feat: add per-model concurrency limits

frontend-web/src/core/system-settings/api.ts

@@ -342,6 +342,18 @@ export interface SkillCompressionUpdate {
   keep_index_in_prompt?: boolean;
 }
 
+export interface ModelConcurrencySettings {
+  enabled: boolean;
+  default_limit: number;
+  model_limits: Record<string, number>;
+}
+
+export interface ModelConcurrencyUpdate {
+  enabled?: boolean;
+  default_limit?: number;
+  model_limits?: Record<string, number>;
+}
+
 export async function loadSkillCompressionSettings(): Promise<SkillCompressionSettings> {
   const response = await apiFetch(
     `${getBackendBaseURL()}/api/system-settings/skill-compression`,
@@ -361,6 +373,53 @@ export async function loadSkillCompressionSettings(): Promise<SkillCompressionSe
   return data;
 }
 
+export async function loadModelConcurrencySettings(): Promise<ModelConcurrencySettings> {
+  const response = await apiFetch(
+    `${getBackendBaseURL()}/api/system-settings/model-concurrency`,
+  );
+  if (!response.ok) {
+    throw new Error(
+      `Failed to load model concurrency settings: ${response.status}`,
+    );
+  }
+  const data = (await response.json()) as ModelConcurrencySettings;
+  writeYLog({
+    optType: OptType.view,
+    optDetail: "CM助手-全量-系统设置-模型并发限制查看",
+    url: "/api/system-settings/model-concurrency",
+    returnParam: data,
+  });
+  return data;
+}
+
+export async function updateModelConcurrencySettings(
+  body: ModelConcurrencyUpdate,
+): Promise<ModelConcurrencySettings> {
+  const response = await apiFetch(
+    `${getBackendBaseURL()}/api/system-settings/model-concurrency`,
+    {
+      method: "PUT",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+    },
+  );
+  if (!response.ok) {
+    const detail = await response
+      .json()
+      .catch(() => ({ detail: "Failed to update model concurrency settings" }));
+    throw new Error(detail.detail ?? "Failed to update model concurrency settings");
+  }
+  const data = (await response.json()) as ModelConcurrencySettings;
+  writeYLog({
+    optType: OptType.edit,
+    optDetail: "CM助手-全量-系统设置-模型并发限制编辑",
+    url: "/api/system-settings/model-concurrency",
+    reqParam: body,
+    returnParam: data,
+  });
+  return data;
+}
+
 export async function updateSkillCompressionSettings(
   body: SkillCompressionUpdate,
 ): Promise<SkillCompressionSettings> {

frontend-web/src/core/system-settings/hooks.ts

@@ -7,6 +7,7 @@ import {
   loadAppearanceSettings,
   loadCitationDisplaySettings,
   loadLeaderboardSettings,
+  loadModelConcurrencySettings,
   loadPrivateSquareStatus,
   loadPromptPrefixSettings,
   loadPublishSquares,
@@ -15,6 +16,7 @@ import {
   updateAppearanceSettings,
   updateCitationDisplaySettings,
   updateLeaderboardSettings,
+  updateModelConcurrencySettings,
   updatePrivateSquareSettings,
   updatePromptPrefixSettings,
   updatePublishSquares,
@@ -22,6 +24,7 @@ import {
   type AppearanceUpdate,
   type CitationDisplayUpdate,
   type LeaderboardSettingsUpdate,
+  type ModelConcurrencyUpdate,
   type PrivateSquareUpdate,
   type PromptPrefixUpdate,
   type PublishSquaresUpdate,
@@ -37,6 +40,7 @@ const APPEARANCE_KEY = ["systemSettings", "appearance"] as const;
 const CITATION_DISPLAY_KEY = ["systemSettings", "citationDisplay"] as const;
 const LEADERBOARD_KEY = ["systemSettings", "leaderboard"] as const;
 const SKILL_COMPRESSION_KEY = ["systemSettings", "skillCompression"] as const;
+const MODEL_CONCURRENCY_KEY = ["systemSettings", "modelConcurrency"] as const;
 
 export function usePromptPrefixSettings() {
   return useQuery({
@@ -238,6 +242,27 @@ export function useUpdateSkillCompressionSettings() {
   });
 }
 
+export function useModelConcurrencySettings() {
+  return useQuery({
+    queryKey: MODEL_CONCURRENCY_KEY,
+    queryFn: () => loadModelConcurrencySettings(),
+    staleTime: 60_000,
+    refetchOnWindowFocus: false,
+    retry: false,
+  });
+}
+
+export function useUpdateModelConcurrencySettings() {
+  const queryClient = useQueryClient();
+  return useMutation({
+    mutationFn: (body: ModelConcurrencyUpdate) =>
+      updateModelConcurrencySettings(body),
+    onSuccess: (data) => {
+      queryClient.setQueryData(MODEL_CONCURRENCY_KEY, data);
+    },
+  });
+}
+
 export function useLeaderboardSettings() {
   return useQuery({
     queryKey: LEADERBOARD_KEY,