Engram-Vercel-Provider

Engram (formerly sglang-mamba) — persistent stateful inference for Mamba/SSM hybrid models on SGLang. Snapshot save/restore, 3-tier memory hierarchy, conversation_id tracking, tested across Granite, Nemotron, and Qwen architectures on H200. Use when working on the Engram repo (formerly sglang-mamba), debugging snapshot pipeline, deploying Mamba models, or running test phases.

Reference for the Vercel AI SDK Language Model Specification V4. This skill should be used when implementing or modifying a custom provider for the Vercel AI SDK, understanding the LanguageModelV4 interface, mapping prompts to provider format, implementing streaming, or handling errors in a provider context. Also covers the upgrade path from V1 to V4.