sefia-litellm

A sefia LLMClient implementation that connects to various LLM providers via LiteLLM.

Install

pip install sefia-litellm

Or through the stack: pip install 'sefios[litellm]'.

Usage

import sefia_litellm

client = sefia_litellm.LiteLLMClient(model="gpt-4o")

Suppressing LiteLLM logging

LiteLLM logs through the standard-library LiteLLM logger — mostly INFO-level messages that become visible whenever your application configures logging at INFO/DEBUG — and, when a request raises, prints debug info including a Provider List: ... banner. LiteLLMClient silences all of this by default.

Real failures still surface: LiteLLM raises exceptions on errors (which this client maps to sefia exceptions), so silencing the logs never hides a failure.

There are two ways to control this:

• Constructor argument suppress_logs (takes precedence)

  # Suppress logs (default)
  client = sefia_litellm.LiteLLMClient(model="gpt-4o")
  client = sefia_litellm.LiteLLMClient(model="gpt-4o", suppress_logs=True)
  
  # Let LiteLLM log as usual
  client = sefia_litellm.LiteLLMClient(model="gpt-4o", suppress_logs=False)
  

• Environment variable SEFIA_LITELLM_SUPPRESS_LOGS (used as the default when suppress_logs is None)

  # Disable suppression (let logs through). 0/false/no/off disable it.
  export SEFIA_LITELLM_SUPPRESS_LOGS=false
  

  An explicit suppress_logs argument overrides the environment variable. When unset, suppression is on.

When suppression is on, the LiteLLM logger is fully silenced (its level is set above CRITICAL) and litellm.suppress_debug_info is set to True. The logger level is applied as soon as sefia_litellm is imported — before LiteLLM itself is imported — so even LiteLLM's import-time warnings (e.g. optional-dependency preload warnings) are suppressed.

On slow imports

Importing LiteLLM is heavy and can take around a second (see BerriAI/litellm#7605).

This package mitigates that as follows:

1. Lazy import — LiteLLM is imported only inside the methods that actually send a request. Importing sefia_litellm alone costs nothing. After the first request the module is cached in sys.modules, so subsequent imports are effectively free. This is the primary mitigation.

2. Local model cost map — LITELLM_LOCAL_MODEL_COST_MAP=True is set before LiteLLM is imported so that it uses its bundled cost map JSON instead of fetching it over the network. This speeds up the import and keeps it working offline.

   If you need up-to-date pricing for the newest models and the bundled map is stale, you can restore the original behavior:

   export LITELLM_LOCAL_MODEL_COST_MAP=False
   

3. (Optional) Warm up at startup — if you also want to hide the first request's latency, import LiteLLM in the background during application startup. A daemon thread works regardless of whether an asyncio event loop is running yet:

   import threading
   
   threading.Thread(target=__import__, args=("litellm",), daemon=True).start()
   

Status

Early development. APIs may change before v1.0.

License

MIT