jianf/ModelRouter
1
0
Fork 0
mirror of https://gitcode.com/JianFeeeee/ModelRouter.git synced 2026-09-20 08:57:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: source-model routing disambiguation (source-model/:/ prefix); same-tier round-robin load balancing; public_base_url for copy config; fmtTok(B/M/K) unit scaling; status column reorder (reachability first); drop emoji from seed-warn modal; fix tests for first-run adapter seeding; install lua5.1 dev lib

Browse Source
This commit is contained in:
root
2026-08-10 12:55:22 +08:00
parent fe06e0861a
commit 41c9b0e14a
9 changed files with 128 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
internal/gateway/chat.go
View File

@ -228,7 +228,7 @@ func (g *Gateway) resolveByModel(model string) ([]*provider.Provider, string) {
if isAuto(model) {
return g.core.Registry().Resolve("AUTO"), ""
}
return g.core.Registry().Resolve(model), model
return g.core.Registry().Resolve(model), g.core.Registry().EffectiveModel(model)
}
// chatOnly keeps providers that expose at least one chat-capable model, so a
@ -616,13 +616,17 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
type autoPlan struct {
p *provider.Provider
model string
tier int
quota int64
win int64
}
// autoPlans builds the schedulable AUTO slots from the persisted rules. A
// slot is schedulable while its model is available and (when quota > 0) the
// tokens used within its reset window are below the quota.
// tokens used within its reset window are below the quota. Slots are returned
// tiered high→low, and within the same tier the order is rotated round-robin
// so concurrent requests spread evenly across equal-priority sources (still
// with failover to the next slot if one errors).
func (g *Gateway) autoPlans() []autoPlan {
rules := g.core.AutoRules()
if len(rules) == 0 {
@ -641,9 +645,35 @@ func (g *Gateway) autoPlans() []autoPlan {
if e.TokenQuota > 0 && g.stats.WindowTokens(e.Model, e.Source, win) >= e.TokenQuota {
continue
}
plans = append(plans, autoPlan{p: p, model: e.Model, quota: e.TokenQuota, win: win})
plans = append(plans, autoPlan{p: p, model: e.Model, tier: e.Tier, quota: e.TokenQuota, win: win})
}
return plans
return g.rotateSameTier(plans)
}
// rotateSameTier reorders the leading plan of each consecutive same-tier run
// using a global round-robin counter, so requests distribute across
// equal-priority sources while preserving tier ordering and in-tier failover.
func (g *Gateway) rotateSameTier(plans []autoPlan) []autoPlan {
if len(plans) < 2 {
// allow single slot without varying
return plans
}
rot := int(g.autoRR.Add(1))
out := make([]autoPlan, 0, len(plans))
for i := 0; i < len(plans); {
j := i
for j < len(plans) && plans[j].tier == plans[i].tier {
j++
}
run := plans[i:j]
if len(run) > 1 {
off := rot % len(run)
run = append(run[off:], run[:off]...)
}
out = append(out, run...)
i = j
}
return out
}
// singleChatAuto runs a non-streaming AUTO request slot by slot: each slot