Revert "fix opt fc1/fc2 layer modules should not be quantized (#118)" (#149)

Qubitium · web-flow · commit 83c002d9fcf2 · 2024-07-03T00:00:09.000+08:00
This reverts commit c9a0688.
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"], disabled: not a good candidate for quantization
-        # ["fc2"], disabled: not a good candidate for quantization
+        ["fc1"],
+        ["fc2"],
     ]

Original file line number	Diff line number	Diff line change
`@@ -15,6 +15,6 @@ class OPTGPTQ(BaseGPTQModel):`
`15`	`15`	`layer_modules = [`
`16`	`16`	`["self_attn.k_proj", "self_attn.v_proj", "self_attn.q_proj"],`
`17`	`17`	`["self_attn.out_proj"],`
`18`		`- # ["fc1"], disabled: not a good candidate for quantization`
`19`		`- # ["fc2"], disabled: not a good candidate for quantization`
	`18`	`+ ["fc1"],`
	`19`	`+ ["fc2"],`
`20`	`20`	`]`