feat: choose model backend automatically.

Author: XuZhang99

README.md

@@ -167,7 +167,6 @@ Run the following command to start xLLM engine:
 ```bash
 ./build/xllm/core/server/xllm \    # launch xllm server
     --model=/path/to/your/llm  \   # model path（to replace with your own path）
-    --backend=llm \                # indicate the LLM backend
     --port=9977 \                  # set service port to 9977
     --max_memory_utilization 0.90  # set the maximal utilization of device memory
 ```

README_zh.md

@@ -168,7 +168,6 @@ python setup.py bdist_wheel
 ```bash
 ./build/xllm/core/server/xllm \    # 启动 xllm 服务器程序
     --model=/path/to/your/llm  \   # 指定模型路径（需替换为实际路径）
-    --backend=llm \                # 指定后端类型为 LLM
     --port=9977 \                  # 设置服务端口为 9977
     --max_memory_utilization 0.90  # 设置最大内存利用率为 90
 ```

docs/en/getting_started/single_node.md

@@ -4,7 +4,6 @@ Start the single-node `xllm` service directly:
 ```bash linenums="1"
 ./build/xllm/core/server/xllm \
     --model=/path/to/your/qwen2-7b  \
-    --backend=llm \
     --port=9977 \
     --max_memory_utilization 0.90
 ```

docs/en/getting_started/start_vlm_service.md

@@ -5,11 +5,11 @@ This document describes how to start a VLM model service based on the xLLM infer
 ## Single Device
 Start the service by executing the following command in the main directory of the `xllm` project:
 ```bash
-ASCEND_RT_VISIBLE_DEVICES=0 ./build/xllm/core/server/xllm --model=/path/to/Qwen2.5-VL-7B-Instruct  --port=12345  --max_memory_utilization 0.90 --backend vlm
+ASCEND_RT_VISIBLE_DEVICES=0 ./build/xllm/core/server/xllm --model=/path/to/Qwen2.5-VL-7B-Instruct  --port=12345  --max_memory_utilization 0.90
 ```
 
 ## Multiple Devices
 Start the service by executing the following command in the main directory of the `xllm` project:
 ```bash
-ASCEND_RT_VISIBLE_DEVICES=0,1 ./build/xllm/core/server/xllm --model=/path/to/Qwen2.5-VL-7B-Instruct  --port=12345  --max_memory_utilization 0.90 --backend vlm
+ASCEND_RT_VISIBLE_DEVICES=0,1 ./build/xllm/core/server/xllm --model=/path/to/Qwen2.5-VL-7B-Instruct  --port=12345  --max_memory_utilization 0.90
 ```

docs/zh/cli_reference.md

@@ -26,7 +26,6 @@ xLLM使用gflags来管理服务启动参数，具体的参数含义如下：
 | `enable_prefix_cache` | bool | true | false | 是否开启prefix cache（DeepSeek暂不支持） |  |
 | `communication_backend` | string | "hccl" | "lccl" | 通信操作采用的后端 |  |
 | `block_size` | int32 | 128 |  | KV Cache存储的block size大小 |  |
-| `backend` | string | "llm" | "vlm" | 模型类型 |  |
 | `task` | string | "generate" | "embed" | 服务类型，生成式或embedding |  |
 | `max_cache_size` | int64 | 0 |  | 可使用的KV Cache大小，单位byte |  |
 

docs/zh/getting_started/single_node.md

@@ -3,7 +3,6 @@
 ```bash linenums="1"
 ./build/xllm/core/server/xllm \
     --model=/path/to/your/qwen2-7b  \
-    --backend=llm \
     --port=9977 \
     --max_memory_utilization 0.90
 ```

docs/zh/getting_started/start_vlm_service.md

@@ -4,11 +4,11 @@
 ## 单卡
 启动服务，在`xllm`工程主目录中执行下面命令：
 ```bash
-ASCEND_RT_VISIBLE_DEVICES=0 ./build/xllm/core/server/xllm --model=/path/to/Qwen2.5-VL-7B-Instruct  --port=12345  --max_memory_utilization 0.90 --backend vlm --devices auto 
+ASCEND_RT_VISIBLE_DEVICES=0 ./build/xllm/core/server/xllm --model=/path/to/Qwen2.5-VL-7B-Instruct  --port=12345  --max_memory_utilization 0.90 --devices auto 
 ```
 
 ## 多卡
 启动服务，在`xllm`工程主目录中执行下面命令：
 ```bash
-ASCEND_RT_VISIBLE_DEVICES=0,1 ./build/xllm/core/server/xllm --model=/path/to/Qwen2.5-VL-7B-Instruct  --port=12345  --max_memory_utilization 0.90 --backend vlm --devices auto
+ASCEND_RT_VISIBLE_DEVICES=0,1 ./build/xllm/core/server/xllm --model=/path/to/Qwen2.5-VL-7B-Instruct  --port=12345  --max_memory_utilization 0.90 --devices auto
 ```

xllm/core/common/global_flags.cpp

@@ -56,10 +56,11 @@ DEFINE_string(model_id, "", "hf model name.");
 
 DEFINE_string(model, "", "Name or path of the huggingface model to use.");
 
-DEFINE_string(backend,
-              "llm",
-              "Choose the backend model type. 'llm' for text-only, "
-              "'vlm' for multimodal (text and images).");
+DEFINE_string(
+    backend,
+    "",
+    "Choose the backend model type. 'llm' for text-only, "
+    "'vlm' for multimodal (text and images), 'dit' for diffusion models.");
 
 DEFINE_string(task,
               "generate",

xllm/models/model_registry.cpp

@@ -71,6 +71,7 @@ void ModelRegistry::register_causallm_factory(const std::string& name,
                                               << " already registered.");
   } else {
     instance->model_registry_[name].causal_lm_factory = factory;
+    instance->model_backend_[name] = "llm";
   }
 }
 
@@ -83,6 +84,7 @@ void ModelRegistry::register_causalvlm_factory(const std::string& name,
                                                << " already registered.");
   } else {
     instance->model_registry_[name].causal_vlm_factory = factory;
+    instance->model_backend_[name] = "vlm";
   }
 }
 
@@ -95,6 +97,7 @@ void ModelRegistry::register_embeddinglm_factory(const std::string& name,
                                                  << " already registered.");
   } else {
     instance->model_registry_[name].embedding_lm_factory = factory;
+    instance->model_backend_[name] = "llm";
   }
 }
 
@@ -107,6 +110,7 @@ void ModelRegistry::register_dit_model_factory(const std::string& name,
                                               << " already registered.");
   } else {
     instance->model_registry_[name].dit_model_factory = factory;
+    instance->model_backend_[name] = "dit";
   }
 }
 
@@ -229,6 +233,11 @@ TokenizerArgsLoader ModelRegistry::get_tokenizer_args_loader(
   return instance->model_registry_[name].tokenizer_args_loader;
 }
 
+std::string ModelRegistry::get_model_backend(const std::string& name) {
+  ModelRegistry* instance = get_instance();
+  return instance->model_backend_[name];
+}
+
 std::unique_ptr<CausalLM> create_llm_model(const ModelContext& context) {
   // get the factory function for the model type from model registry
   auto factory = ModelRegistry::get_causallm_factory(

xllm/models/model_registry.h

@@ -126,8 +126,11 @@ class ModelRegistry {
   static ImageProcessorFactory get_image_processor_factory(
       const std::string& name);
 
+  static std::string get_model_backend(const std::string& name);
+
  private:
   std::unordered_map<std::string, ModelMeta> model_registry_;
+  std::unordered_map<std::string, std::string> model_backend_;
 };
 
 std::unique_ptr<CausalLM> create_llm_model(const ModelContext& context);