On 10 November 2026, UC San Diego’s Hao AI Lab introduced vstack, a five-stage verification workflow for coding agents, and veriserve, a prototype large language model serving engine. The lab reports that applying the workflow to SGLang and vLLM uncovered three classes of concurrency problems and led to proposed upstream fixes.[1]
The work examines serving systems that split prompt processing from token generation and transfer key-value (KV) cache pages between them using remote direct memory access (RDMA). The authors report reproducing races in which a page or transfer identity was reused before an earlier transfer stopped. They say late transfers could read or overwrite another request’s cache while the server returned HTTP 200. They also report a corrupted page entering a shared prefix cache and stalled transfers leaving requests waiting.[1]
Vstack pins specifications, maps model transitions to code, checks model fidelity against buggy and patched behavior, and applies model checking and theorem proving. In veriserve, a design called PdLedger tracks page generations and separates pages awaiting evidence that transfers have stopped from pages quarantined when that evidence is unavailable. The authors report a TLAPS safety proof for the abstract design under stated assumptions, and a mean 0.18% median time-per-output-token overhead across six paired runs. These are the authors’ reported results; the proof does not establish that the Python implementation, GPU kernels or RDMA providers are bug-free.[1]
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
73a88513f54a1ce21fd4a6af5a987e0cea104978301e28580310148a0e474fcc