Text generation · served with vLLM · ↓ 847 downloads a month · View on Hugging Face
codellama/CodeLlama-70b-Instruct-hf needs about 157 GB of GPU memory for conversations of up to 4,096 tokens. It runs on 4x RTX A6000 from $1.86/h right now; best value: 2x RTX PRO 6000D at $2.96/h.
Weights
138 GB
GPU memory needed
about 157 GB
Conversation length
4,096 tokens
Model allows
4,096 tokens
Memory · disk
77 GB · 192 GB
Architecture
LlamaForCausalLM
Licence
llama2
Hardware and deploy
$2.96/h from your balance while it runs, billed by the hour; stop it from its page. GridPup rents the machine, which downloads the model from Hugging Face (usually ready in 10 to 15 minutes), and gives you an https address, keys and a Playground.
Questions
How much GPU memory does CodeLlama-70b-Instruct-hf need?
About 157 GB for conversations of up to 4,096 tokens: 137.95 GB of weights, the cache for one full-length conversation and the runtime's own memory.
What is the cheapest GPU to run CodeLlama-70b-Instruct-hf?
Right now 4x RTX A6000 at $1.86 an hour; 2x RTX PRO 6000D at $2.96 an hour is the best value for its speed. Prices are live offers from GPU clouds, checked every 15 minutes.
Can CodeLlama-70b-Instruct-hf run on an RTX 4090?
Not on one: it needs about 157 GB and an RTX 4090 has 24 GB. Use a bigger GPU, or a smaller (quantized) version of the model.
How do I deploy CodeLlama-70b-Instruct-hf as an API?
Pick a machine on its GridPup page and press Deploy: GridPup rents the GPU, downloads the model from Hugging Face and serves it at an https address with keys and OpenAI's chat API, billed by the hour.
Prices are per hour for the whole machine and its disk, right now. Speed is Vast's DLPerf score of the machine, relative to the cheapest. Best value: the most speed per dollar. Only machines with a Vast reliability above 95% that download at 400 Mbit/s or more are offered: the machine fetches the model itself.