrated in llama.cpp v1.0.0 · b1b770d
Independent community tool — not affiliated with, endorsed by, or sponsored by ggml.org (llama.cpp maintainers) or Hugging Face. Model metadata is fetched live from huggingface.co only when you request it. All trademarks belong to their owners.
© Abdulaziz Almalki · X @almalki_io · LinkedIn