[bot] Instrument SGLang offline engine (574,007 weekly downloads)

Start with py/src/braintrust/integrations/cohere/patchers.py and inspect py/src/braintrust/integrations/, py/src/braintrust/wrappers/, py/pyproject.toml, and py/noxfile.py for existing integration patterns. Read SGLang's Offline Engine API documentation first; done means the listed generate, async_generate, encode, and rerank paths, including streaming and VLM inputs, produce Braintrust spans with project registration and validation coverage.

<!-- library-gap-audit: sglang-offline-engine -->

Summary

sglang is a high-performance LLM/multimodal serving framework with RadixAttention, continuous batching, and speculative decoding, positioned alongside vllm (already tracked as a separate open instrumentation gap in this repo) as one of the two dominant open-source inference engines. This repository has zero instrumentation for it.

• Current version 0.5.20 (released 2026-09-18), under very active development
• Not present in py/pyproject.toml [tool.braintrust.matrix], no noxfile.py session, no directory under py/src/braintrust/integrations/ or py/src/braintrust/wrappers/
• Repo-wide grep for sglang under py/ returns zero matches

What is missing

SGLang's in-process Offline Engine API (sglang.Engine, docs: https://docs.sglang.io/docs/basic_usage/offline_engine_api) is the primary way Python applications call it directly (as opposed to going through its OpenAI-compatible HTTP server, which a user could separately point the already-instrumented openai client at). None of these produce a Braintrust span today:

• Engine.generate() — synchronous text/chat generation, non-streaming and streaming variants
• Engine.async_generate() — asyncio generation entrypoint, non-streaming and streaming variants; this is the method the framework's own example server-wrapping code (examples/runtime/engine/custom_server.py) uses to build custom serving layers
• Engine.encode() — embedding execution (e.g. BERT-base/e5-mistral style models) for retrieval/similarity use cases
• Engine.rerank() — query/document relevance scoring, analogous to the cohere.rerank span type this repo already instruments (py/src/braintrust/integrations/cohere/patchers.py)

This engine also supports vision-language model (VLM) inference via the same generate()/async_generate() surface with image inputs.

Weekly downloads

Weekly downloads: 574,007 (as of 2026-09-21; https://pypistats.org/api/packages/sglang/recent)

Braintrust docs status

not_found — checked https://www.braintrust.dev/docs/integrations/ai-providers and https://www.braintrust.dev/docs/guides/traces/integrations; neither mentions SGLang. (vllm, the closest analog, is also absent from both pages and is tracked separately.)

Upstream sources

• GitHub: https://github.com/sgl-project/sglang
• PyPI: https://pypi.org/project/sglang/ (v0.5.20, released 2026-09-18)
• Offline Engine API docs: https://docs.sglang.io/docs/basic_usage/offline_engine_api

Local repo files inspected

• py/src/braintrust/integrations/ — no sglang/ directory
• py/src/braintrust/wrappers/ — no SGLang wrapper
• py/pyproject.toml [tool.braintrust.matrix] — no sglang entry
• py/noxfile.py — no test_sglang session
• Repo-wide case-insensitive grep for sglang under py/ — zero matches