add deepseekv3 and llama4

vllm-v0.6.2/vllm/transformers_utils/config.py

@@ -13,7 +13,7 @@ from transformers import GenerationConfig, PretrainedConfig
 from transformers.models.auto.image_processing_auto import (
     get_image_processor_config)
 from transformers.models.auto.modeling_auto import (
-    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES)
+    MODEL_FOR_CAUSAL_LM_MAPPING_NAMES, MODEL_MAPPING_NAMES)
 from transformers.utils import CONFIG_NAME as HF_CONFIG_NAME
 
 from vllm.envs import VLLM_USE_MODELSCOPE
@@ -229,13 +229,16 @@ def get_config(
         model_type = MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[config.model_type]
         config.update({"architectures": [model_type]})
 
-    # Some composite config classes (e.g. Llama4Config) may not preserve
-    # the 'architectures' field from config.json. Restore it from the
-    # raw config_dict if needed.
-    if config_format == ConfigFormat.HF:
-        raw_archs = config_dict.get("architectures")
-        if raw_archs and not getattr(config, "architectures", None):
-            config.architectures = raw_archs
+    # Architecture mapping for models without explicit architectures field
+    if not getattr(config, "architectures", None):
+        if config.model_type not in MODEL_MAPPING_NAMES:
+            logger.warning(
+                "Model config does not have a top-level 'architectures' "
+                "field: expecting `hf_overrides={'architectures': "
+                "['...']}` to be passed in engine args.")
+        else:
+            model_type = MODEL_MAPPING_NAMES[config.model_type]
+            config.update({"architectures": [model_type]})
 
     patch_rope_scaling(config)