fix

Author: Jintao-Huang

src/mcore_bridge/model/mm_gpt_model.py

@@ -46,10 +46,11 @@ def forward(_self, input_):
             res = origin_forward(_self, input_)
             _self.reduce_scatter_embeddings = reduce_scatter_embeddings
             packed_seq_params = kwargs.get('packed_seq_params')
-            if self.config.language_model_only:
-                res = self.visual.get_inputs_embeds_language_model(res, **kwargs)
-            else:
-                res = self.visual.get_inputs_embeds(res, **kwargs)
+            if self.visual is not None:
+                if self.config.language_model_only:
+                    res = self.visual.get_inputs_embeds_language_model(res, **kwargs)
+                else:
+                    res = self.visual.get_inputs_embeds(res, **kwargs)
             kwargs.clear()
             if isinstance(res, dict):
                 # compat dict

src/mcore_bridge/model/mm_gpts/gemma4.py

@@ -132,7 +132,7 @@ def get_inputs_embeds(self, inputs_embeds, **kwargs):
             audio_mask_e = audio_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
             inputs_embeds = inputs_embeds.masked_scatter(audio_mask_e, audio_features)
         res['inputs_embeds'] = inputs_embeds
-        res['llm_input_ids'] = input_ids
+        res['llm_input_ids'] = llm_input_ids
         return res