app-store/apps/serge/metadata/description.md
2023-09-29 07:58:18 +02:00

3.3 KiB

Serge - LLaMA made easy 🦙

Serge is a chat interface crafted with llama.cpp for running Alpaca models. No API keys, entirely self-hosted!

  • 🌐 SvelteKit frontend
  • 💾 Redis for storing chat history & parameters
  • ⚙️ FastAPI + LangChain for the API, wrapping calls to llama.cpp using the python bindings

🧠 Supported Models

Category Models
Alpaca 🦙 Alpaca-LoRA-65B, GPT4-Alpaca-LoRA-30B
Chronos 🌑 Chronos-13B, Chronos-33B, Chronos-Hermes-13B
GPT4All 🌍 GPT4All-13B
Koala 🐨 Koala-7B, Koala-13B
LLaMA 🦙 FinLLaMA-33B, LLaMA-Supercot-30B, LLaMA2 7B, LLaMA2 13B, LLaMA2 70B
Lazarus 💀 Lazarus-30B
Nous 🧠 Nous-Hermes-13B
OpenAssistant 🎙️ OpenAssistant-30B
Orca 🐬 Orca-Mini-v2-7B, Orca-Mini-v2-13B, OpenOrca-Preview1-13B
Samantha 👩 Samantha-7B, Samantha-13B, Samantha-33B
Vicuna 🦙 Stable-Vicuna-13B, Vicuna-CoT-7B, Vicuna-CoT-13B, Vicuna-v1.1-7B, Vicuna-v1.1-13B, VicUnlocked-30B, VicUnlocked-65B
Wizard 🧙 Wizard-Mega-13B, WizardLM-Uncensored-7B, WizardLM-Uncensored-13B, WizardLM-Uncensored-30B, WizardCoder-Python-13B-V1.0

Additional weights can be added to the serge_weights volume using docker cp:

docker cp ./my_weight.bin serge:/usr/src/app/weights/

⚠️ Memory Usage

LLaMA will crash if you don't have enough available memory for the model:

Model Max RAM Required
7B 4.5GB
7B-q2_K 5.37GB
7B-q3_K_L 6.10GB
7B-q4_1 6.71GB
7B-q4_K_M 6.58GB
7B-q5_1 7.56GB
7B-q5_K_M 7.28GB
7B-q6_K 8.03GB
7B-q8_0 9.66GB
13B 12GB
13B-q2_K 8.01GB
13B-q3_K_L 9.43GB
13B-q4_1 10.64GB
13B-q4_K_M 10.37GB
13B-q5_1 12.26GB
13B-q5_K_M 11.73GB
13B-q6_K 13.18GB
13B-q8_0 16.33GB
33B 20GB
33B-q2_K 16.21GB
33B-q3_K_L 19.78GB
33B-q4_1 22.83GB
33B-q4_K_M 22.12GB
33B-q5_1 26.90GB
33B-q5_K_M 25.55GB
33B-q6_K 29.19GB
33B-q8_0 37.06GB
65B 50GB
65B-q2_K 29.95GB
65B-q3_K_L 37.15GB
65B-q4_1 43.31GB
65B-q4_K_M 41.85GB
65B-q5_1 51.47GB
65B-q5_K_M 48.74GB
65B-q6_K 56.06GB
65B-q8_0 71.87GB

🧾 License

Nathan Sarrazin and Contributors. Serge is free and open-source software licensed under the MIT License.

💬 Support

Need help? Join our Discord