add Yi-VL-34B model

Former-commit-id: a388cadfc0

src/llmtuner/extras/constants.py

@@ -1215,6 +1215,9 @@ register_model_group(
         "YiVL-6B-Chat": {
             DownloadSource.DEFAULT: "BUAADreamer/Yi-VL-6B-hf",
         },
+        "YiVL-34B-Chat": {
+            DownloadSource.DEFAULT: "BUAADreamer/Yi-VL-34B-hf",
+        },
     },
     template="yi_vl",
     vision=True,

src/llmtuner/model/loader.py

@@ -78,8 +78,15 @@ def load_tokenizer(model_args: "ModelArguments") -> "TokenizerModule":
     patch_tokenizer(tokenizer)
 
     if model_args.visual_inputs:
-        processor = AutoProcessor.from_pretrained(model_args.model_name_or_path, **init_kwargs)
-        setattr(processor, "tokenizer", tokenizer)
+        try:
+            processor = AutoProcessor.from_pretrained(model_args.model_name_or_path, **init_kwargs)
+            setattr(processor, "tokenizer", tokenizer)
+        except Exception:
+            raise ValueError(
+                "This multimodal LLM is not supported.\n"
+                "Download LLaVA-1.5 models from: https://huggingface.co/llava-hf\n"
+                "Download Yi-VL models from: https://huggingface.co/BUAADreamer"
+            )
     else:
         processor = None
 

src/llmtuner/model/utils/visual.py

@@ -58,7 +58,7 @@ class LlavaMultiModalProjectorForYiVLForVLLM(LlavaMultiModalProjectorForYiVL):
         self.linear_2 = torch.nn.LayerNorm(text_hidden_size, bias=True)
         self.linear_3 = torch.nn.Linear(text_hidden_size, text_hidden_size, bias=True)
         self.linear_4 = torch.nn.LayerNorm(text_hidden_size, bias=True)
-        self.act = torch.nn.GELU()
+        self.act = ACT2FN[projector_hidden_act]
 
 
 def autocast_projector_dtype(