How much GPU memory does meditron-70b need?

About 167.8 GB for conversations of up to 8,192 tokens: 137.95 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 meditron-70b?

Right now 8x RTX 3090 at $1.53 an hour; 2x RTX PRO 6000 at $3.47 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.

Can meditron-70b run on an RTX 4090?

Not on one: it needs about 167.8 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.

Models / meditron

epfl-llm/meditron-70b

Text generation · served with vLLM · ↓ 116 downloads a month · View on Hugging Face

epfl-llm/meditron-70b needs about 167.8 GB of GPU memory for conversations of up to 8,192 tokens (an estimate: its settings are behind its licence). It runs on 8x RTX 3090 from $1.53/h right now; best value: 2x RTX PRO 6000 at $3.47/h.

The model's publisher asks users to accept its licence (llama2). GridPup cannot download this one yet: choose an ungated version below.

Weights: 138 GB
GPU memory needed: about 167.8 GB (estimate)
Conversation length: 8,192 tokens
Memory · disk: 77 GB · 192 GB
Architecture: LlamaForCausalLM
Licence: llama2

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 meditron-70b need?: About 167.8 GB for conversations of up to 8,192 tokens: 137.95 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 meditron-70b?: Right now 8x RTX 3090 at $1.53 an hour; 2x RTX PRO 6000 at $3.47 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.
Can meditron-70b run on an RTX 4090?: Not on one: it needs about 167.8 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.

Loading…