llm-quantisation-methods.md

2024-03-27 - TIL - LLM Quantisation Methods

Quantisation methods for LLMs.

TL;DR

• Q4_K_M is recommended for most LLMs (medium, balanced quality)
• Q5_K_S is recommended for low quality loss (large, low quality loss)
• Q5_K_M is recommended for very low quality loss (large, very low quality loss)

Quantisation Methods

Name	Quant method	Bits	Size	Max RAM required	Use case
dolphin-2.2.1-mistral-7b.Q2_K.gguf	Q2_K	2	3.08 GB	5.58 GB	smallest, significant quality loss - not recommended for most purposes
dolphin-2.2.1-mistral-7b.Q3_K_S.gguf	Q3_K_S	3	3.16 GB	5.66 GB	very small, high quality loss
dolphin-2.2.1-mistral-7b.Q3_K_M.gguf	Q3_K_M	3	3.52 GB	6.02 GB	very small, high quality loss
dolphin-2.2.1-mistral-7b.Q3_K_L.gguf	Q3_K_L	3	3.82 GB	6.32 GB	small, substantial quality loss
dolphin-2.2.1-mistral-7b.Q4_0.gguf	Q4_0	4	4.11 GB	6.61 GB	legacy; small, very high quality loss - prefer using Q3_K_M
dolphin-2.2.1-mistral-7b.Q4_K_S.gguf	Q4_K_S	4	4.14 GB	6.64 GB	small, greater quality loss
dolphin-2.2.1-mistral-7b.Q4_K_M.gguf	Q4_K_M	4	4.37 GB	6.87 GB	medium, balanced quality - recommended
dolphin-2.2.1-mistral-7b.Q5_0.gguf	Q5_0	5	5.00 GB	7.50 GB	legacy; medium, balanced quality - prefer using Q4_K_M
dolphin-2.2.1-mistral-7b.Q5_K_S.gguf	Q5_K_S	5	5.00 GB	7.50 GB	large, low quality loss - recommended
dolphin-2.2.1-mistral-7b.Q5_K_M.gguf	Q5_K_M	5	5.13 GB	7.63 GB	large, very low quality loss - recommended
dolphin-2.2.1-mistral-7b.Q6_K.gguf	Q6_K	6	5.94 GB	8.44 GB	very large, extremely low quality loss
dolphin-2.2.1-mistral-7b.Q8_0.gguf	Q8_0	8	7.70 GB	10.20 GB	very large, extremely low quality loss - not recommended

Note: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.

Use Cases

• Quantisation methods for LLMs cheatsheet
• Helpful reference to use when selecting a quantisation method

References

• TheBloke/dolphin-2.2.1-mistral-7B-GGUF · Hugging Face