feat: SGLANG_COMPAT — chat_template_kwargs.enable_thinking=false при -no-think (llama.cpp/sglang)

Dockerfile

@@ -1,5 +1,5 @@
 FROM docker.io/library/eclipse-temurin:17-jre-alpine
-COPY llm-proxy.jar /app/llm-proxy.jar
+COPY build/libs/llm-proxy.jar /app/llm-proxy.jar
 WORKDIR /app
 EXPOSE 8100
 ENTRYPOINT ["java", "-jar", "/app/llm-proxy.jar"]

podman-compose.yaml

@@ -3,15 +3,22 @@
 # моделей с суффиксом "-no-think", проксирует на RouterAI. Ответ — как есть.
 #
 # Запуск на 76.179 (llm-router): podman-compose up -d
+# Образ тянется из нашего реестра (zot): images.binom.pw/llm-proxy:<tag>
 services:
   llm-proxy:
     image: images.binom.pw/llm-proxy:latest
     container_name: llm-proxy
     restart: unless-stopped
-    network_mode: bifrost_default
+    networks:
+      - bifrost
     environment:
       - PORT=8100
       - UPSTREAM_URL=https://routerai.ru/api/v1
       - ROUTER_API_KEY=__SET_FROM_CONFIG_DB__
       - EXCLUDED_PROVIDERS=deepseek
       - THINKING_MODELS=deepseek/deepseek-v4-flash-0731,deepseek/deepseek-v4-flash
+
+networks:
+  bifrost:
+    external: true
+    name: bifrost_default

src/main/kotlin/pw/binom/llmproxy/Main.kt

@@ -103,6 +103,10 @@ private suspend fun handleChat(
         return
     }
 
+    val start = System.currentTimeMillis()
+    val model = try {
+        json.parseToJsonElement(patched).jsonObject["model"]?.jsonPrimitive?.content ?: "?"
+    } catch (e: Exception) { "?" }
     val req = HttpRequest.newBuilder()
         .uri(URI.create(upstream.trimEnd('/') + "/chat/completions"))
         .header("Authorization", "Bearer $apiKey")
@@ -111,6 +115,7 @@ private suspend fun handleChat(
         .build()
 
     val stream = json.parseToJsonElement(raw).jsonObject["stream"]?.jsonPrimitive?.content == "true"
+    try {
         if (stream) {
             // SSE-стрим: транслируем как есть, чанк за чанком.
             val resp = http.send(req, HttpResponse.BodyHandlers.ofInputStream())
@@ -118,10 +123,19 @@ private suspend fun handleChat(
             call.respondOutputStream(ContentType.parse(ct), HttpStatusCode.fromValue(resp.statusCode())) {
                 resp.body().use { input -> input.copyTo(this, 8192) }
             }
+            println("[llm-proxy] chat model=$model status=${resp.statusCode()} в ${System.currentTimeMillis() - start}ms stream=true")
         } else {
             val resp = http.send(req, HttpResponse.BodyHandlers.ofByteArray())
             val ct = resp.headers().firstValue("content-type").orElse("application/json")
             call.respondBytes(resp.body(), ContentType.parse(ct), HttpStatusCode.fromValue(resp.statusCode()))
+            println("[llm-proxy] chat model=$model status=${resp.statusCode()} в ${System.currentTimeMillis() - start}ms stream=false")
+        }
+    } catch (e: Exception) {
+        println("[llm-proxy] chat model=$model ОШИБКА: ${e.message} в ${System.currentTimeMillis() - start}ms")
+        call.respondBytes(
+            """{"error":{"message":"upstream: ${e.message}"}}""".toByteArray(),
+            ContentType.Application.Json, HttpStatusCode.BadGateway,
+        )
     }
 }
 
@@ -141,6 +155,12 @@ internal fun patchBody(raw: String, excluded: List<String>, thinking: List<Strin
         if (root["reasoning"] == null) {
             root["reasoning"] = JsonObject(mapOf("enabled" to JsonPrimitive(false)))
         }
+        // SGLANG_COMPAT: Qwen3.8 глушится только через chat_template_kwargs.enable_thinking=false
+        if (System.getenv("SGLANG_COMPAT") == "true") {
+            val ctk = root["chat_template_kwargs"]?.jsonObject?.toMutableMap() ?: mutableMapOf()
+            ctk["enable_thinking"] = JsonPrimitive(false)
+            root["chat_template_kwargs"] = JsonObject(ctk)
+        }
     }
 
     val provider = root["provider"]?.jsonObject?.toMutableMap() ?: mutableMapOf()