From 5bf289a1466b1fd9e431b65262fc5b2ec4e0a158 Mon Sep 17 00:00:00 2001 From: Qubitium-ModelCloud Date: Sat, 29 Jun 2024 18:24:54 +0800 Subject: [PATCH] fix opt fc1/fc2 layer modules should not be quantized (#118) --- gptqmodel/models/opt.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/gptqmodel/models/opt.py b/gptqmodel/models/opt.py index 1cbf407e..c2e27c05 100644 --- a/gptqmodel/models/opt.py +++ b/gptqmodel/models/opt.py @@ -15,6 +15,6 @@ class OPTGPTQ(BaseGPTQModel): layer_modules = [ ["self_attn.k_proj", "self_attn.v_proj", "self_attn.q_proj"], ["self_attn.out_proj"], - ["fc1"], - ["fc2"], + # ["fc1"], disabled: not a good candidate for quantization + # ["fc2"], disabled: not a good candidate for quantization ]