Llama-3-8B-Instruct-OmniQuant / compressa-config.json
Vasily Alexeev
add asymm quantized model, add two eos in code sample
6758e8a
raw
history blame
732 Bytes
{
"model_name": "resources/models/models/NousResearch_Meta-Llama-3-8B-Instruct",
"quant_config": {
"wbits": 4,
"abits": 16,
"group_size": 128,
"symmetric": false
},
"resume": null,
"start_sample": 0,
"nsamples": 128,
"epochs": 20,
"aug_loss": true,
"eval_ppl": true,
"real_quant": true,
"lwc_lr": 0.01,
"use_lr_scheduler": false,
"cache_dir": "resources/cache",
"output_dir": "resources/models/models/NousResearch_Meta-Llama-3-8B-Instruct_omniquant_asymm_e20/logs",
"save_dir": "resources/models/models/NousResearch_Meta-Llama-3-8B-Instruct_omniquant_asymm_e20/NousResearch_Meta-Llama-3-8B-Instruct",
"config_class": "OmniquantConfig"
}