●Change8

litellm-precall-dev-v1.81.14

📦 litellmView on GitHub →
✨ 12 features🐛 17 fixes🔧 5 symbols

Summary

This release introduces several new features, including custom favicon support, configurable semantic cache dimensions, and new guardrail policies. Numerous bug fixes address issues across Bedrock integration, budget calculation, testing stability, and data normalization.

Migration Steps

  1. If using Bedrock Converse API batch models, verify modelInput format.
  2. If relying on hardcoded timezone maps for budget calculations, note that this is replaced by ZoneInfo.

✨ New Features

  • Show proxy URL in ModelHub.
  • Add custom favicon support for LiteLLM Proxy.
  • Support configurable vector dimensions for semantic-cache.
  • Add optional digest mode for Slack alert types.
  • Enable context-1m-2025-08-07 beta header for vertex_ai provider (after a revert/re-enable cycle).
  • Introduce Guardrail Policy Versioning.
  • Add Noma guardrails v2 based on custom guardrails.
  • Introduce LiteLLM network mock.
  • Implement Guardrail Monitor to measure guardrail reliability in production.
  • Add Singapore guardrail policies (PDPA + MAS AI Risk Management).
  • Add variant parameter to video content download.
  • Add Blog Dropdown in Navbar for the UI.

🐛 Bug Fixes

  • Correct modelInput format for Bedrock Converse API batch models.
  • Only tag selected deployment in access group creation.
  • Prevent double UUID in create_file S3 key for Bedrock.
  • Normalize camelCase thinking param keys to snake_case in utils.
  • Fix timezone config lookup and replace hardcoded timezone map with ZoneInfo in budget module.
  • Add missing return type annotations to iterator protocol methods in streaming_handler.
  • Avoid mutating caller-owned dicts in SpendUpdateQueue aggregation.
  • Fix model cost map for anthropic fast and inference_geo.
  • Add Priority PayGo cost tracking for gemini/vertex ai.
  • Fix StopIteration in prisma self-heal cooldown test.
  • Use absolute path for model_prices JSON validation test.
  • Add INCOMPLETE to interactions status enum expected values in tests.
  • Make RPM limit test sequential to fix race condition.
  • Fix video content download: pass api_key from litellm_params to video remix handlers.
  • Fix Spend Update Queue Aggregation Never Triggers with Default Presets.
  • Fix Anthropic model wildcard access issue.
  • Encode model arns for OpenAI compatible bedrock imported models.

Affected Symbols

streaming_handlerSpendUpdateQueuevertex_aibedrockmodel_prices