Fonctionnement technique
Model weights take parameters × bits per weight ÷ 8 bytes. The KV cache grows with context: 2 × layers × KV heads × head dimension × tokens × batch × bytes per value. The tool adds a runtime overhead of about 10% of the weights plus 0.5 GiB for CUDA context and activations, then compares the total with the VRAM you have.
Mode d’emploi étape par étape
Pick a preset model or choose custom and enter parameters, layers, KV heads and head dimension.
Choose the quantization (FP16, Q8, Q6, Q5, Q4 or Q3 style) and the KV cache precision.
Set the context length and batch size you plan to use.
Enter the VRAM per GPU and the number of GPUs to see whether the model fits and how much is left.
Fonctionnalités & performances
- Presets for Llama 3.1, Mistral 7B, Qwen2.5 and Gemma 2 with their real layer and KV head counts
- Separate lines for weights, KV cache and overhead so you see what to cut first
- Fit check against common GPU sizes and multi-GPU setups
- Runs locally; nothing you type is sent anywhere
Méthodologie, formules & standards
Effective bits per weight for GGUF-style quantizations are averages (Q4_K_M is about 4.85 bits, Q8_0 about 8.5). Real files differ slightly by architecture.
Cas d’usage courants
- Deciding whether a 24 GB card can run a 32B model at 4-bit with a long context.
- Sizing a multi-GPU server before renting hardware.
- Seeing how much memory a longer context window costs.
Exemples concrets & cas limites
Llama 3.1 8B KV cache
Llama 3.1 8B, 8,192 tokens, batch 1, FP16 cache
1.00 GiB of KV cache
128 KiB per token: 2 × 32 layers × 8 KV heads × 128 dims × 2 bytes.
Weights of a 70B model
70B parameters at 16 bits
140 GB of weights
Far beyond one consumer GPU, which is why 70B is usually run at 4-bit or split across cards.
Limites & notes de précision
- • This is an estimate. Inference engines differ in overhead, and activation memory depends on the backend and batch shape.
- • It does not model offloading to system RAM, tensor-parallel duplication or training memory (gradients and optimizer state).
Traitement local dans le navigateur & confidentialité
Les convertisseurs classiques envoient vos documents, extraits ou médias sur des serveurs distants. Calculateur de VRAM pour LLM locaux s’exécute directement dans votre onglet :
Vos textes et fichiers sont traités localement sans aucun envoi vers un serveur de calcul.
Microdeck n’enregistre ni ne stocke les fichiers ou données collés dans les outils.
Les résultats restent dans la mémoire de votre onglet et sont libérés dès que vous le fermez.
Questions fréquentes (FAQ)
Why does the KV cache matter?
It grows linearly with context length and batch size. At very long contexts it can rival the weights, so a model that fits at 4K tokens may not fit at 128K.
Does a 4-bit model lose quality?
Usually a little. Q4_K_M and Q5 variants are widely used because the quality loss is small compared with the memory saved, but it depends on the model and task.
Can I run a model that is larger than my VRAM?
Some runtimes offload layers to system RAM, which works but is much slower. This tool only checks whether everything fits on the GPU.
Outils associés à découvrir
Compteur de Tokens LLM & Estimateur de Coût
Estimez les tokens et coûts API / VRAM des modèles LLM courants
Convertisseur Stockage SI/IEC (Go/Gio) & Temps de Téléchargement
Convertissez Octets, Ko/Kio, Mo/Mio, Go/Gio, To/Tio & Bits et estimez vos durées de transfert avec surcharge TCP
Convertisseur JSON vers Zod, Pydantic & JSON Schema
Transformez vos charges JSON en schémas Zod, Pydantic et JSON Schema