LiteLLM

About the LiteLLM integration

telemetry-dev-litellm instruments litellm.completion, litellm.acompletion, litellm.embedding, litellm.aembedding, streaming CustomStreamWrapper responses, and litellm.Router calls. Because LiteLLM normalizes 100+ providers behind one interface, a single instrument_litellm() call gives you tracing across every provider you route through.

Key features

• One call to instrument everything: instrument_litellm() patches exactly four package attributes, thread-safe and idempotent; uninstrument_litellm() restores them.
• Per-call drop-ins: Import completion, acompletion, embedding, or aembedding from telemetry_dev_litellm when you don't want global monkey-patching — each produces exactly one span even when global instrumentation is active.
• Router support: Router deployment attempts are traced per attempt; wrap_router(router) adds an enclosing Router-level span for retries and fallbacks.
• Streaming: Streamed responses are traced through LiteLLM's stream wrapper.
• Fail-open: Without an API key, LiteLLM calls run unchanged and telemetry is a no-op.

Get started

import telemetry_dev
from telemetry_dev_litellm import instrument_litellm

telemetry_dev.init()  # reads TELEMETRY_DEV_API_KEY
instrument_litellm()

import litellm

litellm.completion(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Say hi"}],
)

Note the import-order rule: references captured before instrumentation (from litellm import completion) keep pointing at the original function. Instrument at process startup, or use the telemetry_dev_litellm drop-ins instead.