red-future/model-gateway
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor(asynch): 重构异步模型配置和队列管理

Browse Source
This commit is contained in:
王利钊
2026-06-02 20:26:45 +08:00
parent c7e9eb889b
commit 52124385a1
18 changed files with 726 additions and 1006 deletions
Download Patch File Download Diff File Expand all files Collapse all files

118
model/dto/model_dto.go
View File

@@ -9,34 +9,36 @@ import (
// CreateModelReq 添加模型配置
type CreateModelReq struct {
g.Meta `path:"/createModel" method:"post" tags:"模型管理" summary:"创建模型配置" dc:"添加新的模型配置"`
ModelName string `p:"modelName" json:"modelName" v:"required#modelName不能为空" dc:"模型名称（唯一标识）"`
ModelType int `p:"modelType" json:"modelType" v:"required#modelType不能为空" dc:"模型类型：1-文本生成 2-图像生成 3-语音 4-视频 5-多模态"`
BaseURL string `p:"baseUrl" json:"baseUrl" v:"required#baseUrl不能为空" dc:"模型服务基础地址（如 gateway(s)://host:port）"`
HttpMethod string `p:"httpMethod" json:"httpMethod" dc:"请求方式：GET/POST（默认POST）"`
HeadMsg string `p:"headMsg" json:"headMsg" dc:"请求头绑定（支持多个，逗号分隔），示例：Authorization:Bearer xxx,Content-Type:application/json"`
IsPrivate *int `p:"isPrivate" json:"isPrivate" v:"in:0,1#私有化参数只能为0或1" dc:"是否私有化：0-私有（默认） 1-公共"`
Enabled *int `p:"enabled" json:"enabled" v:"in:0,1#启用参数只能为0或1" dc:"是否启用：0-禁用，1-启用（默认1）"`
IsChatModel *int `p:"isChatModel" json:"isChatModel" v:"in:0,1#对话模型参数只能为0或1" dc:"是否为对话模型：0-否，1-是（默认0）"`
IsOwner *int `p:"isOwner" json:"isOwner" v:"in:0,1#是否为所有者参数只能为0或1" dc:"是否为所有者：0-否，1-是（默认0）"`
OperatorName string `p:"operatorName" json:"operatorName" v:"required#operatorName不能为空" dc:"运营商名称"`
TokenConfig map[string]any `p:"tokenConfig" json:"tokenConfig" dc:"token计算配置"`
ExtendMapping map[string]any `p:"extendMapping" json:"extendMapping" dc:"附加映射"`
QueryConfig map[string]any `p:"queryConfig" json:"queryConfig" dc:"结果配置"`
ApiKey string `p:"apiKey" json:"apiKey" dc:"调用凭证/密钥，用于模型认证"`
Form []map[string]any `p:"form" json:"form" dc:"动态表单配置（JSON），用于前端渲染配置项"`
RequestMapping map[string]any `p:"requestMapping" json:"requestMapping" dc:"请求映射"`
ResponseMapping map[string]any `p:"responseMapping" json:"responseMapping" dc:"返回映射"`
ResponseBody map[string]any `p:"responseBody" json:"responseBody" dc:"返回主体"`
ResponseTokenField string `p:"responseTokenField" json:"responseTokenField" dc:"响应中消耗token的字段映射"`
MaxConcurrency int `p:"maxConcurrency" json:"maxConcurrency" dc:"最大并发数（默认10）"`
QueueLimit int `p:"queueLimit" json:"queueLimit" dc:"排队队列上限（默认1000）"`
TimeoutSeconds int `p:"timeoutSeconds" json:"timeoutSeconds" dc:"请求超时时间（秒，默认600）"`
ExpectedSeconds int `p:"expectedSeconds" json:"expectedSeconds" dc:"模型预计执行时间（秒，默认600）"`
RetryTimes int `p:"retryTimes" json:"retryTimes" dc:"失败重试次数（默认3）"`
RetryQueueMaxSeconds int `p:"retryQueueMaxSeconds" json:"retryQueueMaxSeconds" dc:"失败重试最大排队时间（秒，默认600）"`
AutoCleanSeconds int `p:"autoCleanSeconds" json:"autoCleanSeconds" dc:"任务完成后自动清理时间（秒，默认86400）"`
Remark string `p:"remark" json:"remark" dc:"备注说明"`
g.Meta `path:"/createModel" method:"post" tags:"模型管理" summary:"创建模型配置" dc:"添加新的模型配置"`
ModelName string `p:"modelName" json:"modelName" v:"required#模型名称不能为空" dc:"模型名称（唯一标识）"`
ModelType int `p:"modelType" json:"modelType" v:"required#模型类型不能为空" dc:"模型类型"`
BaseURL string `p:"baseUrl" json:"baseUrl" v:"required#模型地址不能为空" dc:"模型服务地址"`
HttpMethod string `p:"httpMethod" json:"httpMethod" dc:"请求方式：GET/POST（默认POST）"`
HeadMsg map[string]any `p:"headMsg" json:"headMsg" dc:"请求头JSON结构"`
IsPrivate *int `p:"isPrivate" json:"isPrivate" dc:"是否私有化：0-私有 1-公共"`
Enabled *int `p:"enabled" json:"enabled" dc:"是否启用：0-停用 1-启用"`
IsChatModel *int `p:"isChatModel" json:"isChatModel" dc:"是否为对话模型：0-否 1-是"`
IsAsync *int `p:"isAsync" json:"isAsync" dc:"是否异步：0-同步 1-异步"`
IsStream *int `p:"isStream" json:"isStream" dc:"是否流式：0-非流式 1-流式"`
IsOwner *int `p:"isOwner" json:"isOwner" dc:"是否为所有者：0-否 1-是"`
ApiKey string `p:"apiKey" json:"apiKey" dc:"调用凭证/密钥"`
Form []map[string]any `p:"form" json:"form" dc:"动态表单配置"`
RequestMapping map[string]any `p:"requestMapping" json:"requestMapping" dc:"请求映射"`
ResponseMapping map[string]any `p:"responseMapping" json:"responseMapping" dc:"返回映射"`
ResponseBody string `p:"responseBody" json:"responseBody" dc:"返回主体"`
ResponseTokenField string `p:"responseTokenField" json:"responseTokenField" dc:"响应中消耗token的字段映射"`
OperatorName string `p:"operatorName" json:"operatorName" dc:"运营商名称"`
TokenConfig map[string]any `p:"tokenConfig" json:"tokenConfig" dc:"token计算配置"`
ExtendMapping map[string]any `p:"extendMapping" json:"extendMapping" dc:"附加映射"`
QueryConfig map[string]any `p:"queryConfig" json:"queryConfig" dc:"查询/回调配置"`
StreamConfig map[string]any `p:"streamConfig" json:"streamConfig" dc:"流式输出配置"`
FirstFrame string `p:"firstFrame" json:"firstFrame" dc:"首帧图片参数"`
LastFrame string `p:"lastFrame" json:"lastFrame" dc:"尾帧图片参数"`
MaxConcurrency int `p:"maxConcurrency" json:"maxConcurrency" dc:"最大并发数（默认10）"`
TimeoutSeconds int `p:"timeoutSeconds" json:"timeoutSeconds" dc:"请求超时时间（秒，默认600）"`
RetryTimes int `p:"retryTimes" json:"retryTimes" dc:"失败重试次数（默认3）"`
AutoCleanSeconds int `p:"autoCleanSeconds" json:"autoCleanSeconds" dc:"任务完成后自动清理时间（秒，默认86400）"`
CallbackUrl string `p:"callbackUrl" json:"callbackUrl" dc:"回调地址"`
}
type CreateModelRes struct {
@@ -44,35 +46,37 @@ type CreateModelRes struct {
}
type UpdateModelReq struct {
g.Meta `path:"/updateModel" method:"put" tags:"模型管理" summary:"更新模型配置" dc:"更新指定ID的模型配置"`
ID int64 `p:"id" json:"id" v:"required#id不能为空" dc:"配置ID"`
ModelName string `p:"modelName" json:"modelName" dc:"模型名称（唯一标识）"`
ModelType int `p:"modelType" json:"modelType" dc:"模型类型ID列表（逗号分隔）（可选更新）"`
BaseURL string `p:"baseUrl" json:"baseUrl" dc:"模型服务基础地址"`
HttpMethod string `p:"httpMethod" json:"httpMethod" dc:"请求方式：GET/POST（可选更新）"`
HeadMsg string `p:"headMsg" json:"headMsg" dc:"请求头绑定（可选更新）"`
ApiKey string `p:"apiKey" json:"apiKey" dc:"调用凭证/密钥，用于模型认证（可选更新）"`
Form []map[string]any `p:"form" json:"form" dc:"动态表单配置（JSON）（可选更新）"`
RequestMapping map[string]any `p:"requestMapping" json:"requestMapping" dc:"请求参数映射（可选更新）"`
ResponseMapping map[string]any `p:"responseMapping" json:"responseMapping" dc:"返回参数映射（可选更新）"`
ResponseBody map[string]any `p:"responseBody" json:"responseBody" dc:"返回主体（可选更新）"`
ResponseTokenField string `p:"responseTokenField" json:"responseTokenField" dc:"响应中消耗token的字段映射"`
Enabled *int `p:"enabled" json:"enabled" dc:"是否启用：0-禁用，1-启用（可选更新）"`
IsPrivate *int `p:"isPrivate" json:"isPrivate" v:"in:0,1#私有化参数只能为0或1" dc:"是否私有化：0-私有（默认） 1-公共"`
IsChatModel *int `p:"isChatModel" json:"isChatModel" v:"in:0,1#对话模型参数只能为0或1" dc:"是否为对话模型：0-否，1-是（默认0）"`
IsOwner *int `p:"isOwner" json:"isOwner" v:"in:0,1#是否为所有者参数只能为0或1" dc:"是否为所有者：0-否，1-是（默认0）"`
OperatorName string `p:"operatorName" json:"operatorName" v:"required#operatorName不能为空" dc:"运营商名称"`
TokenConfig map[string]any `p:"tokenConfig" json:"tokenConfig" dc:"token计算配置"`
ExtendMapping map[string]any `p:"extendMapping" json:"extendMapping" dc:"附加映射"`
QueryConfig map[string]any `p:"queryConfig" json:"queryConfig" dc:"结果配置"`
MaxConcurrency int `p:"maxConcurrency" json:"maxConcurrency" dc:"最大并发数（可选更新）"`
QueueLimit int `p:"queueLimit" json:"queueLimit" dc:"排队队列上限（可选更新）"`
TimeoutSeconds int `p:"timeoutSeconds" json:"timeoutSeconds" dc:"请求超时时间（秒）（可选更新）"`
ExpectedSeconds int `p:"expectedSeconds" json:"expectedSeconds" dc:"模型预计执行时间（秒）（可选更新）"`
RetryTimes int `p:"retryTimes" json:"retryTimes" dc:"失败重试次数（可选更新）"`
RetryQueueMaxSeconds int `p:"retryQueueMaxSeconds" json:"retryQueueMaxSeconds" dc:"失败重试最大排队时间（秒）（可选更新）"`
AutoCleanSeconds int `p:"autoCleanSeconds" json:"autoCleanSeconds" dc:"自动清理间隔（秒）（可选更新）"`
Remark string `p:"remark" json:"remark" dc:"备注说明（可选更新）"`
g.Meta `path:"/updateModel" method:"put" tags:"模型管理" summary:"更新模型配置" dc:"更新指定ID的模型配置"`
ID int64 `p:"id" json:"id" v:"required#id不能为空" dc:"配置ID"`
ModelName string `p:"modelName" json:"modelName" dc:"模型名称"`
ModelType int `p:"modelType" json:"modelType" dc:"模型类型"`
BaseURL string `p:"baseUrl" json:"baseUrl" dc:"模型服务地址"`
HttpMethod string `p:"httpMethod" json:"httpMethod" dc:"请求方式：GET/POST"`
HeadMsg map[string]any `p:"headMsg" json:"headMsg" dc:"请求头JSON结构"`
IsPrivate *int `p:"isPrivate" json:"isPrivate" dc:"是否私有化：0-私有 1-公共"`
Enabled *int `p:"enabled" json:"enabled" dc:"是否启用：0-停用 1-启用"`
IsChatModel *int `p:"isChatModel" json:"isChatModel" dc:"是否为对话模型：0-否 1-是"`
IsAsync *int `p:"isAsync" json:"isAsync" dc:"是否异步：0-同步 1-异步"`
IsStream *int `p:"isStream" json:"isStream" dc:"是否流式：0-非流式 1-流式"`
IsOwner *int `p:"isOwner" json:"isOwner" dc:"是否为所有者：0-否 1-是"`
ApiKey string `p:"apiKey" json:"apiKey" dc:"调用凭证/密钥"`
Form []map[string]any `p:"form" json:"form" dc:"动态表单配置"`
RequestMapping map[string]any `p:"requestMapping" json:"requestMapping" dc:"请求映射"`
ResponseMapping map[string]any `p:"responseMapping" json:"responseMapping" dc:"返回映射"`
ResponseBody string `p:"responseBody" json:"responseBody" dc:"返回主体"`
ResponseTokenField string `p:"responseTokenField" json:"responseTokenField" dc:"响应中消耗token的字段映射"`
OperatorName string `p:"operatorName" json:"operatorName" dc:"运营商名称"`
TokenConfig map[string]any `p:"tokenConfig" json:"tokenConfig" dc:"token计算配置"`
ExtendMapping map[string]any `p:"extendMapping" json:"extendMapping" dc:"附加映射"`
QueryConfig map[string]any `p:"queryConfig" json:"queryConfig" dc:"查询/回调配置"`
StreamConfig map[string]any `p:"streamConfig" json:"streamConfig" dc:"流式输出配置"`
FirstFrame string `p:"firstFrame" json:"firstFrame" dc:"首帧图片参数"`
LastFrame string `p:"lastFrame" json:"lastFrame" dc:"尾帧图片参数"`
MaxConcurrency int `p:"maxConcurrency" json:"maxConcurrency" dc:"最大并发数"`
TimeoutSeconds int `p:"timeoutSeconds" json:"timeoutSeconds" dc:"请求超时时间（秒）"`
RetryTimes int `p:"retryTimes" json:"retryTimes" dc:"失败重试次数"`
AutoCleanSeconds int `p:"autoCleanSeconds" json:"autoCleanSeconds" dc:"任务完成后自动清理时间（秒）"`
CallbackUrl string `p:"callbackUrl" json:"callbackUrl" dc:"回调地址"`
}
type UpdateModelRes struct {

32
model/dto/task_dto.go
View File

@@ -18,6 +18,38 @@ type CreateTaskRes struct {
TaskID string `json:"taskId" dc:"任务ID"`
}
type ModelTaskCallbackReq struct {
g.Meta `path:"/modelCallback" method:"post" tags:"异步任务" summary:"模型任务回调通知"`
TaskID string `json:"id" dc:"任务ID"`
Status string `json:"status" dc:"queued/running/succeeded/failed/expired"`
Content map[string]any `json:"content,omitempty" dc:"任务结果内容"`
Usage map[string]any `json:"usage,omitempty" dc:"token用量"`
}
type ModelTaskCallbackRes struct {
Success bool `json:"success" dc:"是否接收成功"`
}
// QueryPendingTasksReq 批量轮询请求
type QueryPendingTasksReq struct {
g.Meta `path:"/queryPending" method:"get" tags:"异步任务" summary:"批量轮询进行中的任务"`
Limit int `p:"limit" json:"limit" dc:"查询数量，默认10"`
}
// QueryPendingTasksRes 批量轮询响应
type QueryPendingTasksRes struct {
Total int `json:"total" dc:"本次查询数量"`
Results []QueryTaskItem `json:"results" dc:"查询结果列表"`
}
// QueryTaskItem 单个任务查询结果
type QueryTaskItem struct {
TaskID string `json:"taskId" dc:"任务ID"`
Status string `json:"status" dc:"任务状态"`
Content map[string]any `json:"content,omitempty" dc:"结果内容"`
Usage map[string]any `json:"usage,omitempty" dc:"token用量"`
}
// GetTaskResultReq 获取结果（只返回 oss 地址）
type GetTaskResultReq struct {
g.Meta `path:"/getTaskResult" method:"get" tags:"任务管理" summary:"获取任务结果" dc:"根据任务ID获取结果（只返回OSS地址）"`

80
model/entity/asynch_model.go
View File

@@ -14,25 +14,25 @@ type asynchModelCol struct {
ResponseMapping string
ResponseBody string
ResponseTokenField string
Prompt string
IsPrivate string
IsChatModel string
IsAsync string
IsStream string
ApiKey string
Enabled string
MaxConcurrency string
QueueLimit string
TimeoutSeconds string
ExpectedSeconds string
RetryTimes string
RetryQueueMaxSecs string
AutoCleanSeconds string
Remark string
IsOwner string
OperatorName string
TokenConfig string
ExtendMapping string
QueryConfig string
StreamConfig string
FirstFrame string
LastFrame string
CallbackUrl string
}
var AsynchModelCol = asynchModelCol{
@@ -47,57 +47,57 @@ var AsynchModelCol = asynchModelCol{
ResponseMapping: "response_mapping",
ResponseBody: "response_body",
ResponseTokenField: "response_token_field",
Prompt: "prompt",
IsPrivate: "is_private",
IsChatModel: "is_chat_model",
IsAsync: "is_async",
IsStream: "is_stream",
ApiKey: "api_key",
Enabled: "enabled",
MaxConcurrency: "max_concurrency",
QueueLimit: "queue_limit",
TimeoutSeconds: "timeout_seconds",
ExpectedSeconds: "expected_seconds",
RetryTimes: "retry_times",
RetryQueueMaxSecs: "retry_queue_max_seconds",
AutoCleanSeconds: "auto_clean_seconds",
Remark: "remark",
IsOwner: "is_owner",
OperatorName: "operator_name",
TokenConfig: "token_config",
ExtendMapping: "extend_mapping",
QueryConfig: "query_config",
StreamConfig: "stream_config",
FirstFrame: "first_frame",
LastFrame: "last_frame",
CallbackUrl: "callback_url",
}
// AsynchModel 异步模型配置
type AsynchModel struct {
beans.SQLBaseDO `orm:",inline"`
ModelName string `orm:"model_name" json:"modelName"`
ModelType int `orm:"model_type" json:"modelType"`
BaseURL string `orm:"base_url" json:"baseUrl"`
HttpMethod string `orm:"http_method" json:"httpMethod"`
HeadMsg string `orm:"head_msg" json:"headMsg"`
Form []map[string]any `orm:"form_json" json:"form"`
RequestMapping map[string]any `orm:"request_mapping" json:"requestMapping"`
ResponseMapping map[string]any `orm:"response_mapping" json:"responseMapping"`
ResponseBody map[string]any `orm:"response_body" json:"responseBody"`
ResponseTokenField string `orm:"response_token_field" json:"responseTokenField"`
Prompt string `orm:"prompt" json:"prompt"`
IsPrivate *int `orm:"is_private" json:"isPrivate"`
IsChatModel *int `orm:"is_chat_model" json:"isChatModel"`
ApiKey string `orm:"api_key" json:"apiKey"`
Enabled *int `orm:"enabled" json:"enabled"`
MaxConcurrency int `orm:"max_concurrency" json:"maxConcurrency"`
QueueLimit int `orm:"queue_limit" json:"queueLimit"`
TimeoutSeconds int `orm:"timeout_seconds" json:"timeoutSeconds"`
ExpectedSeconds int `orm:"expected_seconds" json:"expectedSeconds"`
RetryTimes int `orm:"retry_times" json:"retryTimes"`
RetryQueueMaxSeconds int `orm:"retry_queue_max_seconds" json:"retryQueueMaxSeconds"`
AutoCleanSeconds int `orm:"auto_clean_seconds" json:"autoCleanSeconds"`
Remark string `orm:"remark" json:"remark"`
IsOwner *int `json:"isOwner" orm:"is_owner"`
OperatorName string `orm:"operator_name" json:"operatorName"`
TokenConfig map[string]any `orm:"token_config" json:"tokenConfig"`
ExtendMapping map[string]any `orm:"extend_mapping" json:"extendMapping"`
QueryConfig map[string]any `orm:"query_config" json:"queryConfig"`
StreamConfig map[string]any `orm:"stream_config" json:"streamConfig"`
beans.SQLBaseDO `orm:",inline"`
ModelName string `orm:"model_name" json:"modelName"`
ModelType int `orm:"model_type" json:"modelType"`
BaseURL string `orm:"base_url" json:"baseUrl"`
HttpMethod string `orm:"http_method" json:"httpMethod"`
HeadMsg map[string]any `orm:"head_msg" json:"headMsg"`
Form []map[string]any `orm:"form_json" json:"form"`
RequestMapping map[string]any `orm:"request_mapping" json:"requestMapping"`
ResponseMapping map[string]any `orm:"response_mapping" json:"responseMapping"`
ResponseBody string `orm:"response_body" json:"responseBody"`
ResponseTokenField string `orm:"response_token_field" json:"responseTokenField"`
IsPrivate *int `orm:"is_private" json:"isPrivate"`
IsChatModel *int `orm:"is_chat_model" json:"isChatModel"`
IsAsync *int `orm:"is_async" json:"isAsync"`
IsStream *int `orm:"is_stream" json:"isStream"`
ApiKey string `orm:"api_key" json:"apiKey"`
Enabled *int `orm:"enabled" json:"enabled"`
MaxConcurrency int `orm:"max_concurrency" json:"maxConcurrency"`
TimeoutSeconds int `orm:"timeout_seconds" json:"timeoutSeconds"`
RetryTimes int `orm:"retry_times" json:"retryTimes"`
AutoCleanSeconds int `orm:"auto_clean_seconds" json:"autoCleanSeconds"`
IsOwner *int `json:"isOwner" orm:"is_owner"`
OperatorName string `orm:"operator_name" json:"operatorName"`
TokenConfig map[string]any `orm:"token_config" json:"tokenConfig"`
ExtendMapping map[string]any `orm:"extend_mapping" json:"extendMapping"`
QueryConfig map[string]any `orm:"query_config" json:"queryConfig"`
StreamConfig map[string]any `orm:"stream_config" json:"streamConfig"`
FirstFrame string `orm:"first_frame" json:"firstFrame"`
LastFrame string `orm:"last_frame" json:"lastFrame"`
CallbackUrl string `orm:"callback_url" json:"callbackUrl"`
}