litellm_fix(perf): skip global router_settings in per-request Router creation

[shin-bot-litellm]

Problem

┌─────────────────────────────────────────────────────────────┐
│                     STARTUP                                 │
│  Global router_settings → Applied to shared llm_router ✅   │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                   EVERY REQUEST                             │
│  _get_hierarchical_router_settings() returns global settings│
│           ↓                                                 │
│  Creates NEW Router (redundant - already on shared!) ❌     │
└─────────────────────────────────────────────────────────────┘

Issue: Global router_settings in DB trigger per-request Router creation, but they're already applied to the shared llm_router at startup.

Fix

Remove global settings lookup from _get_hierarchical_router_settings():

- # 3. Try global router_settings
- db_router_settings = await prisma_client.db.litellm_config.find_first(...)
- return db_router_settings.param_value
+ # Skip global - already on shared router from startup
+ return None

Now only key/team-specific settings create per-request Routers.

Related

• Issue [Bug]: Significant performance regression after upgrading from 1.80.5 to 1.81.x (UI + API slowness) #19921
• Follow-up to PR fix(proxy): resolve high CPU when router_settings in DB by avoiding REGISTRY.collect() in PrometheusServicesLogger #20087

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Jan 31, 2026 8:41am

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[shin-bot-litellm]

Closing - need deeper investigation first. The fix may be incomplete (DB query still happens on every request). Will reopen with proper fix after full analysis.