add deepseekv3 and llama4

vllm-v0.6.2/vllm_mlu/vllm_mlu/model_executor/models/llama4.py

@@ -389,6 +389,15 @@ def vllm__llama4__Llama4ForCausalLM__load_weights(
         if "rotary_emb.inv_freq" in name:
             continue
 
+        # Strip language_model. prefix for Llama4ForConditionalGeneration
+        if name.startswith("language_model."):
+            name = name[len("language_model."):]
+        # Skip vision encoder weights
+        elif (name.startswith("multi_modal_projector.")
+              or name.startswith("vision_encoder.")
+              or name.startswith("vision_model.")):
+            continue
+
         # Permute Q/K weights for rotary embedding
         name, loaded_weight = self.permute_qk_weight_for_rotary(
             name, loaded_weight)