How much GPU memory does Fugaku-LLM-13B need?

About 34.2 GB for conversations of up to 8,192 tokens: 26.67 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 Fugaku-LLM-13B?

Right now 4x RTX 3060 at $0.43 an hour; RTX PRO 6000 at $1.20 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.

Can Fugaku-LLM-13B run on an RTX 4090?

Not on one: it needs about 34.2 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.

Models / Fugaku-LLM

Fugaku-LLM/Fugaku-LLM-13B

Text generation · served with vLLM · ↓ 10 downloads a month · View on Hugging Face

Fugaku-LLM/Fugaku-LLM-13B needs about 34.2 GB of GPU memory for conversations of up to 8,192 tokens (an estimate: its settings are behind its licence). It runs on 4x RTX 3060 from $0.43/h right now; best value: RTX PRO 6000 at $1.20/h.

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

Weights: 27 GB
GPU memory needed: about 34.2 GB (estimate)
Conversation length: 8,192 tokens
Memory · disk: 21.3 GB · 70 GB
Architecture: GPT2LMHeadModel
Licence: other

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 Fugaku-LLM-13B need?: About 34.2 GB for conversations of up to 8,192 tokens: 26.67 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 Fugaku-LLM-13B?: Right now 4x RTX 3060 at $0.43 an hour; RTX PRO 6000 at $1.20 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.
Can Fugaku-LLM-13B run on an RTX 4090?: Not on one: it needs about 34.2 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.

Loading…