Llama-3-8B-Instruct-Gradient-4194k-GGUF

Fixing prompt format issues

• Use iMatrix for Llama 3 prompt format on Q4 and below, or try Q4_K_M fixed
• Use ChatML for Q6 and below
• Use Llama 3, see issues

Issues

• Context length is not defined correctly in quant, not sure if this is a llama.cpp issue Use RoPE settings
• Output ends with or other EOS tokens, might be an issue with their training data

This model was converted to GGUF format from gradientai/Llama-3-8B-Instruct-Gradient-4194k using llama.cpp via the ggml.ai's GGUF-my-repo space. Refer to the original model card for more details on the model.

Filename	Quant	Size	Description
llama-3-8b-instruct-gradient-4194k.f16.gguf	f16	14.9GB	Lossless
llama-3-8b-instruct-gradient-4194k.Q8_0.gguf	Q8_0	8.54GB	Extremely high quality
llama-3-8b-instruct-gradient-4194k.Q6_K.gguf	Q6_K	6.60GB	Very high quality, near perfect, recommended.
llama-3-8b-instruct-gradient-4194k.Q5_K_M.gguf	Q5_K_M	5.73GB	High quality
llama-3-8b-instruct-gradient-4194k.Q5_K_S.gguf	Q5_K_S	5.60GB	Even higher quality
llama-3-8b-instruct-gradient-4194k.Q4_K_M.gguf	Q4_K_M	4.92GB	Recommended, medium-high quality
llama-3-8b-instruct-gradient-4194k.Q4_K_M.fixed.gguf	Q4_K_M (fixed)	4.92GB	Fixed version (requanted)
llama-3-8b-instruct-gradient-4194k.Q4_K_S.gguf	Q4_K_S	4.69GB	Recommended, medium quality
llama-3-8b-instruct-gradient-4194k.Q4_0.gguf	Q4_0	4.66GB	Usable, better than Q3 but worse than Q4
llama-3-8b-instruct-gradient-4194k.Q3_K_L.gguf	Q3_K_L	4.32GB	Usable
llama-3-8b-instruct-gradient-4194k.Q3_K_M.gguf	Q3_K_M	4.02GB	Bad quality, use Q4
llama-3-8b-instruct-gradient-4194k.Q3_K_S.gguf	Q3_K_S	3.66GB	Not recommended
llama-3-8b-instruct-gradient-4194k.Q2_K.gguf	Q2_K	2.95GB	Very low quality, would not use on 8b models

Use with llama.cpp

Install llama.cpp through brew.

brew install ggerganov/ggerganov/llama.cpp

Invoke the llama.cpp server or the CLI.

CLI:

llama-cli --hf-repo leafspark/llama-3-8b-instruct-gradient-4194k.Q8_0-GGUF --model llama-3-8b-instruct-gradient-4194k.Q8_0.gguf -p "The meaning to life and the universe is"

Server:

llama-server --hf-repo leafspark/llama-3-8b-instruct-gradient-4194k.Q8_0-GGUF --model llama-3-8b-instruct-gradient-4194k.Q8_0.gguf -c 2048

Note: You can also use this checkpoint directly through the usage steps listed in the Llama.cpp repo as well.

git clone https://github.com/ggerganov/llama.cpp &&             cd llama.cpp &&             make &&             ./main -m llama-3-8b-instruct-gradient-4194k.Q8_0.gguf -n 128