feat: sync deployed v1.0.0-rc.38-custom.67

service/group_degradation.go

@@ -0,0 +1,48 @@
+package service
+
+import (
+	"sync"
+	"time"
+
+	fingerprint "github.com/QuantumNous/new-api/pkg/model_fingerprint"
+	"github.com/gin-gonic/gin"
+)
+
+var groupDegradationSnapshot struct {
+	sync.Mutex
+	updated   time.Time
+	latest    map[string]AstraSamplingHistory
+	benchmark map[string]DetectionHistory
+}
+
+func requestGroupDegraded(c *gin.Context, modelName, group string) bool {
+	policy := RequestRoutingPolicy(c)
+	if policy == nil {
+		return false
+	}
+	skip := policy.SkipDegradedGroups && (modelName == fingerprint.Model || policy.SkipDegradedNonAstra)
+	if policy.Mode == "smart" {
+		skip = !policy.AllowDegraded && (modelName == fingerprint.Model || policy.SkipDegradedNonAstra)
+	}
+	if !skip {
+		return false
+	}
+	if c.GetBool("pool_degradation_unavailable") {
+		return true
+	}
+	latest, found := c.Get("pool_group_fingerprints")
+	benchmark, _ := c.Get("pool_group_benchmarks")
+	if !found {
+		rounds, qualities, err := cachedGroupQuality(c.Request.Context())
+		if err != nil {
+			c.Set("pool_degradation_unavailable", true)
+			return true
+		}
+		latest, benchmark = rounds, qualities
+		c.Set("pool_group_fingerprints", latest)
+		c.Set("pool_group_benchmarks", benchmark)
+	}
+	rounds, _ := latest.(map[string]AstraSamplingHistory)
+	h, ok := rounds[group]
+	return (ok && h.ValidSamples > 0 && h.Closest != "" && h.Closest != fingerprint.Model)
+}