[M2M100] update conversion script (#17916)

src/transformers/models/m2m_100/convert_m2m100_original_checkpoint_to_pytorch.py

@@ -44,7 +44,7 @@ def make_linear_from_emb(emb):
 
 def convert_fairseq_m2m100_checkpoint_from_disk(checkpoint_path):
     m2m_100 = torch.load(checkpoint_path, map_location="cpu")
-    args = m2m_100["args"]
+    args = m2m_100["args"] or m2m_100["cfg"]["model"]
     state_dict = m2m_100["model"]
     remove_ignore_keys_(state_dict)
     vocab_size = state_dict["encoder.embed_tokens.weight"].shape[0]
@@ -69,7 +69,7 @@ def convert_fairseq_m2m100_checkpoint_from_disk(checkpoint_path):
 
     state_dict["shared.weight"] = state_dict["decoder.embed_tokens.weight"]
     model = M2M100ForConditionalGeneration(config)
-    model.model.load_state_dict(state_dict)
+    model.model.load_state_dict(state_dict, strict=False)
     model.lm_head = make_linear_from_emb(model.model.shared)
 
     return model