metadata

base_model: codellama/CodeLlama-70b-Instruct-hf
license: llama2
model_creator: codellama
model_name: CodeLlama 70B Instruct
pipeline_tag: text-generation
quantized_by: Second State Inc.
language:
  - code
tags:
  - llama-2

CodeLlama-70b-Instruct-hf-GGUF

Original Model

codellama/CodeLlama-70b-Instruct-hf

Run with LlamaEdge

LlamaEdge version: v0.2.11 and above

Prompt template

Prompt type: codellama-super-instruct

Prompt string

<s>Source: system\n\n {system_prompt} <step> Source: user\n\n {user_message_1} <step> Source: assistant\n\n {ai_message_1} <step> Source: user\n\n {user_message_2} <step> Source: assistant\nDestination: user\n\n

Reverse prompt: <step> Source: assistant\nEOT: true

Context size: 8192

Run as LlamaEdge service

wasmedge --dir .:. --nn-preload default:GGML:AUTO:CodeLlama-70b-Instruct-hf-Q2_K.gguf llama-api-server.wasm -p codellama-super-instruct -c 1024 --reverse-prompt 'Source: assistant\nEOT: true'

Note that the model only works in the non-streaming mode.

Quantized GGUF Models

Name	Quant method	Bits	Size	Use case
CodeLlama-70b-Instruct-hf-Q2_K.gguf	Q2_K	2	25.5 GB	smallest, significant quality loss - not recommended for most purposes
CodeLlama-70b-Instruct-hf-Q3_K_L.gguf	Q3_K_L	3	36.1 GB	small, substantial quality loss
CodeLlama-70b-Instruct-hf-Q3_K_M.gguf	Q3_K_M	3	33.3 GB	very small, high quality loss
CodeLlama-70b-Instruct-hf-Q3_K_S.gguf	Q3_K_S	3	29.9 GB	very small, high quality loss
CodeLlama-70b-Instruct-hf-Q4_0.gguf	Q4_0	4	38.9 GB	legacy; small, very high quality loss - prefer using Q3_K_M
CodeLlama-70b-Instruct-hf-Q4_K_M.gguf	Q4_K_M	4	41.4 GB	medium, balanced quality - recommended
CodeLlama-70b-Instruct-hf-Q4_K_S.gguf	Q4_K_S	4	39.2 GB	small, greater quality loss
CodeLlama-70b-Instruct-hf-Q5_0.gguf	Q5_0	5	47.5 GB	legacy; medium, balanced quality - prefer using Q4_K_M
CodeLlama-70b-Instruct-hf-Q5_K_M.gguf	Q5_K_M	5	48.8 GB	large, very low quality loss - recommended
CodeLlama-70b-Instruct-hf-Q5_K_S.gguf	Q5_K_S	5	47.5 GB	large, low quality loss - recommended