[bot] LiteLLM: `batch_completion` (and `batch_completion_models`) not instrumented

[braintrust-bot]

Summary

LiteLLM's Python SDK ships a documented batch-inference execution surface — litellm.batch_completion() — which fans a single set of messages (or a list of prompt lists) out to concurrent completion() calls across one or more models and returns the list of responses. The Braintrust LiteLLM integration (py/src/braintrust/integrations/litellm/) instruments 18 top-level LiteLLM entry points (completion/acompletion, text_completion/atext_completion, responses/aresponses, image_generation/aimage_generation, embedding/aembedding, moderation/amoderation, speech/aspeech, transcription/atranscription, rerank/arerank), but has no patcher for batch_completion, batch_completion_models, or batch_completion_models_all_responses. Calls made through these batch entry points produce zero Braintrust spans — no input/output, no token usage, no cost.

What is missing

LiteLLM function	Description	Instrumented?
litellm.completion / .acompletion	Single-call chat completion	Yes
litellm.batch_completion	Fans a list of message lists out to concurrent completion() calls against one model, returns list of responses	No
litellm.batch_completion_models	Parallel calls across multiple models, returns first response	No
litellm.batch_completion_models_all_responses	Parallel calls across multiple models, returns all responses	No

Source of the function list: litellm/batch_completion/main.py and the Batching Completion() docs:

from litellm import batch_completion

responses = batch_completion(
    model="claude-2",
    messages=[
        [{"role": "user", "content": "good morning?"}],
        [{"role": "user", "content": "what's the time?"}],
    ],
)

Each element of messages is dispatched as an independent completion()-shaped call (concurrently, via a thread pool, max_workers configurable), so this is a distinct execution surface from the already-instrumented single-shot completion/acompletion, not just a client-side loop the user writes themselves — the concurrency and result aggregation happen inside LiteLLM.

Braintrust docs status

not_found. The Braintrust LiteLLM integration page states Braintrust "patches LiteLLM's top-level call entry points and creates an LLM-typed span per call," and explicitly lists the entry points covered: completion, acompletion, text_completion, atext_completion, responses, aresponses, image_generation, aimage_generation, embedding, aembedding, moderation, amoderation, speech, aspeech, transcription, atranscription, rerank, arerank. batch_completion and its variants are not mentioned anywhere on the page.

Upstream sources

• LiteLLM batch completion source: https://github.com/BerriAI/litellm/blob/main/litellm/batch_completion/main.py
• LiteLLM "Batching Completion()" docs: https://docs.litellm.ai/docs/completion/batching
• LiteLLM proxy-server batch completion guide (related, separate surface): https://docs.litellm.ai/docs/proxy/user_keys#beta-batch-completions---pass-model-as-list

Local repo files inspected

• py/src/braintrust/integrations/litellm/patchers.py — defines LiteLLMCompletionPatcher through LiteLLMArerankPatcher (18 patchers total); no batch_completion patcher exists. Grep for batch_completion across the entire litellm/ integration directory returns zero matches.
• py/src/braintrust/integrations/litellm/integration.py — LiteLLMIntegration.patchers = _ALL_LITELLM_PATCHERS, sourced entirely from patchers.py.
• py/src/braintrust/integrations/litellm/tracing.py — wrapper implementations for all 18 patched entry points; no batch_completion wrapper.
• py/src/braintrust/integrations/litellm/test_litellm.py — test_litellm_async_parallel_requests exercises concurrent acompletion() calls via asyncio.gather, not the batch_completion API; no test references batch_completion.
• py/pyproject.toml [tool.braintrust.matrix.litellm] (line 398-402): latest = "litellm==1.100.0", floor "1.74.0" = "litellm==1.74.0" — both versions ship batch_completion.