3c65189f30
后端从占位回显变为真实生成:管理员经控制面登记/激活模型,Gateway 经 NATS 下发,Dispatcher 热更新 LLM Pool,Eino 图用 OpenAI 兼容流式真实推理。 - shared: contract.ModelConfig(provider/base_url/api_key/model) + 配置 subjects; bus.RequestModelConfig/ServeModelConfig/Publish/Subscribe ModelConfigUpdated - gateway: store.LLMModel→sundynix_model(AutoMigrate,唯一激活) + admin REST (GET/POST/active/delete/test models, api_key 脱敏) + main ServeModelConfig + 变更广播; 路由 /api/v1/admin/models* - dispatcher: llm.Pool OpenAI 兼容 SSE 流式客户端(ChatStream) + 热更新配置 + 未配置则降级桩; poolModel.Ready()?真实流式:注入记忆的桩; main 取配置+订阅 - 开发期接在线 API 不拉本地模型(见 llm-provider-strategy memory) - 验证: 4 模块 build✓ + e2e PASS; mock OpenAI 服务 live 跑通——登记/测试连接✓/ 激活→NATS 热更新→提交→真实 SSE 流出 mock 回复, mock 日志证明端点被调用且 注入画像(老王)进了模型上下文 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
149 lines
3.7 KiB
Go
149 lines
3.7 KiB
Go
// Package llm 抽象 LLM Pool(vLLM / Ollama / 第三方在线 API)的负载均衡与流式推理。
|
||
package llm
|
||
|
||
import (
|
||
"bufio"
|
||
"bytes"
|
||
"context"
|
||
"encoding/json"
|
||
"fmt"
|
||
"net/http"
|
||
"strings"
|
||
"sync"
|
||
"time"
|
||
|
||
"github.com/sundynix/sundynix-shared/contract"
|
||
)
|
||
|
||
// ChatMessage 是一条对话消息(role: system/user/assistant)。
|
||
type ChatMessage struct {
|
||
Role string `json:"role"`
|
||
Content string `json:"content"`
|
||
}
|
||
|
||
// Pool 维护当前激活的后端配置(由控制面经 NATS 下发,可热更新)。
|
||
type Pool struct {
|
||
mu sync.RWMutex
|
||
cfg *contract.ModelConfig
|
||
hc *http.Client
|
||
}
|
||
|
||
func NewPool() *Pool {
|
||
return &Pool{hc: &http.Client{Timeout: 120 * time.Second}}
|
||
}
|
||
|
||
// SetConfig 热更新后端配置(控制面变更时调用)。
|
||
func (p *Pool) SetConfig(cfg *contract.ModelConfig) {
|
||
p.mu.Lock()
|
||
p.cfg = cfg
|
||
p.mu.Unlock()
|
||
if cfg != nil {
|
||
// 不打印 api_key。
|
||
fmt.Printf("[llm] model config set: provider=%s base=%s model=%s\n", cfg.Provider, cfg.BaseURL, cfg.Model)
|
||
}
|
||
}
|
||
|
||
func (p *Pool) config() *contract.ModelConfig {
|
||
p.mu.RLock()
|
||
defer p.mu.RUnlock()
|
||
return p.cfg
|
||
}
|
||
|
||
// Ready 报告是否已配置可用后端。
|
||
func (p *Pool) Ready() bool { return p.config().Ready() }
|
||
|
||
// ChatStream 以 OpenAI 兼容协议流式推理,逐 token 回调 onToken。
|
||
// 仅在 Ready() 时可用(调用方据此决定真实推理或降级桩)。
|
||
func (p *Pool) ChatStream(ctx context.Context, msgs []ChatMessage, onToken func(string)) error {
|
||
cfg := p.config()
|
||
if !cfg.Ready() {
|
||
return fmt.Errorf("no model configured")
|
||
}
|
||
body, _ := json.Marshal(map[string]any{
|
||
"model": cfg.Model,
|
||
"messages": msgs,
|
||
"stream": true,
|
||
})
|
||
req, err := http.NewRequestWithContext(ctx, http.MethodPost, cfg.BaseURL+"/chat/completions", bytes.NewReader(body))
|
||
if err != nil {
|
||
return err
|
||
}
|
||
req.Header.Set("Content-Type", "application/json")
|
||
if cfg.APIKey != "" {
|
||
req.Header.Set("Authorization", "Bearer "+cfg.APIKey)
|
||
}
|
||
resp, err := p.hc.Do(req)
|
||
if err != nil {
|
||
return fmt.Errorf("llm request: %w", err)
|
||
}
|
||
defer resp.Body.Close()
|
||
if resp.StatusCode >= 400 {
|
||
buf := new(bytes.Buffer)
|
||
_, _ = buf.ReadFrom(resp.Body)
|
||
return fmt.Errorf("llm http %d: %s", resp.StatusCode, strings.TrimSpace(buf.String()))
|
||
}
|
||
|
||
// 解析 OpenAI 兼容 SSE:data: {choices:[{delta:{content}}]} … data: [DONE]
|
||
sc := bufio.NewScanner(resp.Body)
|
||
sc.Buffer(make([]byte, 0, 64*1024), 1024*1024)
|
||
for sc.Scan() {
|
||
line := strings.TrimSpace(sc.Text())
|
||
if !strings.HasPrefix(line, "data:") {
|
||
continue
|
||
}
|
||
payload := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
|
||
if payload == "[DONE]" {
|
||
break
|
||
}
|
||
var chunk struct {
|
||
Choices []struct {
|
||
Delta struct {
|
||
Content string `json:"content"`
|
||
} `json:"delta"`
|
||
} `json:"choices"`
|
||
}
|
||
if json.Unmarshal([]byte(payload), &chunk) != nil {
|
||
continue
|
||
}
|
||
if len(chunk.Choices) > 0 && chunk.Choices[0].Delta.Content != "" {
|
||
onToken(chunk.Choices[0].Delta.Content)
|
||
}
|
||
}
|
||
return sc.Err()
|
||
}
|
||
|
||
// ---- 占位降级(未配置后端时)----
|
||
|
||
// 占位参数:模拟真实后端的 TTFT(首 token 延迟) 与逐 token 间隔。
|
||
const (
|
||
timeToFirstToken = 700 * time.Millisecond
|
||
interTokenDelay = 60 * time.Millisecond
|
||
)
|
||
|
||
// StreamText 按节奏把给定文本流式回调(未配置真实后端时的降级桩)。
|
||
func (p *Pool) StreamText(ctx context.Context, text string, onToken func([]byte)) error {
|
||
select {
|
||
case <-ctx.Done():
|
||
return ctx.Err()
|
||
case <-time.After(timeToFirstToken):
|
||
}
|
||
for _, tok := range tokenize(text) {
|
||
select {
|
||
case <-ctx.Done():
|
||
return ctx.Err()
|
||
default:
|
||
}
|
||
onToken([]byte(tok))
|
||
time.Sleep(interTokenDelay)
|
||
}
|
||
return nil
|
||
}
|
||
|
||
func tokenize(s string) []string {
|
||
out := make([]string, 0, len(s))
|
||
for _, r := range s {
|
||
out = append(out, string(r))
|
||
}
|
||
return out
|
||
}
|