How much GPU memory does Velvet-14B need?

About 36 GB for conversations of up to 8,192 tokens: 28.16 GB of weights, the cache for one full-length conversation and the runtime's own memory (an estimate: its settings are behind its licence).

What is the cheapest GPU to run Velvet-14B?

Right now 2x RTX 3090 at $0.48 an hour; RTX PRO 6000 at $1.14 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.

Can Velvet-14B run on an RTX 4090?

Not on one: it needs about 36 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.

Models / Velvet

AlmawaveLabs/Velvet-14B

Text generation · served with vLLM · ↓ 3.2K downloads a month · View on Hugging Face

AlmawaveLabs/Velvet-14B needs about 36 GB of GPU memory for conversations of up to 8,192 tokens (an estimate: its settings are behind its licence). It runs on 2x RTX 3090 from $0.48/h right now; best value: RTX PRO 6000 at $1.14/h.

The model's publisher asks users to accept its licence (apache-2.0). GridPup cannot download this one yet: choose an ungated version below.

Weights: 28 GB
GPU memory needed: about 36 GB (estimate)
Conversation length: 8,192 tokens
Memory · disk: 22.1 GB · 71 GB
Architecture: MistralForCausalLM
Licence: apache-2.0

Hardware and deploy

The publisher has not let GridPup download this licence-gated model yet. An ungated version of it (above) can be deployed.

Questions

How much GPU memory does Velvet-14B need?: About 36 GB for conversations of up to 8,192 tokens: 28.16 GB of weights, the cache for one full-length conversation and the runtime's own memory (an estimate: its settings are behind its licence).
What is the cheapest GPU to run Velvet-14B?: Right now 2x RTX 3090 at $0.48 an hour; RTX PRO 6000 at $1.14 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.
Can Velvet-14B run on an RTX 4090?: Not on one: it needs about 36 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.

Loading…