Skip to content

Latest commit

 

History

History
73 lines (66 loc) · 8.53 KB

File metadata and controls

73 lines (66 loc) · 8.53 KB

Default OpenVINO quantization configs for INT4 quantization

This page lists the default quantization configurations for OpenVINO LLMs. These configs are used when using optimum-cli with --weight-format int4 without specifying additional quantization parameters:

optimum-cli export openvino -m model_id --weight-format int4 model-id-ov-int4

The table shows settings that are explicitly set. If a cell is empty, the setting is not set. For example, for the HuggingFaceH4/zephyr-7b-beta model, AWQ is set, but scale_estimation is not, and no dataset is specified for AWQ (data-free AWQ will be used). The default config for this model is equivalent to doing

optimum-cli export openvino -m HuggingFaceH4/zephyr-7b-beta --sym --group-size 128 --ratio 0.8 --awq zephyr-7b-beta-ov-int4
model sym group_size ratio quant_method dataset scale_estimation all_layers
baichuan-inc/Baichuan2-13B-Chat False 128 1.0 AWQ
baichuan-inc/Baichuan2-7B-Chat False 128 0.8
bigcode/starcoder2-15b False 64 1.0
bigcode/starcoder2-3b False 128 0.9
bigscience/bloomz-560m True 64 0.8 AWQ wikitext2
databricks/dolly-v2-3b False 128 1.0 wikitext2 True
deepseek-ai/DeepSeek-R1-Distill-Llama-8B False 64 0.8 AWQ
deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B False 32 0.7 AWQ
deepseek-ai/DeepSeek-R1-Distill-Qwen-7B False 128 1.0 AWQ
EleutherAI/gpt-j-6b False 64
facebook/opt-2.7b True 128 0.7
facebook/opt-6.7b False 64 0.8
HuggingFaceH4/zephyr-7b-beta True 128 0.8 AWQ
lmsys/longchat-7b-16k False 128 1.0 AWQ wikitext2 True
lmsys/vicuna-7b-v1.5 False 128 1.0
meta-llama/Llama-2-13b-chat-hf True 64 0.8
meta-llama/Llama-2-7b-chat-hf True 128 1.0 AWQ
meta-llama/Llama-2-7b-hf True 128 0.6
meta-llama/Llama-3.1-8B False 64 0.8 AWQ
meta-llama/Llama-3.1-8B-Instruct False 64 0.8 AWQ
meta-llama/Llama-3.2-1B-Instruct False 128 1.0 AWQ
meta-llama/Meta-Llama-3.1-8B False 64 0.8 AWQ
meta-llama/Meta-Llama-3.1-8B-Instruct False 64 0.8 AWQ
microsoft/phi-2 False 64 1.0 AWQ wikitext2 True
microsoft/Phi-3-mini-4k-instruct False 64 1.0 wikitext2 True
microsoft/Phi-3.5-mini-instruct False 64 1.0 AWQ
microsoft/Phi-4-mini-instruct False 64 1.0 AWQ
mistralai/Mistral-7B-v0.1 True 128 0.9
mistralai/Mixtral-8x7B-v0.1 True 128 0.8
openlm-research/open_llama_3b False 64 True
openlm-research/open_llama_3b_v2 False 64 1.0 AWQ wikitext2
pansophic/rocket-3B True 128 0.8
psmathur/orca_mini_3b True 64 True
Qwen/Qwen-7B-Chat True 128 0.6
Qwen/Qwen2.5-1.5B-Instruct False 128 0.9 AWQ wikitext2 True
Qwen/Qwen2.5-7B-Instruct False 128 1.0 AWQ
Qwen/Qwen2.5-Coder-3B-Instruct False 64 1.0 wikitext2 True
Qwen/Qwen3-1.7B True 64 1.0 AWQ wikitext2 True
Qwen/Qwen3-4B True 128 1.0 AWQ
Qwen/Qwen3-8B False 128 1.0 wikitext2 True
stabilityai/stable-code-3b True 64 0.8
stabilityai/stablelm-3b-4e1t True 64 0.8 AWQ
stabilityai/stablelm-tuned-alpha-3b False 128 0.8
stabilityai/stablelm-tuned-alpha-7b False 64 1.0 wikitext2 True
stabilityai/stablelm-zephyr-3b False 128 0.9 wikitext2 True
THUDM/chatglm2-6b True 128 0.72
tiiuae/falcon-7b-instruct False 64
TinyLlama/TinyLlama-1.1B-Chat-v1.0 False 64 1.0 AWQ
togethercomputer/RedPajama-INCITE-7B-Instruct False 128
togethercomputer/RedPajama-INCITE-Chat-3B-v1 False 128 1.0 wikitext2 True

Source

optimum-cli export openvino documentation