[Bugfix] Weight loading fix for OPT model (vllm-project#9042)

domenVres · dvres · Alvant · commit f7607cebd6c5 · 2024-10-26T10:11:40.000+03:00
Co-authored-by: dvres &lt;dvres@fri.uni-lj.si&gt;
Signed-off-by: Alvant &lt;alvasian@yandex.ru&gt;
diff --git a/vllm/model_executor/models/opt.py b/vllm/model_executor/models/opt.py
@@ -353,7 +353,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         ]
         params_dict = dict(self.named_parameters(remove_duplicate=False))
         for name, loaded_weight in weights:
-            if "lm_head.weight" in name:
+            if "lm_head.weight" in name and self.config.tie_word_embeddings:
                 continue
             if name.startswith("decoder."):
                 name = "model." + name

Original file line number	Diff line number	Diff line change
`@@ -353,7 +353,7 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):`
`353`	`353`	`]`
`354`	`354`	`params_dict = dict(self.named_parameters(remove_duplicate=False))`
`355`	`355`	`for name, loaded_weight in weights:`
`356`		`- if "lm_head.weight" in name:`
	`356`	`+ if "lm_head.weight" in name and self.config.tie_word_embeddings:`
`357`	`357`	`continue`
`358`	`358`	`if name.startswith("decoder."):`
`359`	`359`	`name = "model." + name`