[Models] Support DeepSeek-V4 GGUF weights mapping and architecture fa…

tests/test_plugin.py

@@ -221,6 +221,33 @@ def fake_parse(
     assert config.architectures == ["Qwen3MoeForCausalLM"]
 
 
+def test_gguf_config_parser_fallback_to_config_architectures(tmp_path, monkeypatch):
+    gguf_path = tmp_path / "model.gguf"
+    gguf_path.write_bytes(b"GGUF")
+
+    def fake_parse(
+        self, model, trust_remote_code, revision=None, code_revision=None, **kwargs
+    ):
+        cfg = PretrainedConfig(model_type="custom_or_new_model")
+        cfg.architectures = ["DeepseekV4ForCausalLM"]
+        return {}, cfg
+
+    monkeypatch.setattr(
+        gguf_config_parser_module.HFConfigParser,
+        "parse",
+        fake_parse,
+    )
+    monkeypatch.setattr(
+        gguf_config_parser_module,
+        "maybe_patch_hf_config_from_gguf",
+        lambda model, config: config,
+    )
+
+    config_dict, config = GGUFConfigParser().parse(gguf_path, trust_remote_code=False)
+    assert config_dict["architectures"] == ["DeepseekV4ForCausalLM"]
+    assert config.architectures == ["DeepseekV4ForCausalLM"]
+
+
 def test_register_sets_engine_args_for_gguf_model(monkeypatch):
     register()
     captured = {}

vllm_gguf_plugin/config_parser.py

@@ -46,6 +46,8 @@ def parse(
             and config.model_type in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
         ):
             architecture = MODEL_FOR_CAUSAL_LM_MAPPING_NAMES[config.model_type]
+        if architecture is None and getattr(config, "architectures", None):
+            architecture = config.architectures[0]
         if architecture is None:
             raise RuntimeError(f"Can't get gguf config for {config.model_type}.")
 

vllm_gguf_plugin/weights_adapter/transformers.py

@@ -62,7 +62,7 @@ def build_name_map(
             model_type = "command-r"
         if model_type == "gemma3_text":
             model_type = "gemma3"
-        if model_type in ("deepseek_v3", "deepseek_v2"):
+        if model_type in ("deepseek_v4", "deepseek_v3", "deepseek_v2"):
             model_type = "deepseek2"
             for idx in range(config.num_hidden_layers):
                 gguf_to_hf_name_map[f"blk.{idx}.exp_probs_b.bias"] = (