feat(ai): 调整模型

service-ai/service-ai-chat/src/main/resources/application.yml

@@ -9,7 +9,7 @@ spring:
       api-key: ENC(K+Hff9QGC+fcyi510VIDd9CaeK/IN5WBJ9rlkUsHEdDgIidW+stHHJlsK0lLPUXXREha+ToQZqqDXJrqSE+GUKCXklFhelD8bRHFXBIeP/ZzT2cxhzgKUXgjw3S0Qw2R)
       chat:
         options:
-          model: 'Qwen3-1.7-vllm'
+          model: 'Qwen3/qwen3-1.7b'
   mvc:
     async:
       request-timeout: 3600000

service-ai/service-ai-chat/src/test/java/com/lanyuanxiaoyao/service/ai/chat/TestModel.java

@@ -1,7 +1,9 @@
 package com.lanyuanxiaoyao.service.ai.chat;
 
+import cn.hutool.core.util.StrUtil;
 import java.io.IOException;
 import java.net.http.HttpClient;
+import java.time.Duration;
 import java.util.Arrays;
 import java.util.List;
 import org.noear.solon.ai.rag.Document;
@@ -25,12 +27,12 @@ import org.springframework.web.reactive.function.client.WebClient;
  * @author lanyuanxiaoyao
  * @version 20250514
  */
-public class TestSpringAIToolChat {
+public class TestModel {
   public static void main(String[] args) throws IOException {
-    testChatModel();
+    // testChatModel();
     testVisualModel();
-    // testEmbeddingModel();
+    testEmbeddingModel();
-    // testRerankingModel();
+    testRerankingModel();
   }
 
   private static void testChatModel() {
@@ -41,12 +43,14 @@ public class TestSpringAIToolChat {
         "Qwen3/qwen3-4b-q4km",
         "Qwen3/qwen3-8b-q4km"
     )) {
+      System.out.println(model);
+      long start = System.currentTimeMillis();
       ChatClient client = chatClient(model);
       String content = client.prompt()
-          .user("你好")
+          .user("你好，详细介绍一下是谁，能帮我做什么？")
           .call()
           .content();
-      System.out.println(content);
+      System.out.println(content.length() * 1000.0 / (System.currentTimeMillis() - start));
     }
   }
 
@@ -55,19 +59,17 @@ public class TestSpringAIToolChat {
         "Qwen2.5/qwen2.5-vl-7b",
         "Qwen2.5/qwen2.5-vl-7b-q4km",
         "Qwen2.5/qwen2.5-vl-3b-instruct",
-        "Qwen2.5/qwen2.5-vl-3b-instruct-awq",
         "Qwen2.5/qwen2.5-vl-7b-instruct",
-        "Qwen2.5/qwen2.5-vl-7b-instruct-awq",
         "MiniCPM/minicpm-o-2.6-7.6b",
         "MiniCPM/minicpm-o-2.6-7.6b-q4km"
 
     )) {
       ChatClient client = chatClient(model);
       String content = client.prompt()
-          .user(spec -> spec.text("图片中有什么").media(MimeTypeUtils.IMAGE_PNG, new FileSystemResource("/Users/lanyuanxiaoyao/Pictures/deepseek.png")))
+          .user(spec -> spec.text("根据图片中的内容编一个童话小故事").media(MimeTypeUtils.IMAGE_PNG, new FileSystemResource("/Users/lanyuanxiaoyao/Pictures/deepseek.png")))
           .call()
           .content();
-      System.out.println(content);
+      System.out.println(StrUtil.trim(content));
     }
   }
 
@@ -78,9 +80,7 @@ public class TestSpringAIToolChat {
         "Qwen3/qwen3-embedding-4b-q4km",
         "Qwen3/qwen3-embedding-8b-q4km",
         "BGE/bge-m3",
-        "BGE/bge-m3-q4km",
+        "BGE/bge-m3-q4km"
-        "MiniCPM/minicpm-embedding",
-        "MiniCPM/minicpm-embedding-light"
     )) {
       EmbeddingModel embeddingModel = embeddingModel(model);
       float[] worlds = embeddingModel.embed("Hello world");
@@ -90,12 +90,10 @@ public class TestSpringAIToolChat {
 
   private static void testRerankingModel() throws IOException {
     for (String model : List.of(
-        "BGE/beg-reranker-v2",
-        "MiniCPM/minicpm-reranker",
-        "MiniCPM/minicpm-reranker-light",
         "BGE/beg-reranker-v2",
         "BGE/beg-reranker-v2-q4km"
     )) {
+      System.out.println(model);
       RerankingModel rerankingModel = rerankingModel(model);
       List<Document> list = rerankingModel.rerank(
           "你好",
@@ -149,6 +147,7 @@ public class TestSpringAIToolChat {
     return RerankingModel.of("http://132.121.206.65:10086/v1/rerank")
         .model(model)
         .apiKey("*XMySqV%>hR&v>>g*NwCs3tpQ5FVMFEF2VHVTj<MYQd$&@$sY7CgqNyea4giJi4")
+        .timeout(Duration.ofMinutes(10))
         .build();
   }
 

service-ai/service-ai-chat/src/test/java/com/lanyuanxiaoyao/service/ai/chat/TestSpringAiTools.java

@@ -0,0 +1,73 @@
+package com.lanyuanxiaoyao.service.ai.chat;
+
+import cn.hutool.core.util.StrUtil;
+import java.net.http.HttpClient;
+import org.springframework.ai.chat.client.ChatClient;
+import org.springframework.ai.openai.OpenAiChatModel;
+import org.springframework.ai.openai.OpenAiChatOptions;
+import org.springframework.ai.openai.api.OpenAiApi;
+import org.springframework.ai.tool.annotation.Tool;
+import org.springframework.ai.tool.annotation.ToolParam;
+import org.springframework.http.client.JdkClientHttpRequestFactory;
+import org.springframework.http.client.reactive.JdkClientHttpConnector;
+import org.springframework.web.client.RestClient;
+import org.springframework.web.reactive.function.client.WebClient;
+import reactor.core.Disposable;
+
+/**
+ * @author lanyuanxiaoyao
+ * @version 20250613
+ */
+public class TestSpringAiTools {
+  public static void main(String[] args) {
+    ChatClient client = ChatClient.builder(
+            OpenAiChatModel.builder()
+                .openAiApi(
+                    OpenAiApi.builder()
+                        .baseUrl("http://132.121.206.65:10086")
+                        .apiKey("*XMySqV%>hR&v>>g*NwCs3tpQ5FVMFEF2VHVTj<MYQd$&@$sY7CgqNyea4giJi4")
+                        .restClientBuilder(restClientBuilder())
+                        .webClientBuilder(webClientBuilder())
+                        .build()
+                )
+                .defaultOptions(
+                    OpenAiChatOptions.builder()
+                        .model("Qwen3/qwen3-1.7b")
+                        .build()
+                )
+                .build()
+        )
+        .build();
+    Disposable disposable = client.prompt()
+        .tools(new TestTool())
+        .user("调用submit工具生成一个关于「猪」的笑话")
+        .stream()
+        .content()
+        .subscribe(System.out::println);
+    while (!disposable.isDisposed()) {
+    }
+  }
+
+  private static HttpClient httpClient() {
+    return HttpClient.newBuilder()
+        .version(HttpClient.Version.HTTP_1_1)
+        .build();
+  }
+
+  private static RestClient.Builder restClientBuilder() {
+    return RestClient.builder()
+        .requestFactory(new JdkClientHttpRequestFactory(httpClient()));
+  }
+
+  private static WebClient.Builder webClientBuilder() {
+    return WebClient.builder()
+        .clientConnector(new JdkClientHttpConnector(httpClient()));
+  }
+
+  public static final class TestTool {
+    @Tool(description = "传入任意动物名称，返回一个关于这个动物的笑话")
+    public String submit(@ToolParam(description = "动物名称") String animalName) {
+      return StrUtil.format("{}掉沟里了", animalName);
+    }
+  }
+}

service-ai/service-ai-cli/src/main/java/com/lanyuanxiaoyao/service/ai/cli/LlamaSwapTool.java

@@ -25,39 +25,32 @@ public class LlamaSwapTool extends GeneratorTool {
         Map.of(
             "models", List.of(
                 llamaCppEmbeddingCmd("BGE/bge-m3-q4km", "bge-m3-Q4_K_M.gguf", 20),
-                vllmEmbeddingCmd("BGE/bge-m3", "bge-m3", 20, 5),
+                vllmEmbeddingCmd("BGE/bge-m3", "bge-m3", 5),
                 llamaCppRerankerCmd("BGE/beg-reranker-v2-q4km", "bge-reranker-v2-m3-Q4_K_M.gguf", 20),
-                llamaCppRerankerCmd("BGE/beg-reranker-v2", "bge-reranker-v2-m3", 20),
+                vllmRerankerCmd("BGE/beg-reranker-v2", "bge-reranker-v2-m3", 5),
 
-                vllmCmd("Qwen3/qwen3-0.6b", "Qwen3-0.6B", 35, 5, true),
+                vllmCmd("Qwen3/qwen3-0.6b", "Qwen3-0.6B", 5, true),
-                vllmCmd("Qwen3/qwen3-1.7b", "Qwen3-1.7B", 35, 5, true),
+                vllmCmd("Qwen3/qwen3-1.7b", "Qwen3-1.7B", 5, true),
-                vllmCmd("Qwen3/qwen3-4b", "Qwen3-4B", 35, 8, true),
+                vllmCmd("Qwen3/qwen3-4b", "Qwen3-4B", 8, true),
                 llamaCppCmd("Qwen3/qwen3-4b-q4km", "Qwen3-4B-Q4_K_M.gguf", 35),
                 llamaCppCmd("Qwen3/qwen3-8b-q4km", "Qwen3-8B-Q4_K_M.gguf", 35),
 
-                vllmEmbeddingCmd("Qwen3/qwen3-embedding-0.6b", "Qwen3-Embedding-0.6B", 35, 5),
+                vllmEmbeddingCmd("Qwen3/qwen3-embedding-0.6b", "Qwen3-Embedding-0.6B", 5),
-                vllmEmbeddingCmd("Qwen3/qwen3-embedding-4b", "Qwen3-Embedding-4B", 35, 8),
+                vllmEmbeddingCmd("Qwen3/qwen3-embedding-4b", "Qwen3-Embedding-4B", 8),
                 llamaCppEmbeddingCmd("Qwen3/qwen3-embedding-4b-q4km", "Qwen3-Embedding-4B-Q4_K_M.gguf", 35),
                 llamaCppEmbeddingCmd("Qwen3/qwen3-embedding-8b-q4km", "Qwen3-Embedding-8B-Q4_K_M.gguf", 35),
 
-                vllmRerankerCmd("Qwen3/qwen3-reranker-0.6b", "Qwen3-Reranker-0.6B", 35, 5),
+                // 0.9.1 vllm还未支持
-                vllmRerankerCmd("Qwen3/qwen3-reranker-4b", "Qwen3-Reranker-4B", 35, 8),
+                // vllmRerankerCmd("Qwen3/qwen3-reranker-0.6b", "Qwen3-Reranker-0.6B", 5),
+                // vllmRerankerCmd("Qwen3/qwen3-reranker-4b", "Qwen3-Reranker-4B", 8),
 
                 llamaCppVisualCmd("Qwen2.5/qwen2.5-vl-7b", "Qwen2.5-VL-7B-Instruct-BF16.gguf", 35),
                 llamaCppVisualCmd("Qwen2.5/qwen2.5-vl-7b-q4km", "Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf", 35),
-                vllmCmd("Qwen2.5/qwen2.5-vl-7b-instruct", "Qwen2.5-VL-7B-Instruct", 35, 8, false),
+                vllmCmd("Qwen2.5/qwen2.5-vl-3b-instruct", "Qwen2.5-VL-3B-Instruct", 8, false),
-                vllmCmd("Qwen2.5/qwen2.5-vl-7b-instruct-awq", "Qwen2.5-VL-7B-Instruct-AWQ", 35, 8, false),
+                vllmCmd("Qwen2.5/qwen2.5-vl-7b-instruct", "Qwen2.5-VL-7B-Instruct", 8, false),
-                vllmCmd("Qwen2.5/qwen2.5-vl-3b-instruct", "Qwen2.5-VL-3B-Instruct", 35, 8, false),
-                vllmCmd("Qwen2.5/qwen2.5-vl-3b-instruct-awq", "Qwen2.5-VL-3B-Instruct-AWQ", 35, 8, false),
 
                 llamaCppVisualCmd("MiniCPM/minicpm-o-2.6-7.6b-q4km", "MiniCPM-o-2_6-7.6B-Q4_K_M.gguf", 35),
-                vllmCmd("MiniCPM/minicpm-o-2.6-7.6b", "MiniCPM-o-2_6", 35, 10, false),
+                vllmCmd("MiniCPM/minicpm-o-2.6-7.6b", "MiniCPM-o-2_6", 10, false)
-
-                vllmEmbeddingCmd("MiniCPM/minicpm-embedding", "MiniCPM-Embedding", 20, 8),
-                vllmEmbeddingCmd("MiniCPM/minicpm-embedding-light", "MiniCPM-Embedding-Light", 20, 5),
-
-                vllmEmbeddingCmd("MiniCPM/minicpm-reranker", "MiniCPM-Reranker", 20, 8),
-                vllmEmbeddingCmd("MiniCPM/minicpm-reranker-light", "MiniCPM-Reranker-Light", 20, 5)
             )
         ),
         "config.yaml"
@@ -114,23 +107,23 @@ public class LlamaSwapTool extends GeneratorTool {
     );
   }
 
-  private DockerCmd vllmCmd(String name, String model, Integer thread, Integer cache, Boolean isReasonable) {
+  private DockerCmd vllmCmd(String name, String model, Integer cache, Boolean isReasonable) {
-    return vllmCmd(name, model, thread, cache, false, false, isReasonable, false);
+    return vllmCmd(name, model, cache, false, false, isReasonable);
   }
 
-  private DockerCmd vllmEmbeddingCmd(String name, String model, Integer thread, Integer cache) {
+  private DockerCmd vllmEmbeddingCmd(String name, String model, Integer cache) {
-    return vllmCmd(name, model, thread, cache, true, false, false, false);
+    return vllmCmd(name, model, cache, true, false, false);
   }
 
-  private DockerCmd vllmRerankerCmd(String name, String model, Integer thread, Integer cache) {
+  private DockerCmd vllmRerankerCmd(String name, String model, Integer cache) {
-    return vllmCmd(name, model, thread, cache, false, true, false, false);
+    return vllmCmd(name, model, cache, false, true, false);
   }
 
-  private DockerCmd vllmVisualCmd(String name, String model, Integer thread, Integer cache, Boolean isReasonable) {
+  private DockerCmd vllmVisualCmd(String name, String model, Integer cache, Boolean isReasonable) {
-    return vllmCmd(name, model, thread, cache, false, false, isReasonable, true);
+    return vllmCmd(name, model, cache, false, false, isReasonable);
   }
 
-  private DockerCmd vllmCmd(String name, String model, Integer thread, Integer cache, Boolean isEmbedding, Boolean isReranker, Boolean isReasonable, Boolean isVisual) {
+  private DockerCmd vllmCmd(String name, String model, Integer cache, Boolean isEmbedding, Boolean isReranker, Boolean isReasonable) {
     List<String> arguments = ListUtil.list(
         false,
         StrUtil.format("--model /models/{}", model),
@@ -138,11 +131,13 @@ public class LlamaSwapTool extends GeneratorTool {
         "--port ${PORT}",
         StrUtil.format("--api-key {}", API_KEY),
         "--disable-log-requests",
-        "--uvicorn-log-level error"
+        "--uvicorn-log-level error",
+        "--trust-remote-code"
     );
     if (isEmbedding) {
       arguments.add("--task embedding");
     } else if (isReranker) {
+      arguments.add("--task score");
     } else if (isReasonable) {
       arguments.add("--enable-auto-tool-choice");
       arguments.add("--tool-call-parser hermes");
@@ -150,7 +145,7 @@ public class LlamaSwapTool extends GeneratorTool {
       arguments.add("--reasoning-parser deepseek_r1");
     }
     return new DockerCmd(
-        "vllm-server-cpu:0.9.1",
+        "vllm-server-cpu:0.8.5.post1",
         name,
         model,
         StrUtil.format("http://vllm-{}:${PORT}", displayName(model)),
@@ -158,9 +153,7 @@ public class LlamaSwapTool extends GeneratorTool {
             StrUtil.format("--name vllm-{}", displayName(model)),
             "--privileged=true",
             "--shm-size=4g",
-            StrUtil.format("-e VLLM_CPU_KVCACHE_SPACE={}", cache),
+            StrUtil.format("-e VLLM_CPU_KVCACHE_SPACE={}", cache)
-            StrUtil.format("-e VLLM_CPU_OMP_THREADS_BIND=0-{}", thread - 1),
-            "-e VLLM_CPU_MOE_PREPACK=0"
         ),
         arguments
     );

service-ai/service-ai-cli/src/main/resources/template/llama-swap.ftl

@@ -1,4 +1,4 @@
-healthCheckTimeout: 120
+healthCheckTimeout: 600
 logLevel: warn
 models:
 <#list models as model>

service-ai/service-ai-knowledge/src/main/resources/application.yml

@@ -9,10 +9,10 @@ spring:
       api-key: ENC(K+Hff9QGC+fcyi510VIDd9CaeK/IN5WBJ9rlkUsHEdDgIidW+stHHJlsK0lLPUXXREha+ToQZqqDXJrqSE+GUKCXklFhelD8bRHFXBIeP/ZzT2cxhzgKUXgjw3S0Qw2R)
       chat:
         options:
-          model: 'Qwen3-1.7-vllm'
+          model: 'Qwen3/qwen3-1.7b'
       embedding:
         options:
-          model: 'Bge-m3-vllm'
+          model: 'Qwen3/qwen3-embedding-4b'
     vectorstore:
       qdrant:
         host: 132.121.206.65