fix opt fc1/fc2 layer modules should not be quantized (ModelCloud#118)

DeJoker · Jun 29, 2024 · 5bf289a · 5bf289a
1 parent 47a9f4b
commit 5bf289a
Showing 1 changed file with 2 additions and 2 deletions.
diff --git 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
     ]